182K - views

The Cost of Annoying Ads Daniel G

Goldstein Microsoft Research NYC 102 Madison Ave 12th Floor New York NY 10016 dggmicrosoftcom R Preston McAfee Google Strategic Technologies 1600 Amphitheatre Parkway Mountain View CA 94043 prestonmcafeecc Siddharth Suri

Embed :
Pdf Download Link

Download Pdf - The PPT/PDF document "The Cost of Annoying Ads Daniel G" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.

The Cost of Annoying Ads Daniel G






Presentation on theme: "The Cost of Annoying Ads Daniel G"— Presentation transcript:

etarycostofannoyingads.The rstandmaincontributionofthisworkisthatwemeasurethecompensatingwagedif-ferentialofannoyingads.Thatis,wemeasurehowmuchmoreonemustpayausertodothesameamountofworkinthepresenceofannoyingadscomparedtoinnocuousadsornoads.Thecompensatingdi erentialisimportanttomeasurebecauseitcapturessomeofthenegativee ectsofadvertising,whichpublishersneedtoheedasalowerboundwhensettingthepricetorunanad.Inatwo-experimentinvestigation,wecomputethecom-pensatingdi erentialforannoyingads.Inthe rstexper-imentusersrandomlyratedeitherananimatedadoritsstaticcounterpart.Thisdesignshowsthatanimationhasanegativeimpactonuserratings.Forthoseadsthatusersrateasannoyingweaskthemtoexplaintheirthinking.Ananalysisoftheratingsandcommentsyieldsabetterofun-derstandingofwhatusers ndannoyingabouttheseads.Thisanalysiswillalsoexhibithowannoyingadsnegativelya ectuserperceptionsofadvertisers.Theseanalysesareadditionalcontributionsofthiswork.Inthesecondexperiment,weusethoseadsidenti edasmoreorlessannoying,alongwiththerecentmethodolog-icalinnovationofToomimetal.[18],toestimatethepayrateincreasenecessarytogenerateanequalnumberofpageviewsinthepresenceofannoyingads,comparedtoinnocu-ousadsornoads.Thisestimateisthecostofannoyingadsinourexperiment.Wechosecategorizingemailsasthetasktoproxyforusingapublisher'ssitebecauseusersei-therimplicitlyorexplicitlyneedtocategorizetheiremailsasspamornotspaminthepresenceofadswhenusingfreeweb-basedemailservicessuchasYahoo!Mail,GMail,andMail.com.Finally,weprovideatheoreticalmodelofhowourempirical ndingscoulda ectthedisplayadvertisingindustry,whichisthethirdcontributionofthiswork.2.RELATEDWORKAsmentionedinSection1,weusethemethodologicalin-novationofToomimetal.[18]forcomputingcompensatingdi erentials.Toomimetal.conductedaMechanicalTurkexperimentinwhichparticipantsrandomlyexperiencedaneasy,medium,orhardversionofataskatarandomlyas-signedpayrate.Thisallowedtheauthorstocomputehowmuchmoreonewouldhavetopayaworkertodothehardtaskoverthemediumandeasytasks.Theauthorsalsoex-hibitedthistechniqueinanexperimentinwhichparticipantswererandomlyassignedtouseeitheran\ugly"ora\pretty"interfacetodoatask.Wewillusethistechniquetoisolatethee ectoftheadqualityonuserabandonment.Next,wedescribepriorexperimentalworkwhichstudiestheimpactofadqualityonbehavior.DrezeandHussherr[4]conductedanexperimentontheef-fectivenessofdisplayadvertisementsusingeye-trackingtech-nology.Theirconclusion,thatusersrarelyfocusdirectlyonbannerads,isoftenreferredtoasbannerblindness,atermcoinedbyBenway[1].Burkeetal.[2]hadparticipantsper-formvisualsearchtasksinthepresenceofnoads,astaticdisplayad,orananimateddisplayad.Theyfoundthatadsdidreducesearchtime,however,therewasnosigni cantdi erencebetweenanimatedandstaticads.Perhapsevenmoresurprisingly,theydidaposthoctestwhichfoundthatanimatedadswererememberedlessfrequentlythanstaticads.YooandKim[21]askedasimilarresearchquestion.Theyconductedalarger-scalelaboratoryexperimentinwhichpar-ticipantswererandomlyexposedtowebpageswithadswithnoanimation,slow-movinganimationorfast-movinganima-tion.Theyfoundthatmoreanimationdidincreaseattentiontoads.Moreover,moderateanimationincreasedadrecogni-tionratesandbrandattitudes.Highlyanimatedads,how-ever,decreasedrecognitionratesandbrandattitudes.ThisresultcomplementstheresultsofBurkeetal.[2].YooandKim[21]concludethat,\Webadvertisersshouldbeawareofthepossibilitythatexcessiveanimationcanback reagainsttheoriginalintentionofe ectivecommunication."GoldfarbandTucker[5]conducteda eldexperimentinwhichtheyfoundthatadsthatmatchedthesite'scontentoradsthatwereintrusiveincreasedparticipant'sself-reportedintenttopurchase.However,adsthatwerebothintru-siveandmatchedthewebsite'scontentreducedintenttopurchase.Adswereconsideredintrusiveif,forexample,theyproducedapopupwindow,tookoverthewholescreen,playedmusic,orobscuredthewebpagetext.Theauthorssuggestthatthereasonforthisinteractione ectisthatusersaremoresensitivetotargetedandintrusiveadswhentheproductadvertisedisprivacysensitive.Inthecontextofsponsoredsearch,Buscheretal.[3]foundthatadsthatarerelevanttothesearchtermsreceivedmorevisualatten-tionthanadsthatwerelessrelevant.ThiscomplementstheresultsofGoldfarbandTucker[5]whichwerefoundinthedomainofdisplayadvertising.Takenasawhole,thesestudiessuggesttheremaybeben-e tstoasmalldegreeofanimationorintrusivenessinad-vertising,butthattoomuchanimationorintrusivenesscanhaveadetrimentalimpactontheade ectiveness.3.RATINGTHEQUALITYOFADSWenextdescribeourexperiments,bothofwhichwereconductedonAmazon'sMechanicalTurk1,anonlinelabormarket.Sinceitwasoriginallybuiltforjobsthataredif- cultforcomputersbutareeasyforhumans(e.g.,imagerecognition),jobsonMechanicalTurkarecalledHumanIn-telligenceTasksorHITs.TherearetwotypesofpeopleonMechanicalTurk:requestersandworkers.RequesterscanpostHITsandworkerscanchoosewhichHITstodoforpay.AfteraworkersubmitsaHIT,therequestercaneitheracceptorrejecttheworkbasedonitsquality.ThefractionofHITsthataworkersubmitswhichareacceptedisthatworker'sapprovalrating.Thisfunctionsasarep-utationmechanism.TheAmazonAPIgiveseachworkeraccountaunique,anonymousidenti er.BytrackingtheIDsoftheworkerswhoacceptedourHITs,wecouldenforcethatparticipantswereonlyallowedtoparticipateinoneofthetwoexperiments,andtheywereonlyallowedtodothatexperimentonetime.ThereisaburgeoningliteratureonconductingbehavioralexperimentsonMechanicalTurk[12,11,16,6,7,20,13,9,17].Inthissetting,theexperimentertakesontheroleoftherequesterandtheworkersarethepaidparticipantsoftheexperiment.MasonandSuri[10]provideahow-toguideforconductingbehavioralexperimentsonMechanicalTurk.Wenowdescribethedesignandresultsofour rstexperiment,whichservedtoidentifysetsofmoreandlessannoyingads 1http://www.mturk.com Figure1:Thetoppanelranksadsbyannoyingnessandshowsthatthe21mostannoyingadswereanimatedandthe24leastannoyingadswerestatic.Thebottompanelrankspairsofadsbytheannoyingnessoftheanimatedvariant.Thestaticvariantstendtofallbelowtheiranimatedversions,suggestingthatanimationincreasesannoyingness,evenwhentheadvertiserandproductareheldconstant.Errorbarsare1standarderror. Figure4:Screenshotoftheemailcategorizationtaskshowingthebadadscondition.Atthebottomofeachemailclassi cationpage,partici-pantswereshownhowmanyemailstheyhadrated,theirpayrate,andareviewoftheinstructions.Thefooterincludedtwobuttons:oneallowingthemtosubmitandrateanotheremail,andasecondallowingthemtostopcategorizingandcollecttheirpayment.Participantswereallowedtoclassifyupto1000emails.4.2ResultsLetanimpressionbeoneparticipantviewingoneemail(anditsaccompanyingads,ifany),regardlessofwhethertheparticipantclassi estheemailorquitsbeforeclassifyingit.Sinceanemailispresentedassoonastheuseracknowledgestheinstructions,eachofthe1223participantsgeneratedatleastoneimpression.Theoveralldistributionofimpressionsperpersonisskewedwithameanof61,amedianof25and rstandthirdquartilesof6and57.Beingboundedby1frombelowande ectivelyunboundedfromabove(onlytwoparticipantsreachedtheupperlimit),theseimpressionsconstitutecountdata.Inparticular,theyareoverdispersedcountdatarelativetothePoisson(observedvariance/the-oreticalPoissondatavarianceis228.7,p.0001)andthuswellsuitedtoanegativebinomialgeneralizedlinearmodel(GLM)[19].Model1inTable2providesthecoecientsofanegativebinomialGLMofimpressionsonpayrateanddummyvariablesforthepresenceof\goodads"ornoads,relativetothebaselineof\badads".Relativetoabaseline Model1Model2 (Intercept)3:433:43(0:12)(0:12)Goodads0:17(0:10)Noads0:22(0:10)Payrate26:4726:61(4:80)(4:80)Goodadsornoads0:19(0:08) AIC12158:5712156:85BIC12184:1212177:29LogLikelihood�6074:29�6074:43Deviance1481:001481:04Numberofobservations12231223 ***p0:01,**p0:05,*p0:1Table2:NegativebinomialGLMofimpressionsonadconditionandpayrate.Badadsleadtofewerimpressionsthangoodadsornoads.Coecientsareexpressedinlogimpressions;predictedvaluesaredisplayedinFigure5.Payrateisindollarsper veimpressions(.01,.02,.03).Standarderrorsareinparentheses.of\badads",boththe\goodads"conditionandthenoadsconditionledtosubstantiallymoreimpressions(19%and25%moreimpressions,respectively).Model2isthesameasModel1butreplacesthetwoaddummieswithonenewdummyrepresentingthe\goodads"andnoadsconditionscombinedandresultsinasimilarconclusion.Astheco-ecientsinTable2areexpressedinlogterms,thee ectsoftheconditionsonrawimpressionsismosteasilyseeninFigure5,whichalsomakesclearthatthedi erenceinim-pressionsbetweenthe\goodads"and\noads"conditionsisnotsigni cant.ThemodelexpressedinTable2andFigure5canbeusedtoestimatethecompensatingdi erentialofannoyingadsinthisexperiment.Sincethecurvesareslightlynon-linear,arangeofcompensatingdi erentialscouldbecalculatedacrossthepayrateandadconditions.Togetasimple,sin-gleapproximationweusethemiddle,\goodads"conditiontoestimatethee ectofpayraises.Wetaketheaverageofthe.2to.4and.4to.6centdi erences,givinganestimatedincreaseof16.58impressionsresultingfroma.2centperimpressionpayraise.Whensummarizingthee ectofadquality,weusethenumberofimpressionsatthe.4centpayrate.Movingfrom\badads"tonoads,impressionsincreaseby12.68.Thepayraiserequiredtoachievea12.68impres-sionincreaseis.153centsperimpression(=:212:68=16:58)or$1.53CPM(costperthousandimpressions).Thatis,inthisexperiment,aparticipantinthe\badads"conditionwouldneedtobepaidanadditional$1.53perthousandim-pressionstogenerateasmanyimpressionsasapersonintheconditionwithoutads.Similarly,movingfromthe\badads"conditiontothe\goodads"conditionresultedinanadditional9.52impressionsperperson.Itwouldrequireapayraiseof.115centsperimpression(=:29:52=16:58)togenerate9.52additionalimpressions,meaningthatpeoplein Figure7:PhasediagramrelatingmarketsharetouserutilityasdescribedbyTheorem1Inaddition,ifxx,thenuserutilityu&#x-383;uandisde-creasingovertime.Ifx&#x-383;x,uuandisincreasingovertime.TheproofofthistheoremisgiventheinAppendix.ThesolutionisillustratedinthephasediagramgiveninFigure7.Theequilibriumforanystartingmarketsharexinvolvesthepathpointingtoward(x;u).Thevalueofuadjuststoputthepublisheronthispath.Startingwithalowmarketshare,thepublishersetsahighuserutilitywhichisacombinationoflowadvertisingandhighcontentquality,andthengradu-allydegradesuserutilityandincreasesadvertisingintensity.Incontrast,apublisherwhostartswithahighmarketsharewillsetaverylowcontentqualityandhighadvertisingin-tensity,andthengraduallyimprovetheuserexperience.Anincreaseintheinterestratedecreasesx,theasymptoticmarketshare.Anincreaseinthecompetitiveleveluin-creasesxwhenislog-convexandvice-versa.Thereareseveralconclusionsonecandrawfromthismodel.First,sincetheterminalmarketsharepredictedinTheo-rem1dependson,whichdependsonAandu,themodeljusti estheratiooftherevenuetousercostasthekeymetricforadvertisingselection.Second,inacompetitiveadvertis-ingmarket,alladswillsellforaconstanttimestheusercost.Annoyingadswillrunonlywhentheirrevenueisveryhighorthepublisherisextremelywillingtosacri ceuserexpe-rienceforrevenue.Third,alegacypublisher,whosemarketshareislargebecausetheyinitiallyfacedlittlecompetition,willstartwithaloweruserexperienceinvolvingbothmoreadsandworsecontentthananentrant.Thiswillresultinthelegacypublisherseeingafastdeclineinuserbase.Thelegacypublisherscontentwillgraduallyimproveuntilasta-blepointisreached.Finally,ifconsumersreactsucientlyslowlytochangesincontent(thatis,issmall),alegacypublisherwillgraduallygoextinctratherthano erabetteruserexperience.6.CONCLUSIONThe rststudyreportedhereshowedthatpeople ndani-matedadvertisementsmoreannoyingthanstaticones,hold-ingallelseconstant.Thisstudyalsoidenti ed vecategoriesofcomplaintsaboutannoyingadsprovidinga rstpassatidentifyingundesirablefeatures.Weusedthegoodandbadadsfromthisstudytomeasurethecompensatingwagedif-ferentialinthesecondstudy.Themainresultofthispa-peristhatannoyingadsleadtositeabandonmentandthusfewerimpressionsthangoodadsornoads.Inwhatmightbeseenasgoodnewsforpublishers,goodadsandnoadsledtoroughlyequalnumbersofimpressions.Annoyingadsimpairedpeople'sabilitytocarryoutanemailclassi cationtask,suggestingthatannoyingadshavearealcosttousersbeyondmereannoyance.Finally,weprovidedatheoreticalmodelthatcomputesadynamicequilibrium,whichpermitsstudyingnotonlypropertiesofthesteadystate,butthead-justmenttothatstateaswell.Thismodelcanbeusedtounderstandthebehavioroflegacypublishers,whoinheritedalargemarketshare,inthefaceofcompetitionfromnewentrants.Wecalculatedthecompensatingwagedi erentialinourexperimentofbadadstonoadstobe$1.53CPM,badadstogoodadstobe$1.15,andgoodadstonoadstobe$.38CPM.Somecaremustbetakenininterpretingthesenumbers.Whilewepickedatask|classifyingemails|thatshouldbefamiliarandcommonformostinternetusers,thistaskmaynotberepresentativeofotherinternettaskslikereadingnewsstoriesorsearchingforproductstopurchase.Abandonmentratesmaydi erwithdi erenttasksandthee ectsofadvertisingmayvaryaswell.Whilevirtuallyev-erywebservicefeaturescompetition,theswitchingcostsvaryfromverylowinconsumingnewstorelativelyhighinchangingemailservices.BecauseourusersonMechanicalTurkhaveanoutsideoptionofworkingonanalternativeHIT,weexpectourresultstobemostapplicabletosit-uationsinvolvinglowerswitchingcosts.Nevertheless,weexpectthatour ndingthatannoyingadscosttheuseratleast$1CPMovermorepleasantadswillbeobtainedinsomeotherenvironments.Forthesereasons,wesuggestfurtherstudiesbedoneonMechanicalTurk,as eldexperiments,andinlaboratoriestomeasurethisdi erentialonsimilaranddi erenttasks.Ifstudiesacrossvariousdomainswithavarietyoftasksandoutsideoptionsarriveatsimilardi erentials,morecredencecanbeplacedonthesenumbers.Weviewthisworkasa rststepinthisdirection.Iffutureworkarrivesatsimilarestimatesacrossavarietyofpublishers,suchestimatescouldserveasausefullowerboundforwhatapublishershouldchargetoruntheseads.Moreover,itwillbevaluabletousethecompensatingdi erentialsapproachtopricethevariousbadaspectsofads,includinganimationandpooraesthetics.Thisworkalsosuggestsavarietyofpolicyrecommenda-tions.Mostdirectly,the$1CPMusercostofbadadshaspracticalconsequencesforpublishers,especiallyasbadadsoftencommandlowerCPMs.Itisareasonthatpublishersshouldinsistonasubstantialpremiumforannoyingadver-tisements.Moreover,apublishercouldrandomizewhichusersseewhichadsandtrackbothtimespentonthepageandthefrequencywithwhichusersreturntothesite.Thistypeofexperimentationwouldcapturelongerterme ectsofannoyingadsthanthosestudiedhere.Also,publisherscouldgiveusersanoptiontocloseorreplaceanad.Areplace-menteventwouldallowthepublishertoinferthatauserwouldpreferarandomadovertheadcurrentlyshown.Ad-vertiserswithahighclosurerateshouldbechargedmore.Furthermore,itwouldbereasonabletoassumethatmore annoyingadswouldbeclosedorreplacedfasterthanlessannoyingads.Adreplacementwouldhelptheuserbyre-movingtheannoyingadandthepublisherbymakingitpos-sibletochargefortwoimpressions.7.ACKNOWLEDGMENTSWethankRandallA.Lewis,JustinM.Rao,andDavidH.Reileyforhelpfulconversations.APPENDIXInthissectionwegivetheproofofTheorem1.Proof.De ney=logx 1�x.Note,y0=x0 x+x0 1�x=x0 x(1�x),andx=ey 1+ey.Furthermore,1+ey=1+x 1�x=1 1�x.Thuswecanreformulatethepublisher'soptimizationproblemasthatofmaximizingR10e�rtey 1+ey�u+1 y0dt.LetF(y;y0;t)=e�rtey 1+ey�u+1 y0.TheEulerequationforthisproblemis0=@F @y�d dt@F @y0=e�rtey (1+ey)2u+1 y0�1 d dte�rtey 1+ey0u+1 y0=x e�rt[(1�x)(u)+r0(u)�0(u)(u�u)(1�x)�00(u)u0]Thus,00(u)u0=(1�x)(u)+r0(u)�(u�u)(1�x)0(u):Asteadystateofthesystemholdswhenx0=u0=0,oru=uand0=(1�x)(u)+r0(u).Thisisequivalentto1�x=�r 0(u) (u):If�r 0(u) (u)1,alloptimalpathsinvolvex!0asthereisnointernalsteadystate.When�r 0(u) (u)1,thereisaninteriorsteadystate.Theu0=0curveoccurswhen0=(1�x)(u)+r0(u)�(u�u)(1�x)0(u):Thus,near(x;u),du dx u0=0=(u)�(u�u)0(u) (1�x)0(u)+r00(u)�(1�x)0(u)�(u�u)(1�x)000(u)=(u)�(u�u)0(u) +r00(u)�(u�u)(1�x)000(u)(u) r00(u)0:Wecanobtaininsightaboutthepathsnearthissolutionbya rstorderTaylorapproximation.Thestrategylookslikethis.Writex0u0= (u�u)x(1�x)(1�x)(u) 00(u)+r0(u) 00(u)�(u�u)(1�x)0(u) 00(u)!=g(x;u)h(x;u):x0u0 @g/@x@g/@u@h/@x@h/@u! (x;u)=(x;u)x�xu�u.Locallythebehaviorofthegeneralsystemisapproximatedbythebehaviorofthelinearsystem.Theonlychallengingterminthematrixis@h @u u=u;x=x=@ @u(1�x)(u) 00(u)+r0(u) 00(u)�(u�u)(1�x)0(u) 00(u) u=u;x=x=�(1�x)(u)000(u) 00(u)2+r1�0(u)000(u) 00(u)2=rThus,x0u00x(1�x)�(u) 00(u)rx�xu�uTheeigenvaluesofthelinearsystemaredeterminedbyso-lutionsto0=det�x(1�x)�(u) 00(u)r�0=2�r+2x(1�x)(u) 00(u)solvingforgives,=1 2 rs r2�42x(1�x)(u) 00(u)!Because00(u)0,thereisonepositiveandonenegativeeigenvalueandbotharereal.Thus,thebehaviorofthesystemisasaddle,asillustratedinFigure7.Therearein nitelymanypathsconsistentwithequilibriumgivenbythedi erentialequations.Whichoneistherightone?Inthecasewhen�r 0(u) (u)1,allpathsthatdon'tviolatetransversalityleadtox=0.Supposexisacandidatelimit.Considersettingu=u+fortunitsoftime.The rmearns Zt0e�rsdsx(u+)+Zt0e�rsds(x+x(1�x)t)(u)y=1 r�1�e�rtx(u+)+1 re�rt(x+x(1�x)t)(u)1 t@ @ =0=1 t�1 r�1�e�rtx0(u)+ re�rtx(1�x)t(u)= rx(u)1�e�rt t0(u) (u)+e�rt(1�x)= rx(u) r0(u) (u)+(1�x)Thus,itpaystoincreaseaconvergentxifandonlyifx1�r 0(u) (u),implyingthatthisisonlycandidateforconvergentpathswhen�r 0(u) (u)1.