Вы находитесь на странице: 1из 32

Do Display Ads Influence

Search? Attribution and


Dynamics in Online
Advertising
Pavel Kireyev
Koen Pauwels
Sunil Gupta

Working Paper
13-070

February 9, 2013

Copyright 2013 by Pavel Kireyev, Koen Pauwels, and Sunil Gupta


Working papers are in draft form. This working paper is distributed for purposes of comment and
discussion only. It may not be reproduced without permission of the copyright holder. Copies of working
papers are available from the author.

DoDisplayAdsInfluenceSearch?

AttributionandDynamicsinOnlineAdvertising

PavelKireyev

KoenPauwels

SunilGupta1

February9,2013

1
PavelKireyevisaPh.D.studentandSunilGuptaistheEdwardCarterProfessorofBusinessAdministrationatthe
HarvardBusinessSchool,andKoenPauwelsisProfessoratOzyeginUniversity,Istanbul,Turkey.
DoDisplayAdsInfluenceSearch?

AttributionandDynamicsinOnlineAdvertising

Abstract

Asfirmsincreasinglyrelyononlinemediatoacquireconsumers,marketingmanagers
feelcomfortablejustifyinghigheronlinemarketingspendbyreferringtoonlinemetricssuchas
clickthroughrate(CTR)andcostperacquisition(CPA).However,thesestandardonline
advertisingmetricsareplaguedwithattributionproblemsanddonotaccountfordynamics.
Theseissuescaneasilyleadfirmstooverspendonsomeactionsandthuswastemoney,and/or
underspendinothers,leavingmoneyonthetable.

Wedevelopamultivariatetimeseriesmodeltoinvestigatetheinteractionbetween
paidsearchanddisplayads,andcalibratethemodelusingdatafromalargecommercialbank
thatusesonlineadstoacquirenewcheckingaccountcustomers.Wefindthatdisplayads
significantlyincreasesearchconversion.Bothsearchanddisplayadsalsoexhibitsignificant
dynamicsthatimprovetheireffectivenessandROIovertime.Finally,inadditiontoincreasing
searchconversion,displayadexposurealsoincreasessearchclicks,therebyincreasingsearch
advertisingcosts.Afteraccountingforthesethreeeffects,wefindthateach$1investedin
displayandsearchleadstoareturnof$1.24fordisplayand$1.75forsearchads,which
contrastssharplywiththeestimatedreturnsbasedonstandardmetrics.Weusetheseresults
toshowhowoptimalbudgetallocationmayshiftdramaticallyafteraccountingforattribution
anddynamics.Althoughdisplaybenefitsfromattribution,thestrongdynamiceffectsofsearch
callforanincreaseinsearchadvertisingbudgetsharebyupto36%inourempiricalcontext.

1
Introduction

Firmsaremotivatedtospendmoreoftheirmarketingbudgetonlineasconsumers
increasinglyuseonlinemediatofindinformation.Worldwidedigitaladvertisingspendingin
2012was$103billion,orabout20%oftotalmoneyspentonadvertising,andisexpectedto
increaseto$163billion,or25%oftotaladvertisingspend,bytheendof2016(eMarketer2013).
In2012,almosthalfofalldigitaladdollarsworldwidewerespentonpaidsearch,and38%were
usedfordisplayads(ZenithOptimedia2012).

Theintroductionofonlinemetricssuchasclickthroughrate(CTR)andcostper
acquisition(CPA)byGoogleandotheronlineadvertisershasmadeiteasyformarketing
managerstojustifytheironlineadspendincomparisontothebudgetsusedfortelevisionand
othermedia.However,thesemetricssufferfromthefundamentalproblemofattribution,since
theygivecredittothelastclickandignoretheimpactofotheradformatsthatmayhavehelped
aconsumermovedowntheconversionfunnel.Consider,forexample,aconsumersearching
onlineforabanktoopenanewcheckingaccount.Duringthissearch,theconsumerseesapaid
searchadforaparticularbank,clicksonit,andconverts,recallingthatshesawdisplayadsof
thesamebankafewweeksearlier.Howshouldsearchanddisplayadsbecreditedforthe
conversion,andtowhatextent?

Mostmanagersrecognizetheattributionproblem,andintuitivelybelievethatdisplay
andsearchadsinteracttoinfluenceconsumers.Recently,analyticalfirmsandadagencieshave
startedaddressingthisproblem,butmostoftheirsolutionstendtobeadhoc.Forexample,
someindustrymodelsgiveequalweightorcredittoalladexposuresreceivedbyaconsumerin,
say,atwoweekperiod;othersgivemoreweighttorecentadexposuresandexponentially
lowerweighttopastads(HavasDigital2010).

Asfirmsspendmoreoftheiraddollarsononlinesearchanddisplay,managersand
researchersalikerecognizeaneedformorecarefulattributionadjustmentthattakesinto
accountthejourneyconsumersfollowbeforeconversion.Unfortunately,theeffectsdifferent
advertisingmediahaveonconsumersatdifferentstagesofconsiderationarenotyetwell

2
understood(MarketingScienceInstitute2012).Inthisresearchweusetimeseriesmodelsto
infertheinteractionbetweensearchanddisplayads.Specifically,weaddressthefollowing
questions:

Dodisplayadsinfluencepaidsearchandviceversa?
Ifso,howlargearetheseeffectsandwhatdynamicpatternsdotheyfollow?
Whataretheimplicationsforonlinemarketingmetricsandoptimalbudgetallocation?

Thisresearchdrawsbroadlyontwostreamsofliteratureonlineadvertising
effectiveness(specificallytoourcontext,displayandsearch)andthespillovereffectsofonline
advertising.

Inthecontextofdisplayads,researchershavestudiedtheimpactofadexposureon
clickthroughbehavior(Chatterjeeetal.2003),longtermbrandawareness(DrzeandHussherr
2003),andrepurchasedecisions(Manchandaetal.2006).Researchhasalsoexploredthe
potentialoftargeteddisplayadvertising(ShermanandDeighton2001,Shamdasanietal.2001,
Mooreetal.2005)andtheconsequencesofitsintrusiveness(Edwardsetal.2002,Goldfarband
Tucker2011a,2011b).LewisandRiley(2011)usearandomizedexperimenttomeasurethe
causaleffectofonlinedisplayadvertisingonofflineretailsales.

Inthecontextofpaidsearch,researchershavefocusedonunderstandingoptimal
advertisingstrategyincomplexsearchengineenvironments.GhoseandYang(2009)andRutz
etal(2012)adoptakeywordspecificapproachtounderstandtheperformanceofindividual
keywordsandguideoptimalkeywordinvestmentdecisions.Furtherworkexaminedspillover
withinsearch.YangandGhose(2010)identifycomplementaritiesacrossorganicandpaid
searchlistings,andRutzandBucklin(2011)findspillovereffectsfromgenericsearchtobranded
search.Wieseletal(2011)modelconsumerprogressionthroughthepurchasefunnel,and
explainhowonlineadvertisingmaydrivesalesintheofflinechannel.Incontrasttotheabove
study,whichmainlyexaminesadvertisingeffectivenesswithinaparticularonlinechannel,we
studyadeffectivenesstakingintoaccounttheinteractionandfeedbackbetweenbothsearch
anddisplaychannels.Furthermore,wefocusontheroleofsearchanddisplayadvertisingin

3
customeracquisitioninthecommercialbankingindustry,whereconsumerdecisionprocess
tendstobelongerandmoreinvolved,andtheattributionproblemismoresevere.

Severalstudiesconsiderspilloversandsynergiesinonlineandofflineconsumer
behavior.NaikandPeters(2009)proposeahierarchicalmodeltocapturesynergieswithinthe
offlinechannelandacrossonlineandofflinechannels.Theirmodelbuildsonearlierwork(Naik
andRaman2003),andarguesthatinvestinginofflineandonlineadvertisingsimultaneously
generatesgreaterrevenuesthaninvestingineachchannelindividually.

Asmallnumberofstudiesexaminetheinteractionbetweenpaidsearchanddisplay.
Papadimitrouetal(2012)conductafieldexperimenttoexploretheimpactofdisplayexposure
onsearchqueries.Theyfindthatexposuretoadisplayadincreasesthenumberofrelevant
searchqueriessubmittedby5%25%.LewisandNguyen(2011)conductafieldexperimentto
exploretheimpactofdisplayadvertisingonadvertiserandcompetitorbrandedsearchqueries.
Usingaveryshorttimewindow(10minutes)theyfinda27%45%liftinsearchesattributable
tothedisplayadvertisingexposure.Anumberofstudiesconductedbyindustryresearchersalso
exploretheimpactofdisplayadvertisingonpaidsearch.AsurveyconductedbyiProspect
(2009)findsthatabout50%ofallinternetusersreacttoadisplayadbyconductingasearch
relatedtothebrandorproductdescribedinthead.Thestudyfindsthat14%ofusersmakea
purchaseafterconductingthesearch.AstudybycomScore(FulgoniandMom,2008)findsa
38%liftinbrandedsearchactivityforconsumersexposedtoadisplayad.Thestudytracked
individualconsumersexposedtoadisplayad,andcomparedtheirbehaviorwithagroupof
similarconsumersnotexposedtodisplayadvertising.AniCrossingstudy(MalmandHamman
2009)findsa14%changeinsearchvisitsafteracompanyactivateditsdisplayadvertising
campaign.Inourapplicationforthebank,thebanksadagencyalsoconductedanexperiment
tofindthatdisplayadsimprovesearchadconversionby1520%.

Asthesestudiessuggest,displayadsappeartoinfluencetheeffectivenessofsearchads.
Howevertheylacktwoimportantelementsthatweconsiderinourstudy.First,almostallof
thesestudiesignorethedynamiceffectsofadvertising,wherebydisplayadsmayimpact
consumerssearchbehaviorovertime.Studiesthatattempttoincorporatedynamicsdosoin

4
anadhocfashion.Forexample,theadagencyforourbankdecidedtouseatwoweekperiod
(anadhocassumption)toexamineitseffect.Inourapplication,weshowthatthesedynamic
effectsareverystrongandmaylastseveralweeks.Ignoringthemcanleadtosignificant
underestimationoftheeffectivenessofonlineads.Second,mostofthepreviousstudiesused
clickthroughratesorsimilarmetricstomeasuretheimpactofdisplayadsonsearch.In
contrast,weexaminehowdisplayadsinfluencesearchclicks,conversionandultimatelythe
profitabilityofthefirm.Thisallowsustodetermineappropriatebudgetallocationbetween
searchanddisplay.

Theremainderofthisarticleisorganizedasfollows.First,wepresentourconceptual
frameworkbyexplainingtherelationbetweenattributionandconsumerfunnelprogression.
Then,wepresentthedata,modelingmethodology,andempiricalanalysis.Toconclude,we
provideasetofattributionanddynamicsadjustedmarketingmetrics,anddiscussmanagerial
implications.

ConceptualFramework

Theconsumerjourneycanbeconceptualizedasaconversionfunnel.Aconsumermay
beexposedtoabrandthroughdisplayads,shemayclickontheseadstogetmoreinformation,
andmayeventuallyconvert.Thisisthedirectimpactofdisplayadsonconversionthatmost
studiesfindtobeverysmall.Alternatively,aconsumercouldbeactivelysearchingforaproduct
online,wheresheencountersasearchad,clicksonit,andconverts.Thisisthedirecteffectof
searchads,whichisusuallybiggerthanthedirecteffectofdisplayads.Itiscommonto
measurethesedirecteffectsofdisplayandsearchusingonlinemetricssuchasCTR,CPC,and
CPA.

Besidesdirectconversion,passiveformsofadvertisingexposuremayalsoinfluence
consumersconsiderationsets,andsubsequentactiveengagementwiththefirmmoves
consumersdownthefunneltowardsconversion.Inthisscenario,displayadsmayinfluence
consumersatthetopormiddleofthepurchasefunnelwhilesearchadsmayhavemoreimpact
atthebottomofthefunnel.

5
Figure1showshowthefirmsonlineadvertisingstrategymayinfluenceconsumers
purchasebehaviorandthefirmsbudgetallocation.Inthisframework,thefirmallocatesa
budgetbetweensearchanddisplayadsthatdeterminesthenumberofadimpressionsto
consumers.Theseinturnaffectdisplayorsearchclicks,andeventually,conversion.Two
importantaspectsofourframeworkshouldbenoted.First,weexpectstronginteraction
betweensearchanddisplayadimpressionsandclicks.Theempiricalresultswillshowifdisplay
adsindeedinfluencesearchadeffectiveness,andifso,byhowmuch.Second,thesystem
explicitlyrecognizesendogeneity,wherebythefirmsadvertisingbudgetinfluencesconsumers
exposureandpurchasebehavior,whichinturnaffectshowmuchthefirmspendson
advertising.

Figure1:ImpactofOnlineAdvertisingonConsumerBehavior

6
DataDescription

WeusedatafromalargecommercialbankthatoperatesmainlyinthesouthernU.S.
Afterthefinancialcrisis,advertisingtoacquirenewconsumersbecameincreasinglyimportant
givenreducedmarginsanddecliningconsumerconfidence.Thebankinvestsheavilyinboth
paidsearchanddisplayadstoacquirecustomersforitscheckingaccount.

Forthecalendaryear2010,thebankanditsadvertisingagencyprovidedusweeklydata
onthebanksonlinemarketingexpenditure,searchanddisplayimpressionsandclicks,andthe
numberofonlineapplicationscompletedbyconsumersforanewcheckingaccount.Thebank
investedabout$1millioninonlineadvertising,almostequallysplitbetweensearchand
display.2Therearetwolimitationsofourdataset.First,thebankdoesnottrackifonline
advertisinginfluencesconsumerstoopenacheckingaccountinitsretailbranch.Thismeans
thatwecannotinvestigatetheimpactofonlineadvertisingonofflinebehaviorandviceversa.
Second,ourdatasetconsistsofonlyaggregatelevelsofconsumerbehavior.Althoughthelack
ofindividualleveldataisalimitationforourstudy,managersroutinelyuseaggregatedatato
assesstheperformanceoftheironlinecampaigns.

Paidsearchdatacaptureweeklyspend,clicks,impressions,andthenumberof
applicationscompletedthroughthepaidsearchadslandingpage.Thedisplaydataalsocontain
informationonweeklyspend,clicks,andimpressions.Usinginternetcookies,applications
completedwereattributedtodisplayadvertisingifaconsumerhadseenadisplayadatleast
onemonthbeforeconvertingthroughthedisplayadnetworkslandingpageusingorganic
searchoradirectlink.3However,thedisplayadvertisingdataexcludeddisplaydrivenpaid
searchconversions,asthepaidsearchcampaignsareoverseenbyplatformsmaintainedby
searchenginesunrelatedtothedisplayadnetworks.

2
Tomaintaintheconfidentialityoftheclientbank,wehavedisguisedsomeofthedatawhilemaintainingthe
relationshipbetweenthevariablesofinterest.
3
Thisisanotherformofadhocattributionbetweendisplayadsandotheronlinemedia.However,wedonot
investigatethisinourstudyduetothelackofindividualleveldataavailabletous.

7
Thebankinvestedinfivesearchenginesandelevenadnetworks.Weaggregateover
searchenginesandadnetworkstotheweekleveltoavoidoverparameterizationasour
primaryinterestliesintheinterplayofsearchanddisplayadvertising,asopposedtothe
performanceofindividualsearchenginesordisplayadnetworks.Furthermore,thebanks
limitedinvestmentoveranumberofsmalleradnetworksandsearchenginesmakesitdifficult
toestimatetheimpactatthelevelofasearchengineoradnetwork.

Table1presentsacorrelationmatrixofthevariablesinourdata.Thenotationsusedfor
thevariablesareindicatedbelow:

:Checkingaccountapplicationscompletedthroughpaidsearchinweekt.
:Checkingaccountapplicationscompletedafterexposuretoadisplayad.
:Paidsearchadimpressions.
:Paidsearchadclicks.
:Weeklyexpenditureonpaidsearchadvertising.
:Displayadimpressions.
:Displayadclicks.
:Weeklyexpenditureondisplayadvertising.

Table1:CorrelationMatrix

SA DA SI SC SE DI DC DE
SA
DA 0.76
SI 0.80 0.60
SC 0.66 0.56 0.65
SE 0.80 0.75 0.88 0.68
DI 0.71 0.89 0.59 0.44 0.75
DC 0.67 0.88 0.58 0.41 0.74 0.98
DE 0.56 0.84 0.54 0.34 0.69 0.94 0.94

Table1showsthatmanyvariablesarehighlycorrelated,especiallythoserelatedtothe
samemarketinginstrument.Displayimpressionsexhibita0.98correlationwithclicksanda
0.94correlationwithspend.Therefore,weexcludeddisplayclicksandspendfromtheanalysis.

8
Inthecaseofpaidsearch,spendexhibitsahighcorrelation(0.88)withimpressions,sowe
excludesearchadvertisingspendfromtheanalysistominimizepossiblecollinearity.

Table2:Summarystatistics(perweek)
Variable SA DA SI SC DI
Mean 143 139 230,927 9,778 6,023,264
Median 125.5 150.5 194,013 9,478 6,176,496
Maximum 278 304 745,911 21,073 14,885,122
Minimum 43 0 45,565 2,685 242
Std Dev 66 97 134,968 4,060 5,008,892

Figure1:WeeklyTrend

9
Table2providessummarystatisticsofourdataandFigure2illustratestheweeklytrend
oftheresultingseries.Thedecreasingtrendpresentinallvariablesarisesasaconsequenceof
thebankexhaustingitsadvertisingbudget,andhencedecreasingitsinvestmentsovertimeto
avoidoverspending.Weincorporatethistrendasanondeterministiccomponentofthemodel,
allowingfortheotherendogenousvariablestoexplainit.

MethodologyandAnalysis

Weusepersistencemodelingtechniquestocapturethecomplexdynamic
interdependenciesinonlineadvertising(DekimpeandHanssens1999).Persistencemodeling
extendsmultivariatetimeseriesmethodsintothedomainofmarketing,therebyenabling
researcherstomodeltheeffectsofspilloverandfeedbackdynamicsthroughasystemof
equationsinvolvingmarketingactionsandconsumerresponse.Persistencemodelingis
particularlyrelevantinthecontextofonlineadvertisingastheassociatedmultivariatetime
seriestechniquesrequirenostringentapriorirestrictionsonmodelstructureandallowall
variablesofinteresttoaffecteachother.

Persistencemodelinginvolvesseveralsteps.Aseriesoftestsareusedtodeterminethe
correctmodelspecification.Grangercausalitytestsareusedtoidentifywhichvariablesenter
thesystemendogenously.Unitroottestsaredonetodeterminewhichoftheendogenous
variablesexhibitnonstationarybehaviorandshouldenterthemodelindifferences.Next,
cointegrationtestsareusedtoidentifystationarylinearcombinationsofnonstationary
endogenousvariablesthatmustbeconsideredinthespecificationtocorrectfortemporary
deviationsawayfromtheimpliedlongrunequilibria.

Grangercausalitytests,conductedpairwiseforvariablelaglengthsrangingfrom1to
20,suggestthatallvariablesshouldenterthesystemendogenously.Figure3presentsa
schematicoftheGrangercausalityresults.Forexample,anarrowfrom to indicatesthat
isfoundtoGrangercause foratleastoneofthelaglengthsconsidered.Interestingly,no
arrowexistsfromdisplayimpressionstosearchapplications,implyingthatifdisplaydoesaffect
search,theeffecttravelsthroughsearchimpressionsandsearchclicks.Thecomplexnatureof

10
interdependenciesdepictedinFigure3pointstotheappropriatenessofusingaflexible
approach,suchaspersistencemodeling,tocapturecrossadspilloversandonlineadvertising
dynamics.

Figure3:Grangercausalitygraph

SA DA

SC DI

SI

WeconductAugmentedDickeyFuller(ADF)andKwiatkowskiPhillipsSchmidtShin
(KPSS)unitrootteststodetermineiftheendogenousvariablesareevolvingorstationary.Table
3summarizestheresultingstatisticsoftheunitroottests.TheKPSStestidentifiesallseriesas
evolving,whereastheADFtestidentifiesallbut asevolving.Wechoosetoinclude asan
evolvingvariablefollowingtheoutcomeoftheKPSStesttoallowforrichercointegration
possibilities.

Table3:Summaryofunitroottestresults

Test\Variable SA DA SI SC DI
ADF 2.006 0.428 4.444 2.657 1.984
KPSS 0.872 0.874 0.813 0.587 0.850
Note:Boldnumbersindicatesignificantevidenceofnonstationarity

11
TheJohansencointegrationtracetestidentifiesthreecointegratingrelations.These
relationscanbeinterpretedaslongrunequilibriumconditionswhichmayariseasaresultof
firmbudgetingrulesorconsumerdecisionprocesses.BasedontheoutcomesoftheGranger
causality,unitrootandcointegrationtests,wespecifyavectorerrorcorrectionmodel(VEC)
withallvariablesasendogenous.TheinterpretationofVECmodelsisparticularlyinteresting
fromasubstantiveperspective.Bothmanagersandresearchersmayexpectalongterm
equilibriumlinkingsearchanddisplayapplicationstoacombinationoffirmcontrolvariables
(e.g.searchanddisplayadimpressions)andconsumeractions(e.g.clicksontheseads).
EconomistWalterEndersstatesthateconomictheoryaboundswithequilibriumtheorieswhich,
iftheyinvolvenonstationaryvariables,requiretheexistenceofacombinationofthevariables
thatisstationary(2010,p.356).Withinmarketing,researchersoftenvoiceopinionsaboutthe
necessaryintricaterelationbetweenfirmactivity,consumeractivity,andpurchaseaction.Error
correctionmodelshavebeenusedtostudythelongrunimpactofaproductharmcrisis(Van
Heerdeetal2007),marketsharecannibalizationbynewinnovations(VanHeerdeetal2010),
andlongrunsalessensitivitytopricechanges(Foketal2006).

ThegeneralformoftheVECmodelwith lagsisgivenbyequation1,

, (1)

~ 0, ,

Inequation(1), isthevectorofendogenousvariablesattime , and are
vectorsofdeterministiccomponents(e.g.intercept,trend), isamatrixofcointegrating
relations, , , , , and areparametermatricestobeestimated,and isthe
covariancematrixofthemultivariatenormallydistributederrorterms .Thecoefficientsin
,, capturetheeffectsofpastchangesintheendogenousvariablesontheircurrent
deviations.Thecoefficientsin reflectthespeedofadjustmentoftheendogenousvariables
towardstheequilibriumcointegratingrelationsdefinedin .Werefinethemodelfurtherby

12
allowingforaninterceptinboth and .Theinterceptinthemodelspecificationallowsfor
thepossibilityofadeterministictimetrendtoexistconcurrentlywiththestochasticone
impliedbytheerrorcorrectionmodel.Theinterceptterminthecointegratingvectoris
includedtoaccountfortheinitialvaluesoftheendogenousvariables.TheBayesianInformation
Criterionidentifiesalaglengthof1asoptimal.Theresultingmodelspecificationisindicatedin
equation(2):






1 (2)
,
1 0 0 ,
0 1 0 ,
0 0 1 ,
,

Theparametersarerecoveredintwosteps.First,Johansensprocedureisusedto
estimatethecointegratingvectors.Then,thefirstdifferencesoftheendogenousvariablesare
regressedonanintercept,theirlagsandthecointegratingvectorstorecovertheremainderof
thecoefficients.Notallthecoefficientsinthismodelareidentified.Inparticularstandard
errorscannotberecoveredfor , and .Furthermore,anarbitrarynormalizationis
requiredtoidentifytheremainingcoefficientsofthe matrix.

Table4summarizesthefullsetofparameterestimatesandasymptoticstandarderrors.
Themodelexhibitsgoodfitforamodelindifferences,withindividualequation statistics
rangingfrom0.27to0.45.Portmaneautestsfailtofindsignificantevidenceofresidual
autocorrelationandnormalitytestsfailtorejectnormalityoftheresiduals.Furthermore,
generalizedfluctuationtestsforstructuralchangefailtofindsignificantevidenceofparameter
instability.

13
Table4:VECparameterestimates(asymptotictstatisticsinparentheses)

Cointegrating Eq: , , ,

1.000000 0.000000 0.000000

0.000000 1.000000 0.000000

0.000000 0.000000 1.000000

-0.026943 0.001345 -17.58090


[-6.19270] [ 0.58142] [-5.00585]

1.77E-06 -2.14E-05 -0.009488


[ 0.51097] [-11.6546] [-3.39954]

109.1041 -26.13430 5200.822

Error Correction:

, -0.091170 -0.175613 79.30134 5.523423 7495.433


[-1.52498] [-2.74302] [ 0.66615] [ 0.89771] [ 1.21460]

, -0.107651 -0.139443 -429.4406 -29.77596 32198.84


[-1.01232] [-1.22448] [-2.02803] [-2.72068] [ 2.93331]

, -0.000157 0.000195 -0.402342 0.002561 3.103370


[-2.85132] [ 3.31444] [-3.67005] [ 0.45207] [ 0.54608]

-0.443194 -0.138795 -721.8211 -3.728723 -7399.273


[-2.78235] [-0.81368] [-2.27574] [-0.22745] [-0.45002]

0.234133 -0.433164 87.36913 -2.568378 -11868.29


[ 1.90658] [-3.29385] [ 0.35729] [-0.20322] [-0.93627]

0.000189 -4.02E-05 0.327024 -0.014157 1.069115


[ 2.63937] [-0.52406] [ 2.29089] [-1.91883] [ 0.14448]

-0.002645 -0.000102 -4.195442 -0.127549 249.0512


[-1.62591] [-0.05828] [-1.29520] [-0.76186] [ 1.48318]

-3.72E-06 5.22E-06 -0.009798 -0.000447 0.144877


[-2.05150] [ 2.68708] [-2.71173] [-2.39148] [ 0.77352]

-1.594333 -6.117405 -7426.962 -356.0680 -249869.3


[-0.45454] [-1.62862] [-1.06336] [-0.98637] [-0.69013]

0.396648 0.381397 0.453215 0.370934 0.266994

Itisdifficulttodirectlyinterprettheparametersofpersistencemodels,soweproceed
toderiveimplicationsbyimpulseresponseanalysis.

14
TheEffectsofSearchandDisplayAds

Asrecommendedformultivariatetimeseriesmodels(Sims1980),weuseimpulseresponse
functionstoanalyzetheimpactofsearchanddisplayadvertising,andassesssignificanceby
applyingaonestandarderrorbandtotheimpulseresponsecoefficients4(SimsandZha1999,
DekimpeandHanssens1999).PesaranandShin(1998)provideaderivationofthegeneralized
impulseresponsefunction,whichcapturestheimpactofanunexpectedshocktothe
endogenousvariablesinaVECmodelbyconstructingtwoforecastsandtakingtheirdifference.
Oneforecasttakestheshockintoconsideration,whiletheotherdoesnot.Thedifferenceofthe
twoforecastsprovidestheincrementalimpactoftheshock.Impulseresponsefunctionstrace
theimpactofashocktooneendogenousvariablethroughotherendogenousvariables,thereby
providingacumulativeviewofalldynamicinteractionsthattakeplace.

Forsearchclicksanddisplayimpressions,anunexpectedshockrepresentsan
investmentinjectionbythefirm.Inthecaseofsearchanddisplayapplications,andsearch
impressions,itwouldimplyascenarioinwhichwewouldobserveanunexpectedincreasein
applicationsorsearchviewership,holdingdisplayexposureandsearchclicksunchanged.Itis
commoninpracticetomakebudgetingdecisionsbasedonsearchclicksandCPC,anddisplay
impressionsandcostperthousand(CPM)impressions.Therefore,weusesearchclicksand
displayimpressionsasthemarketingvariablesofinterestandinterprettheforecaststhatresult
fromtheirshocksastheeffectsofincreasesinmarketinginvestment.Weapplyonestandard
deviationshockstothemarketingvariablesandstudytheirsustenance,implicationsfor
performance,andinteractionbetweensearchanddisplay.

Figure4presentsthesustenancelevelsofsearchclicksanddisplayimpressions.
Sustenancemeasurestheresponseofavariabletoaonestandarddeviationshocktoitself.The

4
Wecalculateconfidencebandsfortheimpulseresponsefunctionsbysimulating1000randomdrawsfroma
multivariatenormaldistributionwithmeanzeroandcovariancematrixequaltotheresidualcovariancematrixof
themodel,usingthesedrawstoperturbthedata,andestimatingtheimpulseresponsefunctions1000timeson
theresultingsimulateddatasets.Quantilesofthedistributionsofcoefficientsprovideanindicationoftheaccuracy
oftheimpulseresponsefunctions.Wetakethe16thand84thpercentilesoftheempiricaldistributionto
approximateaonestandarderrorband.

15
plotssuggestthatpersistentinvestmentandcomplexconsumertransitionsbetweendifferent
channelsoftheconversionfunnelleadtosustainedlevelsoflongrunexposuretomarketing.
Panel4ashowsthatashockof4,000searchclickswearsinafter810weeksandstabilizesat
about900clicksperweekinthelongrun.Displayimpressionsfollowasimilarpattern
accordingtopanel4b.Ashockof5millionimpressionswearsinoveraperiodof78weeksand
stabilizesatasustainedlevelof1.4millionimpressionsperweek.

Figure4:Sustenancelevelsofmarketingvariables

TheplotsinFigure5showtheperformanceimpactofmarketing.Thetoprowcaptures
theimpactofinitialshocksandpersistenceinmarketingexposureonsearchapplications.Panel
5ashowstheimpactofsearchclicksonsearchapplications.Ashockof4,000clicksgenerates
15searchapplicationsinitially.Afterawearinperiodof4weeks,900clicks(Figure4a)
generate26applicationsperweek(Figure5a).Asmallernumberofsearchclicksisrequiredto
maintainahigherlevelofsearchapplicationsinthelongrun,suggestingthattheeffectiveness
ofaninjectiontosearchadvertisingincreasesasitpersistsovertime.

16
Figure5:Performanceimpactofmarketingvariables

Panel5bshowstheimpactofdisplayimpressionsonsearchapplications.Asexpected,
aninitialincreaseindisplayimpressionsdoesnotgenerateanysearchapplications.However,
afteraperiodoftwoweeks,displayimpressionspositivelyimpactsearchapplications.A
sustainedlevelof1.4millionimpressions(Figure4b)generatesabout20searchapplications
perweek(Figure5b).Consistentwithourconceptualframework,displayexposureappearsto
driveconsumerstopaidsearchovertime.

ThebottomrowofFigure5capturestheimpactofonlineadsondisplayapplications.A
shocktosearchclicksdoesnotaffectdisplayapplications(Figure5c),exceptfortheinitial
period,whichmaypointtoconsumerswhowouldhaveappliedthroughdisplaysubstituting
intothesearchchannel.Panel5dshowsthattheeffectofdisplayimpressionsondisplay

17
applicationsispowerfulandimmediate.Ashockof5millionimpressions(Figure4b)generates
34applicationsimmediately.Afteroneweek,displayapplicationsdipandthenstabilizeat28
applicationsper1.4millionimpressions(Figure4b)inthelongrun.

Tofurtherunderstandtheimpactofadvertising,weconsidertheinteractionbetween
searchanddisplayadstoseehowincreasedlevelsofdisplayadsmaydrivesearchimpressions
andclicks.Figure6plotstheimpactofashockindisplayimpressionsonsearchimpressionsand
searchclicks.Intheshortrun,weobserveadecreaseinbothsearchimpressionsandclicks,
whichmaybedrivenbyconsumersubstitutionacrosschannels.Inthelongrun,asustained
increaseindisplayimpressionsdrivesasignificantincreaseinsearchimpressionsandclicks,
suggestingthatdisplayexposurenotonlyincreasesconversionthroughsearch,butalsodrives
searchvisitationandsearchclicks.Thisfinding,togetherwiththelackofdirectGranger
causalitybetweendisplayimpressionsandsearchapplications,suggeststhatdisplayadvertising
drivessearchapplicationsthroughsearchimpressionsandclicks.Hence,incalculatingthe
overallimpactofdisplayadvertising,wemusttakeintoaccountthepotentialassociated
increasesincostsfromsearchadvertising.

Figure6:Impactofdisplayadvertisingonsearchfunnelprogression

18
Weexploredthesensitivityoftheimpulseresponseanalysistoourmodeling
assumptions.Inparticular,weinvestigatedhowstronglytheresultsdependonthenon
stationarityandcointegrationofthedataseries.Weestimatedabasicvectorautoregressive
simultaneousequationsmodel,withtheendogenousvariablesenteredinlevels.Althoughsuch
aspecificationisknowntoyieldbiasedestimates,itisinformativetoseehowsignificantly
thesebiasesaffectourresults.TheVARspecificationyieldedqualitativelysimilarfindingsfor
theshorttermandwearinperiod.Asimpliedbythisspecification,thelongruneffectswere
notpersistent.Hence,nonstationarityandcointegrationonlydrivethelongrunbehaviorof
theimpulseresponsefunctions.Thisshowsthatonlythepartofourmodelthatisattributable
tononstationaritydependsonit,andhenceinvokesconfidenceintheshortrunandwearin
impulseresponseestimates.

Wealsoestimatedvariancedecompositionsofforecasterrorstoconfirmthatdisplay
advertisingindeeddrivessearchbehaviorasimpliedbytheGrangercausalitytestsandimpulse
responseanalysis.Variancedecompositionsshowedthatanexogenousshocktodisplay
impressionsexplained40%oftheforecasterrorvarianceinsearchimpressions,17%insearch
clicks,and16%insearchapplications,suggestingthatdisplayimpressionsindeedmove
consumersthroughsearchmedia.

ManagerialImplications

StandardonlinemetricssuchasCPAandROIarestaticmeasuresthatignoreattribution
ordynamiceffects.Asimpliedbytheimpulseresponseanalysis,shockstodisplayadvertising
increasebothexposurestosearchmarketingandsearchapplications.Moreover,the
performanceeffectsarenonstationaryandstabilizeonly24weeksaftertheinitialmarketing
shock,implyingthatmarketingmetricsshouldtakeintoaccountnotonlyattribution,butalso
thedynamiceffectsofmarketing.

CPAandROIofSearchAds

Webeginwiththeimplicationsforsearchmetrics,whereonlydynamicsneedtobe
takenintoaccount,sincesearchclicksdonotimpactdisplayapplications(Figure5c).Althougha

19
complicatedpricingandbiddingsystemdrivescostperclick(CPC),asasimplificationwe
assumethatCPCremainsconstantovertheimpulseresponseforecast.

Table5contrastsCPAandROIcalculatedinastandardfashionwiththeirdynamic
counterpartsasimpliedbytheimpulseresponseanalysis.CPAiscalculatedasthetotalsearch
expendituredividedbythesumofallsearchapplications.Thestandardapproachisastatic
measureofcostperacquisition(orapplicationinourcontext)thatiscommonlyusedbymost
marketingmanagersandsearchengineslikeGoogle.Incontrast,dynamicCPAincorporatesthe
longruneffectsimpliedbytheimpulseresponsefunctions.

Table5:CPAandROIforsearchads

Standard Dynamic Dynamic Dynamic %Change:


(immediate) (wearin) (longrun) Standardvs.
Dynamic
CPA $73 $296 $52 $38 47.5%
ROI $1.27 $0.46 $1.52 $1.75 +37.7%
Numbersinthistablehavebeenrounded.

Usingthebanksannualexpenditureonsearchadsandthetotalnumberofsearch
applications,wefindthestandardCPAtobe$73,anumberthatthebankanditsadagency
usedtoassesstheperformanceofitssearchads.Forestimatingthedynamiceffectofsearch,
recallthatashockof4,000searchclicksgeneratesabout15applicationsintheshortrun.Using
theaverageCPCrateof$1.07fromourdata,theimmediatedynamicCPAisthen$296.
However,asdiscussedearlier,searchclicksincreaseby900perweekandgenerateabout26
searchapplicationsperweekinthelongrun,implyingalongrunCPAof$38,48%lowerthan
thestandardCPA.

Table5alsoprovidesameasureofROIthatshowsthereturnforevery$1investedin
searchads.Accordingtothebank,about80%ofthecustomerswhocompleteonline
applicationsareapprovedforacheckingaccount,andtwothirdoftheapprovedcustomers
actuallyfundtheaccount(i.e.,putsomemoneyintheiraccountwithinamonth).Inother
words,80%*67%=53.6%oftheapplicationsbecomeactivecustomers.Inaddition,bothsearch

20
anddisplayadsweregenerallyaccompaniedbyapromotionalofferthatincludedafreeiPod
Nano,iPodTouchor$100$150cash.Thebankestimatesthatonaveragetheeffectivecostof
thesepromotionsisabout$100foreachnewactivecustomeracquiredthroughtheonline
channel.Thebankfurtherestimatestheaveragecustomerlifetimevalue(CLV)tobe$300for
everyactiveaccount.

UsingthisinformationwecalculatedtheROIforthestandardandthedynamic
approaches.Forexample,thestandardCPAis$73,buttheeffectivecostofgettinganactive
accountis$[(73/0.536)+100]=$236,andthebenefitofthisaccountinthelongrunisitsCLVof
$300.ROIisthensimplythebenefit($300)dividedbytheeffectivecostofanactiveaccount
($236),or1.27forthestandardapproach.

TheresultsshowthataccountingforlongrundynamiceffectsreducessearchCPAby
48%andincreasesitsROIby38%comparedtothestandardmetricsthatignorethesedynamic
effects.Inotherwords,thefirmmaybeunderinvestinginsearchbyrelyingonstandardmetrics.

CPAandROIofDisplayAds

Resultsforimpulseresponsefunctionsshowthatsearchadsdonotaffectdisplay
applications,butdisplayimpressionsinfluencebothsearchanddisplayapplications.Therefore,
inadditiontodynamics,displaymetricsshouldincorporateattribution.Moreover,shocksto
displayadvertisingnotonlyincreasessearchapplications,butalsoincreasessearchclicks,
whichmayleadtogreatersearchcost.Displayattributionmusttakeintoaccountnotonlythe
benefit,butalsothisadditionalcostofspilloverintothesearchchannel.

Tomakethispointsalient,Table6presentsacomparisonofthreemethodsfor
calculatingdisplayCPAandROIwithoutattributiontosearch,withattributiontosearch
applications,butwithoutaccountingforadditionalsearchcost,andwithadditionalsearch
applicationsandsearchcostbothconsidered.Forthesecalculationsweusetheaveragecost
perthousand(CPM)impressionsof$2.05fromourdata.ThecalculationsforCPAandROI
followthesamelogicasbefore,exceptthatnowwealsoincludetheimpactofdisplay
impressionsonsearchapplicationsandsearchcost.

21
Table6:CPAandROIfordisplayads

Standard Dynamic Dynamic Dynamic %


(immediate) (wearin) (longrun) Change
CPA $88 Noattribution $298 $120 $99 +12.7%
Attributionto $258 $71 $57 35.2%
searchapplications
only
Attributionto $233 $94 $76 14.1%
searchapplications
andclicks
ROI $1.14 Noattribution $0.46 $0.92 $1.05 7.3%
Attributionto $0.52 $1.29 $1.45 +28.0%
searchapplications
only
Attributionto $0.56 $1.09 $1.24 +9.6%
searchapplications
andclicks

ThelongrunCPAfordisplayis35%lowerthaninthestandardapproachwhenwe
accountforitsimpactonsearchapplicationsbutignoretheadditionalcostitmaydrive.
However,evenafteraccountingfortheadditionalcost,longrunCPAfordisplayis14%lower
thaninthestandardapproach.ROIofdisplayimpressionsexhibitsasimilarpatternitis28%
higherthanthestandardROIwhenonlyattributiontosearchapplicationsisconsidered,andis
about10%higherwhenbothadditionalsearchapplicationsandextrasearchcostsdueto
displayadsareincluded.

22
BudgetAllocation

Itisclearfromthepreviousanalysisthatsearchadsaremoreeffectivethandisplayads
evenwhenweaccountforattributioneffectsofdisplayads.Thisisduetothefactthatsearch
adsshowasignificantdynamiceffect,whichisperhapsreasonableinthecontextofabank
checkingaccount,whereconsumersarelikelytotakeseveralweeksbeforemakingadecision.
Theseresultshavedirectimplicationsforbudgetallocation.Howshouldthefirmallocateits
onlineadvertisingbudgetbetweensearchanddisplayandhowdoesthisallocationcompareto
thefirmscurrentallocation?

Inanonstationaryscenario,thefirmshouldallocatebudgetaccordingtothelongrun
effectivenessofmarketinginstruments(DekimpeandHansenns1999).Optimalbudgeting
wouldthenallocatesharesaccordingtotheratioofdisplayandsearchadvertisingelasticities.
Table7presentstheadvertisingelasticities5ofmarketingactionsaftertakingintoaccount
attributionanddynamics,andTable8showstheactualandproposedbudgetallocation.

Table7:Advertisingelasticities

AdElasticity AdElasticity AdElasticity


(immediate) (wearin) (longrun)
Search 0.12 0.71 0.96
Display 0.19 0.46 0.57

Consistentwithourpreviousresults,wefindthatsearchelasticitiesaresignificantly
higherthanthedisplayelasticities,suggestingthatthefirmwouldbebetteroffspendingmore
onsearchthanitscurrent50%budgetallocation.Displayadvertisingyieldsalowerelasticity
evenafteraccountingforattribution.Wieseletal(2011)similarlyfindahighsearchadvertising

5
Thelongrunelasticityreflectthepercentchangeinthetotalnumberofdisplayandsearchapplicationsfroma
1%changeininvestmentforaparticularmarketinginstrument,takingintoaccountanyadditionalcostsitmay
drive.Weusesamplemeansinsteadofthelastobservationintheseriesasourdataexhibitadecreasingtendency,
whereasimpulseresponsefunctionsarecalculatedasaveragesoverthedatarange.Therefore,usingthelast
observationintheseriesmayyieldinconsistentelasticityestimates.

23
elasticityof4.35inthecontextofamultichannelfurnitureretailer.Dinneretal(2011)finda
longrunsearchelasticityof0.49andadisplayelasticityof0.15.Manchandaetal(2006)finda
displayelasticityofaround0.02withrespecttorepeatpurchasebehavior.Whileourelasticity
estimatesarewithinthebroadrangeoftheestimatesfoundinthepreviousstudies,itis
importanttonotethatourcontextofbankapplicationsandthe$100incentiveofferedbythe
bankmakesdirectcomparisonsacrossstudiessomewhatdifficult.

Giventhecurrentadvertisingbudgetofthefirm,theoptimalallocationbetweensearch
anddisplayadsisgivenbytheratiooftheirelasticities.Table8showstheactualandproposed
budgetallocation.

Table8:Actualandproposedbudgetallocations

ActualBudget ProposedBudget %Change

Search $542,000 $739,000 +36%


Display $639,000 $442,000 31%

Thefirmiscurrentlyallocating54%ofitsonlineadbudgetondisplayadvertisingeven
thoughthestandardmetricsusedbythefirmshowthesearchCPA($73)tobeabout20%lower
thanthedisplayCPA($88).Thebankanditsadagencymadethisallocationrecognizingthatthe
standardmetricsdonotaccountforattribution.Giventhenatureoftheproductcategorythey
expecteddisplaytohavesignificantimpactonsearchapplications.Totestthishypothesisthe
adagencyconductedafieldexperimentandfoundthatsearcheffectivenessimprovedbyabout
20%whenitwasfollowedbydisplayads.Thisfactoredintotheirbudgetallocation.

However,ourmodelsuggeststhatsearchshouldhave63%ofthetotalbudget,or
almost36%higherthanthebudgetcurrentlyallocatedbythefirm,anddisplaybudgetshould
bereduced.Itmayseemcounterintuitivetoreducethebudgetfordisplayadsafteraccounting
foritsattribution(somethingthatthefirmisalsotryingtodothroughitsexperiment),buta
simpleattributionanalysisignorestwoimportantaspects.First,itignorestheadditionalcostof
searchclicksthatareaccompaniedbythesearchapplicationsgeneratedbydisplay.Second,

24
andperhapsmoreimportantinourapplication,thefirmisignoringdynamiceffectsthatare
particularlystrongforsearchads.

Conclusions

Ourgoalinthisstudywastofindoutifonlinedisplayadsinfluencesearch(attribution
problem),ifonlineadvertising,moregenerally,exhibitsdynamiceffects,andifso,what
implicationsthishasforthefirmsbudgetallocation.Weusedpersistencemodelingondata
fromabankthatusedonlineadvertisingtoacquirenewcustomersforitscheckingaccount.We
foundthatdisplayadshaveasignificantimpactonsearchapplications,aswellasclicks.The
majorityofthisspilloverwasnotinstant,buttookeffectonlyaftertwoweeks.Ontheother
hand,searchadvertisingdidnotleadtoanincreaseindisplayapplications.

Ourfindingssuggestthatsimplestaticmetrics,commonlyusedintheindustry,maynot
accuratelymeasuretheeffectivenessofonlineadvertising.Weproposedynamicversionsofthe
classicmetricsandfindthatsearchCPAis48%lowerthanthestaticCPA,whilesearchROIis
38%higherthanthestaticROI.Similarpatternemergesfordisplayadvertising,wherewealso
accountforattribution.ThismadedisplayCPA14%lowerandROI10%higherthantheir
standardcounterparts.Finally,weshowthattheserevisedmeasuresofadeffectivenessleadto
averydifferentbudgetallocationthantheoneusedcurrentlybythefirm.Specifically,wefind
thateventhoughourproposedallocationgivescredittodisplayduetoitseffectonsearch
applications,searchadbudgetshouldbeincreasedby36%fromitscurrentlevelduetoits
strongdynamiceffects,anddisplayadbudgetshouldbedecreasedby31%.

Ourstudyhasseverallimitationsthatcanprovideavenuesforfutureresearch.Wedo
notconsiderspilloverseffectsofsearchanddisplayintootherchannels.Futureresearchmay
examinetheeffectsofonlineadsonconversionsandfunnelprogressioninmobileandoffline
channels.Weuseaggregatedatathatdoesnotallowustountanglethemechanismthatmay
bedrivingconsumerdecisions.Usingdisaggregatedata,futureresearchcouldprovidericher
insightsintotheconsumerjourneyandprogression,andthedifferentialimpactofvarious

25
marketinginstrumentsatvariousstagesoftheconversionfunnel.Futurestudiesmayalsowish
togeneralizeourfindingsbyexaminingmultipleproductsandcontexts.

Overall,ourresearchsuggeststhatmanagersshouldcarefullyconsidertheinteraction
anddynamiceffectsofsearchanddisplayadvertising.Ourresultsshowthatclassicmetrics
usedinpracticearehighlybiasedsincetheydonotaccountfortheseeffects.Asaresultfirms
maybemakingsuboptimalbudgetallocationdecisions.

26
References

Chatterjee,Patrali,DonnaL.HoffmanandThomasP.Novak(2003),ModelingtheClickstream:
ImplicationsforWebBasedAdvertisingEfforts,MarketingScience,22(4),520541.

Dekimpe,M.G.,&Hanssens,D.M.(1999),Sustainedspendingandpersistentresponse:Anew
lookatlongtermmarketingprofitability,JournalofMarketingResearch,36(4),397412.

Dinner,IsaacM.,HaraldJ.vanHeerdeandScottA.Neslin(2011),DrivingOnlineandOffline
Sales:TheCrossChannelEffectsofDigitalversusTraditionalAdvertising,WorkingPaper.

Drze,Xavier,andFranoisXavierHussherr(2003),"InternetAdvertising:IsAnybody
Watching?"JournalofInteractiveMarketing,17(4),823.

Edwards,S.M.,Li,H.,andLess,J.H.(2002),ForcedExposureandPsychologicalReactance:
AntecedentsandConsequencesofthePerceivedIntrusivenessofPopUpAds,Journalof
Advertising,31(3),8395.

eMarketer(2013),DigitaltoAccountforOneinFiveAdDollars,Retrievedfrom
http://www.emarketer.com/Article/DigitalAccountOneFiveAd
Dollars/1009592#kPpaVi4w0A2DpGPA.99

Enders,W.(2008).Appliedeconometrictimeseries.JohnWiley&Sons.

Fok,D.,Horvth,C.,Paap,R.,&Franses,P.H.(2006),AHierarchicalBayesErrorCorrection
ModeltoExplainDynamicEffectsofPriceChanges,JournalofMarketingResearch,43(3),
443461.

Fulgoni,G.M.,&Morn,M.(2008),Howonlineadvertisingworks:Whithertheclick,comScore.
comWhitepaper.

Ghose,A.,S.Yang(2009),Anempiricalanalysisofsearchengineadvertising:Sponsoredsearch
inelectronicmarkets,ManagementScience,55(10)16051620.

27
Goldfarb,AviandCatherineTucker(2011a),OnlineDisplayAdvertising:Targetingand
Obtrusiveness,MarketingScience,30(3),389404.

Goldfarb,AviandCatherineTucker(2011b),SearchEngineAdvertising:ChannelSubstitution
WhenPricingAdstoContext,ManagementScience,57(3),45870.

HavasDigital(2010),ArtemisAttributionWeighting,[Whitepaper].Retrievedfrom
http://www.havasdigital.com/wpcontent/uploads/2011/02/HD_Insight_AttributionDM.pdf

iProspect(2009),SearchEngineMarketingandOnlineDisplayAdvertisingIntegrationStudy,
[Whitepaper].Retrievedfromhttp://www.iprospect.com/wp
content/uploads/2011/11/iProspectStudy_May2009_SearchEngineMarketingandOnline
DisplayAdvertisingIntegrationStudy.pdf

Lewis,R.,&Nguyen,D.(2012),WasntThatAdforaniPad?DisplayAdvertisingsImpacton
AdvertiserandCompetitorBrandedSearch,WorkingPaper.

Lewis,R.,&Reiley,D.(2011),Doesretailadvertisingwork:Measuringtheeffectsofadvertising
onsalesviaacontrolledexperimentonYahoo,Workingpaper.

Malm,K.,&Hamman,D.(2009),Theeffectsofdisplaymediaonsearchtraffic,Whitepaper.
Retrievedfromhttp://www.icrossing.com/sites/default/files/effectsofdisplayonsearch
traffic.pdf

Manchanda,P;Dube,JP;Goh,KY;etal.(2006),TheeffectofbanneradvertisingonInternet
purchasing,JournalofMarketingResearch,43(1):98108.

MarketingScienceInstitute(2012).20122014ResearchPriorities.
http://www.msi.org/MSI_RP1214.pdf

Moore,RobertS.,ClaireAllisonStammerjohan,RobinA.Coulter(2005),BannerAdvertiser
WebSiteContextCongruityandColorEffectsonAttentionandAttitudes,Journalof
Advertising,34(2),7184.

28
Naik,P.A.,&Peters,K.(2009),Ahierarchicalmarketingcommunicationsmodelofonlineand
offlinemediasynergies,JournalofInteractiveMarketing,23(4),288299.

Naik,P.A.,&Raman,K.(2003),Understandingtheimpactofsynergyinmultimedia
communications,JournalofMarketingResearch,375388.

Papadimitriou,P.,GarciaMolina,H.,Krishnamurthy,P.,Lewis,R.A.,&Reiley,D.H.(2011),
Displayadvertisingimpact:Searchliftandsocialinfluence,InProceedingsofthe17th
ACMSIGKDDinternationalconferenceonKnowledgediscoveryanddatamining(pp.1019
1027).ACM.

Pesaran,H.H.,&Shin,Y.(1998),Generalizedimpulseresponseanalysisinlinearmultivariate
models,EconomicsLetters,58(1),1729.

RutzO.J.,BucklinR.E.,SonnierG.P.(2012),Alatentinstrumentalvariablesapproachto
modelingkeywordconversioninpaidsearchadvertising,JournalofMarketingResearch,
49(3):306319.

RutzO.J.,BucklinR.E.(2011),Fromgenerictobranded:Amodelofspilloverinpaidsearch
advertising,JournalofMarketingResearch,48(1),87102.

Shamdasani,P.N.,AndreaJ.S.StanalandandJulianaTan(2001),Location,Location,Location:
InsightsforAdvertisingPlacementontheWeb,JournalofAdvertisingResearch,41(4),721.

Sherman,L.,&Deighton,J.(2001),Banneradvertising:Measuringeffectivenessandoptimizing
placement,JournalofInteractiveMarketing,15(2),6064.

Sims,C.A.(1980),MacroeconomicsandReality,Econometrica,48(1),148.

Sims,C.A.,&Zha,T.(1999),Errorbandsforimpulseresponses,Econometrica,67(5),1113
1155.

Trusov,M.,Bucklin,R.E.,&Pauwels,K.(2009),EffectsofWordofMouthVersusTraditional
Marketing:FindingsfromanInternetSocialNetworkingSite,JournalofMarketing,73,90
102.

29
VanHeerde,H.J.,Helsen,K.,&Dekimpe,M.G.(2007),Theimpactofaproductharmcrisison
marketingeffectiveness,MarketingScience,26(2),230245.

VanHeerde,H.J.,Srinivasan,S.,&Dekimpe,M.G.(2010),Estimatingcannibalizationratesfor
pioneeringinnovations,MarketingScience,29(6),10241039.

Wiesel,T.,Pauwels,K.,&Arts,J.(2011),PracticePrizePaperMarketing'sProfitImpact:
QuantifyingOnlineandOfflineFunnelProgression,MarketingScience,30(4),604611.

Yang,S.,A.Ghose(2010),Analyzingtherelationshipbetweenorganicandsponsoredsearch
advertising:Positive,negative,orzerointerdependence?MarketingScience,29(4)602
623.

ZenithOptimedia(2012),AdvertisingExpenditureForecastsDecember2012,[Whitepaper].

30

Вам также может понравиться