ANA BABIĆ ROSARIO, FRANCESCA SOTGIU, KRISTINE DE VALCK, and TAMMO

H.A. BIJMOLT*

The increasing amount of electronic word of mouth (eWOM) has signif-

icantly affected the way consumers make purchase decisions. Empirical
studies have established an effect of eWOM on sales but disagree on which
online platforms, products, and eWOM metrics moderate this effect. The
authors conduct a meta-analysis of 1,532 effect sizes across 96 studies
covering 40 platforms and 26 product categories. On average, eWOM is
positively correlated with sales (.091), but its effectiveness differs across
platform, product, and metric factors. For example, the effectiveness of
eWOM on social media platforms is stronger when eWOM receivers can
assess their own similarity to eWOM senders, whereas these homophily
details do not influence the effectiveness of eWOM for e-commerce
platforms. In addition, whereas eWOM has a stronger effect on sales for
tangible goods new to the market, the product life cycle does not moderate
the eWOM effectiveness for services. With respect to the eWOM metrics,
eWOM volume has a stronger impact on sales than eWOM valence. In
addition, negative eWOM does not always jeopardize sales, but high
variability does.

Keywords: electronic word of mouth, online platforms, social media, eWOM

metrics, meta-analysis

Online Supplement: http://dx.doi.org/10.1509/jmr.14.0380

The Effect of Electronic Word of Mouth on

Sales: A Meta-Analytic Review of Platform,
Product, and Metric Factors

In marketing, word of mouth (WOM) is the act of
consumers providing information about goods, services,
brands, or companies to other consumers. Such information
communicated through the Internet (through, e.g., reviews,
tweets, blog posts, “likes,” “pins,” images, video testimonials)
is called “electronic word of mouth” (eWOM), and it
*Ana Babić Rosario is a doctoral candidate in Marketing, HEC Paris
(e-mail: ana.babic-rosario@hec.edu). Francesca Sotgiu is Assistant Professor
of Marketing, Vrije Universiteit Amsterdam (e-mail: f.sotgiu@vu.nl). Kristine
de Valck is Associate Professor of Marketing, HEC Paris (e-mail: devalck@
hec.fr). Tammo H.A. Bijmolt is Professor of Marketing Research, Faculty of
Economics and Business, University of Groningen (e-mail: t.h.a.bijmolt@rug.
nl). This project is part of the first author’s doctoral dissertation. The authors
gratefully acknowledge the support of the HEC Foundation (F1307) and
Labex ECODEC (Investissements d’Avenir ANR-11-IDEX-0003/Labex
Ecodec/ANR-11-LABX-0047) and thank Katrijn Gielens and the partici-
pants of the 2014 HEC-INSEAD-ESSEC research seminar for their con-
structive feedback. V. Kumar served as associate editor for this article.
represents one of the most significant developments in con-
temporary consumer behavior. With more than three billion
consumers and seven billion devices connected to the Internet
(International Telecommunication Union 2014), eWOM has
become ubiquitous and accessible, turning consumers into
“web-fortified” decision makers (Blackshaw and Nazzaro
2006). Inducing, collecting, and displaying eWOM have be-
come priorities of many companies as part of their efforts to
stimulate sales. According to Bain & Company (Barry et al.
2011), the average billion-dollar company spends $750,000 a
year on earned media, with some early adopters such as Dell and
American Express investing significantly more. Although the
market relevance of eWOM is recognized, many professionals
have not yet determined how to manage eWOM successfully. A
Forrester survey (Elliott et al. 2012) of interactive marketers
shows that assessing the return on investment of eWOM-related
efforts is considered one of the greatest challenges interactive
marketers face today.

© 2016, American Marketing Association Journal of Marketing Research

ISSN: 0022-2437 (print) Vol. LIII (June 2016), 297–318
1547-7193 (electronic) 297 DOI: 10.1509/jmr.14.0380
298
The topic of assessing eWOM’s impact on firm per-
formance has also garnered a great amount of academic
interest. In the past 15 years, more than 100 studies have
investigated whether and to what extent eWOM is linked to
the bottom line. Yet the number of studies addressing the
effectiveness of eWOM has decreased in recent years (for
an overview of the studies and effect sizes from 1999 to
2013, see the Web Appendix), suggesting that a full un-
derstanding of the phenomenon has been reached; none-
theless, two key debates remain unsettled (see Table 1).
The first inconclusive area of investigation pertains to the
moderating role of platform and product characteristics on
the effect of eWOM on sales. Prior studies have mostly relied
on a single sample (and, consequently, one platform and/or one
product type) and thus have not been able to investigate the
moderating effects of these factors. As Forman, Ghose, and
Wiesenfeld (2008, p. 291) note, “[The] research in this arena is
fragmented [and] we have yet to understand why, how, and
what aspects of online consumer-generated product reviews
influence sales.”
The second inconclusive area involves eWOM metrics.
Although most research has concluded that eWOM has a
significant monetary effect on sales beyond other marketing-
mix effects (Chen, Wang, and Xie 2011; Chevalier and
Mayzlin 2006; Moe and Trusov 2011), there is disagreement
about which particular metrics of eWOM representing dif-
ferent aspects (e.g., reach, preference, consumer [dis]agree-
ment) drive the effect.1 Whereas some researchers find
evidence that the number of online reviews predicts product
sales (e.g., Duan, Gu, and Whinston 2008; Gu, Park, and
Konana 2012; Ho-Dac, Carson, and Moore 2014; Liu 2006;
Xiong and Bharadwaj 2014), others posit that the main pre-
dictor is not the volume of eWOM, but rather its valence (e.g.,
Chintagunta, Gopinath, and Venkataraman 2010; Dellarocas,
Zhang, and Awad 2007), variance (e.g., Sun 2012), mere
existence (e.g., Davis and Khazanchi 2008), or specific content
(e.g., Onishi and Manchanda 2012). In addition, findings on
the impact of negative eWOM are inconclusive. Some studies
have shown that negative eWOM is detrimental and even more
powerful in decreasing sales than positive eWOM is in in-
creasing it (Chevalier and Mayzlin 2006; Sun 2012); con-
versely, other studies have shown that the presence of negative
eWOM increases product evaluations and sales (e.g., Doh and
Hwang 2009; Hiura et al. 2010; Kikumori and Ono 2013).
These varying results hinder the development of syste-
matic insight that can help marketers make informed decisions
about eWOM management. The current study discusses a
meta-analysis that offers a comprehensive synthesis across
studies. This is important for two reasons. First, the studies
differ substantively in terms of online platform, industry, pro-
duct, geographic region, and eWOM metrics investigated.
Second, the studies differ in methodological approaches.
Meta-analytic methods enable researchers to (1) obtain em-
pirical generalizations about a specific effect size across
varying substantive and methodological conditions and (2)
examine whether and how these conditions affect the focal
effect size (see Farley, Lehmann, and Sawyer 1995). There-
fore, our meta-analysis contributes to understanding of the
1In the remainder of the article, we use the term “eWOM metrics” (e.g.,
“valence”). We acknowledge that it is the underlying aspects of eWOM (e.g.,
“preference”) that drive sales and not the metrics per se.
JOURNAL OF MARKETING RESEARCH, JUNE 2016
influence of platform characteristics, product characteristics,
and eWOM metrics on the effect of eWOM on sales.
First insights along these lines come from the work of
Floyd et al. (2014) and You, Vadakkepatt, and Joshi (2015).
Floyd et al. (2014) examine 26 studies investigating the
effect of online reviews on firm performance. Although
consumer-generated online reviews are an important cat-
egory of eWOM, the phenomenon of eWOM cannot be
reduced to online reviews only. You, Vadakkepatt, and
Joshi’s (2015) meta-analysis focuses on the effects of
consumer-generated information on firm performance,
extending the analysis to 51 papers and a few online
platforms (i.e., blogs, discussion forums, and Twitter).
Although these two studies provide important insights
into the impact of eWOM on sales, their generalizations are
limited in terms of the number of platforms, products, and
metrics investigated, leaving the aforementioned debates
unresolved. Our study provides answers to these debates.
We outline the main differences between our study and
previous meta-analyses in Table 2 and the Web Appendix.
This study makes several contributions to the literature
on eWOM and interpersonal influence. First, we offer fur-
ther insight into the moderating effects of platform and
product characteristics (across 40 platforms and 26 product
categories). To do so, we compile a unique data set by
collecting supplementary time-varying information that re-
flects the nature of each platform and product category at
the time of the original data collection. This is critical in
meta-analytic work because it is often impossible to re-
trieve time-specific information that matches the primary
studies. Using the Wayback Machine (https://archive.org/
web), an Internet archive that provides access to past
platform interfaces, we tracked the evolution of each
platform over time and the type of information provided
about the eWOM senders and messages. Moreover, we
employed expert coders to code additional product char-
acteristics at the time of the original data collection. With
the addition of this primary data set, we could investigate
dimensions of eWOM that were not part of the primary
studies and that have been neglected as possible moderators
of the effect of eWOM on sales. Thus, our meta-analysis
moves beyond a summary of extant literature. Our em-
pirical results reveal how both platform and product
characteristics influence the effect of eWOM on sales.
Among other things, we find that the effectiveness of
eWOM on social media platforms is stronger for platforms
that enable eWOM receivers to assess their own similarity
to eWOM senders on the basis of username, avatar, profile
page, and geographic location. However, for e-commerce
platforms, these homophily details do not influence
the effectiveness of eWOM. Instead, we find that on
e-commerce platforms, eWOM increases sales more when
it is immediately visible (i.e., no scrolling is required to
access eWOM information) and when it is less structured
(e.g., no summaries are provided). We also find that dif-
ferent moderators influence eWOM effectiveness across
products. For example, eWOM has a stronger effect on
sales for tangible goods new to the market. In contrast, the
product life cycle does not moderate the eWOM effec-
tiveness for services and hedonic products.
Second, we examine the interplay between platform-
and product-level moderators of eWOM effectiveness to
The Effect of Electronic Word of Mouth on Sales
CONCEPTUAL FRAMEWORK TO ADDRESS UNSETTLED DEBATES
RQ1: For Which Platform and Product Characteristics Is eWOM More Effective?
Platforms
Unsettled debate Insufficient evidence due to single-platform Insufficient evidence due to single-product
studies category studies
Key characteristics • Platform type
- Social media platform
- Review platform
- E-commerce platform
- Other
• eWOM sender
- Homophily
- Trustworthiness
• eWOM message
- Time stamp
- Helpfulness rating
- Visibility
- Structured display
• Platform maturity
• eWOM posting costs
• Product characteristics
• Operationalization of eWOM metrics
Methodological • Coding of time-varying information about
approach 40 platforms at the time of the original
data collection
• Overall assessment of the relative
importance of different platform types
and characteristics (weighted random-
effect HiLMA)
• Platform-specific analysis of moderating
effects (weighted random-effect split-
sample HiLMA)
identify not only for which platform and product charac-
teristics but also for which platform–product combinations
eWOM more effectively increases sales. For example,
when a product is new to the market, eWOM is more ef-
fective when appearing on an e-commerce platform or a
review platform rather than on a social media platform.
Third, we investigate other representations of eWOM
(e.g., variance) important in the marketplace, whose effects
on sales remain unclear. Thus, we go beyond analyzing
volume and valence metrics to investigate the differential
effects of four key metrics of eWOM (i.e., volume, valence,
composite valence–volume, and variance) comprising 12
submetrics categorized on the basis of their operationa-
lizations. By considering this larger set of metrics, we
capture variation within all metrics (e.g., how the effect of
positive valence differs from the effect of average valence
and positive volume), thereby gaining more detailed insight
299
Table 1
RQ2: Which eWOM Metric Is
Products More Effective?
• Volume increases sales more than valence
(e.g., Liu 2006)
• Volume increases sales less than valence
(e.g., Chintagunta et al. 2010)
• Negative eWOM decreases sales (e.g.,
Chevalier and Mayzlin 2006)
• Negative eWOM increases sales (e.g.,
Doh and Hwang 2009)
• Unclear effect of variance (e.g., Sun 2012)
• Functional risk • Operationalization of eWOM metrics
- Tangible vs. digital vs. service - eWOM volume
- Hedonic vs. utilitarian product - eWOM valence
- New vs. mature product - eWOM composite valence–volume
• Financial risk - eWOM variance
- High vs. low risk • Platform characteristics
• Platform characteristics • Product characteristics
• Operationalization of eWOM metrics
• Coding of time-varying information about • Inclusion of metrics other than volume
26 product categories at the time of the and valence (e.g., variance; 12 submetrics
original data collection in total)
• Overall assessment of the relative • Differentiation between the absolute and
importance of different product types and relative effect of valence: introduction
characteristics (weighted random-effect of a new metric (composite
HiLMA) valence–volume)
• Product-specific analysis of moderating • Overall assessment of the relative
effects (weighted random-effect split- importance of the most common metrics
sample HiLMA) (weighted random-effect HiLMA)
• Metric-specific analysis of moderating
effects (weighted random-effect split-
sample HiLMA)
in the mechanisms underlying the impact of eWOM on
sales. Importantly, we show that failing to distinguish
between the submetrics of eWOM may lead to a mis-
interpretation of the relative effect of volume versus va-
lence. In particular, we introduce a new variable, composite
valence–volume (e.g., total number of five-star ratings),
which has been commonly labeled as valence (e.g., per-
centage of five-star ratings). We argue that these two
metrics are conceptually different in that they represent
different underlying aspects, and we show that using a
more precise measure of valence—uncontaminated by
volume—reverses the conclusion that valence has a stronger
effect on sales than volume, as previous meta-analyses have
shown. We find this result consistently across platforms and
products, with the exception of services, for which the effect of
volume metrics on sales is not significantly different from the
effect of positive valence.
300 JOURNAL OF MARKETING RESEARCH, JUNE 2016

Table 2
KEY DIFFERENCES ACROSS META-ANALYSES ON eWOM

Sample
Number Number of
Focus
of Number of Number of Product
Study Studies Effects Platforms Categories Metrics Platform Characteristics Product Characteristics
Floyd et al. (2014) 26 443 16 13 Volume Product involvement
Valence

You, Vadakkepatt, 51 339 (volume) 15 18 Volume Independent review sites, Privately consumed,
and Joshi 271 (valence) Valence specialized review sites, trialability, category
(2015) community-based sites competition, durables

This study 96 1,532 40 26 Volume Social media platforms, review Tangible, digital, service,
589 (volume) Valence platforms, e-commerce utilitarian, hedonic, financial
596 (valence) Composite platforms, other platforms, risk, new, mature
108 (composite valence–volume eWOM sender homophily,
valence–volume) Variance eWOM sender trustworthiness,
137 (variance) Other eWOM message time stamp
102 (other) and helpfulness rating, eWOM
visibility, eWOM structured
display, platform maturity,
eWOM posting costs

Fourth, to better monitor and manage eWOM, it is crucial
to understand not only which eWOM metric contributes
more to a sales increase (e.g., volume vs. valence) but also
which metric can significantly decrease sales. We are the
first to reconcile evidence of the effects of negative eWOM
on sales by identifying platform and product characteristics
for which negative eWOM (negative valence and negative
volume) does not jeopardize sales and those for which
it does. Furthermore, we show that another submetric that
is often overlooked—variability (consumer disagreement
about product quality)—decreases sales on e-commerce
platforms and for tangible, utilitarian, new, and high-
financial-risk products.
In the following sections, we first discuss our con-
ceptual framework and then describe the collection and
coding of 1,532 effect sizes from 96 studies across 40
platforms and 26 product categories. We present results
from two meta-analytic procedures—namely, a Hedges–
Olkin meta-analysis (HOMA; Hedges and Olkin 1985)
and a hierarchical linear meta-analysis (HiLMA; Lipsey
and Wilson 2001). Using split samples, we also show how
the moderators of the link between eWOM and sales differ
across platforms, products, and metrics. Finally, we dis-
cuss the implications of our findings for marketers and
researchers and provide directions for further research.
CONCEPTUAL FRAMEWORK
As the market environment has become more saturated
with products and marketer-generated information, it has
become increasingly difficult for consumers to know and
process all alternatives. Electronic word of mouth helps
consumers minimize uncertainty (Dichter 1966; Roselius
1971), and thus further insights into the specific circum-
stances in which eWOM is a more powerful risk-reducing
tool are required. In the following sections, we provide a
framework that describes the key characteristics of the
platforms, products, and eWOM metrics that may moderate
the impact of eWOM on sales (see Table 1).
For Which Platform and Product Characteristics Is eWOM
More Effective?
Platform characteristics. It is important to account for
the characteristics of the channel in which eWOM is dis-
played (Berger and Iyengar 2013; Schweidel and Moe
2014). We examine five characteristics of the eWOM plat-
form. First, we account for the different types of plat-
forms: (1) social media platforms (e.g., Facebook, blogs,
discussion forums), (2) review platforms (e.g., Epinions,
Yahoo!Movies), (3) e-commerce platforms (e.g., Amazon.
com, eBay), and (4) other platforms (e.g., Internet overall).
Second, we acknowledge that consumers often evaluate the
value of online platforms as information channels on the
basis of additional information provided about the eWOM
sender. Of particular importance are signals of homophily
and trustworthiness (Fogg et al. 2003; Hung, Li, and Tse
2011). Prior research has shown that messages coming from
similar others are more persuasive (Brown and Reingen
1987; Forman, Ghose, and Wiesenfeld 2008; Reichelt,
Sievert, and Jacob 2014). Electronic word of mouth is also
more effective when receivers trust the sender and can be
confident about their good intentions (McGinnies and
Ward 1980). Consequently, information about the sender
aids receivers in assessing whether the eWOM is relevant
to them and whether the sender is trustworthy, potentially
leading to a higher correlation with sales. Third, extant
studies have shown that additional information pro-
vided about the eWOM message, such as time stamp and
helpfulness rating, increases sales (Berger 2014; Robins,
Holmes, and Stansbury 2010). Similarly, when eWOM in-
formation is immediately visible and displayed in a more
The Effect of Electronic Word of Mouth on Sales
structured way (e.g., with summaries of the most represen-
tative positive and negative reviews), it may have a greater
impact on the bottom line. Fourth, the reputation of a
platform as a valuable information channel needs time to
develop. In earlier stages of development, platforms attract
fewer visitors, and they experience smaller network effects
than more mature platforms. Therefore, eWOM displayed
on platforms introduced more recently may have a lesser
impact on sales than eWOM appearing on more mature
platforms. Fifth, we differentiate platforms according to
posting policies. Specifically, eWOM senders may incur
posting hurdles. For example, they may need to have
purchased the product or registered as a member to create
or disseminate a review, upload a video, and so on
(Mayzlin, Dover, and Chevalier 2014). Previous research
has shown that such costs decrease the prevalence of fake
reviews and thus increase the value of eWOM (Ott, Cardie,
and Hancock 2012). Therefore, the effect of eWOM on
sales may be greater for platforms imposing eWOM posting
costs.
Product characteristics. The topic of eWOM pertains to
goods, services, brands, and companies. These goods and
services have different levels of functional and financial
risk (Wangenheim and Bayón 2004). To reduce uncer-
tainty about perceived risk, consumers consult with others
(Roselius 1971). We examine four product characteristics
moderating eWOM effectiveness, for which the expected
differences depend on the importance of functional and
financial risks.
Functional risk is linked to uncertainty about a product’s
performance, and research has demonstrated that it is higher
for services (vs. tangible goods), hedonic (vs. utilitarian)
products, and new (vs. mature) products. In particular, the
performance quality of services (e.g., hotel stays, restaurant
dinners) is usually more difficult to assess before purchase
than that of goods (Murray and Schlacter 1990; Zeithaml
1981). Because eWOM may replace information obtained
through sampling or purchase, consumers may rely more on
eWOM for services than for goods to reduce perceived
functional risk (Murray 1991). Similarly, functional risk is
higher for hedonic products, which are pleasant and en-
joyable and appeal to the senses (e.g., perfume) (Dhar and
Wertenbroch 2000), and consumers experience and assess
them more subjectively than utilitarian products (which
are useful and practical, e.g., vacuum cleaners). Electronic
word of mouth can reduce this risk by helping “consumers
identify the products that best match their idiosyncratic
preferences” (Moe and Trusov 2011, p. 444). Functional
risk is also higher for newly introduced products because
it is difficult to anticipate their performance. For these
products, WOM plays a particularly important role in building
product awareness and providing information to consumers
(Mahajan, Muller, and Kerin 1984).
Financial risk reflects the money it takes for consumers to
make the product work properly or to replace it if it fails or
does not meet expectations (Roselius 1971). In the case of
high financial risk, consumers rely more heavily on eWOM
(Lin and Fang 2006).
Which eWOM Metric Is More Effective?
Electronic word of mouth has been measured in prac-
tice and operationalized in extant academic literature in
301
multiple ways to capture different aspects. We distinguish
among the following eWOM metrics: volume, valence,
composite valence–volume, variance, and other. Electronic
word of mouth volume measures “the total amount of
eWOM interaction” (Liu 2006, p. 75)—that is, the total
number of eWOM units sent about a particular object.
Because eWOM volume inherently delivers information
about how many other people have experienced or used the
product and how popular the product is in the market, it can
increase consumers’ awareness of and reduce their un-
certainty about the product, ultimately leading to an in-
crease in sales (Chen, Wang, and Xie 2011; Chintagunta,
Gopinath, and Venkataraman 2010; Park, Gu, and Lee
2012). The underlying dynamic is the bandwagon effect
(e.g., Van den Bulte and Lilien 2001), in which the mere
availability of other consumers’ opinions has an influence
on other consumers, regardless of whether these opinions
are positive or negative (Godes and Mayzlin 2009; Xiong
and Bharadwaj 2014). The main idea behind the reliance on
the amount of peer-generated information in consumers’
decision-making process is rooted in herding behavior and
social impact theory, according to which people tend to
follow the previous behavior of others to reduce risk in the
environment (Banerjee 1992; Latané 1981). In addition, the
more consumers discuss a product, the greater the chance
that other consumers will become aware of it, because
message repetition attracts people’s attention to the topic of
the message (Cacioppo and Petty 1989; Tellis 1988).
Valence is “the idea that eWOM can be either positive,
negative, or neutral” (Liu 2006, p. 75). It is also called the
“favorability,” “sentiment,” or “polarity” of eWOM and
refers to both objective information found in the eWOM
message (e.g., “The hotel room was infested with cock-
roaches”) and the affect expressed therein (e.g., “I hated
that movie!”). In this case, consumers’ preferences for the
product are formed, reinforced, or altered from the expo-
sure to (un)favorable eWOM, which is indicative of a
product’s reputation and expected product quality (Kim
and Gupta 2012; Liu 2006). Research refers to this as the
persuasion effect or informational influence of eWOM (Godes
and Mayzlin 2009; Rui, Liu, and Whinston 2013). Valence
has been captured in various ways, some of which may be
theoretically contaminated with measures of eWOM volume.
Whereas operationalizations using relative terms, such as the
ratio of positive tweets (e.g., Hennig-Thurau, Wiertz, and
Feldhaus 2014) or the percentage of one-star ratings (e.g.,
Chen, Wang, and Xie 2011), are a nonconfounded repre-
sentation of sentiment (positive or negative valence), oper-
ationalizations using absolute terms, such as the number of
positive tweets (e.g., Hennig-Thurau, Wiertz, and Feldhaus
2014) or the number of one-star ratings (e.g., Chevalier and
Mayzlin 2006), indicate both volume and valence.
Thus, we introduce a new metric labeled “composite
valence–volume” (with two submetrics: positive volume
and negative volume). This composite measure represents
the combined influence of persuasion and bandwagon ef-
fects. For example, seeing that a product received 500
Facebook “likes” informs a consumer about the sentiment
toward the product while also providing an indication about
the actual number of people who formed an opinion about
it. This represents both valence and volume. In contrast,
the metric percentage of one-star ratings indicates that
302
some consumers share negative sentiment, but it does not
communicate how many consumers share an opinion. This
is a measure of valence.
A less commonly investigated eWOM metric is vari-
ance, which represents “a natural measure to capture the
heterogeneity in consumer opinions [such that] upon see-
ing a high variance, consumers infer that the product is a
niche one that some people love and others hate” (Sun
2012, p. 697; emphasis added). Low variance of eWOM
means that consumers agree that the product is either good
or bad, which explains why the effect of eWOM on sales
can be either positive or negative.
Finally, a range of other metrics exists, including the
mere presence of consumer-generated information on a
particular platform (eWOM existence) and specific words
and phrases, such as the word “award” in movie-related
eWOM (eWOM content). In summary, we expect that
platform characteristics, product characteristics, and met-
rics, as well as their interplay, moderate the impact of
eWOM on sales.
METHODOLOGY
Collection and Coding of Studies
To identify the empirical studies investigating the effects of
eWOM on sales, we conducted a rigorous and thorough lit-
erature search. We checked several online scientific databases,
including SSR, ABI Inform, and EconLit, and carried out an
issue-by-issue search of the top journals in three streams of re-
search: (1) marketing, (2) economics and management, and (3)
information systems and computer science. We also searched
for unpublished work, including dissertations and working
papers in databases, such as the ProQuest Dissertation
Express, MSI, SSRN, EconPapers, REPEC, and AISeL, as well
as papers in the proceedings of the most prominent conferences
in the specified research fields. Next, we scanned the Internet
using Google Scholar. Because eWOM has a myriad of aliases,
we used the following keywords in the search process: “buzz,”
“consumer-generated content,” “electronic word of mouth,”
“eWOM,” “online opinion,” “online rating,” “online recom-
mendation,” “online review,” “online word of mouth,” “online
WOM,” “peer recommendation,” “user comments,”
“user-generated content,” “user ratings,” “user review,” and
“social earned media.” We applied a snowballing procedure, in
which we examined the references in the publications obtained
to find additional studies. Moreover, we corresponded with the
researchers represented in the original data set, requesting
additional information and inviting related studies on the topic.
Finally, we posted messages calling for additional studies on
the electronic mailing list ELMAR.
After completing the search process on December 1,
2014, we excluded theoretical papers, qualitative investiga-
tions, and quantitative studies that did not report findings
on the outcomes of eWOM but investigated only its an-
tecedents. We further restricted the focus of our analysis to
work examining the impact of eWOM on objective mea-
sures of performance, such as the number of product units
sold and revenues from sales of firm-created products
and services, while excluding those assessing consumer
product evaluations, purchase intentions, television view-
ership, free online music sampling, sales distributions,
JOURNAL OF MARKETING RESEARCH, JUNE 2016
growth rates, and forecasts. The final data set consists of
1,532 effect sizes, retrieved from 96 studies (see the Web
Appendix).
The average number of effect sizes reported per study is
16, with a minimum of 1 and maximum of 260. The division
of effect sizes over the three major research streams is as
follows: information systems and computer science (52%),
marketing (36%), and economics and management (12%).
Our sample includes 55 articles (57%) published in 25
journals, as well as 20 conference proceedings papers
(21%), 16 working papers (17%), and 5 doctoral disser-
tations (5%), all published between 2004 and 2014 (with
data collected between 1999 and 2013).
We developed a transparent and replicable coding pro-
tocol containing a detailed coding manual with descrip-
tions of each variable. A single coder was trained to code all
the studies. To ensure the reliability of coding, another
coder independently coded a subsample of 920 effect sizes
and related variables. The Cohen’s kappa coefficient of
interrater reliability is .97, which is satisfactory (Landis
and Koch 1977), and disagreements were solved through
discussion.
Operationalization of sales. The sales variable was
operationalized in six ways in the primary studies: sales
rank (60.5%), number of customers or product units sold
(11%), total sales (12.5%; i.e., sum of online and offline
sales), offline sales (7.5%), online sales (5.5%), and per-
centage of opening-weekend revenues of total revenues
(3%). When sales rank was used as the dependent variable,
we changed the sign of the effect size to account for
the inverse relationship of this measure to actual sales
(Brynjolfsson, Hu, and Smith 2003) unless corrections
were already undertaken in the primary studies.2
Platforms. In the primary studies, eWOM was collected
on more than 40 different online platforms: Allocine.fr,
Amazon (U.S., U.K., and German versions), Apple and
Google app stores, Autobytel, Barnes & Noble, Car and
Driver, CNET, ConsumerReports.org, Ctrip, Dangdang,
Delicious, Digg, DpReview, eBay, Edmunds.com,
Epinions, Facebook, GameSpot (i.e., VideoGames.com),
Gforums, HowardForums.com, Internet Movie Database,
Kiva friends, MSN, MySpace, Naver, Netflix, Plurk,
Rotten Tomatoes, Taobao, Tongcheng (17u), TStore,
Twitter, Yahoo!Movies, Yelp, YouTube, several un-
named e-commerce platforms (in Asia, Europe, and North
America), various blogs, and the Internet in general. The
most frequently investigated platforms are Amazon (44%),
Barnes & Noble (12%), and Yahoo!Movies (10%). We
capture variation across all the platforms further using a
set of platform characteristics discussed subsequently.
2Because Amazon sales rank rates the best-selling products with a lower
number, this measure represents an inverse of actual sales. When Amazon
sales rank was used as the dependent variable, we searched for information
about potential corrections; if we were unable to find any, we changed the
sign of the effect size to account for the inverse relationship of this measure to
actual sales. Some authors already use available approximations of (online)
sales based on Amazon sales rank (log[sales] = b1 + b2 × log[sales rank] + e)
(Brynjolfsson, Hu, and Smith 2003; Chevalier and Goolsbee 2003; Ghose,
Smith, and Telang 2006), in which cases we treat these as online sales and do
not correct the effect sizes in the aforementioned way. We provide several
examples of effect size computations in the Web Appendix.
The Effect of Electronic Word of Mouth on Sales
Products. Electronic word of mouth collected in the
primary studies is related to 26 product categories: audio
players, apparel, books, cars, cellular phone devices,
cellular phone services (e.g., prepaid cards), computer
memory, digital cameras, electronics, financial services
(microloans), furniture, garden products, green tea, hotel
stays, houseware, Internet services, mobile applications,
movies, music albums, perfume, restaurant services, soft-
ware, video cassettes and DVDs (rental and purchase),
video games, and video players. The majority of the effects
are related to outcomes in three main product categories:
books (39%), movies (20%), and digital cameras (8%).
Services represent 24% of the products in our data set.
Hedonic products make up 39% of the sample. While
products were analyzed at different stages of their life cycle,
as many as 37% of the effects are for newly introduced
goods and services; 21% are products that carried high fi-
nancial risk at the time of the primary data collection.
Metrics. Most researchers captured eWOM mainly by
eWOM volume (used in 88% of the studies) and valence
(used in 81% of the studies) while paying limited attention
to variance (used in 18% of the studies) and other eWOM
metrics (i.e., mere existence of eWOM or specific words
contained in a textual post were used in 16% of the studies).
In addition, we find that eWOM was operationalized as a
composite variable containing elements of both valence and
volume—namely, as the number of positive/negative posts
containing eWOM (used in 14% of the studies). So far, the
primary studies labeled such a variable as either valence
or volume, even though it represents a combination of
these two measures and is coded as a composite variable
in this article. Furthermore, 12% of the studies investigated
eWOM by focusing on one metric only, while the vast
majority of the studies used multiple metrics for eWOM.
We find great variation in the way extant literature oper-
ationalized each of these variables, and we classify (1) each
volume measure as an average, an incremental, or a cumulative
measure; (2) each valence and composite valence–volume
measure as positive or negative; and (3) each variance measure
as agreement and precision, variability, or incremental.3
Control variables. Multiple methodological and study
characteristics could also moderate the effects of eWOM
on sales, including the type of endogeneity controls,
operationalization of the dependent variable, and other
study characteristics. Furthermore, omitting from the es-
timation models other variables that are known to affect
sales, such as marketing-mix controls, may lead to erro-
neous results (for a review of relevant confounding vari-
ables, see Van den Bulte and Lilien 2001). Therefore, we
include these as control variables in our study.
Primary Data Collection
To investigate the moderation effects of platform char-
acteristics, we collected primary data by visiting the plat-
forms at the time of the original data collection using the
Wayback Machine. We trained three coders to collect in-
formation on all online platforms from our sample and to
evaluate the platform characteristics in terms of sender
3To provide a complete overview of the relative importance of the different
metrics used in the literature, in the HOMA we also include two other eWOM
submetrics (i.e., existence and content).
303
details, message details, platform maturity, and eWOM
posting costs. For sender details, we created vectors for
homophily and trustworthiness (Metzger, Flanagin, and
Medders 2010). Electronic word of mouth sender homo-
phily captures the presence of cues that can help receivers
assess their similarity to eWOM senders according to their
username, avatar, profile page, and geographic location.
Sender trustworthiness is based on real names, duration of
platform membership, and contact information. For mes-
sage details, we created a vector that consists of two var-
iables: the proportion of eWOM with a time stamp and the
proportion of eWOM with a helpfulness rating. In addition,
we coded eWOM visibility (the number of scrolls needed to
access eWOM), structured display of eWOM (e.g., having
summary sections), the year of platform introduction, and
eWOM posting costs (when purchase or registration is
required before a consumer can provide eWOM) (for de-
tails, see the Web Appendix).4
In addition to these platform characteristics, we code
information related to products’ perceived risk. Three
marketing experts classified each product category in
our sample as hedonic or utilitarian. Furthermore, they
assessed the product categories’ financial risk at the time
of the original data collection. As a result, these product-
related variables are also time varying (e.g., digital cameras
were coded differently for the year of their market in-
troduction than for more recent years because the financial
risk related to their purchase had decreased). Table 3 provides
the operationalization of all variables, and the Web Appendix
reports the correlation matrix and the descriptive statistics.
Meta-Analytic Calculations
Computation of effect sizes. To measure the effect size of
eWOM on sales, we use (bivariate and partial) correlations,
which is a common approach for meta-analytic reviews
in marketing and management (e.g., Carney et al. 2011;
De Matos and Rossi 2008; Heugens and Lander 2009). In
contrast, Floyd et al. (2014) and You, Vadakkepatt, and
Joshi (2015) use elasticities. Conducting a meta-analysis
using correlations instead of elasticities offers two ad-
vantages. First, the interpretation of correlations is in-
dependent of the measurement scale (Eisend 2006; Lipsey
and Wilson 2001). This is important when examining the
effects of various eWOM metrics on sales because both the
dependent measures (e.g., Amazon sales rank vs. box-office
revenues) and the eWOM metrics used in primary studies
are diverse in both nature and scale. In particular, eWOM
volume metrics are usually measured in absolute terms,
whereas eWOM valence metrics are usually measured on a
five-point rating scale. As Van Heerde, Gupta, and Wittink
(2003) and Van Heerde (2005) show, elasticity may not be
comparable across variables (e.g., X1 and X2), because a
4When eWOM was collected from multiple platforms, unspecified plat-
forms (e.g., blogs), or the Internet overall, we proceed in the following way:
For multiple platforms (e.g., Ctrip and Tongcheng), we code the platform
characteristics for each individual platform separately at the moment of data
collection in the primary study and for the specific product category. We then
average their values. For unspecified platforms or the Internet in general, we
use the mean of our sample as a missing value imputation. In addition, we
rerun the HiLMA model, excluding these cases altogether (174 cases in total)
and using an alternative missing value imputation (i.e., the median instead of
the mean). All three approaches yield the same results.
304 JOURNAL OF MARKETING RESEARCH, JUNE 2016

Table 3
VARIABLES IN THE HiLMA

Variable Description and Operationalization

Platforms
Platform Typea
e-Commerce sites Dummy variable that assumes the value of 1 if the online platform from which eWOM was collected is a e-commerce site
(e.g., Amazon, CNET, eBay), and 0 otherwise.
Review sites Dummy variable that assumes the value of 1 if the online platform from which eWOM was collected is a noncommercial
review site (e.g., Epinions, Gamespot, Yahoo!Movies), and 0 otherwise.
Other platforms Dummy variable that assumes the value of 1 if the online platform from which eWOM was collected is not specified or
Internet overall, and 0 otherwise.
Amazon Dummy variable that assumes the value of 1 if the online platform from which eWOM was collected is Amazon (Amazon.
com, Amazon.de, Amazon.uk), and 0 otherwise.
Details Related to. . .
eWOM sender: homophily Presence of cues pertaining to the eWOM sender’s similarity to the eWOM receiver is operationalized as a vector
(continuous variable, mean-centered) of four continuous variables capturing the relative number of instances (1) when
eWOM senders’ geographic location is displayed, (2) when eWOM senders have a profile page, (3) when eWOM senders’
usernames are displayed, and (4) when eWOM senders have an avatar. These instances are relative to the number of eWOM
messages available on the first page accessed through Wayback Machine for a given product category at the moment of the
data collection in the primary study (e.g., Amazon.de books in March 2008, Amazon.uk music albums in December 2006,
Amazon.com digital cameras in March 2007). If, in four out of six book reviews available on the landing product page on
Amazon, the reviewer’s geographic location is specified, variable 1 is coded as .67; values for variables 1–4 are summed to
create the vector.
eWOM sender: trustworthiness Presence of cues pertaining to the eWOM sender’s trustworthiness is captured as the sum of three continuous
variables indicating how often (1) eWOM senders’ real names are displayed, (2) the duration of eWOM senders’
memberships within the platform is displayed, and (3) it is possible to contact eWOM senders through e-mail or
private message. The number of instances is relative to the number of eWOM messages available on the first page
accessed through Wayback Machine for a given product category at the moment of the data collection in the primary
study.
eWOM message recency and Presence of additional information about the eWOM message. It is operationalized as a vector (continuous variable; mean-
helpfulness rating centered) of two variables indicating (1) whether a time stamp is displayed for eWOM on this platform, and (2) whether
a helpfulness rating is displayed for eWOM on this platform (based on Wayback Machine).
eWOM visibility Number of scrolls needed with a computer mouse to visualize eWOM information (continuous variable, mean-centered)
(based on Wayback Machine).
eWOM structured display Dummy variable indicating whether on a particular platform eWOM is organized into categories, provided with titles or
summary sections (based on Wayback Machine).
Other Characteristics
Platform maturity Number of years since the online platform’s introduction at the time of the data collection. It is calculated by
subtracting the year of data collection from the year of the online platform introduction (continuous variable; mean-
centered).
eWOM posting costs Dummy variable indicating whether eWOM senders incur posting costs on a particular online platform (based on Wayback
Machine).
Products
Product Typeb
Service Dummy variable indicating whether the product whose sales is examined is an intangible, perishable good that is
inseparable from its provider.
Digital product Dummy variable indicating whether the product whose sales is examined is an intangible good that exists in digital
form.
Other Product Characteristics
Financial risk Variable on a 1–5 Likert scale that assumes the value of 1 if the product carries very low financial risk (e.g., mobile apps) and
5 if it carries very high financial risk (e.g., cars) (mean-centered). Products are classified on the basis of their historical price
at the time of the primary data collection using one of the following sources: (1) average prices reported in primary studies; (2)
representative prices through the Wayback Machine for the product, time, and geographic location of the data collection (e.g.,
Amazon.uk, Edmunds.com); and (3) other sources (e.g., Tomshardware.com). Then, three coders classified all products on a 1–5
Likert scale according to the relative prices in our sample—1 for lowest prices (e.g., mobile apps) and 5 for highest prices (e.g.,
cars). Agreement was reached through discussion.
Hedonic product Dummy variable that assumes the value of 1 if the product is predominantly hedonic and 0 if it is predominantly utilitarian.
Three coders classified all products, and agreement was reached through discussion.
Stage in the Product Life Cycle
New product Dummy variable that assumes the value of 1 if the product whose sales is examined is reported in the primary study as newly
introduced at the time of the original data collection, and 0 otherwise.
The Effect of Electronic Word of Mouth on Sales
Variable
Metrics
Volume
Cumulative volume
Incremental volume
Average volume
Valence
Average valence
Incremental valence
Positive valence
Negative valence
Composite Valence–Volume
Positive volume
Negative volume
Variance
Agreement and precision
Variability
Incremental variance
Study Characteristics
Endogeneity Controlsc
Simultaneous equations
First-difference model
Instrumental variables
Marketing Controls
Price control
Price promotion control
Other Methodological Controls
Year of data collection
Number of parameters
Lagged DV
Sales rank
Top-tier publication
Standard error
aRef= social media sites.
bRef= tangible good.
cRef= no endogeneity correction.
Notes: DV = dependent variable.
percentage change in X1 is often not comparable to a
percentage change in X2, which is an important limitation of
using elasticities in a broad meta-analysis. We argue that
when comparing metrics on the basis of very different
305
Table 3
CONTINUED
Description and Operationalization
Dummy variable indicating whether eWOM was operationalized as the total amount of eWOM available at a particular
time, including past periods and, in some cases, the current period (reference in the model).
Dummy variable indicating whether eWOM was operationalized as the difference in the total or average amount of eWOM
between two periods.
Dummy variable indicating whether eWOM was operationalized as average number of eWOM per product.
Dummy variable indicating whether eWOM was operationalized as an average aggregate measure.
Dummy variable indicating whether eWOM was operationalized as the difference in the average ratings between two
periods or between two online platforms.
Dummy variable indicating whether eWOM was operationalized as the (1) polarity (i.e., the ratio of positive to negative
eWOM), (2) subjectivity (i.e., the ratio of positive to neutral eWOM), or (3) the ratio of positive eWOM to total eWOM.
Dummy variable indicating whether eWOM was operationalized as the (1) polarity (i.e., the ratio of negative to positive
eWOM), (2) subjectivity (i.e., the ratio of negative to neutral eWOM), or (3) the ratio of negative eWOM to total eWOM.
Dummy variable indicating whether eWOM was operationalized as amount of positive eWOM.
Dummy variable indicating whether eWOM was operationalized as amount of negative eWOM.
Dummy variable indicating whether eWOM was operationalized as the inverse of the variance in numerical ratings (i.e.,
precision of the ratings).
Dummy variable indicating whether eWOM was operationalized as average variance, or standard deviation in numerical ratings.
Dummy variable indicating whether eWOM was operationalized as a difference in standard deviation of or average variance
of numerical ratings.
Dummy variable that indicates with 1 if endogeneity was accounted for by using a Granger causality test or simultaneous
equations model, and 0 otherwise.
Dummy variable that indicates with 1 if endogeneity was accounted for by using a first-difference model, and 0 otherwise.
Dummy variable that indicates with 1 if endogeneity was accounted for by using instrumental variables or a generalized-
method-of-moments approach, and 0 otherwise.
Dummy variable that assumes the value of 1 if product price was controlled for, and 0 otherwise.
Dummy variable that assumes the value of 1 if price promotion was controlled for, and 0 otherwise.
Year of the data collection. If the data collection spans over the course of several years, we consider the mean year
(continuous variable; mean-centered).
Number of variables in the response model (continuous variable; mean-centered).
Dummy variable indicating whether a lagged term of sales was included in the response model.
Dummy variable that assumes the value of 1 if sales was operationalized as sales rank, and 0 otherwise. Effect sizes signs
were inverted when needed. Effect sizes from studies that converted sales rank into sales using the formula by Brynjolfsson,
Hu, and Smith (2003) are coded as 0.
Dummy variable indicating whether the primary study has been published in a top-tier academic journal (Journal of
Marketing, Journal of Marketing Research, Marketing Science, International Journal of Research in Marketing,
Management Science, or Information Systems Research).
Standard errors of the Fisher-transformed effect sizes (continuous variable; mean-centered).
measurement scales, correlations enable a more informative
and objective comparison.
The second advantage stems from the observation
that elasticity cannot be computed for a large number of
306
studies in our sample, because necessary effect size statis-
tics (typically, the averages of the dependent and explan-
atory variables) were not reported in the primary study and
could not be obtained from the authors. As Peterson and
Brown (2005) note, the inclusion of effect sizes based on
partial correlations reduces both sampling errors because of
the increased number of effect sizes and nonsampling errors
because of the inclusion of a broader array of research
designs. Thus, by using correlations (i.e., a scale-free
measure that can be computed on the basis of a wide
range of statistics), we broaden the scope of the meta-
analysis to 1,532 effect sizes obtained from 96 studies
covering 40 platforms and 26 product categories.
The impact of eWOM on sales is captured by bivariate
Pearson product-moment correlations (r) and partial cor-
relation coefficient effect sizes (rxy,z; Lipsey and Wilson
2001; Rosenthal 1988),5 where 89% of the effects were
based on partial correlations. The partial correlations are
based primarily on regression-type models that assess sales
as the dependent variable, using a variety of explanatory
variables, including eWOM metrics.6
We transform all correlation coefficients into Fisher’s Z
effect sizes (zr) because they are easy-to-interpret scale-free
measures that have the desirable statistical properties of
being approximately normally distributed with a sample
variance that depends only on sample size and not on the
population correlation itself.7 Furthermore, because the
studies in our sample vary in the number of observations,
we weight each effect size by its inverse variance to give
more weight to more accurate measures (Lipsey and Wilson
2001; Shadish and Haddock 2009).8
HOMA. In line with meta-analytic standards, we first
summarize the overall effects of eWOM using a random-
effects HOMA for combining study estimates (Carney et al.
2011; Geyskens et al. 2009). To estimate mean effects, we
account for differences in the precision of the retrieved
effect sizes by using weights (w; Hedges and Olkin 1985).
We also use these weights to calculate the standard error
and confidence interval of the mean effect.
5We use the following formula to compute the partial correlation co-
efficient effect size (Rosenthal 1988, p. 25): rxy,z = t/√(t2 + d.f.), where t is
the t-value associated with the regression parameter that captures the effect of
eWOM on sales and d.f. are the degrees of freedom of the reported regression
model.
6Bivariate correlations and partial correlations are potentially different
because the latter are computed while controlling for other explanatory
variables. Therefore, as a robustness check, we include a dummy variable to
capture the mean difference between both types of correlations and include
several moderators indicating whether specific other explanatory variables
were controlled for when computing the partial correlation. These variables
were not significant, so we removed them from the model. The full set of
results is available on request.
7We transform average effect sizes (HOMA) and regression estimates
(HiLMA) back into a standard correlational form (r) for ease of interpretation as
well as to avoid overestimation of the population value of z (Silver and Dunlap
1987). We use the following formulae for Fisher’s Z: (1) transformation: zr =
1/2 log ([1 + r]/[1 – r]) (Rosenthal 1988, p. 27) and (2) back-transformation
to correlation units: r = (e2z – 1)/(e2z + 1) (Lipsey and Wilson 2001, p. 64).
8We calculate the weight, w, as follows: w = 1=ðse2 + b vq Þ, where se is the
i zr
standard error of the effect size, which we calculate as sezr = –1/√(n – 3), and
b
vu is the random-effects variance component. We calculate the meta-analytic
mean effect size as follows: zr =  (w × zr)/w, where its standard error
is sezr = √(1/w) and its 95% confidence interval is computed as: lower CI =
zr – 1.96 (sezr Þ and upper CI = zr + 1.96 ðsezr Þ.
JOURNAL OF MARKETING RESEARCH, JUNE 2016
HiLMA. Systematic attenuating statistical artifacts
other than sample size are corrected for during HiLMA
procedures (Lipsey and Wilson 2001). The HiLMA is pre-
ferable to more conventional subgroup moderator analyses
for its use of a multivariate, regression-based format
(Carney et al. 2011; Geyskens et al. 2009; Tellis 1988). This
procedure enables us to filter out the effects of important
moderators that were or were not part of the primary
studies. In our case, we collected platform and product data
from additional sources to explain heterogeneity across
effect sizes. In the HiLMA, we consider four compre-
hensive sets of moderating variables to explain the variation
in the correlation between eWOM and sales: (1) platform-
related factors, (2) product-related factors, (3) eWOM metrics,
and (4) characteristics of the studies in our sample. As a rule,
we include in our analysis only variables that are employed in
at least seven regression models.
The model contains ten variables representing plat-
form characteristics. These include three dummies for plat-
form type (for review platforms, e-commerce platforms, and
other, with social media platforms used as reference),
eWOM sender details (homophily and trustworthiness
vectors), eWOM message details (vector for time stamp and
helpfulness rating), eWOM message visibility, structured
display of eWOM, platform maturity, and the imposition of
eWOM posting costs.9 Five variables are related to pro-
duct characteristics (two dummies for services and digital
products, with tangible goods as the reference; a dummy for
hedonic products; a five-item Likert scale indicating the
level of financial risk; and a dummy for newly introduced
products). We use 11 dummies to capture eWOM sub-
metrics (average volume, incremental volume, average
valence, incremental valence, positive valence, negative
valence, positive volume, negative volume, agreement and
precision, variability, and incremental variance, with cu-
mulative volume as the reference). The model also includes
12 variables that account for the differences in methodo-
logical choices made in the primary studies. Table 3 de-
scribes the operationalization of all variables. To account for
the statistical dependencies among effect sizes based on the
same subject samples, we follow Bijmolt and Pieters (2001)
and estimate a hierarchical random-effects meta-analytic
model to control for within-study correlation.
RESULTS
HOMA Results
The 1,532 back-transformed Z effect sizes span across a
large range, from highly negative (–.69) to highly positive
(.98; M = .08, SD = .18). We observe 52.5% significant and
positive effects (807 effects), 11% significant and negative
effects (165 effects), and 36.5% nonsignificant effects (560
effects). These mixed results highlight the great variation
in the effects of eWOM on sales and call for a formal
analysis to assess the overall impact of eWOM (HOMA)
(see the Web Appendix and Table 4). From the random-
effects HOMA, we conclude that there is an overall sig-
nificant and positive relationship between eWOM and
9In addition, because more than 40% of observations in our sample are
from Amazon, we filter out potential platform-specific effects by including an
Amazon dummy variable in the moderator analysis (HiLMA).
The Effect of Electronic Word of Mouth on Sales 307

Table 4
HOMA RESULTS

Weighted Average
Variable k Number of Studies N + and Significant – and Significant Random-Effect r (SE) Q I2
Overall 1,532 96 2,391,602 807 165 .091 (.006)*** 666,138*** .998
Platforms
Social media platforms 275 24 151,385 172 22 .132 (.009)*** 16,351*** .983
Review platforms 237 29 175,852 132 31 .121 (.013)*** 29,235*** .992
E-commerce platforms 1,001 55 2,176,362 493 110 .071 (.008)*** 618,496*** .998
Products
Tangible good 1,027 53 2,023,557 517 112 .070 (.008)*** 608,260*** .998
Service 368 33 271,479 195 47 .146 (.011)*** 49,488*** .993
Digital product 109 14 59,040 73 2 .108 (.012)*** 3,440*** .969
Utilitarian product 939 52 1,918,803 453 105 .064 (.008)*** 589,672*** .998
Hedonic product 593 52 483,964 354 60 .136 (.008)*** 76,162*** .992
High financial risk 320 36 633,557 226 28 .149 (.022)*** 427,431*** .999
Low financial risk 1,212 71 1,760,050 581 137 .074 (.004)*** 185,834*** .993
Mature product 966 57 1,541,559 509 95 .071 (.005)*** 180,193*** .995
New product 566 41 850,043 298 70 .127 (.015)*** 453,680*** .999
Metrics
eWOM Volume 589 84 2,277,093 399 52 .141 (.014)*** 527,364*** .999
Average 7 3 117,734 6 0 .360 (.074)*** 5,536*** .999
Incremental 144 14 106,831 48 16 .059 (.016)*** 8,764*** .984
Cumulative 438 70 2,053,045 345 36 .161 (.017)*** 509,939*** .999
eWOM Valence 596 78 2,264,176 248 63 .049 (.003)*** 30,702*** .981
Average rating 312 62 1,583,339 189 30 .075 (.005)*** 23,997*** .987
Incremental rating 139 7 539,918 22 5 .021 (.006)*** 461*** .700
Positive valence 91 23 698,721 30 4 .036 (.006)*** 1,823*** .951
Negative valence 54 16 662,214 7 24 −.013 (.005)* 780*** .932
eWOM Composite Valence–Volume 108 13 56,620 41 19 .061 (.020)** 4,612*** .977
Positive volume 66 13 56,620 37 2 .140 (.026)*** 3,999*** .984
Negative volume 42 9 10,865 4 17 −.064 (.019)*** 227*** .819
eWOM Variance 137 17 944,404 59 19 .061 (.009)*** 11,068*** .988
Agreement and precision 13 2 5,908 1 8 −.023 (.073) 741*** .984
Variability 42 13 937,484 30 5 .117 (.016)*** 9,935*** .996
Incremental 82 2 1,012 28 6 .041 (.010)*** 251*** .677
Other eWOM Measures 102 15 150,000 60 12 .102 (.014)*** 6,753*** .985
Existence 27 8 76,078 18 5 .105 (.025)*** 1,807*** .986
Content 75 8 87,684 42 7 .101 (.017)*** 4,737*** .984
*p < .05.
**p < .01.
***p < .001.
Notes: The table reports back-transformed Fisher’s Z correlations; positive and negative valence measures can be both average (e.g., five-star or one-star rating)
and incremental (e.g., the change of the percentage of positive messages). k = number of effect sizes; N = total sample size; +/– and significant = number of positive/
negative and significant effect sizes; Q = Cochran’s homogeneity test statistic; I2 = scale-free index of heterogeneity. The significance level of effect sizes is based
on t-values (for partial correlations) and p-values (for bivariate correlations).

sales (ro = .091, p < .001), which is nonnegligible (Aloe
and Thompson 2013). This result is consistently positive
across the different platforms and products.
Overall, we find more positive and statistically significant
effect sizes for volume than for valence (26% vs. 16%), as well
as higher weighted average random-effects correlations (for the
overall measures: ro = .141 vs. .049; for the different submetrics
[e.g., average volume vs. average rating]: ro = .360 vs. .075).
Although a formal model is required to filter out potential
confounding effects and assess whether the effects across dif-
ferent metrics are significantly different, the HOMA results
provide first evidence that the volume of consumer-generated
content (ro = .141, p < .001) is more strongly related to sales than
all other eWOM operationalizations: valence (ro = .049, p <
.001), composite valence–volume (ro = .061, p < .01), variance
(ro = .061, p < .001), and other eWOM measures (ro = .102,
p < .001). In addition, the HOMA results provide evidence that
negative eWOM is not critical, given the low correlations with
sales for negative valence (ro = –.013, p < .05). For negative
volume (ro = –.064, p < .001), however, the negative effects
are more pronounced, underscoring the importance of
differentiating between relative and absolute measures of
eWOM.
Furthermore, the significant Cochran’s Q-test of homoge-
neity and the high scale-free index of homogeneity I2 confirm a
substantial amount of heterogeneity, implying that the vari-
ability of the effect sizes is larger than would be expected from
subject-level sampling error alone. Overall, the results of the
HOMA show that eWOM significantly affects sales, but the
direction, size, and statistical significance of the average effects
differ between and within the main eWOM metrics, calling for a
moderator analysis.
308
HiLMA Results
In this section, we show how the link between eWOM
and sales differs across platforms, products, eWOM met-
rics, and other study characteristics. Because we find
significant differences across platform and product char-
acteristics in the overall-sample analysis, we conduct
additional moderator analyses by platform and product
characteristics (moderated-moderation or three-way in-
teraction) to further investigate possible interaction effects.
Because adding interaction terms to the model leads to
high levels of multicollinearity, we conduct a series of
split-sample analyses by running the model separately for
(1) social media platforms, (2) review platforms, (3)
e-commerce platforms, (4) tangible goods, (5) services,
(6) hedonic products, (7) utilitarian products, (8) new
products, (9) mature products, and, finally, products with
(10) high and (11) low financial risk. We report the results
in Table 5 and summarize the key insights in Table 6.
We fill the gaps in the literature and address inconsis-
tencies by conducting a moderator analysis on 1,430 effects
across the different platforms, products, and the four key
eWOM metrics (93% of our total sample).10 The model fit is
satisfactory (pseudo-R2 = .26) and in line with prior meta-
analyses (e.g., Bijmolt, Van Heerde, and Pieters 2005).
Overall, multicollinearity does not severely affect the model.
The highest reported variance inflation factor (VIF) is 8.37
for the year of data collection (average VIF = 3.07, median
VIF = 2.37), and the results remain unaltered when re-
moving this control variable. Moreover, the analysis of the
correlation matrix (see the Web Appendix) indicates that
the highest correlation is –.73 between e-commerce plat-
forms and services. We detail our robustness checks, as
well as our approaches to dealing with publication bias, in
the Web Appendix. These checks confirm the stability of
our results. Table 5 shows the back-transformed estimates
(b) of the HiLMA, as well as the resulting back-transformed
predicted correlations (r) computed by setting all other
variables at their sample means and, in the case of con-
tinuous variables, their upper and lower quartiles. In the
next section, we discuss the key findings from Table 5 row
by row.
Platform characteristics amplifying the impact of eWOM
on sales: e-commerce platforms. Overall, Table 5 shows that
the impact of eWOM on sales is stronger for e-commerce
platforms (bo = .100; p < .05; ro = .052) than social media
platforms, while eWOM effectiveness on review platforms
does not significantly differ from that on social media or
e-commerce platforms.11 We argue that this result can be
explained by the nature of these different platforms and
10We exclude effect sizes of eWOM content and existence on sales from
the HiLMA model because there are not enough observations for these
submetrics and because the results are too specific (e.g., the frequency of the
word “advertising” in Japanese blogs) for generalization.
11This result is not driven by the platform-specific sales measures. Among
effect sizes collected on e-commerce platforms, 13% are based on gross
wholesale sales, whereas 18% of the effect sizes collected on social media
platforms are based on site-specific sales. Notably, there is no statistical
difference between the effect of eWOM on site-specific sales and gross
wholesale sales among the e-commerce platforms (t-value = .42, p = .68) or
the social media platforms (t-value = .85, p = .40), suggesting that the
dominance of eWOM on e-commerce platforms is not driven by the criterion
variable used in primary studies.
JOURNAL OF MARKETING RESEARCH, JUNE 2016
how they are commonly used by consumers. E-commerce
and review platforms are primarily designed to support
consumers’ decision journeys, whereas social media sites
help maximize social exchanges (Schweidel and Moe
2014). Moreover, for new products (bn = .263; p < .0001;
rn = .253) and, marginally, for utilitarian products (bu =
.105; p < .10; ru = .081) eWOM displayed on e-commerce
platforms is more effective than on social media platforms.
Platforms with more details that enable consumers to
better assess their similarity to the eWOM sender exhibit a
higher link to sales (bo = .048; p < .05; rolow = .002;
rohigh = .050), especially when eWOM appears on social
media platforms (bsm = .061; p < .0001; rsm low = .138;
rsm
high = .166). In addition, eWOM displayed on platforms
that offer more homophily details is particularly impactful
on the sales of hedonic products (bh = .050; p < .10;
rhlow = .082; rhhigh = .132), new products (bn = .124; p < .01;
rnlow = .027; rnhigh = .171), and both high and low financial
risk products (bhf = .120; p < .10; rhf hf
low = .106; r high =
.168; b = .033; p < .05; rlow = .066; rhigh = .099). Thus, for
lf lf lf
hedonic products and newly introduced products, homo-
phily information reduces uncertainty about functional
product performance. These results also suggest that
homophily details will amplify the effect of eWOM at any
price level. eWOM sender trustworthiness details amplify
eWOM effectiveness only for review platforms (br = .193;
p < .10; rrlow = .027; rrhigh = .219). Overall, these findings
reveal a stronger weight of homophily details than that of
trustworthiness details influencing the effectiveness of
eWOM.
Across the board, eWOM message details (time stamp
and helpfulness rating) do not significantly moderate the link
between eWOM and sales. Instead, eWOM visibility is of
general importance for the entire sample (bo = –.056; p < .01;
rolow = .062; rohigh = .005) and, in particular, for e-commerce
platforms (be = –.054; p < .05; relow = .104; rehigh = –.004),
utilitarian products (bu = –.081; p < .01; rulow = .120;
ruhigh = –.042), mature products (bm = –.045; p < .05;
rm m
low = .102; rhigh = .013), and high-financial-risk products
(b = –.256; p < .0001; rhf
hf hf
low = .276; rhigh = .108).
Overall, the structured display of eWOM information is
linked to lower sales (bo = –.077; p < .01; ro = –.044).
This finding also emerges in the split-sample analyses for
e-commerce platforms (be = –.093; p < .01; re = –.016),
tangible products (bt = –.090; p < .01; rt = –.079), utilitarian
products (bu = –.099; p < .01; ru = –.024), and new products
(bn = –.085; p < .05; rn = .055). In general (across all of
Table 5), platform maturity and eWOM posting costs are
not significant, though eWOM posting costs moderate the
impact of eWOM on sales for products with higher financial
risk (bhf = .216; p < .05; rhf = .166).
Product characteristics amplifying the impact of eWOM
on sales. Overall, we find no significant differences in the
effectiveness of eWOM across tangible goods, services, and
digital products, or between hedonic and utilitarian products.
This is a notable finding that underscores the importance of
measuring and managing eWOM for a broad range of prod-
ucts. The only exception is for services: eWOM in social media
is particularly impactful on the sales of services (bsm = .255;
p < .01; rsm = .287).
 Table 5
HiLMA RESULTS

Overall and by Platforms and Productsa

Overall Social Media Platforms Review Platforms e-Commerce Platforms Tangible Goods Services
sm sm r r e t t s
bo ro b r b r b re b r b rs

Intercept .176** .023 .118 .164 .225* .146 .295*** .067 .185* .068 .074 .164
Platforms
Social media platforms (ref) −.049 −.020 .156
Review platformsc .043 −.006 .032 .012 .023 .179
E-commerce platformsc .100* .052 .103 .083 −.123 .034
Other platformsc −.124 −.173 −.160 −.180
eWOM sender homophily detailsb .048* .002 .050 .061*** .138 .166 −.076 .178 .128 .045 .070 .094 .002 .162 .164
eWOM sender trustworthiness detailsb −.001 .023 .022 .193† .027 .219 −.011 .068 .062 −.039 .073 .053 .113 .096 .165
eWOM message time stamp and helpfulness ratingb −.018 .024 .016 .061 .106 .167 .038 .070 .080 −.006 .068 .066 −.020 .177 .158
eWOM visibility (scrolls)b −.056** .062 .005 −.050 .180 .147 −.054* .104 −.004 −.045 .071 .036 −.076 .215 .166
eWOM unstructured display (ref) .034 .174 .077 .099 .010 .175
eWOM structured displayd −.077** −.044 −.087 .088 −.093** −.016 −.090** .079 −.053 .124
Platform maturityb −.001 .024 .020 −.009 −.011
No eWOM posting costs (ref) .038 .084 .097
eWOM posting costse −.026 .012 −.035 .049 −.064 .034
The Effect of Electronic Word of Mouth on Sales

Products
Tangible good (ref) .022 .035 .061
Servicef −.012 .010 .255** .287 −.030 .031
Digital productf .053 .075 .003 .038 .084 .144
Utilitarian (ref) .024 .065 .063 .064
Hedonicg −.004 .020 .013 .077 .027 .090 .106 .169
Financial riskb .054* .012 .119 −.018 .167 .132 −.068 .159 .092 .058† .055 .112 .047 .059 .151
Mature product (ref) −.003 .194 .049 .042 .135
New producth .069* .067 −.043 .152 .098* .147 .136* .177 .034 .169
eWOM Metrics
Cumulative volume (reference) .114 .214 .167 .173 .162 .219
Average volumei −.031 .083
Incremental volumei −.036 .079 −.107** .109 .161 .319 −.027 .147 −.019 .143 −.001 .218
Average valencei −.220*** −.109 −.059** .157 −.069* .099 −.248*** −.077 −.242*** −.083 −.095*** .127
Incremental valencei −.078** .036 −.078** .096 −.075** .088
Positive valencei −.148*** −.034 −.070*** .145 −.034 .134 −.158*** .016 −.152*** .010 −.022 .198
Negative valencei −.181*** −.069 −.076 .092 −.188*** −.015 −.185*** −.023 −.111* .111
Positive volumei −.007 .107 −.048 .167 −.119 .055 −.056 .108 −.138 .083
Negative volumei −.153 −.040 −.297** −.089 −.120 .055 −.128 .035 −.336† −.126
Agreementi −.192 −.080
Variabilityi −.271*** −.163 −.292*** −.125 −.289*** −.133
Incremental variancei −.008 .107 −.007 .166 .002 .164
Study Characteristics
Not Amazon (ref) .015 .048
Amazonj .018 .032 .033 .081
Year of data collectionb .001 .021 .025 .025 .050 .124
No price control (ref) .118 .122 .106 .113 .157
Price controlk −.162*** −.046 .257** .367 −.049 .057 −.058† .055 .180*** .328
No promotion control (ref) .024 .076 .077
Promotion controll −.040† −.016 −.209*** −.135 −.211*** −.136
No lagged DV (ref) .029 .076 .078 .151
†
Lagged DVm −.067 −.038 −.100* −.025 −.101* −.023 .225 .364
Sales DV (ref) .049 .093 .097
309

Sales rankn −.044 .005 −.031 .062 −.036 .061

 Table 5
310

CONTINUED

Overall and by Platforms and Productsa

Overall Social Media Platforms Review Platforms e-Commerce Platforms Tangible Goods Services

bo ro bsm rsm br rr be re bt rt bs rs

No endogeneity control (ref) .052 .180 .177 .117 .109 .197

Simultaneous equation modelo −.003 .049 −.321*** −.150 −.007 .169 .029 .145 .025 .134 −.124† .075
First-difference modelo −.118*** −.066 −.130*** −.014 −.118*** −.009
IV or GMMo .015 .067 −.002 .178 −.180* −.003 .018 .135 .044 .153 −.199*** −.001
Number of parametersb −.005** .043 .008 −.012*** .212 .128 −.015*** .235 .132 .065 .068 1.4e-04 .069 .068 −.017*** .261 .116
Not a top-tier publication (ref) .005 .148 .049 .051 .179
†
Top-tier publicationp .072 .077 .085 .230 .076 .125 .066 .116 −.060 .120
Standard errorb .512 .007 .029 .861 .129 .232 .886 .115 .158 −.117 .070 .065 −.010 .069 .068 .398 .156 .187
k (N) 1,430 (96) 220 (24) 225 (26) 940 (55) 961 (53) 338 (32)
−2 residual log-likelihood 2,244.5 −99.6 233.8 1,602.5 589.4 339.1

By Productsa
Hedonic Products Utilitarian Products New Products Mature Products High Financial Risk Low Financial Risk
h h u u n n m m hf hf
b r b r b r b r b r blf rlf

Intercept .078 .126 .250* .064 −.063 .122 .129† .071 .131 .101 .181*** .080
Platforms
Social media platforms (ref) −.025 −.011 .071 .070
Review platformsc .020 −.005 .135* .124 .030 .101 .030 .099
e-Commerce platformsc .105† .081 .263*** .253 .091 .161 .009 .078
Other platformsc −.178 −.203 .020 .091
eWOM sender homophily detailsb .049† .089 .137 .024 .065 .078 .124** .027 .171 .036 .065 .091 .120† .106 .168 .033* .066 .099
eWOM sender trustworthiness detailsb .023 .112 .131 .002 .064 .065 −.074 .167 .093 .071* .063 .081 −.126† .136 .094 .025 .073 .090
eWOM message time stamp and helpfulness ratingb −.055 .161 .107 .015 .064 .070 −.049 .153 .105 −.034 .071 .059 −.051† .080 .062
eWOM visibility (scrolls)b −.001 .126 .126 −.081** .120 −.042 −.046 .153 .123 −.045* .102 .013 −.256*** .276 .108 −.012 .089 .077
eWOM unstructured display (ref) .119 .075 .139 .070 .079
eWOM structured displayd .036 .155 −.099** −.024 −.085* .055 .022 .092 −.091 .039 .012 .091
Platform maturityb .001 .119 .127 .004 .063 .078 −.005 .147 .114 −2e-05 .080 .080
No eWOM posting costs (ref) .084 −.036 .090 −.052 .096
eWOM posting costse −.046 .038 .205 .170 −.028 .062 .216* .166 −.028 .068
Products
Tangible good (ref) .090 .055 .119 .073 .080
Servicef .047 .144 .129 .182
Digital productf −.024 .074 .073 .128 .046 .164 −.013 .059 .006 .085
Utilitarian (ref) .019 .095 .082
Hedonicg .148† .167 −.034 .062 −.004 .078
Financial riskb −.003 .127 .124 .070 .051 .188 .091† .104 .280 −.010 .073 .054
Mature product (ref) .082 .073
New producth .067 .148 .098 .170
eWOM Metrics
Cumulative volume (reference) .146 .169 .311 .088 .303 .125
Average volumei
Incremental volumei .114** .256 −.077 .094 −.155 .164 .031 .119 .013 .315 .004 .129
Average valencei −.098*** .049 −.248*** −.082 −.463*** −.177 −.051*** .038 −.447*** −.167 −.088*** .037
Incremental valencei −.082** .088 −.066*** .023 −.071*** .054
Positive valencei −.010 .136 −.157 .013 −.107† .212 −.139*** −.052 −.298* .005 −.136*** −.012
Negative valencei −.076* .071 −.191*** −.022 −.203* .116 −.168*** −.081 −.282* .023 −.170*** −.046
Positive volumei .144** .284 −.133 .037 −.147 .173 .066 .153 −.086 .223 .026 .150
JOURNAL OF MARKETING RESEARCH, JUNE 2016
 Table 5
CONTINUED
By Productsa
Hedonic Products Utilitarian Products New Products Mature Products High Financial Risk Low Financial Risk

bh rh bu ru bn rn bm rm bhf rhf blf rlf

Negative volumei −.057 .090 −.167 .002 −.240 .077 −.060 .028 −.147 .163 −.104 .021
Agreementi −.415* −.119 −.094 .031
Variabilityi −.076** .070 −.315*** −.154 −.564*** −.306 −.024 .064 −.533*** −.275 −.051** .075
Incremental variancei −.035 .135 .001 .089 −.003 .121
Study Characteristics
Not Amazon (ref) .071 .097
Amazonj −.007 .065 .011 .108
Year of data collectionb .003 .066 .077
No price control (ref) .100 .107 .076 .096
Price controlk .149*** .246 −.052 .056 −.022 .054 −.028 .069
No promotion control (ref) .126 .072 .139 .071 .115 .081
Promotion controll .014 .139 −.214*** −.144 −.473*** −.358 −.013 .059 −.166 −.052 −.057* .025
No lagged DV (ref) .148 .070 .111 .096 .127 .082
Lagged DVm −.174*** −.027 −.069 3.4E-04 .165 .272 −.134*** −.039 −.137 −.010 −.026 .056
Sales DV (ref) .092
Sales rankn −.034 .058
The Effect of Electronic Word of Mouth on Sales

No endogeneity control (ref) .144 .119 .146 .070 .152 .120

Simultaneous equation modelo −.094 .051 .021 .140 −.106 .041 −.032 .038 −.040 .081
First-difference modelo −.150 −.006 −.135*** −.015 −.276 −.135 −.139*** −.070 −.186 −.035 −.123*** −.003
IV or GMMo −.059† .086 −.017 .102 −.119 .028 .029 .099 −.085 .067 −.053† .068
Number of parametersb −.013*** .188 .090 .003 .055 .076 −.018*** .212 .105 −.002 .078 .062 .001 .096 .107 −.005*** .102 .064
Not a top-tier publication (ref) .119 .059 .104 .066 .074
Top-tier publicationp .027 .145 .021 .080 .097 .199 .025 .091 .033 .107
Standard errorb .002 .126 .126 .206 .058 .066 −.376 .132 .100 .849 .036 .084 −.990 .184 .046 .513 .067 .086
k (N) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1,161 (71)
−2 residual log-likelihood 399.0 1,493.8 1,077.0 593.2 1,654.1 339.4

†p < .10.
*p < .05.
**p < .01.
***p < .001.
aBack-transformed unstandardized Fisher-Z transformed regression parameters are presented; interpret as correlations of eWOM with sales relative to the reference category. We removed from the analysis (1)
variables with less than seven observations and (2) highly collinear variables. Results on all the variables (together with their VIF values) are available on request.
bAll continuous variables are mean-centered.
cRef = social media platforms.
dRef = eWOM unstructured display.
eRef = eWOM posting costs.
fRef = tangible good.
gRef = utilitarian good.
hRef = mature product.
iRef = cumulative volume.
jRef = not Amazon.
kRef = no price control.
lRef = no promotion control.
mRef = no lagged DV.
nRef = sales DV.
oRef = no endogeneity control.
pRef = not a top-tier publication.
Notes: Two-sided tests of significance. k = number of effect sizes; N = number of studies; DV = dependent variable; IV = independent variable; GMM = generalized method of moments. The number of effect sizes and
311

number of studies per variable are available in the Web Appendix.

312
In general, eWOM has a stronger link to the sales of new
products than that of mature products (bo = .069; p < .05;
ro = .067), further demonstrating the importance of moni-
toring (and potentially stimulating) eWOM in the early stages
of the product life cycle. The relevance of eWOM for newer
products is of particular importance on e-commerce platforms
(be = .098; p < .05; re = .147) and for tangible products (bt =
.136; p = .01; rt = .177). Managers should pay particular at-
tention to eWOM about products with higher financial risk
(bo = .054; p < .05; rolow = .012; rohigh = .119), especially on
e-commerce platforms (be = .058; p < .10; relow = .055;
rehigh = .112) and for new products (bn = .091; p < .10;
rnlow = .104; rnhigh = .280), though these effects are marginally
significant.
eWOM metric amplifying effectiveness on sales: volume.
The predicted correlations confirm a positive impact
of the volume of eWOM on sales (rocum:volume = .114;
roavg:volume = .083; roincr:volume = .079), providing supporting
evidence for the bandwagon effect. A formal test of the
overall volume and overall valence metrics lends further
support to the conclusion that eWOM volume has a stronger
impact on sales than eWOM valence (t-value = 5.59, p < .001).
Our results overturn the finding of Floyd et al. (2014) and You,
Vadakkepatt, and Joshi (2015) that volume is less effective
than valence. We explain this difference in results by a
combination of our conceptual and methodological choices.
First, we offer a conceptualization of eWOM metrics that
disentangles the often mixed-up effects of “valence only” and
“valence plus volume,” which are commonly labeled together
as “valence.” To empirically verify whether the difference is
driven primarily by the more careful operationalization of
the eWOM metrics, we incorporated our composite valence–
volume metric into valence, as is more commonly done in the
literature, and tested for differences between this confounded
valence metric and volume. In this case, we no longer find a
significant difference (t-value = –.04, p = .964, k = 1,430).
This is an important finding because it illustrates that an
incorrect classification of valence not only is conceptually
wrong but also drastically changes empirical results and
managerial recommendations.
The second difference between our results and those of the
prior meta-analyses pertains to the use of different effect sizes
(partial correlations vs. elasticities). In contrast with elastici-
ties, our effect sizes are fully independent of measurement
scale and more comparable across different metrics. Using
effect sizes based on elasticities, we tested whether the stronger
effect of valence over volume, as reported by Floyd et al.
(2014) and You, Vadakkepatt, and Joshi (2015), would hold
when a nonconfounded metric of valence is used. We find
that using a confounded measure of valence leads to a higher
effect of valence over volume, whereas with a nonconfounded
valence metric, the difference between volume and valence
disappears (t-value = –1.44, p = .152, k = 697), again high-
lighting the importance of separating our composite valence–
volume metric from valence.
We find that cumulative volume is the most impactful
metric for review platforms (rr = .167; p < .0001), utili-
tarian products (ru = .169; p < .0001), new products (rn =
.311; p < .0001), and products with high financial risk (rhf =
.303; p < .01). In many instances, cumulative volume exerts
an impact on sales similar to that of incremental volume
(p > .10): for e-commerce platforms (recum:volume = .173;
JOURNAL OF MARKETING RESEARCH, JUNE 2016
reincr:volume = .147), tangible products (rtcum:volume = .162;
rtincr:volume = .143), services (rscum:volume = .219; rsincr:volume = .218),
mature products (rm m
cum:volume = .088; rincr:volume = .119), and pro-
ducts with low financial risk (rcum:volume = .125; rlfincr:volume =
lf
.129).
Moreover, we identify platforms and products for which
the effects of eWOM volume submetrics are not significantly
different from the effects of composite valence–volume
submetrics (p > .10): for social media (rsm cum:volume = .214;
rsm
pos:volume = .167), hedonic products (rhincr:volume =
.256; rhpos:volume = .284), and mature products
(rm m
cum:volume = .088; r pos:volume = .153). This means that for
these platforms and products, it is not only the mere volume
of eWOM that counts most but also the combination of the
bandwagon and persuasion dynamics represented by the
amount of positive eWOM. In only one case (services) is
positive valence as effective (p > .10) as cumulative vol-
ume and incremental volume (rscum:volume = .219;
rsincr:volume = .218; rspos:valence = .198).
Furthermore, we find that not all positive eWOM metrics
are linked to an increase in sales (ropos:valence = –.034, p > .10;
ropos:volume = .107, p < .01). Although previous studies have
combined positive volume and positive valence into one
metric, we find evidence that their effects are not identical and
that positive volume is more strongly correlated with sales (the
only exception identified in our sample is for services). The
composite metric is an absolute number that effectively
summarizes volume and valence at once by representing how
many consumers expressed a positive (or negative) opinion
about the product, whereas the valence metric expresses only
consumers’ relative sentiment about the product. Together
with cumulative volume and incremental volume, positive
volume is the most effective metric of eWOM. Table 6 pro-
vides an overview of the most important metrics per platform
type and product characteristic.
eWOM metric attenuating effectiveness on sales:
variability. The effects of negative eWOM are small and not
significantly different from zero, which suggests that,
on average, negative eWOM does not jeopardize sales
(roneg:valence = –.069, p = .938; roneg:volume = –.040, p = .789).
Negative valence hurts sales only in the later stages of the
product life cycle (rm = –.081, p < .05) and for low-
financial-risk products (rlf = –.046, p < .05). Our overall
finding contrasts with You, Vadakkepatt, and Joshi’s
(2015) results. We argue that this may be due to the different
samples in terms of products covered. Moreover, we are the
first to demonstrate that large heterogeneity among consumers’
product evaluations attenuates the effectiveness of eWOM, as
shown by the negative correlation between variability and
sales (ro = –.163, p < .01). When a product’s evaluation is
polarized, risk and uncertainty increase, thus leading con-
sumers to avoid the product. This finding emerges also from
the split-sample analyses for e-commerce platforms (re =
–.125, p < .05), tangible goods (rt = –.133, p < .01), utili-
tarian products (ru = –.154, p < .001), and products with high
financial risk (rhf = –.275, p < .05). This finding is in line with
the cue diagnosticity theory, which suggests that consumers
rely less on eWOM when the variance of ratings is large
because they may find the information nondiagnostic
(Feldman and Lynch 1988; Li 2015).
Research methodology findings. Though not the main
focus of our study, an important part of our meta-analysis
The Effect of Electronic Word of Mouth on Sales 313

Table 6
IMPLICATIONS

A: For Platform Managers

Implications for Social Media Implications for Review Implications for E-Commerce
Platform Managers Platform Managers Platform Managers
eWOM is more effective for . . . • Social media platforms with more • E-commerce platforms with greater eWOM visibility
sender homophily details • E-commerce platforms with less
structured display of eWOM
• Services • New products
• Cumulative volume, positive volumea • Cumulative volume • Cumulative volume, incremental volumea
• High eWOM variability harms sales
B: For Product Managers
Implications for
Managers Based on
Implications for Implications for Implications for Financial Risk
Managers Based on Managers Based Managers Based on
High
Product Type on Hedonic Score Product Life Cycle Stage
Financial Low Financial
Tangible Service Hedonic Utilitarian New Mature Risk Risk
eWOM is • Platforms • Platforms • E-commerce • Platforms with • Platforms • Platforms with
more with less with higher platforms, more sender with higher more sender
effective structured eWOM review trustworthiness eWOM homophily
for . . . display of visibility platforms details visibility details
eWOM • Platforms • Platforms • Platforms with • Platforms
with less with more higher eWOM with higher
structured sender visibility eWOM
display of homophily posting
eWOM details costs
• Platforms
with less
structured
display of
eWOM
• New products
• Cumulative • Cumulative • Positive • Cumulative • Cumulative • Cumulative • Cumulative • Incremental
volume, volume, volume, volume volume volume, volume volume,
incremental incremental incremental • High • High incremental • High positive
volumea volume, volumea variability variability volume, variability volume,
• High positive harms sales harms sales positive harms sales incremental
variability valencea volumea variancea
harms sales • Negative • Negative
valence harms valence harms
sales sales
aWhen multiple metrics are listed, they are not significantly different from each other.
Notes: This table is based on the HiLMA results on the split samples displayed in Table 5. Results displayed here are significant at p < .05 (two-sided). Cells are
empty if no significant differences are found between platforms, products, or metrics.

involves investigating the moderating impact of research
methodological choices on the relationship between eWOM
and sales. We base our conclusions on the “Overall” column in
Table 5 because the split-sample analyses do not allow us to
draw generalizable conclusions about these methodological
controls (Carney et al. 2011). We made several noteworthy
discoveries. We find that failing to control for the effects of
promotions leads to an overestimation of eWOM effectiveness
(bo = –.040; p < .10; ro = –.016). Not including the lagged term
of the dependent variable in the response model (bo = –.067;
p < .10; ro = –.038) also leads to lower estimations of the
effects of eWOM on sales. We also observe that accounting for
endogeneity using first-difference models leads to lower es-
timates of the impact of eWOM (bo = –.118; p < .001;
ro = –.066), but using other types of endogeneity controls
(e.g., simultaneous equations, instrumental variables) does not
dampen the effectiveness of eWOM. Controlling for endo-
geneity varies across research streams, such that studies in
marketing tend to correct for endogeneity more (66% of the
marketing studies vs. 60% of economics and 57% of in-
formation technology studies). Whereas the most frequent
endogeneity correction method in marketing and information
technology is a first-difference approach, in economics it tends
to be the use of instrumental variables. In a separate model,
we explore interaction effects between these three research
streams and endogeneity by using only one dummy variable
for endogeneity controls to avoid multicollinearity. We find
that, compared with studies from the marketing literature,
the impact of eWOM is lower for economics studies (b =
–.141, p < .01) and not significantly different for in-
formation technology studies (b = –.052, p > .10). After
endogeneity has been accounted for, the effect of eWOM is
314
higher for economics studies (b = .349, p < .01) and similar
for both marketing and information technology studies (b =
.072, p > .10) (see the Web Appendix). We observe that
studies published in top-tier journals record a greater effect
of eWOM on sales overall (bo = .072; p < .10; ro = .080).
Finally, we do not find significant differences for studies
with higher precision of the effect sizes. This last result
demonstrates the absence of publication bias (Stanley and
Doucouliagos 2012).
DISCUSSION
In the last 15 years, many studies have advanced the
understanding of the impact of eWOM. Overall, research has
demonstrated that consumers use eWOM because it reduces
their uncertainty and helps them choose the best offering,
which affects the bottom line. However, prior studies have
mostly relied on a single sample and thus have not been able to
investigate platform- and product-related factors that moderate
the effectiveness of eWOM. In addition, researchers have
disagreed on which metric among the multiple eWOM metrics
best captures this effect on sales. Although studies have
attempted to synthesize earlier work (Floyd et al. 2014; You,
Vadakkepatt, and Joshi 2015), the current study is the first
to systematically examine the overall effect of consumer-
generated information on sales across a large body of litera-
ture (96 studies)—covering 40 platforms, 26 product categories,
and 11 countries, and spanning 15 years (1999–2013)—and to
detail the differential effects of numerous operationalizations of
eWOM (12 submetrics), all while considering various meth-
odological designs. Our unique primary data collection through
the Wayback Machine also enabled us to capture variation
across platforms and products over time. Overall, we find a
positive correlation of .091 between eWOM and sales. This
finding implies that marketers should actively monitor eWOM,
and it justifies the allocation of resources to eWOM manage-
ment. Furthermore, this effect has not changed systematically in
the last 15 years, which suggests that marketers should include
eWOM in their long-term strategic decisions.
We set out to address two debates related to the contextual
factors influencing the effectiveness of eWOM as well as to
provide recommendations on methodological choices for
further research. First, for which platform and product char-
acteristics is eWOM more effective? In answering this
question, we identify the characteristics of the platforms and
their influence on eWOM effectiveness, thereby extending
prior experimental findings (e.g., Berger and Iyengar 2013)
and empirical work (e.g., Schweidel and Moe 2014) that
highlights the role of channel characteristics in WOM com-
munication. We find that the effectiveness of eWOM is not
necessarily “symmetrical”: that is, a product’s eWOM may be
more effective on a given platform, but for that very same
platform, it may be that eWOM about other products leads to
higher sales. This underscores the importance of specifying the
perspective taken in academic studies (i.e., that of a platform
manager [Table 6, Panel A] or that of a product manager
[Table 6, Panel B]). Platform managers can influence the
effectiveness of eWOM by accounting for the following two
platform characteristics:
• To increase the effectiveness of eWOM spread on social media
platforms, managers should encourage consumers to provide
more information about themselves so that eWOM receivers
JOURNAL OF MARKETING RESEARCH, JUNE 2016
can gauge homophily. This result is in line with a long line of
research on tie strength and WOM persuasiveness in personal
networks (e.g., Brown and Reingen 1987).
• For eWOM spread on e-commerce platforms, platform man-
agers’ efforts might focus on bringing eWOM information to the
forefront without overstructuring it. Given an abundance of other
product details available on this type of platform, it is crucial for
eWOM to be prominent to have a strong impact on sales.
From the perspective of product managers who want to in-
crease eWOM-driven sales, it is important to assess the most
salient characteristics of their products (i.e., tangibility, he-
donic score, stage in the product life cycle, and level of fi-
nancial risk). With that in mind, Table 6 (Panel B) offers
specific recommendations:
• For managers of tangible goods, platforms that display eWOM
information in a less structured way host the most influential
eWOM. In addition, eWOM about tangible goods is more
effective in the early stages of the product life cycle when
uncertainty is high and eWOM can be used to reduce functional
risk.
• Managers of utilitarian products should keep in mind that
eWOM is more effective on platforms where it is immediately
visible and less structured.
• For new products, eWOM increases sales when it appears on e-
commerce platforms and review platforms as well as on platforms
with less structured display of eWOM, because these platform
characteristics have been found to amplify the effectiveness of
eWOM. Moreover, early in the product life cycle, consumers may
be more concerned with whether they have something in com-
mon with the eWOM sender, potentially to reduce the risk of
purchasing a product that does not fit their needs.
• For mature products, eWOM is more effective when appearing
on platforms with greater eWOM visibility. In these later stages
of the product life cycle, it becomes more important to assess
whether the eWOM sender can be trusted. Because the product
has been around for a while, there may be less uncertainty about
its performance. However, uncertainty about the honest in-
tentions of eWOM senders may increase over time because of
the practices of review manipulation or fake-review spreading
(Anderson and Simester 2014).
• For products with higher financial risk, eWOM has a stronger
impact on the bottom line on platforms that display it more
prominently or impose higher posting costs, whereas for products
with lower financial risk, eWOM effectiveness is amplified on
platforms with more homophily details.
These results imply that platform and product managers’
perspectives may not be aligned, but they also highlight pos-
sible win-win scenarios. For example, we find a perfect match
between e-commerce platforms and newly introduced prod-
ucts: eWOM on e-commerce platforms is especially effective
for new products; for new products, in turn, eWOM displayed
on e-commerce platforms is linked to the greatest impact
on sales. In addition, we find that on e-commerce sites, it is
eWOM’s visibility and less structured display that increase
effectiveness. These platform characteristics increase eWOM
effectiveness also for utilitarian products, highlighting a mutual
interest between e-commerce platform managers and utilitarian
product managers. Similarly, we show that for social media
platforms, eWOM effectiveness is particularly boosted when
sender homophily details are provided. This is also the case
for products with low financial risk. Thus, in these instances,
platform and product managers’ interests can be aligned,
leading to a win-win situation for both parties.
The Effect of Electronic Word of Mouth on Sales
The second debate addressed in this study centers on the
following question: What are the differential effects of eWOM
metrics on sales? In our research, we move beyond the simple
comparison between volume and valence metrics (Floyd et al.
2014; You, Vadakkepatt, and Joshi 2015) to analyze four key
metrics of eWOM. The composite valence–volume metric is
a new measure that we introduce herein to better distinguish
eWOM sentiment measured in absolute terms (e.g., total
number of positive eWOM) from eWOM sentiment measured
in relative terms (e.g., ratio of positive eWOM). One key
contribution of our research is the insight that volume and
composite valence–volume are the most important metrics
linked to sales. This finding extends the theory of interpersonal
influence and provides new insights into the relative impor-
tance of and interplay between the bandwagon and persuasion
dynamics that underlie the link between eWOM and sales. In
particular, we demonstrate the dominance of the bandwagon
effect over the persuasion effect (with volume submetrics
being more effective than valence submetrics) as the dynamic
that best explains eWOM effectiveness. The persuasion effect
is important, especially in combination with the bandwagon
effect (as demonstrated by the large significant effect of
the positive volume vs. positive valence submetric). Con-
sequently, future studies should differentiate between com-
posite valence–volume and valence to better represent how
eWOM works in the marketplace. Failing to do so could
result in an overestimation of the effectiveness of valence
relative to that of volume.
We reconcile extant literature by demonstrating that
negative eWOM is not linked to a decrease in sales, except
for mature products and products with low financial risk (e.g.,
books, DVDs). The finding that positive eWOM metrics,
overall, have a greater effect on sales than negative eWOM
metrics underscores a positivity bias (Zhang, Craciun, and Shin
2010). This finding lends support to the notion that in the online
context, favorable information produces greater effects than
unfavorable information. This is in line with prior research
demonstrating that consumers prefer and are more influenced
by positive eWOM because they suspect that negative eWOM
likely comes from a company’s competitor (Ong 2011).
Another important contribution of our study is the insight
that eWOM variability negatively affects sales. Neither of
the two prior meta-analyses on eWOM effectiveness takes
this metric into account, thus ignoring its influence on firm
performance. We find that greater consumer consensus lowers
functional risk, consequently boosting sales. In contrast, di-
vergent opinions and polarized sentiment increase consumers’
uncertainty about a product’s performance and thus negatively
affect the bottom line. Our results highlight the relevance of
monitoring and managing heterogeneity among consumers’
product evaluations, especially when eWOM is spread on e-
commerce sites. Furthermore, the split-sample analyses in-
dicate when variability may be less of a concern (i.e., for
services, hedonic products, mature products, and products with
low financial risk).
Finally, the present study offers important implications
for researchers. First, the approach used to control for
endogeneity can influence results. In line with Chintagunta,
Gopinath, and Venkataraman (2010), the impact of eWOM
does not change substantially when using instrumental var-
iables or the generalized method-of-moments approach to
control for endogeneity. However, the use of first-difference
315
models can result in lower parameter estimates. Furthermore,
we find that it is necessary to control for product prices and
promotions, because leaving them out may lead to biased
coefficients. We also find weaker effects of eWOM when the
model includes the previous level of sales. Future models
should avoid omitted variable bias, because a smaller number
of parameters may lead to an overestimation of the effec-
tiveness of eWOM.
LIMITATIONS AND DIRECTIONS FOR FURTHER
RESEARCH
Meta-analyses have limitations as well as strengths.
First, the factors we examine are constrained to variables for
which sufficient primary data are available. Thus, our frame-
work should be considered a summary of the most commonly
studied contextual factors related to the eWOM environment,
not an exhaustive list. Second, we could not investigate the role
of eWOM senders’ and receivers’ characteristics, such as prior
knowledge, product involvement, opinion leadership, and the
stage in the consumer decision-making process, because doing
so requires individual-level data. Third, we could only provide
empirical generalizations about platforms and products covered
in our sample. The majority of data points in our meta-analysis
come from Amazon.com and relate to books and movies. These
platforms and product categories were obvious first choices to
examine the phenomenon of eWOM because data are easily
accessible. However, we encourage researchers to enlarge the
scope of eWOM research in terms of platforms and product
categories. Similarly, most studies on eWOM “use a narrow set
of metrics such as numerical ratings or volume, ignoring the
information content of text in these reviews, which is rich in
consumer expressions” (Tirunillai and Tellis 2012, p. 199).
Consequently, we suggest that researchers consider different
eWOM operationalizations that capture formats other than
textual posts and ratings (e.g., “pins,” images, videos, audio
recordings). These other formats may require more initial
qualitative work to set the ground for future quantitative
analysis. In general, more attention should be devoted to the
analysis of the content of eWOM, which so far has been fairly
limited. Fourth, we encourage scholars to use multiple eWOM
measures to capture the different aspects of eWOM, as one
measure alone cannot fully depict such a heterogeneous and
complex variable. Fifth, more insight is necessary into the way
consumers respond to eWOM that they have actually read,
seen, or heard versus eWOM that was merely present on the
platform but was never received. Sixth, we recognize the lack of
empirical studies investigating the effect of external eWOM
(i.e., eWOM about a competing product, brand, or firm).
Additional research is warranted in this area.
In conclusion, this article makes important contributions
to the understanding of the impact of eWOM on sales and the
factors influencing this relationship. We find that the effec-
tiveness of eWOM is dependent on both the online environ-
ment in which it is displayed and the product to which it
pertains. This means that additional eWOM research should
account for the context of eWOM and that managers should
differentiate their eWOM strategies according to the particular
platforms and products. Finally, it is important to monitor and
measure multiple eWOM metrics while paying particular at-
tention to volume and variability.
316
REFERENCES
Aloe, Ariel M. and Cristopher G. Thompson (2013), “The Synthesis
of Partial Effect Sizes,” Journal of the Society for Social Work and
Research, 4 (4), 390–405.
Anderson, Eric T. and Duncan I. Simester (2014), “Reviews
Without a Purchase: Low Ratings, Loyal Customers, and De-
ception,” Journal of Marketing Research, 51 (June), 249–69.
Banerjee, Abhijit V. (1992), “A Simple Model of Herd Behavior,”
Quarterly Journal of Economics, 107 (3), 797–817.
Barry, Chris, Rob Markey, Eric Almquist, and Chris Brahm (2011),
“Putting Social Media to Work,” Bain & Company, (September
12), (accessed January 20, 2015), [available at http://www.bain.
com/publications/articles/putting-social-media-to-work.aspx].
Berger, Jonah (2014), “Word-of-Mouth and Interpersonal Com-
munication: An Organizing Framework and Directions for
Future Research,” Journal of Consumer Psychology, 24 (4),
586–607.
——— and Raghuram Iyengar (2013), “Communication Channels
and Word of Mouth: How the Medium Shapes the Message,”
Journal of Consumer Research, 40 (3), 567–79.
Bijmolt, Tammo H.A. and Rik G.M. Pieters (2001), “Meta-Analysis
in Marketing When Studies Contain Multiple Measurements,”
Marketing Letters, 12 (2), 157–69.
———, Harald J. van Heerde, and Rik G.M. Pieters (2005), “New
Empirical Generalizations on the Determinants of Price Elas-
ticity,” Journal of Marketing Research, 42 (May), 141–56.
Blackshaw, Pete and Mike Nazzaro (2006), “Consumer-Generated
Media (CGM) 101: Word-of-Mouth in the Age of the Web-
Fortified Consumer,” white paper, Nielsen Buzz Metrics,
(accessed January 20, 2015), [available at http://www.nielsen-
online.com/downloads/us/buzz/nbzm_wp_CGM101.pdf]
Brown, Jacqueline J. and Peter H. Reingen (1987), “Social Ties and
Word-of-Mouth Referral Behavior,” Journal of Consumer Re-
search, 14 (3), 350–62.
Brynjolfsson, Erik, Yu Hu, and Michael D. Smith (2003), “Con-
sumer Surplus in the Digital Economy: Estimating the Value of
Increased Product Variety at Online Booksellers,” Management
Science, 49 (11), 1580–96.
Cacioppo, John T. and Richard E. Petty (1989), “Effects on Message
Repetition on Argument Processing, Recall, and Persuasion,”
Basic and Applied Social Psychology, 10 (1), 3–12.
Carney, Michael, Eric R. Gedajlovic, Pursey P.M.A.R. Heugens,
Marc Van Essen, and J. (Hans) Van Oosterhout (2011),
“Business Group Affiliation, Performance, Context, and Strat-
egy: A Meta-Analysis,” Academy of Management Journal, 54
(3), 437–60.
Chen, Yubo, Qi Wang, and Jinhong Xie (2011), “Online Social
Interactions: A Natural Experiment on Word of Mouth Versus
Observational Learning,” Journal of Marketing Research,
48 (April), 238–54.
Chevalier, Judith A. and Austan Goolsbee (2003), “Measuring Prices
and Price Competition Online: Amazon.com and BarnesandNoble.
com,” Quantitative Marketing and Economics, 1 (2), 203–22.
——— and Dina Mayzlin (2006), “The Effect of Word of Mouth on
Sales: Online Book Reviews,” Journal of Marketing Research,
43 (August), 345–54.
Chintagunta, Pradeep K., Shyam Gopinath, and Sriram Venkataraman
(2010), “The Effects of Online User Reviews on Movie Box Office
Performance: Accounting for Sequential Rollout and Aggregation
Across Local Markets,” Marketing Science, 29 (5), 944–57.
Davis, Alanah and Deepak Khazanchi (2008), “An Empirical
Study of Online Word of Mouth as a Predictor for Multi-Product
Category e-Commerce Sales,” Electronic Markets, 18 (2),
130–41.
JOURNAL OF MARKETING RESEARCH, JUNE 2016
De Matos, Celso Augusto and Carlos Alberto Vargas Rossi (2008),
“Word-of-Mouth Communications in Marketing: A Meta-
Analytic Review of the Antecedents and Moderators,” Journal
of the Academy of Marketing Science, 36 (4), 578–96.
Dellarocas, Chrysanthos, Xiaoquan (Michael) Zhang, and Neveen
F. Awad (2007), “Exploring the Value of Online Product Reviews
in Forecasting Sales: The Case of Motion Pictures,” Journal of
Interactive Marketing, 21 (4), 23–45.
Dhar, Ravi and Klaus Wertenbroch (2000), “Consumer Choice
Between Hedonic and Utilitarian Goods,” Journal of Marketing
Research, 37 (February), 60–71.
Dichter, Ernest (1966), “How Word-of-Mouth Advertising Works,”
Harvard Business Review, 44 (6), 147–66.
Doh, Sun-Jae and Jang-Sun Hwang (2009), “How Consumers
Evaluate eWOM (Electronic Word-of-Mouth) Messages,”
Cyberpsychology & Behavior, 12 (2), 193–97.
Duan, Wenjing, Bin Gu, and Andrew B. Whinston (2008), “The
Dynamics of Online Word-of-Mouth and Product Sales: An
Empirical Investigation of the Movie Industry,” Journal of Re-
tailing, 84 (2), 233–42.
Eisend, Martin (2006), “Two-Sided Advertising: A Meta-
Analysis,” International Journal of Research in Marketing,
23 (2), 187–98.
Elliott, Nate, Emily Riley, Sarah Glass, and James McDavid (2012),
“Social Media Metrics that Matter: Communicating the Right
Data to Different Internal Stakeholders,” Forrester Research
Report, (June 26), (accessed January 20, 2015), [available at
https://www.forrester.com/Social+Media+Metrics+That+Matter/
fulltext/-/E-res58444].
Farley, John U., Donald R. Lehmann, and Alan Sawyer (1995),
“Empirical Marketing Generalizations Using Meta-Analysis,”
Marketing Science, 14 (3), G36–46.
Feldman, Jack M. and John G. Lynch (1988), “Self-Generated
Validity and Other Effects of Measurement on Belief, Attitude,
Intention, and Behavior,” Journal of Applied Psychology, 73 (3),
421–35.
Floyd, Kristopher, Ryan Freling, Saad Alhoqail, Hyun Young Cho, and
Traci Freling (2014), “How Online Product Reviews Affect Retail
Sales: A Meta-Analysis,” Journal of Retailing, 90 (2), 217–32.
Fogg, B.J., Cathy Soohoo, David R. Danielson, Leslie Marable,
Julianne Stanford, and Ellen R. Tauber (2003), “How Do Users
Evaluate the Credibility of Web Sites? A Study with over
2,500 Participants,” research report, Association for Comput-
ing Machinery, (accessed January 20, 2015), [available at http://
htlab.psy.unipd.it/uploads/Pdf/lectures/captology/p1-fogg.pdf].
Forman, Chris, Anindya Ghose, and Batia Wiesenfeld (2008),
“Examining the Relationship Between Reviews and Sales: The
Role of Reviewer Identity Disclosure in Electronic Markets,”
Information Systems Research, 19 (3), 291–313.
Geyskens, Inge, Rekha Krishnan, Jan-Benedict E.M. Steenkamp,
and Paulo V. Cunha (2009), “A Review and Evaluation of Meta-
Analysis Practices in Management Research,” Journal of Man-
agement, 35 (2), 393–419.
Ghose, Anindya, Michael Smith, and Rahul Telang (2006), “Internet
Exchanges for Used Books: An Empirical Analysis of Product
Cannibalization and Welfare Impact,” Information Systems Re-
search, 17 (1), 3–19.
Godes, David and Dina Mayzlin (2009), “Firm-Created Word-of-
Mouth Communication: Evidence from a Field Test,” Marketing
Science, 28 (4), 721–39.
Gu, Bin, Jaehong Park, and Prabhudev Konana (2012), “The
Impact of External Word-of-Mouth Sources on Retailer
Sales of High-Involvement Products,” Information Systems
Research, 23 (1), 182–96.
The Effect of Electronic Word of Mouth on Sales
Hedges, Larry V. and Ingram Olkin (1985), Statistical Methods for
Meta-Analysis. Orlando: Academic Press.
Hennig-Thurau, Thorsten, Caroline Wiertz, and Fabian Feldhaus
(2014), “Does Twitter Matter? The Impact of Microblogging
Word of Mouth on Consumers’ Adoption of New Movies,”
Journal of the Academy of Marketing Science, 43 (3), 375–94.
Heugens, Pursey P.M.A.R. and Michel W. Lander (2009),
“Structure! Agency! (and Other Quarrels): Meta-Analyzing In-
stitutional Theories of Organization,” Academy of Management
Journal, 52 (1), 61–85.
Hiura, Kazuki, Mai Kikumori, Keitaro Kishimoto, Naoko Matsumoto,
Miho Nakagawa, Munetoshi Ujita, et al. (2010), “Does Negative
e-WOM Affect Consumer Attitude Negatively or Positively?”
Journal of eBusiness, 10 (1), 38–42.
Ho-Dac, Nga N., Stephen J. Carson, and William L. Moore (2014),
“The Effects of Positive and Negative Online Customer Reviews:
Do Brand Strength and Category Maturity Matter?” Journal of
Marketing, 77 (November), 37–53.
Hung, Kineta, Stella-Yiyan Li, and David K. Tse (2011), “In-
terpersonal Trust and Platform Credibility in a Chinese Multibrand
Online Community,” Journal of Advertising, 40 (3), 99–112.
International Telecommunication Union (2014), “The World
in 2014: ICT Facts and Figures,” research report, Interna-
tional Telecommunication Union, (accessed May 15, 2015),
[available at www.itu.int/en/ITU-D/Statistics/Documents/facts/
ICTFactsFigures2014-e.pdf].
Kikumori, Mai and Akinori Ono (2013), “Paradoxical Relationship
Between the Amount of Negative eWOM Messages and Positive
Consumer Attitude,” working paper, Graduate School of Busi-
ness and Commerce, Keio University.
Kim, Junyong and Pranjal Gupta (2012), “Emotional Expres-
sions in Online User Reviews: How They Influence Con-
sumers’ Product Evaluations,” Journal of Business Research,
65 (7), 985–92.
Landis, J. Richard and Gary G. Koch (1977), “The Measurement of
Observer Agreement for Categorical Data,” Biometrics, 33 (1),
159–74.
Latané, Bibb (1981), “The Psychology of Social Impact,” American
Psychologist, 36 (4), 343–56.
Li, Xitong (2015), “How Does Online Reputation Affect Social
Media Endorsements and Product Sales? Evidence from
Regression Discontinuity Design,” working paper, HEC
Paris.
Lin, Tom M.Y. and Cheng-Hsi Fang (2006), “The Effects of Per-
ceived Risk on the Word-of-Mouth Communication Dyad,”
Social Behavior and Personality, 34 (10), 1207–16.
Lipsey, Mark W. and David B. Wilson (2001), Applied Social
Research Methods Series Practical Meta-Analysis. Thousand
Oaks, CA: Sage Publications.
Liu, Yong (2006), “Word of Mouth for Movies: Its Dynamics and
Impact on Box Office Revenue,” Journal of Marketing, 70 (July),
74–89.
Mahajan, Vijay, Eitan Muller, and Roger A. Kerin (1984), “In-
troduction Strategy for New Products with Positive and
Negative Word-of-Mouth,” Management Science, 30 (12),
1389–404.
Mayzlin, Dina, Yaniv Dover, and Judith A. Chevalier (2014), “Pro-
motional Reviews: An Empirical Investigation of Online Review
Manipulation,” American Economic Review, 104 (8), 2421–55.
McGinnies, Elliott and Charles D. Ward (1980), “Better Liked Than
Right: Trustworthiness and Expertise as Factors in Credibility,”
Personality and Social Psychology Bulletin, 6 (3), 467–72.
Metzger, Miriam J., Andrew J. Flanagin, and Ryan B. Medders
(2010), “Social and Heuristic Approaches to Credibility Evalu-
ation Online,” Journal of Communication, 60 (3), 413–39.
317
Moe, Wendy W. and Michael Trusov (2011), “The Value of Social
Dynamics in Online Product Ratings Forums,” Journal of
Marketing Research, 48 (June), 444–56.
Murray, Keith B. (1991), “A Test of Services Marketing Theory:
Consumer Information Acquisition Activities,” Journal of
Marketing, 55 (January), 10–25.
——— and John L. Schlacter (1990), “The Impact of Services
Versus Goods on Consumers’ Assessment of Perceived Risk
and Variability,” Journal of the Academy of Marketing Sci-
ence, 18 (1), 51–65.
Ong, Beng S. (2011), “Online Shopper Reviews: Ramifications for
Promotion and Website Utility,” Journal of Promotion Man-
agement, 17 (3), 327–44.
Onishi, Hiroshi and Puneet Manchanda (2012), “Marketing Ac-
tivity, Blogging and Sales,” International Journal of Research in
Marketing, 29 (3), 221–34.
Ott, Myle, Claire Cardie, and Jeff Hancock (2012), “Estimating the
Prevalence of Deception in Online Review Communities,” in
Proceedings of the 21st International Conference on World Wide
Web. Lyon, France: Association for Computing Machinery, 201–10.
Park, JaeHong, Bin Gu, and HoonYoung Lee (2012), “The Re-
lationship Between Retailer-Hosted and Third-Party Hosted
WOM Sources and Their Influence on Retailer Sales,” Electronic
Commerce Research and Applications, 11 (3), 253–61.
Peterson, Robert A. and Steven P. Brown (2005), “On the Use of
Beta Coefficients in Meta-Analysis,” Journal of Applied Psy-
chology, 90 (1), 175–81.
Reichelt, Jonas, Jens Sievert, and Frank Jacob (2014), “How
Credibility Affects eWOM Reading: The Influences of Expertise,
Trustworthiness, and Similarity on Utilitarian and Social Func-
tions,” Journal of Marketing Communications, 20 (1), 65–81.
Robins, David, Jason Holmes, and Mary Stansbury (2010),
“Consumer Health Information on the Web: The Relationship of
Visual Design and Perceptions of Credibility,” Journal of the
American Society for Information Science and Technology,
61 (1), 13–29.
Roselius, Ted (1971), “Consumer Rankings of Risk Reduction
Methods,” Journal of Marketing, 35 (January), 56–61.
Rosenthal, Robert (1988), Applied Social Research Methods Series
Meta-Analytic Procedures for Social Research. London: Sage
Publications.
Rui, Huaxia, Yizao Liu, and Andrew Whinston (2013), “Whose and
What Chatter Matters? The Impact of Tweets on Movie Sales,”
Decision Support Systems, 55 (4), 863–70.
Schweidel, David and Wendy W. Moe (2014), “Listening in on Social
Media: A Joint Model of Sentiment and Venue Format Choice,”
Journal of Marketing Research, 51 (August), 387–402.
Shadish, William R. and C. Keith Haddock (2009), “Combining
Estimates of Effect Sizes,” in The Handbook of Research Syn-
thesis and Meta-Analysis, H. Cooper, L.V. Hedges, and J.C.
Valentine, eds. New York: The Russell Sage Foundation, 257–77.
Silver, N. Clayton and William P. Dunlap (1987), “Averaging
Correlation Coefficients: Should Fisher’s Z Transformation Be
Used?” Journal of Applied Psychology, 72 (1), 146–48.
Stanley, T.D. and Hristos Doucouliagos (2012), Meta-Regression
Analysis in Economics and Business. New York: Routledge.
Sun, Monic (2012), “How Does the Variance of Product Ratings
Matter?” Management Science, 58 (4), 696–707.
Tellis, Gerard J. (1988), “The Price Elasticity of Selective Demand:
A Meta-Analysis of Economic Models of Sales,” Journal of
Marketing Research, 25 (November), 331–41.
Tirunillai, Seshadri and Gerard J. Tellis (2012), “Does Chatter
Really Matter? Dynamics of User-Generated Content and Stock
Performance,” Marketing Science, 31 (2), 198–215.
318
Van den Bulte, Christophe and Gary L. Lilien (2001), “Medical
Innovation Revisited: Social Contagion Versus Marketing Ef-
fort,” American Journal of Sociology, 106 (5), 1409–35.
Van Heerde, Harald J. (2005), “The Proper Interpretation of Sales
Promotion Effects: Supplement Elasticities with Absolute Sales
Effects,” Applied Stochastic Models in Business and Industry,
21 (4/5), 397–402.
———, Sachin Gupta, and Dick R. Wittink (2003), “Is 75% of the
Sales Promotion Bump Due to Brand Switching? No, Only 33%
Is,” Journal of Marketing Research, 40 (November), 481–91.
Wangenheim, Florian v. and Tomás Bayón (2004), “The Effect of
Word of Mouth on Service Switching,” European Journal of
Marketing, 38 (9/10), 1173–85.
JOURNAL OF MARKETING RESEARCH, JUNE 2016
Xiong, Guiyang and Sundar Bharadwaj (2014), “Prerelease Buzz
Evolution Patterns and New Product Performance,” Marketing
Science, 33 (3), 401–21.
You, Ya, Gautham G. Vadakkepatt, and Amit M. Joshi (2015), “A
Meta-Analysis of Electronic Word-of-Mouth Elasticity,” Journal
of Marketing, 79 (March), 19–39.
Zeithaml, Valarie A. (1981), “How Consumer Evaluation Processes
Differ Between Goods and Services,” in Marketing of Services,
J.H. Donnelly and W.R. George, eds. Chicago: American
Marketing Association, 186–90.
Zhang, Jason Q., Georgiana Craciun, and Dongwoo Shin (2010), “When
Does Electronic Word-of-Mouth Matter? A Study of Consumer
Product Reviews,” Journal of Business Research, 63 (12), 1336–41.
 Web Appendix

The Effect of Electronic Word of Mouth on Sales:

A Meta-Analytic Review of Platform, Product, and Metric Factors

Ana Babić

Francesca Sotgiu

Kristine de Valck

Tammo H.A. Bijmolt

1
 Web Appendix A

Impact of eWOM on Sales: Number of Studies, Number of Effects, and Average Effects
over Time

Number of Studies Over Time

25

20

15

10

0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
Year of data collection in primary studies

Note: Empirical studies on the effect of eWOM on sales are represented over time. In several studies, the data
collection spanned two or more years. In many cases, more than one effect size was harvested from a single study.

Average Effect and Number of Effects Over Time

1.000 450
.900 400
.800 350
.700 Number of Effects
Average Effect

300
.600
250
.500
200
.400
150
.300
.200 100

.100 50

.000 0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
Year of Data Collection in Primary Studies

Note: The primary vertical axis shows the average effect (i.e., correlation) in bars, and the secondary vertical axis
shows the number of effects with a line.

2
 Web Appendix B
Key Take-Aways from Meta-Analyses on eWOM: Differences and Main Insights
Floyd, Freling, Alhoqail, Cho, and Freling You, Vadakkepatt, and Joshi (2015)
Scope This study
(2014)
Number of studies (effect sizes) 26 (443) 51 (610) 96 (1532)
Unpublished studies 8% 18% 22%
Type of effect size Elasticity Elasticity Fisher Z-transformed correlations
Canada, China, France, Germany, Japan, Malaysia, Singapore, South Korea,
Countries represented China, Germany, South Korea, USA Japan, South Korea, USA
UK, USA, undisclosed EU country
eWOM volume, number of customers, relative Number of friends, relative measures of sales
Dependent variables in the
measures of sales (market share, sales growth (market share), sales rank, TV viewership, Number of customers, sales rank, observed sales
primary studies
rate), sales rank, TV viewership, observed sales observed sales
Consumer-generated online reviews and online
Content source Consumer-generated eWOM Consumer-generated eWOM
critics’ reviews
Blogs and microblogs, discussion forum, review Blogs and microblogs, discussion forums, review sites, social bookmarking
Platform types Review sites
sites, social network sites, social networks, video sharing site
Audio players, bath-fragrance-beauty, books,
Audio players, apparel, books, cars, cellphones, cellphone services, computer
Audio players, beer, books, camcorders, digital camcorders, cars, cellphones, digital cameras,
memory, digital cameras, DVDs, e-books, electronics, fragrance, furniture,
cameras, DVDs, e-books, electronics, hotel stays, DVDs, e-books, microloans, movies, music
Product categories garden products, green tea, hotel stays, houseware, Internet services,
movies, music albums, video games, video albums, online music albums, software, TV
microloans, mobile applications, movies, music albums, restaurant services,
players, TV shows shows, video cassettes, video games, video
software, video cassettes, video games, video players
players
Volume (cumulative, incremental, and average), valence (average rating,
Volume (cumulative and incremental), valence incremental rating, positive val., negative val.), composite valence-volume
eWOM metrics Volume, valence (average rating)
(average rating, positive ratings, negative ratings) (positive vol., negative vol.), variance (agreement, variability, incremental),
content, existence
Effect sizes from Amazon / 40% 42% 44%
control No Amazon control No Amazon control Amazon control
Sales rank control No No Yes
Product and platform variables coded by experts; Platform and product variables retrieved from Wayback Machine and coded by
Additional data collection Product variables coded by experts
industry variables retrieved from archives experts
Unweighted mean comparison, Hierarchical Linear Unweighted mean comparison, Hierarchical Weighted Hedges-Olkin Meta-Analysis (HOMA)
Meta-analytic procedures
Meta-Analysis (HiLMA) Linear Meta-Analysis (HiLMA) Hierarchical Linear Meta-Analysis (HiLMA)
Weights HiLMA No No Yes
Pooled and separate by platform type and product type (by metric type
Analysis Pooled Separate by metric
available in Web Appendix)
Lagged DV control No Yes Yes
DV operationalization control No No Yes
Outliers N/A Excluded Included
2 publication controls included in the model – 1 publication control included in the model – 2 publication controls included in the model – published/unpublished & top-
Publication bias corrections
published/unpublished & top-tier/others published/unpublished tier/others; standard error control included in model; fail-safe N computed
Minimum observations required
N/A N/A 7
for inclusion of variable

3
 Web Appendix C
Primary Studies Included in the Meta-Analysis on the Effects of eWOM on Sales
Volume (Issue), Number of
Authors Year Publication Outlet Research Context eWOM and Platform
Pages Effect Sizesa
International Journal of
Amblee and Bui 2012 16 (2), 91-113 Short books on Amazon.com Ratings on Amazon.com 1
Electronic Commerce
Archak, Ghose, and
2011 Management Science 57 (8), 1485-1509 Digital cameras on Amazon.com Ratings on Amazon.com 78
Iperiotis
Asur and Huberman 2010 Conference proceedings Movies in the U.S. Tweets 4
Ratings and reviews at an undisclosed online retailer’s
Awad and Zhang 2007 Conference proceedings Houseware 9
website
Books, DVDs, videos, and music albums on
Bao and Chang 2014 Decision Support Systems Forthcoming Ratings on Amazon.com 15
Amazon.com
Bao and Chang 2014 Decision Support Systems Forthcoming Books on Amazon.com Ratings on Amazon.com 9
Basuroy and Ravid 2010 Working paper Movies in the U.S. Reviews on Yahoo! Movies 14
Basuroy and Ravid 2013 Working paper Movies in the U.S. Ratings on Rotten Tomatoes 22
Ratings and comments on an undisclosed hotel
Brandes, Nolte, and Nolte 2011 Working paper Hotel room booking in Germany 8
booking website
Chen, De, and Hub 2014 Information Systems Research Forthcoming Music albums on Amazon.com Ratings on Amazon.com 6
Chen, Dhanasobhon, and
2008 Working paper Books on Amazon.com Ratings on Amazon.com 18
Smith
Journal of Interactive Ratings on Autobytel, Car and Driver, Epinions,
Chen, Fay, and Wang 2011 25 (2), 85-94 Cars in the U.S. 4
Marketing MSN, and Yahoo
Chen, Wang, and Xie 2011 Journal of Marketing Research 48 (2), 238-254 Digital cameras on Amazon.com Ratings on Amazon.com 15
Chen, Wu, and Yoon 2004 Conference proceedings Books on Amazon.com Ratings on Amazon.com 11
Chevalier and Mayzlin 2006 Journal of Marketing Research 43 (3), 345-354 Books on Amazon.com Ratings on Amazon.com and Barnes and Noble 102
Chintagunta, Gopinath,
2010 Marketing Science 29 (5), 944-957 Movies in the U.S. Ratings and reviews on Yahoo! Movies 4
and Venkataraman
64 (11), 1183- Cell phones in the U.S., U.K., Canada, Singapore,
Chung 2011 Journal of Business Research Total internet mentions 4
1189 and Malaysia
Colvin 2013 Dissertation Kindle books on Amazon.com Ratings on Amazon.com 28
International Journal of Consumer electronics and video games on
Cui, Lui, and Guo 2012 17 (1), 39-57 Ratings on Amazon.com 14
Electronic Commerce Amazon.com
Images, ratings, and reviews at a large online
Davis and Khazanchi 2008 Electronic Markets 18 (2), 130-141 Multiple product categories in the U.S. 4
retailer’s website
Dellarocas, Zhang, and Journal of Interactive
2007 21 (4), 23-45 Movies in the U.S. Ratings and reviews on Yahoo! Movies 1
Awad Marketing
Dewan and Ramaprasad 2012 Information Systems Research 23 (3), 1056-1073 Music albums on Amazon.com Ratings on Amazon.com 3
Journal of Interactive
Dhar and Chang 2009 23 (4), 300-307 Music albums on Amazon.com Blog chatter and ratings on Amazon.com 14
Marketing
Duan, Gu, and Whinston 2008 Journal of Retailing 84 (2), 233-242 Movies in the U.S. Ratings and reviews on Yahoo! Movies 13
Ehrmann and Schmale 2008 Conference proceedings Books in Germany Ratings on Amazon.com.de 13
43 (30), 4699-
Elliott and Simmons 2011 Applied Economics Music albums on Amazon.com and Amazon.co.uk Ratings on Amazon.com and Amazon.co.uk 36
4705
Fang, Zhu, and Zhang 2011 Conference proceedings Books in China Reviews and ratings on Dangdang 5
Journal of Interactive
Feng and Papatla 2011 25 (1), 75-84 Cars in the U.S. Reviews on Edmunds.com and ConsumerReports.org 2
Marketing

4
 Volume (Issue), Number of
Authors Year Publication Outlet Research Context eWOM and Platform
Pages Effect Sizesa
Forman, Ghose, and
2008 Information Systems Research 19 (3), 291-313 Books on Amazon.com Ratings on Amazon.com 18
Wiesenfeld
Frick, Tsekouras, and Li 2014 Working paper Music albums on Amazon.com Facebook likes and Tweets 6
IEEE Transactions on
23 (10), 1498- Digital cameras, audio and video players on
Ghose and Ipeirotis 2011 Knowledge and Data Ratings on Amazon.com 9
1512 Amazon.com
Engineering
Gopinath, Chintagunta,
2013 Management Science 59 (12), 2635-54 Movies in the U.S. Blog posts, ratings and reviews on Yahoo! Movies 10
and Venkataraman
Gopinath, Thomas, and
2014 Marketing Science 33 (2), 241-58 Cell phone devices in the U.S. Posts on HowardForums.com 6
Krishnamurthi
Ratings on Cnet, DpReview, Epinions, and
Gu, Park, and Konana 2012 Information Systems Research 23 (1), 182-96 Digital cameras on Amazon.com 28
Amazon.com
Gu, Tang, and Whinston 2013 Decision Support Systems 56 (1), 474-81 Books on Amazon.com Ratings on Amazon.com 8
Hao, Li, Ye, and Zou 2008 Conference proceedings Movies in the U.S. Ratings and reviews on Yahoo! Movies 16
Hennig-Thurau, Wiertz,
2012 Working paper Movies in North America Ratings on Netflix and Tweets 30
and Feldhausc
Hennig-Thurau, Wiertz, Journal of the Academy of Ratings on Netflix, IMDb, and Yahoo! Movies, and
2014 43 (3), 375-94 Movies in North America 23
and Feldhaus Marketing Science Tweets
Ho-Dac, Carson, and
2013 Journal of Marketing 77 (6), 37-53 Blu-ray and DVD players on Amazon.com Ratings on Amazon.com 4
Moore
Hu, Bose, Koh, and Liu 2012 Decision Support Systems 52 (3), 674-84 Books on Amazon.com Ratings on Amazon.com 6
Hu, Koh, and Reddy 2014 Decision Support Systems 57 (1), 42-53 Books on Amazon.com Ratings on Amazon.com 13
Hu, Liu, and
2011 Decision Support Systems 50 (3), 614-26 Books, DVDs, and videos on Amazon.com Ratings on Amazon.com 13
Sambamurthy
Information Technology and
Hu, Liu, and Zhang 2008 9 (3), 1-33 Books, DVDs, and videos on Amazon.com Ratings on Amazon.com 4
Management
d
Hu, Pavlou, and Zhang 2007 Working paper Books, DVDs, and videos on Amazon.com Ratings on Amazon.com 19
Hu, Tian, Liu, Liang, and IEEE Transactions on
2012 59 (2), 189-200 Books on Amazon.com Ratings on Amazon.com 9
Gao Engineering Management
Comments on Facebook and microblogs such as
Twitter and Plurk; reviews on IMDb, Rotten
Huang, Boh, and Goh 2013 Working paper Movies in the U.S. Tomatoes, Yahoo! Movies, The Rolling Stone, and 16
other consumer rating websites; postings on
discussion forums
Journal of Product Innovation
Jang and Chung 2014 Forthcoming Mobile applications in South Korea Posts on T Store 1
Management
Joeckel 2007 Conference proceedings Video games in the U.S. Ratings on GameSpot 2
International Journal of
Karniouchina 2011 28 (1), 62-74 Movies in the U.S. Ratings and reviews on Yahoo! Movies 11
Research in Marketing
Kim and Hanssense 2014 Working paper Movies in the U.S. Blog posts and ratings on Yahoo! Movies 13
Kim, Ma, and Kim 2013 Conference proceedings Books (paper and digital) on Amazon.com Ratings on Amazon.com 9
Koh 2011 Dissertation Books on Amazon.com Ratings on Amazon.com 47
Recherche et Applications en
Larceneux 2008 22 (3), 43-62 Movies in France Ratings on Allocine.fr 5
Marketing
Groupon and Living Social deals for restaurants in
Li 2013 Working paper Ratings on Yelp 6
the U.S.
Li and Hitt 2008 Information Systems Research 19 (4), 456-74 Books on Amazon.com Ratings on Amazon.com 4

5
 Volume (Issue), Number of
Authors Year Publication Outlet Research Context eWOM and Platform
Pages Effect Sizesa
Groupon deals for multiple product categories in the
Li and Wu 2013 Working paper Tweets, Facebook “friends” 17
U.S.
Liu 2006 Journal of Marketing 70 (3), 74-89 Movies in the U.S. Ratings and reviews on Yahoo! Movies 27
Liu 2012 Working paper Movies in the U.S. Tweets 6
Lu, Ba, Huang, and Feng 2013 Information Systems Research 24 (3), 596-612 Restaurant reservations in China Ratings on a Chinese restaurant review website 6
Luan and Neslin 2009 Working paper Video games in the U.S. Tweets 1
Luca 2011 Working paper Restaurant reservations in the U.S. Ratings on Yelp 2
Minnema, Bijmolt, Ratings and reviews on the website of a major
2014 Working paper Electronics and furniture in Europe 6
Gensler, and Wieself European online retailer g
Mishne and Glance 2006 Conference proceedings Movies in the U.S. Blog posts 6
Moe and Trusov 2011 Journal of Marketing Research 48 (3), 444-56 Bath, fragrance, and beauty products Ratings at several retailers’ websites 9
48 (11/12), 2176-
Moon, Park, and Kim 2014 European Journal of Marketing Movies in the U.S. Ratings and reviews on IMDb 4
97
Nam and Kannan 2014 Journal of Marketing 78 (4), 21-40 Various consumer goods and services in the U.S. Blog posts and social tags on Delicious 4
International Journal of
Onishi and Manchanda 2012 29 (3), 221-34 Movies, green tea, and cell phone service in Japan Blog posts and Tweets 36
Research in Marketing
Pai 2008 Dissertation Movies in the U.S. Blog posts, ratings on Yahoo! Movies and IMDb 4
Electronic Commerce Research
Park, Gu, and Lee 2012 11 (3), 253-21 Digital cameras on Amazon.com Ratings on Amazon.com and Cnet 14
and Applications
Pathak, Garfinkel, Gopal, Journal of Management
2010 27 (2), 159-88 Books on Amazon.com Ratings on Amazon.com 16
Venkatesan, and Yin Information Systems
Qu, Wang, Wang, and
2013 European Journal of Marketing 47 (8), 1190-1212 Women’s apparel in China Seller feedback scores and “friends” on Taobao 4
Zhang
Ratings and comments on an undisclosed hotel
Quan, Wenyin, and Douf 2011 Conference proceedings Hotel room booking in China 2
booking website e
Rui, Liu, and Whinston 2013 Decision Support Systems 55 (4), 863-70 Movies in the U.S. Tweets 4
Sadikov, Parameswaran,
2009 Conference proceedings Movies in the U.S. Blog posts 3
and Venetis
Journal of Information,
Sharma, Morales-Arroyo, Videos uploaded to YouTube, blog posts, MySpace
2011 Information Technology, and 6, 41-61 Music albums in the U.S. 27
and Pandey “friends”
Organizations
Song, Park, and Kim 2013 Conference proceedings Mobile applications Ratings on Apple and Google app stores 1
Sonnier, McAlister, and
2011 Marketing Science 30 (4), 702-16 Technology durable goods Total internet mentions 18
Rutz
Stephen and Galak 2012 Journal of Marketing Research 49 (5), 624-39 Financial services (microloans) Blog posts and messages on the forum Kiva Friends 8
Sun 2012 Management Science 58 (4), 696-707 Books on Amazon.com and Barnes and Noble Ratings on Amazon.com and Barnes and Noble 12
Sussan 2005 Dissertation Movies in the U.S. Ratings and reviews on Yahoo! Movies 24
Comments on Facebook and YouTube; ratings on
Tang, Fang, and Wang 2014 Journal of Marketing 78 (4), 41-58 Cars and movies in the U.S. 10
Yahoo! Movies
Triche, Cao, and
2013 Conference proceedings Software Ratings on Cnet 6
Thompson
Wang and Jiang 2008 Conference proceedings Multiple product categories Ratings on Amazon.com 13
Wu 2012 Conference proceedings Books on Amazon.com Tweets, ratings on Amazon.com, blog posts 84
Wu and Ye 2008 Conference proceedings Digital cameras in China Seller feedback scores on Taobao 8
Postings on blogs and forums (e.g., Blogger,
Xiong and Bharadwaj 2014 Marketing Science 33 (3), 401-21 Video games in the U.S. 2
Blogspot, Digg, Gforums, MySpace)
Yang and Mai 2010 Journal of Business Research 63 (9/10), 1050-7 Online video games in the U.S. Ratings on GameSpot 1

6
 Volume (Issue), Number of
Authors Year Publication Outlet Research Context eWOM and Platform
Pages Effect Sizesa
Yang, Kim, Amblee, and
2010 European Journal of Marketing 46 (3), 1523-38 Movies in South Korea Reviews and ratings on Naver 6
Jeong
Perfume, computer memory, and prepaid cell phone Seller feedback scores on global.eBay.com and
Ye, Cheng, and Fang 2013 Decision Support Systems 56 (1), 502-12 24
service cards in China Taobao
Zhang 2006 Dissertation Movies in the U.S. Ratings and reviews on Yahoo! Movies 30
ACM Transactions on
Zhang, Li, and Chen 2012 Management Information 3 (1), 1-23 Books on Amazon.com Ratings on Amazon.com 260
Systems
Zhang, Ma, and
2013 European Journal of Marketing 47 (7), 1115-28 Digital cameras on Amazon.com Ratings on Amazon.com 14
Cartwright
Zhou and Duan 2010 Working paper Software Ratings on Cnet 12
Zhou and Duan 2012 Conference proceedings Software Ratings on Amazon.com and Cnet 2
Zhou and Duan 2012 Conference proceedings Software Ratings on Amazon.com 3
Ratings and reviews on Ctrip and Tongcheng (17u),
Zhu and Lai 2009 Conference proceedings Hotel room booking in China 3
blog posts

a
The table includes both bivariate and partial correlations for all five eWOM metrics used in the HOMA (1532 observations from 96 studies); a subset of these effects (1430
observations from 96 studies) is used in the HiLMA, in which eWOM volume, eWOM valence, eWOM composite valence–volume, and eWOM variance effects are considered.
The partial correlations are based on models that assess the main impact of eWOM on sales, without considering interaction effects with eWOM variables.
b
We include two versions of this study: the conditionally accepted manuscript from 2014 and a working paper from 2013, as some additional results are reported in the more
recent version.
c
We include two versions of this study: from 2011 and 2012, as some additional results are reported in the more recent version.
d
A shorter version of this study was published in Communications of the ACM (52 (10), 144-7). We use the working paper because additional information and results are
reported in the unpublished version.
e
We include two versions of this study: from 2013 and 2014, as some additional results are reported in the more recent version.
f
The retailer asked to remain anonymous, but we were able to obtain necessary information from the authors to assess platform characteristics.

7
 Web Appendix D
Computation of Effect Sizes

Procedure Computation Reference

Computation of the effect size r = t / √(t2 + d.f.) Rosenthal (1988)

Change of sign for effect sizes where Amazon sales rank was used as the dependent
measure
-ln (Amazon sales rank) =…
Representative papers: Cui, Lui, and Guo (2012); Park, Gu, r = t / √(t2 + d.f.)
and Lee (2012); Wang and Jiang (2008) Brynjolfsson, Hu,
ln (Amazon sales rank) =… and Smith (2003);
Representative papers: Bao and Chang (2014); Chen, Wang, Chevalier and
and Xie (2011); Forman, Ghose, and Wiesenfeld (2008); Gu, r = - (t / √(t2 + d.f.)) Goolsbee (2003);
Park, and Konana (2012); Sun (2012); Zhang, Ma, and Ghose, Smith, and
Cartwright (2013) Telang (2006)
ln (1/Amazon sales rank) =…
r = t / √(t2 + d.f.)
Representative paper: Ho-Dac, Carson, and Moore (2013)
ln (sales approximated from Amazon sales rank) =…
r = t / √(t2 + d.f.)
Representative paper: Li and Hitt (2008)

Fisher Z-transformation of effect size = ½ log [(1+r)/ (1-r)] Rosenthal (1988)

= 1/inverse of Cooper and Hedges

Computation of the effect size weight variance (1994); Hedges and
= n-3 Olkin (1985)
Cooper and Hedges
(1994); Hedges and
Computation of the standard error for the Fisher Z- = 1/√ (weight)
Olkin (1985);
transformed effect size = 1/√ (n-3)
Stanley and
Doucouliagos (2012)
Computation of the back-transformed effect size (interpreted Lipsey and Wilson
= (e2z – 1)/(e2z + 1) (2001)
as a correlation)

8
 Web Appendix E
Primary Data Collection: Screenshots from the Wayback Machine

We compiled a truly unique dataset, collecting time-varying information that reflect the nature of each platform at the time of the
original data collection. This enabled us to track changes in platform characteristics over time and include them in our moderator analysis to
examine how they influence the effectiveness of eWOM on sales. We will illustrate the collection of these information using the example of
Chen, Wu, and Yoon (2004).

We first retrieved the time of the original data collection from the paper. Chen, Wu, and Yoon (2004) state on page 715 that the data were
collected from Amazon.com in December 2003. Next, we accessed a digital archive of the Internet called the Wayback Machine (available at
https://archive.org/web/). We typed in the URL link of the online platform from which eWOM was collected in the primary study in our sample
(i.e., Amazon.com in this case).

9
Next, we selected the year that was reported in the paper as the period of data collection.

10
We then selected the month and exact date (whenever specified in the paper and captured by the Wayback Machine—represented by the
blue circle on the calendar, as shown on the screenshot below). For December 2003, Amazon.com was captured multiple times, and we
choose the one closer to the middle of the month (December 15).

Note: In the case of panel data, we repeat the entire process for three month-year combinations, namely for the beginning
of data collection, mid-point of the data collection, and the end of the data collection period.

11
Next, we explore the content of the page and record the presence and characteristics of any consumer-generated information. In addition,
we explore the pages of specific product categories investigated in the primary study—books, in the case of Chen, Wu, and Yoon (2004).

12
On the first page [product page, not platform homepage], we code the following platform characteristics:

1. eWOM visibility (number of scrolls needed to access eWOM, e.g. average rating or textual reviews)
2. eWOM structure (whether eWOM is structured, e.g. organized in summary sections or distributed across categories, such as
‘shipping’ and ‘book difficulty’)
eWOM sender homophily:
3. the proportion of eWOM senders who have a username
4. the proportion of eWOM senders who have an avatar eWOM sender details that allow the receiver to assess
5. the proportion of eWOM senders who have a profile page his/her similarity to the sender

6. the proportion of eWOM senders who have disclosed their location

7. the proportion of eWOM senders who disclosed their real name eWOM sender trustworthiness:
8. whether eWOM senders’ membership s disclosed (e.g., how many years eWOM sender details that allow the receiver to assess
they have been registered at the platform) his/her trustworthiness
9. the proportion of eWOM senders page contact information
eWOM message time stamp and helpfulness:
10. the proportion of eWOM messages (ratings, posts, etc.) displayed with a time stamp
11. the proportion of eWOM messages rated as helpful (at least 1 vote as ‘helpful’) eWOM message details that include the time stamp and
helpfulness rating

We also coded the following information which is not currently included in our moderator analysis due to high multicollinearity:

12. the proportion of eWOM senders with a badge (e.g., ‘top 500 reviewer’)
13. the proportion of eWOM senders for whom the total number of [their] eWOM
(e.g., ratings or posts) is available on this platform eWOM sender expertise:
14. the proportion of eWOM senders for whom a link was present to their other eWOM eWOM sender details that allow the receiver to assess
(e.g., a link leading to their other book reviews) his/her expertise
15. the proportion of eWOM senders for whom the number of friends/ followers/
subscribers is displayed

13
For a book on Amazon.com in 2003  on the product page, we can see eWOM, but in the middle of the page versus on top.

eWOM visibility (scrolling needed to access eWOM)

14
Same example as above [book on Amazon.com in 2003]. After scrolling even further on the first page, we can see customer reviews.

eWOM helpfulness

x
eWOM recency

eWOM sender
eWOM real name
sender or username
avatar
[absent] eWOM sender location

eWOM sender profile page

15
 Importantly, we observe how the online platforms in our sample have changed over time.
Below we include examples of noticeable changes regarding the display of eWOM for Amazon.com and Yelp.

Amazon.com

In 2012, we notice several noticeable differences on an Amazon.com’s book page compared to a page in 2003. Primarily, eWOM is placed
on top of the page (readily visible without scrolling).

eWOM visibility

16
Next, we can see that Amazon.com in 2012 displays eWOM in a much more structured way – it now displays a distribution of numerical
ratings and emphasizes certain content.

eWOM structure

17
Yelp

In 2005, Yelp displayed eWOM in the form of ratings and textual reviews, but it did not structure them or emphasize certain eWOM over
another.

18
In 2011, Yelp highlights certain eWOM, making it more structured.

eWOM structure

19
 Web Appendix F
Distribution of Effect Sizes (Overall Sample)

Notes: All effect sizes represented here were transformed using Fisher’s Z transformation and then transformed back
into correlations for ease of interpretation.

20
 .Web Appendix G
Correlation Matrix (Variables in the HiLMA)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
M .31 .10 .00 .22 .10 .06 .04 .05 .03 .01 .03 .06 .17 .66 .02 .00 .00 .00 .00 .14 .00 .58 .24 .06 .00 .38 .37 .02 .26 .11 .58 .03 .10 .59 .44 .24 .00 .00 .00
SD .46 .30 .07 .41 .30 .24 .19 .21 .17 .09 .17 .23 .37 .47 .13 .88 .47 .51 .78 .33 3.81 .40 .43 .25 .71 .49 .48 .15 .44 .31 .49 .18 .29 .49 .50 .43 7.07 2.71 .03
1. Cumulative volume
2. Incremental volume -.22
3. Average volume -.05 -.02
4. Average valence -.35 -.18 -.04
5. Incremental valence -.22 -.11 -.02 -.17
6. Positive valence -.17 -.09 -.02 -.14 -.09
7. Negative valence -.13 -.07 -.01 -.10 -.07 -.05
8. Positive volume -.15 -.07 -.02 -.12 -.07 -.06 -.04
9. Negative volume -.12 -.06 -.01 -.09 -.06 -.05 -.03 -.04
10. Agreement and precision -.06 -.03 -.01 -.05 -.03 -.02 -.02 -.02 -.02
11. Variability -.12 -.06 -.01 -.09 -.06 -.05 -.03 -.04 -.03 -.02
12. Incremental variance -.16 -.08 -.02 -.13 -.08 -.06 -.05 -.05 -.04 -.02 -.04
13. Review platforms .11 -.07 -.03 .12 -.15 -.03 .04 -.06 -.06 .20 -.02 -.11
14. e-Commerce platforms -.23 .07 .01 .07 .23 -.09 -.01 -.09 -.09 -.12 .07 .18 -.61
15. Other platforms .02 -.04 -.01 -.07 -.04 .01 -.03 .13 .18 -.01 -.02 -.03 -.06 -.18
16. eWOM sender homophily .03 .07 -.06 -.01 .18 -.11 -.08 -.24 -.20 -.05 .06 .16 -.20 .24 .01
17. eWOM sender trustworthiness .01 .04 -.07 -.05 .15 -.07 -.03 -.09 -.07 -.12 .02 .12 -.14 .17 .00 .50
18. eWOM message time stamp and helpfulness rating -.04 -.05 -.07 .10 .19 -.16 -.11 -.18 -.10 .07 .07 .17 -.29 .39 .00 .58 .34
19. eWOM visibility (scrolls) -.22 .19 -.06 -.26 .49 -.07 -.03 -.12 -.07 -.07 -.12 .41 -.36 .41 .00 .41 .25 .42
20. eWOM structured display .08 -.06 .12 .13 -.11 -.04 .01 -.06 -.02 -.04 .02 -.11 .24 -.11 .00 .02 .01 .10 -.35
21. Platform maturity -.07 .02 -.14 -.06 .27 -.08 -.11 -.10 -.01 -.11 -.01 .24 -.37 .39 .06 .36 .26 .50 .53 -.09
22. eWOM posting cost .26 -.20 .06 .26 -.44 .03 .03 .01 -.02 .10 .10 -.36 .27 -.30 .00 -.18 -.21 -.23 -.69 .20 -.30
23. Service .09 -.07 .10 .05 -.19 .06 .04 .01 .06 .15 -.05 -.14 .68 -.73 -.07 -.39 -.26 -.40 -.40 .13 -.43 .34
24. Digital product .15 -.08 .02 .03 -.09 .04 -.05 .00 -.05 -.03 -.03 -.06 -.09 .06 -.03 .06 -.03 .01 -.02 -.03 .27 .15 -.15
25. Financial risk -.08 .01 .08 .00 -.06 .00 .03 .09 .11 -.03 .16 -.07 -.03 .00 .27 .01 .05 -.05 -.19 .22 -.21 .09 -.04 -.40
26. Hedonic product .15 -.11 .07 .09 -.24 .01 .02 .05 .01 .11 .08 -.19 .49 -.60 -.09 -.37 -.12 -.35 -.58 .14 -.48 .46 .67 -.16 .07
27. New product .17 -.12 -.01 .08 -.23 .09 .02 .00 .04 .13 -.05 -.19 .39 -.55 -.04 -.25 .01 -.23 -.39 .16 -.12 .36 .58 .00 -.11 .49
28. Endogeneity: simult. equat. .04 .00 -.01 .00 -.05 .02 .04 -.03 -.03 -.01 .03 -.04 .03 -.10 -.02 .01 -.03 .00 -.01 -.04 -.07 .03 .10 -.04 -.02 .03 .07
29. Endogeneity: first-diff. -.38 .33 -.04 -.31 .55 -.06 .00 -.06 -.01 -.06 -.10 .41 -.26 .42 -.08 .23 .23 .25 .70 -.11 .36 -.71 -.34 -.16 -.05 -.41 -.37 -.09
30. Endogeneity: IV or GMM .06 -.06 .01 .05 -.11 -.04 -.05 .02 .03 -.03 .22 -.09 .13 -.13 .12 .11 .02 .00 -.24 .15 -.06 .15 .07 -.07 .39 .21 -.08 -.05 -.21
31. Price control -.15 .10 -.04 -.11 .27 -.02 .00 -.08 -.07 -.11 .00 .21 -.44 .64 -.15 .42 .26 .41 .52 -.11 .42 -.35 -.62 .08 .03 -.64 -.50 -.06 .50 .01
32. Promotion control .09 -.06 -.01 -.01 -.06 .00 .05 .00 .02 -.02 .02 -.04 -.05 .11 -.02 .05 .06 .04 .03 .06 -.02 .15 -.08 -.02 .06 -.03 .00 -.03 -.11 -.06 .09
33. Lagged DV control -.06 .10 -.02 .00 -.08 -.03 .01 .11 .03 -.03 .11 -.08 -.09 -.02 -.04 .08 .06 .02 -.12 .00 -.05 .15 -.07 -.05 .23 .07 -.07 .13 -.11 .25 .07 -.06
34. Sales rank DV -.14 .11 -.08 -.02 .27 -.10 -.02 -.18 -.11 -.10 .06 .21 -.48 .72 -.15 .42 .29 .49 .53 -.07 .55 -.40 -.68 .07 -.07 -.59 -.42 -.08 .48 -.05 .71 .03 .02
35. Amazon -.03 -.02 -.06 .13 .05 -.09 .03 -.14 -.09 -.07 .11 .03 -.40 .64 -.11 .29 .22 .40 .14 -.05 .31 .05 -.51 .03 -.02 -.33 -.28 -.03 .11 -.04 .45 .12 .11 .60
36. Top-tier journal -.04 .06 -.04 -.06 -.10 .08 .17 .03 .09 -.05 .11 -.13 .04 -.05 .16 -.12 .01 -.17 -.13 .12 -.24 -.09 .01 -.13 .37 .04 -.11 .05 .01 .27 .06 .02 .07 .04 -.05
37. Number of parameters -.07 .07 -.06 -.18 .07 -.01 .03 .15 .17 -.07 .01 .05 -.08 -.07 .51 .08 .09 -.01 .16 -.16 .06 -.16 -.11 -.06 .25 -.07 -.23 .04 .14 .33 .19 -.01 .16 .04 -.06 .33
38. Year of data collection -.01 .05 -.02 -.17 .20 -.01 -.16 .08 -.01 -.11 -.07 .19 -.36 .04 .01 .24 .12 .23 .33 -.08 .61 -.17 -.24 .34 -.13 -.20 -.04 -.10 .19 -.01 .18 -.09 .04 .12 -.04 -.42 .06
39. Standard error (precision) -.03 .09 .05 -.19 .10 -.01 -.04 .11 .13 -.01 -.13 .07 .05 -.15 .04 -.25 .10 -.36 -.02 -.12 -.07 -.05 .21 -.06 -.04 .24 .22 -.05 .15 -.21 -.21 -.12 -.19 -.18 -.16 -.10 .00 .02

Note: all continuous variables are mean-centered.

21
 Web Appendix H
Correlation Matrix for Top Three Online Platforms
1 2 3
1. Amazon 1.00 -.35 -.32
2. Barnes and Noble -.35 1.00 -.14
3. Yahoo! Movies -.32 -.14 1.00
4. Books .36 .43 -.32
5. Movies -.46 -.20 .64
6. Digital cameras .10 -.11 -.10
7. Cumulative volume -.03 -.23 .07
8. Incremental volume -.01 .19 -.05
9. Average volume -.06 -.03 -.03
10. Average valence .13 -.18 .06
11. Incremental valence .04 .34 -.12
12. Positive valence -.09 -.04 .00
13. Negative valence .03 .00 .05
14. Positive volume -.14 -.05 -.05
15. Negative volume -.09 -.02 -.04
16. Agreement and precision -.07 -.04 .25
17. Variability .11 -.07 -.01
18. Incremental variance .03 .27 -.09
19. Platform maturity .32 .20 -.35
20. Review platforms -.39 -.17 .80
21. e-commerce platforms .64 .28 -.50
22. Other platforms -.10 -.05 -.05
23. eWOM sender homophily .29 .13 -.21
24. eWOM sender trustworthiness .22 .13 -.11
25. eWOM message time stamp and helpfulness rating .40 .13 -.24
26. eWOM posting cost .06 -.56 .30
27. eWOM visibility (scrolls) .14 .51 -.30
28. eWOM structured display -.04 -.17 .18
29. New product -.28 -.30 .40
30. Service -.51 -.22 .63
31. Digital product .03 -.10 -.09
32. Financial risk -.03 -.11 -.13
33. Hedonic product -.33 -.31 .46
34. Endogeneity: simultaneous equations -.03 -.06 .07
35. Endogeneity: first-difference model .11 .57 -.21
36. Endogeneity: IV or GMM -.04 -.14 .02
37. Price control .44 .33 -.42
38. Promotion control .12 -.07 -.06
39. Lagged DV control .11 -.12 -.05
40. Sales rank DV .60 .32 -.43
41. Top-tier journal -.05 .06 .05
42. Number of parameters -.06 .14 -.13
43. Year of data collection -.04 .07 -.39
44. Standard error (precision) -.17 .05 .10

Note: all continuous variables are mean-centered.

22
 Web Appendix I

Correlation Matrix for Top Three Product Categories

1 2 3
1. Books 1.00 -.46 -.26
2. Movies -.46 1.00 -.14
3. Digital cameras -.26 -.14 1.00
4. Cumulative volume -.18 .06 .02
5. Incremental volume .15 -.08 -.01
6. Average volume -.06 .01 -.02
7. Average valence -.10 .06 .05
8. Incremental valence .34 -.17 -.05
9. Positive valence -.05 .07 -.02
10. Negative valence -.03 .03 .01
11. Positive volume -.13 .04 -.01
12. Negative volume -.07 .07 .01
13. Agreement and precision -.09 .17 -.03
14. Variability -.07 -.05 .06
15. Incremental variance .27 -.13 -.07
16. Platform maturity .47 -.34 .02
17. Review platforms -.40 .65 -.03
18. e-commerce platforms .54 -.68 .14
19. Other platforms -.12 -.07 -.04
20. eWOM sender homophily .34 -.40 .07
21. eWOM sender trustworthiness .18 -.31 -.02
22. eWOM message time stamp and helpfulness rating .38 -.32 .00
23. eWOM posting cost -.49 .31 .07
24. eWOM visibility (scrolls) .64 -.38 -.18
25. eWOM structured display -.28 .12 .35
26. New product -.42 .59 -.04
27. Service -.52 .90 -.16
28. Digital product .10 -.13 -.07
29. Financial risk -.37 -.14 .46
30. Hedonic product -.71 .65 -.01
31. Endogeneity: simultaneous equations -.03 .07 -.04
32. Endogeneity: first-difference model .55 -.30 .01
33. Endogeneity: IV or GMM -.29 .07 .34
34. Price control .65 -.60 .13
35. Promotion control -.04 -.09 -.01
36. Lagged DV control -.24 -.09 .10
37. Sales rank DV .63 -.61 .19
38. Top-tier journal -.09 -.03 .24
39. Number of parameters .07 -.12 .01
40. Year of data collection .16 -.22 -.02
41. Standard error (precision) -.16 .17 -.03

Note: all continuous variables are mean-centered.

23
 Web Appendix J
Number of Studies and Effect Sizes for the HiLMA Results (Per Variable)
Social Media Review e-Commerce Hedonic Utilitarian Mature High Financial Low Financial
Overall Tangible Goods Services New Products
Platforms Platforms Platforms Products Products Products Risk Risk

Intercept 1430 (96) 220 (24) 225 (26) 940 (55) 961 (53) 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1161 (71)
Platforms
Social media platforms (ref) 229 (24) 220 (24) NO NO 70 (9) 115 (14) NO 64 (6) 163 (16) 64 (8) NO 195 (17)
Review platforms (ref=soc. media) 236 (29) NO 225 (26) NO 21 (5) 207 (21) NO 21 (5) 187 (19) 49 (10) NO 201 (20)
e-Commerce platforms (ref=soc. media) 946 (55) NO NO 940 (55) 849 (41) 16 (5) NO 782 (43) 163 (14) 774 (42) NO 766 (40)
Other platforms (ref=soc. media) 23 (3) NO NO NO 23 (3) NO NO 22 (2) NO 18 (1) NO NO
eWOM sender homophily detailsb 1430 (96) 220 (24) 225 (26) NO 961 (53) 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1161 (71)
eWOM sender trustworthiness detailsb 1430 (96) NO 225 (26) 940 (55) 961 (53) 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1161 (71)
eWOM message time stamp and help. ratingb 1430 (96) NO 225 (26) 940 (55) 961 (53) 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) NO 1161 (71)
eWOM visibility (scrolls) b 1430 (96) NO 225 (26) 940 (55) 961 (53) 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1161 (71)
eWOM unstructured display (ref) 1077 (61) NO 129 (10) 808 (41) 771 (36) 200 (18) 344 (29) 730 (37) 321 (25) 754 (37) 109 (15) 966 (49)
eWOM structured display (ref=not structured) 353 (50) NO 96 (16) 132 (14) 190 (17) 138 (14) 190 23) 155 (15) 197 (16) 147 (20) 248 (21) 195 (22)
Platform maturityb 1430 (96) NO NO NO 961 (53) NO 534 (52) 885 (52) 518 (41) NO NO 1161 (71)
No eWOM posting costs (ref) 399 (6) NO NO 387 (5) 398 (5) NO NO 399 (6) 9 (2) 390 (4) 35 (3) 364 (3)
eWOM posting costs (ref=no posting costs) 1001 (90) NO NO 553 (50) 563 (48) NO NO 486 (46) 509 (39) 511 (53) 222 (33) 797 (68)
Products
Tangible good (ref) 989 (53) 69 (9) NO 849 (41) 961 (53) NO 183 (18) 775 (41) 179 (17) 778 (37) NO 742 (32)
Service (ref=tangible) 349 (33) 110 (14) NO 16 (5) NO 338 (32) 327 (32) 16 (2) NO NO NO NO
Digital product (ref=tangible) 92 (14) 21 (2) NO 69 (12) NO NO NO 84 (11) 33 (3) 58 (11) NO 90 (13)
Utilitarian (ref) 887 (52) NO NO 782 (43) 775 (41) 16 (2) NO NO 160 (13) 724 (39) NO 718 (32)
Hedonic (ref=utilitarian) 543 (52) NO NO 158 (18) 186 (18) 322 (31) 534 (52) NO 358 (30) 177 (23) NO 443 (38)
Financial riskb 1430 (96) 220 (24) 225 (26) 940 (55) NO 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1161 (71)
Mature product (ref) 906 (57) 62 (8) NO 774 (42) 778 (37) 43 (7) 176 (23) NO NO 901 (57) 183 (25) NO
New product (ref=mature) 524 (41) 158 (16) NO 166 (14) 183 (17) 295 (25) 358 (30) NO 518 (41) NO 74 (10) NO
eWOM Metrics
Cumulative volume (reference) 438 (70) 114 (18) 99 (21) 219 (37) 250 (42) 132 (23) 213 (35) 225 (39) 215 (33) 223 (37) 87 (22) 351 (51)
Average volume (ref=cum. volume) 7 (3) NO NO NO NO NO NO NO NO NO NO NO
Incremental volume (ref=cum. volume) 144 (14) 22 (4) 12 (2) 110 (8) 109 (7) 23 (4) 32 (6) 112 (8) 27 (5) 117 (10) 19 (4) 125 (10)
Average valence (ref=cum. volume) 312 (62) 10 (4) 78 (20) 225 (42) 198 (33) 88 (20) 143 (32) 169 (33) 138 (26) 174 (36) 61 (17) 251 (46)
Incremental valence (ref=cum. volume) 139 (7) NO NO 139 (7) 139 (7) NO NO 136 (6) NO 135 (6) NO 134 (5)
Positive valence (ref=cum. volume) 91 (23) 33 (9) 11 (3) 45 (10) 50 (12) 31 (10) 37 (14) 54 (12) 49 (14) 42 (10) 21 (8) 70 (17)
Negative valence (ref=cum. volume) 54 (16) NO 13 (4) 35 (9) 35 (9) 8 (6) 23 (10) 31 (9) 23 (8) 31 (9) 13 (6) 41 (11)
Positive volume (ref=cum. volume) 66 (13) 25 (6) NO 31 (7) 36 (9) 18 (4) 32 (8) 34 (8) 24 (6) 42 (8) 22 (6) 44 (9)
Negative volume (ref=cum. volume) 42 (9) 16 (3) NO 18 (5) 26 (7) 16 (3) 17 (4) 25 (7) 20 (4) 22 (6) 16 (5) 26 (5)
Agreement (ref=cum. volume) 13 (2) NO NO NO NO NO NO NO 13 (2) NO NO 13 (2)
Variability (ref=cum. volume) 42 (13) NO NO 36 (10) 36 (10) NO 25 (7) 17 (8) 9 (4) 33 (10) 18 (4) 24 (10)
Incremental variance (ref=cum. volume) 82 (2) NO NO 82 (2) 82 (2) NO NO 82 (2) NO 82 (2) NO 82 (2)
Study Characteristics
Not Amazon (ref) 797 (68) NO NO NO 381 (29) NO NO NO NO 405 (34) 156 (28) NO
Amazon (ref=not Amazon) 633 (37) NO NO NO 580 (33) NO NO NO NO 496 (29) 101 (9) NO
Year of data collectionb 1430 (96) NO 225 (26) NO NO NO NO NO NO 901 (57) NO NO
No price control (ref) 597 (71) NO 204 (24) 178 (33) 215 (30) 325 (31) 441 (44) 150 (31) NO 206 (36) NO 493 (50)
Price control (ref=no price control) 833 (42) NO 21 (5) 762 (35) 746 (36) 13 (3) 93 (12) 735 (34) NO 695 (34) NO 668 (27)
No promotion control (ref) 1384 (92) NO NO 897 (51) 920 (50) NO 520 (50) 853 (49) 501 (40) 872 (54) 235 (32) 1137 (64)
Promotion control (ref=no prom. control) 46 (9) NO NO 43 (8) 41 (7) NO 14 (3) 32 (7) 17 (2) 29 (7) 22 (4) 24 (5)
No lagged DV (ref) 1293 (85) NO NO 854 (51) 866 (47) 318 (27) 467 (44) 815 46) 484 (35) 800 (50) 207 (30) 1076 (58)
Lagged DV (ref=no lagged DV) 137 (17) NO NO 86 (8) 95 (11) 20 (5) 67 (10) 70 (9) 34 (7) 101 (11) 50 (7) 85 (11)
Sales DV (ref) 585 (64) NO NO 144 (23) 189 (26) NO NO 161 (24) NO NO NO NO
Sales rank (ref=sales) 845 (34) NO NO 796 (34) 772 (29) NO NO 724 (29) NO NO NO NO
No endogeneity control (ref) 866 (82) 186 (18) 177 (23) 484 (49) 479 (45) 273 (27) 400 (43) 465 (46) 444 (37) 422 (46) 135 (27) 731 (60)
Simult. eq. model (ref=no end. control) 32 (6) 11 (3) 8 (2) 11 (2) 15 (4) 17 (3) 15 (3) 17 (4) 19 (3) 13 (3) NO 26 (5)
First-difference model (ref=no end. control) 377 (9) NO NO 376 (8) 376 (8) NO 15 (3) 359 (7) 21 (3) 352 (7) 28 (3) 344 (6)
IV or GMM (ref=no end. control) 155 (18) 23 (7) 39 (6) 75 (6) 97 (9) 47 (8) 104 (13) 50 (6) 40 (7) 114 (12) 94 (8) 60 (11)
Number of parametersb 1430 (96) 220 (24) 225 (26) 940 (55) 961 (53) 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1161 (71)
Not a top-tier publication (ref) 1080 (74) 177 (17) NO 723 (45) 705 (39) 255 (24) 392 (39) 680 (40) 423 (32) 648 (42) NO 953 (55)
Top tier publication (ref=not top-tier) 350 (22) 43 (7) NO 217 (10) 256 (14) 83 (8) 142 (13) 205 (12) 95 (9) 253 (15) NO 208 (12)
Standard errorb 1430 (96) 220 (24) 225 (26) 940 (55) 961 (53) 338 (32) 534 (52) 885 (52) 518 (41) 901 (57) 257 (36) 1161 (71)

Note: NO = the variable is not estimated in this particular model due to insufficient observations or high multicollinearity.

24
 Web Appendix K
Robustness Checks

Issue Conclusion
Missing variables:

Regarding our missing values for the platform

characteristics (eWOM sender and message details), we
tackled the problem as follows. First of all, we contacted
the authors and obtained the name(s) of the platforms that
were unknown in the primary studies (unless the authors
were bounded by confidentiality agreements). This step
already reduced the number of missing values. The
remaining missing values (174 observations, i.e. 12% of
our new HiLMA sample) occur when the primary studies
looked at multiple platforms or when the primary studies
harvested eWOM information on the Internet in general.
In the first case, i.e., for more than one platform (e.g.,
Ctrip and Tongcheng), we coded the platform
characteristics for each individual platform separately—
for the specific product category and at the moment of
data collection—and then averaged their values. In the
second case, i.e., for the Internet in general or multiple
unspecified platforms, we used the mean of our sample as
a missing value imputation.

To verify the stability of our results, we rerun the model
excluding these 174 cases altogether. Moreover, we use
an alternative missing value imputation, i.e., using the
median instead of the mean.
Excluding the 174 observations from the analyses, as well
as using an alternative missing value imputation, leads to
the same substantial results. When we remove these
observations from the analysis, we also drop the variable
‘other platforms’ from the model.

Potential multicollinearity:

We identified variables with at least one correlation
greater than .4 and omitted them one at the time.
This does not alter the findings regarding the other
variables, indicating that the degree of multicollinearity is
not affecting the results. Therefore, we keep all variables
in the model to prevent potential omitted variables biases
(Bijmolt, Van Heerde, and Pieters 2005).

The highest reported variance inflation factor (VIF) is

We checked the variance inflation factor and removed 8.37 for the year of data collection (average VIF = 3.07,
(VIF) variables with a VIF higher than 10. median VIF = 2.37), and the results are unaltered when
removing this control variable.

Sensitivity analysis to further rule out ‘study effect’:

We omitted each of the studies in our sample one at the The results were robust throughout, suggesting that the
time. sample is stable. Please note that in our model we use
random effects for the studies, a set of controls for the
methodological choices in each study, and four dummies
capturing publication type (unpublished, top-tier, stream
IT, stream Economics and Management).

25
 Issue
Sensitivity analysis to further rule out ‘platform
effect’:
We omitted each of the platforms in our sample one at the
time.
We also excluded ‘other platforms’ from the analysis
(N=23).
We added an Amazon-specific dummy variable to control
for potential confounding effects of this dominating
platform.
We checked a full correlation matrix between all the
platforms in our sample and all the variables in our model.
We performed split-sample analyses by platform type.
Sensitivity analysis to further rule out ‘product effect’:
We omitted each of the products in our sample one at the
time.
We checked a full correlation matrix between all the
products in our sample and all the variables in our model.
We performed split-sample analyses by product type.
Sensitivity to eWOM sender operationalization:
We ran the HiLMA model with a different
operationalization of eWOM sender details (by entering
them as 11 different variables or adding them together to
form different combinations of variables).
In addition to the two eWOM-sender-related vectors
(homophily and trustworthiness), we consider a vector
related to sender expertise.
26
Conclusion
The results were robust throughout, suggesting that the
sample is stable. Please note that in our model we include
variables capturing both platform type, as well as time-
varying platform characteristics.
We note that the same results are obtained when excluding
this variable from the model.
The highest correlation between the platforms
(operationalized by dummies, and not used in our model,
with the exception of the Amazon dummy which is
included) and platform characteristics is .98 (for ‘Internet
in general’ and ‘other platforms’, which is in line with
how the variable ‘other platforms’ is operationalized),
while the highest correlation between the platforms and
product characteristics is .80 (for Yahoo! Movies and non-
commercial platforms).
See results of split-sample analyses in Table 5 in the
paper.
The results were robust throughout, suggesting that the
sample is stable. Please note that in our model we control
for product types and product characteristics.
Service is highly correlated to movies (.90), and
commercial platforms (-.73). We, therefore, account for
the fact that the effect of our service variables is mainly
driven by one product category, i.e., movies. (When we
remove movies from the analysis, we find the same
substantial results.)
See results of split-sample analyses in Table 5 in the
paper.
These variables, although presenting the same sign, are
mostly not significant. However, they significantly
increase multicollinearity (VIF>45).
We are not able to include sender expertise in our analysis
due to high multicollinearity (VIF>30).
Sensitivity to the operationalization of eWOM message
details:

We use two dummy variables for message recency and for Similar conclusions are reached when using only a
helpfulness, instead of our vector. We enter both of them dummy variable for message recency and/or for
in the model, as well as only one of them. helpfulness.

Sensitivity to the data points used to code platform

information:

Three data points were used for each paper to count a site
as having a feature or not: at the beginning of the data
collection period in the primary study, mid-point of the
data collection, and end of the data collection period.
We verified that no changes occurred within the collection
period of each paper. For only two papers out of 96 we
find changes in platform characteristics within the time
period used by the authors in their primary study. In these
cases, we use the mid-point of the data collection for all
the platform characteristics variables. We verified that
using the first and last time period does not affect the
results.

Sensitivity to the operationalization of endogeneity

and the sources of endogeneity:

We augment the model to include the interaction effects
between the three research streams and the three
endogeneity controls.
We use a general endogeneity dummy instead of the three
endogeneity controls and add the interactions between this
dummy and the three research streams.
As this inclusion results in extremely high VIF values, for
the purpose of the robustness check we use an
endogeneity control indicator instead of the three
dummies.
We find that the impact eWOM is lower for econ papers
(β = -.14, p <.05) than for marketing studies and not
significantly different for information technology studies
(β = .36, p <.0001). Moreover, the effect of eWOM once
endogeneity has been accounted for is higher for
economics studies (β = .35, p <.01) and similar for both
marketing and IT papers (β = .07, p >.10).

Inclusion of other DVs than sales:

To be able to include seminal studies in our analysis (e.g., The results are stable also to the inclusion of these other
Godes and Mayzlin, 2004) we extend the focus to include DVs.
also other DVs than sales, for example, TV viewership.

Inclusion of geographic location:

We accounted for the fact that the samples were drawn While our key findings are still robust, we do not find
from 11 countries across three continents (Canada, China, significant results across these variables, while our key
France, Germany, Japan, Malaysia, Singapore, South findings remain substantially the same.
Korea, the United Kingdom, the United States, and one
undisclosed European country), with approximately 75%
of the sales data taken from the United States. In more
than 95% of the cases, the studies investigated the impact
of eWOM on a national level within one country.
However, a smaller number of studies investigated the
role of spatial disaggregation by examining sales in up to
253 local markets. We used (1) one dummy variable
capturing if the data comes from the US, or (2) three

27
dummies: Europe, Asia, and Mixed location.

Alternatively, we used Hofstede’s measures capturing
individualism and uncertainty avoidance for the 11
countries in our dataset, as they would offer a better
understanding of the reason behind potential difference on
how eWOM affects sales across countries.
Similarly, we do not find significant results for the
Hofstede’s measures (even not when we use only one or
two of them at a time).

Inclusion of other variables:

We included ten dummy variables capturing (1) Our key findings remain robust when these variables are
economics research stream and (2) IT research stream excluded one by one and all together. Results are available
(marketing stream used as reference), (3) geographic upon request.
location, (4) advertising controls in primary studies, (5)
critics’ reviews controls in primary studies, (6)
competition controls in primary studies, (7) type of data
used in primary studies (cross-sectional vs. panel data),
(8) temporal aggregation of eWOM, and (9) type of effect
size (bivariate vs. partial correlation). The definition and
operationalization of these variables is available upon
request.

28
 Web Appendix L
Publication Bias
Publication bias

To minimize publication bias, we adopt several approaches at various stages of data collection and
analysis. First, we invest significant effort into identifying and retrieving unpublished work across
multiple research streams. As a result, we include a large proportion of unpublished studies, namely
working papers and doctoral dissertations (22% of our sample; see Web Appendix C) (Eisend and Tarrahi
2014; Hopewell, Clarke, and Mallett 2005). In addition, the set of observed effect sizes contains
substantial proportions of positive and negative, as well as significant and non-significant effects (see
Table 5 in the paper), which provides two additional arguments that our meta-analysis does not suffer
from a substantial publication bias.
Second, in our HiLMA model we control for publication outlet (top-tier journal vs. other) (Eisend
and Tarrahi 2014).
Third, we control for publication bias directly in the model by adding the standard error as a control
variable (Doucouliagos, Stanley, and Giles 2012).
Fourth, we calculate the fail-safe sample size (Nfs) for significant results using Orwin’s (1983)
method to assess the possibility of publication bias or the “file-drawer” problem. The larger the Nfs, the
greater is the confidence in the results obtained. We estimate that 603 studies with an effect size of zero
would need to exist to render the effect of eWOM non-significant at the .05 level. We also used an
alternative computation proposed by Rosenthal (1988, p. 109), in which the number of filed studies
required to bring the new overall p to any desired level is computed as follows: Nfs = 19s – n, where 19 is
the ratio of total number of non-significant (at p > .05) results to the number of significant (at p < .05)
results expected when the null hypothesis is true. Using this method, we estimate that as many as 11376
unpublished studies would be required. It is highly unlikely that so many of such studies exist.
Finally, in earlier versions of our models, we controlled for (1) publication status (published vs.
unpublished study), and (2) research streams (marketing, economics and management, and IT). We found
no significant differences.

29
Copyright of Journal of Marketing Research (JMR) is the property of American Marketing
Association and its content may not be copied or emailed to multiple sites or posted to a
listserv without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.