TwitterPower:TweetsasElectronicWordofMouthBernardJ.JansenandMimiZhangC - PDF document

TwitterPower:TweetsasElectronicWordofMouthBernardJ.JansenandMimiZhangCollegeofInformationSciencesandTechnology,PennsylvaniaStateUniversity,UniversityPark,PA18802.E-mail:jjansen@acm.org;mzhang@ist.psu.eduKateSobelSmealCollegeofBusinessAdministration,PennsylvaniaStateUniversity,UniversityPark,PA18802.E-mail:kas5229@psu.eduAbdurChowduryTwitter,Inc.,SanFrancisco,CA94107.E-mail:abdur@ir.iit.edu ReceivedFebruary28,2009;revisedMay7,2009;acceptedMay11,2009©2009ASIS&TPublishedonline6July2009inWileyInterScience(www.interscience.wiley.com).DOI:10.1002/asi.21149servicesprovideconstantconnectivityamongpeoplethatispreviouslyunparalleled.Therearenumerousopenquestionsconcerningtheoverallimpactofthesesocialcommunica-tionplatforms.Inthisstudyweinvestigatetheeffectsofservicesinthecommercialsector,namely,theimpacton becomeuncontrollablebecauselittleornotoolsareavail-abletomanagethecontent”ow(Ennew,Banerjee,&Li,2000).However,brandmanagementistransformingascom-municationtechnologychanges.Althoughsimilartoearlierformsofword-of-mouth,eWOMoffersavarietyofmeanstoexchangeinformation,manytimesanonymouslyorcon“-dentially,aswellastoprovidegeographicalandtemporalfreedom;moreover,eWOMhasatleastsomedegreeofpermanence(Gelb&Sundaram,2002;Kiecker&Cowles,2001).Assuch,eWOMisseenasincreasinglyimportantbybusinessesandorganizationsconcernedwithreputationman-agement.CorporationsandotherorganizationsarewrestlingwithhoweWOMbrandingwillaffectexistingprocesses,suchastrademarks(Goldman,2008).OnepotentiallynewformofeWOMmarketingismicrobloggingusingWebsocialcommunicationservicessuchasTwitter.Oneparadigmforstudyingthecon-stantconnectivityofmodernsocialnetworkingservicesinthecommercialareaiscalledtheattentioneconomy(Davenport&Beck,2002),wherebrandsconstantlycom-petefortheattentionofpotentialcustomers.Inthisattentioneconomy,microbloggingisanewformofcommunicationinwhichuserscandescribethingsofinterestandexpressatti-tudesthattheyarewillingtosharewithothersinshortposts(i.e.,microblogs).Thesepostsarethendistributedbyinstantmessages,mobilephones,email,ortheWeb.Givenitsdis-tinctcommunicationcharacteristics,microbloggingdeservesseriousattentionasaformofeWOM.Microblogsareshortcommentsusuallydeliveredtoanetworkofassociates.Microbloggingisalsoreferredtoasmicro-sharing,micro-updating,orTwittering(fromTwitter,byfarthemostpopularmicrobloggingapplication).Tweets(shortposts)mayenterourlexiconjustasXeroxhasforcopyingandGooglehasforsearching.Forthispaper,wewillrefertothisphenomenonasmicroblogging.MicrobloggingdirectlyimpactseWOMcommunicationbecauseitallowspeopletosharethesebrand-affectingthoughts(i.e.,senti-ment)almostanywhere(i.e.,whiledriving,gettingcoffee,orsittingattheircomputer)toalmostanyoneconnectedŽ(e.g.,Web,cellphone,IM,email)onascalethathasnotbeenseeninthepast.Whiletheshortnessofthemicroblogkeepspeoplefromwritinglongthoughts,itispreciselythemicropartthatmakesmicroblogsuniquefromothereWOMmedi-ums,includingfullblogs,WebPages,andonlinereviews.Astandardmicroblogisapproximatelythelengthofatypi-calnewspaperheadlineandsubhead(Milstein,Chowdhury,Hochmuth,Lorica,&Magoulas,2008),whichmakesiteasytobothproduceandconsume.Themessageisalsoasyn-chronousnoninvasive,sinceonecanchoosewhotoreceiveupdatesfrom.TheyarealsoarchivalinthesensethatthesemicroblogspermanentlyexistandaresearchableviaWebsearchenginesandotherservices.Sincetheyareonline,theyarealsotypicallyaccessiblebyanyonewithanInter-netconnection.Inshort,thesemicro-brandingcommentsareimmediate,ubiquitous,andscalable.ForeWOM,thesemicroblogsofferimmediatesentimentandprovideinsightinaffectivereactionstowardproductsatcriticaljunctionsofthedecision-makingandpurchasingprocess.Inthisstudyweexaminetheexpressionsofbrandattitudesinmicroblogpostings.ReviewoftheLiteraturePriorresearchhasshownthatWOMhasparticularlysig-ni“cantin”uencesonnewconsumerpurchasesofproductsorservices(Engel,Blackwell,&Kegerreis,1969;Katz&Lazarsfeld,1955).eWOMisaformofthiscommunication,de“nedasa:statementmadebypotential,actual,orfor-mercustomersaboutaproductorcompany,whichismadeavailabletoamultitudeofpeopleandinstitutionsviatheInternetŽ(Hennig-Thurau,Gwinner,Walsh,&Gremle,2004,p.39).eWOMmaybelesspersonalinthatitisnotface-to-face(ormaybejustpersonalinadifferentwaythaninthepast),butitismorepowerfulbecauseitisimmediate,hasasigni“cantreach,iscrediblebybeinginprint,andisacces-siblebyothers(Hennig-Thurauetal.,2004,p.42).IntermsofimmediacyofeWOMbranding,microbloggingcanoccurverynearthepurchasedecisionorevenduringthepurchaseprocess(Barton,2006).Thus,microblogginghassigni“cantimplicationsforthesuccessofadvertisers,businesses,andproductsasaneweWOMcommunications,andunderstand-ingtherami“cationsofmicrobloggingiscriticalforthesestakeholders.OnecanconceptuallyvieweWOMexpressionsasutter-ances.Grice(1969)theorizedthatonecoulddeducemeaningincommentsbyexaminingtheunderlyingintentions.Theintentionsmightbetoshareinformation,seekinformation,offeropinions,etc.ThisrelatestotheworkofAllen&Perrault(1986),whopostulatedthattheworldŽisasetofpropositionsinvolvingactions,plans,andspeech.Speechiscomposedofutterances.Theseutterancescouldinform,warn,assert,orpromise.Sundar(2008)statedthatmanypeo-pleexperiencetheworldthroughtheirownself-expressionandtheexpressionsoftheirpeers,whichblursthetradi-tionalboundarybetweeninterpersonalandmasscommu-nication.Asmediabecomesmoreinteractive,multimodal,andnavigable,thereceivertendstobecomethesourceofAlthoughtherearenostudiesatthistimeonmicroblog-gingaseWOMcommunication,priorworkhasexaminedotherformsofeWOMexchanges.Montoya-Weiss,Voss,andGrewal(2003)examinedwhatdrovecustomerstouseanonlinechannel(i.e.,aWebsite)inamultichannelenvironmentthatincludedof”inechannels(i.e.,brickandmortarstore).TheresearchersconcludedthatWebsitedesigncharacteristicsaffectedcustomerevaluationsofonlinechannelservicequal-ityandrisk,which,inturn,droveonlinechanneluse.Nardi,Schiano,Gumbrecht,&Swartz(2004)investigatedwhatcausedpeopletoexpressthemselvesonline.Theyreported“vemajormotivationsforblogging:documentingoneslife,providingcommentaryandopinions,expressingdeeplyfeltemotions,articulatingideasthroughwriting,andformingandmaintainingcommunityforums.2170JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi Goldsmith&Horowitz(2006)investigatedtheconsumermotivationsforonlineopinionseeking.Theresearchersreporteddistinctfactors,includingriskreduction,popularity,loweringcosts,easyinformation,accident,perception,inspi-rationfromof”ineinputssuchasTV,andprepurchaseinformationacquisition.Thorson&Rodgers(2006)testedtheeffectsofaninteractiveblogonattitudestowardaparticu-larWebsite.Focusingoninteractivity,perceivedinteractivity,andsocialinteraction,participantsinthehighinteractivityconditionreportedsigni“cantlyhigherscoresonperceivedinteractivity,demonstratingthatperceptionsofinteractivitywerein”uencedbythepresenceofasingleinteractiveele-mentonaWebsite.TheresearchersconcludedthattherewerepersuasiveeffectsresultingfromprovidinganopportunityforWebsitevisitorstosharetheirthoughtsandopinions.DavisandKhazanchi(2008)evaluatedtheimpactofeWOMattributesandfactorsone-commercesalesusingreal-worlddatafromamultiproductretaile-commerce“rm.TheresearchersvalidatedaconceptualmodelofeWOManditsimpactonproductsales.TheirresearchshowedthattheinteractionsamongeWOMpostings,productcategory,volumeofpostings,andproductwerestatisticallysigni“-cantinexplainingchangesinproductsales.Cheung,Lee,&Rabjohn(2008)examinedtheextenttowhichpeoplewerewillingtoacceptandadoptonlineconsumerreviewsandthefactorsthatencouragedadoption.Theresearch“nd-ingsreportedcomprehensivenessandrelevancetobethemosteffectivecomponentsofonlinepostings.Park&Lee(2009)reportedthatnegativeeWOMhadagreatereffectthanpositiveeWOM.RelatedtoeWOMcommunicationissentimentanalysisoropinionmining,Zhang,Yu,&Meng(2007)statedthatopinionminingrequiredtheretrievalofrelevantdocumentsandthenrankingthosedocumentsaccordingtoexpressedopinionsaboutaquerytopic.Certainly,though,onecouldbeinterestedinaspectsotherthanarankedlist.Liu,Hu,&Cheng(2005)developedanapplicationforanalyzingandcomparingconsumeropinionsforasetofcompetingproducts.Wijaya&Bressan(2008)leveragedthePageRankalgorithmtomea-suremoviesbasedonuserreviews.Theirresultscomparedfavorablywiththeactualboxof“cerankings.Lee,Jeong,&S.Lee(2008)presentedasurveyofthevarioustechniquesforopinionmining.Focusingonblogs,Conrad,Leidner,&Schilder(2008)developedmethodsfordetectingtheauthor-ityofthosemakingopinions.Archak,Ghose,&Ipeirotis(2007)examinedonlineproductreviewsinordertoidentifyspeci“cproductcharacteristicsandthenweighteachintermsofimportancetocustomers.AlthoughcertainlyrelatedtoprioreWOMresearch,therehasbeenlimitedpublishedworkinthemicrobloggingarea.McFedries(2007)presentsashortoverviewofmicroblog-ging,commentingthatonegoalmaybetoenhanceonescyberspacepresence.Java,Song,Finin,&Tseng(2007)studiedthetopologicalandgeographicalpropertiesofTwit-terssocialnetwork.Theresearchersfoundthatpeopleusedmicrobloggingtotalkabouttheirdailyactivitiesandtoseekorshareinformation.Milsteinetal.(2008)reviewedgeneralbackgroundonTwitterandmicroblogging.Ebner&Schiefner(2008)andGrosseck&Holotescu(2008)exam-inedmicroblogginginaneducationalsetting.Therearenumerouspopularpressarticlesonusingmicrobloggingapplications,mainlyTwitter,forbrandingandrelatedpur-poses(cf.,Brogan,2008;Postman,2008;Thompson,2008).Focusingonthesocialnetworkingaspects,Huberman,Romero,&Wu(2009)focusedonascarcityofattentionanddailyactivitiesthatchanneledpeopleintointeractingwithonlyafewpeople,whichtheirstudyofTwitterboreout.Fromareviewoftheliterature,itisapparentthateWOMisanimportantaspectofaconsumerexpressionofbrandsat-isfactionandmayhavecriticalimpactonabrandsimageandawareness.eWOMshowsallthesignsofbecomingevenmoreimportantinthefutureasthesesocialnetworkingappli-cationsbecomemorewidespread.ItisalsoapparentthatmuchofthefocusofprioreWOMresearchhasbeenonblogs,customerreviewsites,andWebPages.ThesearecertainlyaspectsofeWOM,buttherehasbeenlittlepriorworkinthemicrobloggingarea.Microbloggingisbecomingincreas-inglyimportantduetoitsimmediacytotheproducteventandtheincreasinguseofmicrobloggingbyawidergroupofpotentialcustomers.Assuch,microbloggingwillprobablyhaveincreasingin”uenceoneWOMbrandingefforts.Assuch,thereareseveralfundamentalyetunansweredquestionsconcerningmicrobloggingaseWOM.Howpreva-lentarebrandingmicroblogs?Howdopeoplestructurethesemicroblogs?Whattypesofbrandingsentimentdothesemicroblogsexpress?Whataretheireffectsononlinerep-utationmanagement?Whataretheimplicationsforbrandmanagers?Thesearethequestionsthatmotivateourresearch.Whyarecompaniesconcernedabouttheseonlineformsofexpression?Itisbecausetheycanaffect(perhapsbothposi-tivelyandnegatively)thebrandimageofcompany.Figure1presentsaclassicalmodelofbranding.InFigure1ageneralmodelofbrandingEsch,Langner,Schmitt,&Geus(2006)isshownalignedwiththerea-sonableeffectsofeWOMmicroblogs.Eschetal.(2006)evaluatedabrandingmodelintheonlinebrandingenviron-ment.Theyreportedthatcurrentpurchaseswereaffectedbrandimagedirectlyandbybrandawarenessindirectly.ThesetwocomponentsofbrandknowledgeseemtobetheprimaryareaswhereeWOMmicrobloggingwouldhaveadirectin”uence.Eschetal.(2006)alreadyfoundthatbrandknowledgeaffectedfuturepurchasesviaabrandrelationship(whichincludesbrandsatisfaction,brandtrust,andbrandattachment).Itwouldagainappearthatmicrobloggingcouldhaveanimportantin”uenceinthisarea,requiringbrandman-agerstoactivelyengageinthemicrobloggingspace,giventhatWOMcommunicationhasbeenassociatedwithbrandsatisfaction(Brown,Barry,Dacin,&Gunst,2005).Eschetal.(2006)conjecturedthatconsumersengagedinrela-tionshipswithbrandsinamannersimilartothepersonalrelationshiptheyformedwithpeople.Thesebrandrelation-shipsmaybetheresultofparticipationinbrandcommunities(Muniz&OGuinn,2001).Similartosocialnetworks,itseemsthatmicrobloggingapplicationscanhavepositiveandJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092171DOI:10.1002/asi BrandAwareness BrandImage BrandTrust BrandSatisfaction BrandAttachment CurrentPurchase FuturePurchases BrandKnowledgeBrandRelationshipBehavioralOutcomes Micro-blogpostings may effect brand knowledge(directly brandimage) WOM communications are typically associated with brand satisfaction.A companys monitoring and responses to posts and management of their micro-blog accounts may effect brand relationship (specifically brand satisfaction and trust) FIG.1.Generalmodelofbrandingcomponentsandrelationshiptomicroblogging.negativeimpactsasconsumersengageinthebrandcommu-nities.ThepossibleeffectofmicrobloggingviaeWOMonthebrandknowledgeandbrandrelationshipisthetheoreticalunderpinningfortheimportanceofourresearch.ResearchQuestionsWiththisstimulus,ourresearchquestionsare:WhataretheoveralleWOMtrendsofbrandmicroblogging?Toaddressthisresearchquestion,weselected50brandsandanalyzedthemicroblogsthatmentionedthesebrandsover13consecutiveweeks.Wealgorith-micallyanalyzedtheexpressionsorsentimentsofthesemicroblogsandcategorizedthemtodetermineaggre-gatecharacteristicsofbrandmicrobloggingexpression.Wealsoselectedasampleofthesemicroblogsandqual-itativelycodedthesentimentinordertodetermineanaccuracylevelforourautomatedmethods.Thisseriesofanalysisprovidedinsightintotheoverallsentimenttypesandtrendsinbrandmicroblogging.Whatarethecharacteristicsofbrandmicroblogging?Toaddressthisresearchquestion,wequantitativelyanalyzedthemicroblogsfromthe50selectedbrandstodeterminetheirbrandingeWOMsentiment,characteris-tics,andstructureinordertoshedlightontheirunderlyingaffective,cognitive,andcontextualaspects.Weexaminetweetsatthetermandterm-pairlevel,aswellasthestrengthoftermassociationusingthemutualinforma-tionstatistic.Suchananalysiswillprovideinsightintothestructurecharacteristicsofmicroblogs.Whatarepatternsofmicrobloggingcommunicationsbetweencompaniesandcustomers?Toaddressthisresearchquestion,weselectedandcloselyexaminedhowonecompanyusesitscorporateTwitteraccounts.Speci“cally,weanalyzedthecharacteristicsofhowthecompanycommunicatedwithcustomersthroughTwitterandemployedtheseaccountsasbrandmanagementandeWOMtools.Workingfromanexploratoryapproach,weanalyzed,bothqualitativelyandquantitatively,1,907microblogspostedbythecompanyortweetsaddressedtoitaswellasanalyzedthesocialnetworkoftheresultingcommunications.ResearchDesignToinvestigateourresearchquestions,weusedtheTwit-tersocialcommunicationplatform,oneofthemostpopularmicrobloggingservices.However,allmicrobloggingappli-cationsshareasetofsimilarcharacteristics:(1)shorttextmessages,(2)instantaneousmessagedelivery,and(3)sub-scriptionstoreceiveupdates.So,althoughweusedTwitterinthisresearch,weexpectourresultstobeapplicabletoothermicrobloggingapplications.TwitterLaunchedonJuly13,2006,Twitterisamicrobloggingservicewhereuserssendupdates(a.k.a.,tweets)toanetworkofassociates(a.k.a.,followers)fromavarietyofdevices.Tweetsaretext-basedpostsofupto140charactersinlength.Thedefaultsettingfortweetsispublic,whichpermitspeopletofollowothersandreadeachotherstweetswithoutgivingmutualpermission.EachuserhasaTwitterpagewherealltheirupdatesareaggregatedintoasinglelist(hencethenameTweetsarenotonlydisplayedonauserspro“lepage,buttheycanbedelivereddirectlytofollowersviainstantmessag-ing,ShortMessageService(SMS),ReallySimpleSyndica-tion(RSS),email,orothersocialnetworkingplatforms,suchasTwitterri“corFacebook.TheTwitterapplicationprogram2172JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi TABLE1.Brandsandproductsbyindustrysector. IndustrysectorMajorbrandProduct/serviceCompetitorKnownbrandOther ApparelH&MBananaRepublicTopShopAutomotiveToyotaPriusHondaSMARTForTwoMiniClubmanComputerhardwareDellLenovoAveratecMacBookAir,iPhoneComputersoftwareMicrosoftWindowsVistaLeopardWindows7ConsumerelectronicsSonyBRAVIAToshibaMagnavoxNintendo,WiiFitEnergyExxonSunocoFastfoodStarbucksDriveThroughMcDonaldsArbysFoodKelloggsSpecialKCheeriosMalt-O-MealInternetserviceGoogleGmailYahoo!KartOOAmazon,FacebookPersonalcareOral-BOral-BTriumphCrestAquafreshSportinggoodsAdidasAdidasOriginalsReebokSauconyTransportationFedExDHLForeverStamp interface(API)alsoallowstheintegrationofTwitterwithotherWebservicesandapplications.Asthelargestoneofthemicrobloggingservice,Twittersuserbasehasgrown,andithasattractedattentionfromcorporationsandothersinterestedincustomerbehaviorandservice.Givenitsrobust-ness,Twitterisincreasinglyusedbynewsorganizationstoreceiveupdatesduringemergenciesandnaturaldisasters.AnumberofbusinessesandorganizationsareusingTwitterorsimilarmicrobloggingservicestodisseminateinforma-tiontostakeholders(seewww.socialbrandindex.comforapartiallistofbrandsemployingTwitter).Twittersgrowthrateissubstantial,withseveralmillionsusersasof2008(Bausch&McGiboney,2008).WebapplicationssuchasTweetrush(tweetrush.com)estimatetraf“catapproximatelyamilliontweetsaday.Puttingthosenumbersinperspec-tive,fromAugust2006toAugust2008,Twitteruserscreated100,000booksworthofcontent,140charactersatatime(Milsteinetal.,2008).Asthelargest,mostwell-known,andmostpopularofthemicrobloggingsites,TwitterisanidealcandidateforourstudyofmicrobloggingsimpactintheeWOMarea.Identi“cationofBrandGiventhatwewereinterestedinmicrobloggingforbrand-ingpurposes,wehadtoselectkeybrandsforinvestigation.WeexploredseverallistsofbrandsontheWebincludingAmericanCustomerSatisfactionIndex,BusinessWeekTopBrand100,andBrandZTop100MostPowerfulBrandsToensureagoodcross-segmentsample,weemployedanindustryclassi“cationmethodusedbyBusinessWeektomakesurethebrandsspreadoutacrossmajorindustries,butwealsokeptthecategoriescloselyrelatedwithitemsindailylife,undertheassumptionthatthesebrandswouldbemostlikelymentionedinandaffectedbymicroblogging.Wealsocoun-terbalancedthiscross-industryapproachbytryingtoselect TheAmericanCustomerSatisfactionIndexhttp://tinyurl.com/24c8hjBusinessWeeksTopBrand100http://tinyurl.com/2e6ntjBrandZTop100MostPowerfulBrandsRankinghttp://tinyurl.brandsthatprovidedsimilarproductsorservicesinordertomakethemmorecomparable.Foreachindustrysector,weincludedonemajorbrand,oneproductfromthemajorbrand,onecompetitivebrand,onecomparablylesscompetitivebrand,andsomeothernews-worthybrands.Themajorbrandandtheproductfromthemajorbrandallowedustoexploretherelationshipbetweenbrandmanagementandproductmanagement.Themajorbrandandthecompetitivebrandenabledustocomparecompetitivebrandsandidentifypotentialmeanstousecom-petitorsbrandsentimentchangestodeveloporenhancethemajorbrandmanagement.Thelesscomparativemajorbrandcouldbetomorrowsmajorcompetitorsandcouldalsobeasourceoffailure,fromwhichwecouldlearntoavoid.Intheend,wesettledon50brands.ThebrandswiththeindustrysectionsandassociatedinformationareshowninTable1.DataCollectionandAnalysisForour“rstresearchquestion(WhataretheoveralleWOMtrendsofbrandmicroblogging?),wewereinterestedintweetsthatmentionedabrandnameandespeciallytheexpressionofopinionorbrandsentiment.WecollectedthesetweetsusingtheSummizetool.Summize4wasapopularserviceforsearchingtweetsandkeepingupwithemergingtrendsinTwitterinrealtime.LikeTwitter,SummizeoffersanAPI,sootherproductsandservicescan“ltertheconstantqueueofupdatesinavarietyofways.TheSummizeser-viceanalyzedtweetsentimentandgavethequeriedbrandanoverallsentimentratingforagivenperiodusinga“ve-pointLikertscaleandlabelingeachpointfromlowesttohighestaswretched,bad,so-so,swell,andgreat.Wecollecteddatafora13-weekperiod,fromApril4,2008toJuly3,2008.Thisgaveus650reportingepisodes(13reportingperiods50brands).Tocollectthedate,foreachbrandinTable1,wesubmittedqueriestoSummizeintheformat:[brandname]since:[startdate]until:[enddate]toretrievethesentimentforthatweekperiod.UsingBanana SummizewasacquiredbyTwitterinAugustof2008andisnolongeravailableasanindependentservice.Seehttp://tinyurl.com/56j2zxJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092173DOI:10.1002/asi FIG.2.SampleofSummizesgraphicalpresentationofthebrandsentiment.Republicasanexample,thequeryfortheweekfromApril4,2008toApril10,2008,isBananaRepublicsince:2008-04-04until:2008-04-10.Werepeatedthisprocessforthesamebrandthefollowingweekuntilwecollecteddataforallbrandsfromall13-weekperiods.Foreachbrandwethencalculatedtheclassi“cationofeachtweet.Summizeusesalexiconof200,000uni-gramsandbi-gramsofwordsandphrasesthathaveaprobabilitydis-tributiontodeterminethesentimentofthebrandforagivenperiod.Summizetraineditsclassi“erwithnearly15mil-lionviewsontopicsrangingfrommoviestoelectronics,etc.Theobjectivewastodeterminehowpeopleuseadjectivesinonlineutterances.Eachfeature(word)hassomeprob-abilityofbeingusedeitherpositivelyornegativelyandisclassi“edintooneofthe“veLikertscaleclasses.Theclassi-“erisamultinominalBayesmodeltodeterminetheoverallsentimentofeachsetofTwitterposts.ThemultinominalBayescodepickstheclasswiththegreatestprobabilityinawinner-take-allscenario.Inresponsetoagivenquery,theclassi“erprocessesthelatest125…150tweetsfromtheresultset.Duplicatetweetsareignoredandnon-Englishtweetsare“lteredout.Theclassi“eranalyzesthewholetweettodetermineitssentiment.Summizepresentsagraphicalpre-sentationofthebrandsentimentasshowninFigure2.Each Foreachofthesebrandsanddatacollectionperiods,weretrievedtheactualtweetsfromtheTwitterAPIusingasimilarapproachduetotheacquisitionofSummizebyTwitterduringthedatacollectionperiod.coloristhepredominantsentimentforagivensentenceweareInordertoverifytheaccuracyofthealgorithmicclassi-“er,wemanuallycodedtweetsfrom“vebrands(i.e.,BananaRepublic,SMARTForTwo,WiiFit,Google,andForeverStamp).Todothis,wedownloadedthemostrecent250tweetsforthese“vebrands.Wecomparedthesentimentofthe“rstsetof125tweetstothesecond125tweetsforeachbrand.Wealsocomparedtheresultsofthemanualclassi“cationsofthelatest125totheautomaticclassi“cations.Inordertoconductthisanalysis,wedevelopedthemanualcodingschemebelowbyfollowingGlaser&Strausss(1967)codingdevelopmentstrategy:NoSentiment:Tweethasnoemotionwordsorspecialpunc-tuation,ismatter-of-factsounding,orcontainsjustabrandmention(e.g.,WonderingwhattimetheBananaRepublicstoreatthemallcloses).Wretched:Tweetispurelynegativeoverallfeelingsoronlyallowedaslightlypositiveword.Foraproduct,wouldntbuyitagain,Žwouldntrecommendit,Žorhadhorribletimewithit,Ž(e.g.,ScrewyouGooglemaps.ItsagoodthingIhavethiscompassandsharpstick).Bad:Tweetcontainsmainlynegativephrasesandwords,withadisappointedtone.Theremaybeafewpositivestatements,butthenegativefeelingsoutweighthepositiveones(e.g.,SittingnexttoasmartcarŽintraf“c.Thesethingsjustlookweird.Aboutaslongasarickshaw).So-so:Thetweetisamediocreorbalancedsentiment.Thepositiveandnegativestatementsseemtobalanceeachother,2174JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi oritisneitherpositivenornegativeoverall.Eveniftherearemorenegativephrases,thepositiveonesuseastrongerlanguagethanthenegativeones(e.g.,Wii“tis“ne,justleaveenoughroomaroundyoutowaveyourarms!).Swell:Thetweetismainlypositiveterms,suchasgoodornice.Theremaybesomenegativephrases;however,thepos-itiveonesarestrongerandoutweighthenegativeones(e.g.,Youmighthavethoseforeverstampsthatareallgoodnomatterthepriceofacurrentstamp).Great:Purelypositiveintoneandwordinginthetweetexpressingstrongaf“rmativefeelingswithnocomplaints.Itmayhavethesmallestnegativeword,butthetweethasmostlygreat-soundingphrases.Foraproduct,thecommentsare:wouldde“nitelyrecommendit,Žuseagainit,Ž(e.g.,Heavenonearth,theBananaRepublicoutletstore40%offsale).Asabaselineforcomparison,wealsocollected14,200randomtweetsfromTwitterandevaluatedthesetweetsformentionofbrand.Toaddressresearchquestiontwo(Whatarethecharacter-isticsofbrandmicroblogging?),weperformedalinguisticanalysisofthetweets.Weuploadedalltweetsintoarela-tionaldatabaseandgeneratedatermtableandalsoatermco-occurrencetableforeachsetoftweets.Thetermtablecontained“eldsforterms,thenumberofthattermsoccur-renceinthecompletedataset,andtheprobabilityofthattermsoccurrence.Theco-occurrencetablecontains“eldsfortermpairs,thenumberoftimesthatpairoccurswithinthedatasetirrespectiveoforder,andthemutualinformationstatistic(Church&Hanks,1990).Tocalculatethemutualinformationstatistic,wefollowedtheprocedureoutlinedbyWang,Berry,&Yang(2003).Themutualinformationformulameasurestermassociationanddoesnotassumemutualindependenceofthetermswithinthepair.Wecalculatedthemutualinformationstatisticforalltermpairswithinthedataset.Manytimesarelativelylow-frequencytermpairmaybestronglyassociated(i.e.,ifthetwotermsalwaysoccurtogether).Themutualinformationstatisticidenti“esthestrengthofthisassociation.Themutualinformationformulausedinthisresearchis: )areprobabilitiesestimatedbyrelativefrequenciesofthetwowordsand)istherelativefrequencyofthewordpair(orderisnotcon-sidered).Relativefrequenciesareobservedfrequencies(normalizedbythenumberofthequeries: Q;P(w2)=F2 Q;P(w1,w2)=F12 Boththefrequencyoftermoccurrenceandthefrequencyoftermpairsaretheoccurrenceofthetermortermpairwithinthesetofqueries.However,sinceaone-termquerycannothaveatermpair,thesetofqueriesforthefrequencybasediffers.Thenumberofqueriesforthetermsisthenumberofnonduplicatequeriesinthedataset.Thenumberofqueriesfortermpairsisde“nedas:isthenumberofquerieswithwords(1),andisthemaximumquerylength.So,queriesoflengthonehavenopairs.Queriesoflengthtwohaveonepair.Queriesoflengththreehavethreepossiblepairs.Queriesoflengthfourhave“vepossiblepairs.Thiscontinuesuptothequeriesofmaximumlengthinthedataset.Theformulaforqueriesoftermpairs()accountsforthistermpairing.Forresearchquestionthree(Whatarepatternsofmicrobloggingcommunicationsbetweencompaniesandcus-tomers?),weconductedacasestudyonaspeci“ccompanyandthoroughlyanalyzedthetweetsbetweenthecompanyTwitteraccountsandthefollowersoftheseaccounts.ThecompanyweselectedfromTable1wasStarbucks,whichhasproductsandservicescloselyrelatedwitheverydaylife(i.e.,coffeeandpastries)andhasactivetwitteraccounts.Starbucksisaworld-famouscoffeehousechain(Wikipedia,2009).EnglishteacherJerryBaldwin,historyteacherZevSiegel,andwriterGordonBowkerfoundedStarbucksin1971,withits“rststoreinPikePlaceMarketinSeattle,Washington(McGrawHillHigherEducation,n.d.)Thecom-panyaimstobethepremierpurveyorofthe“nestcoffeeintheworldŽ(Starbucks,2008).Itsproductsincludecoffee,handcraftbeverage,freshfood,coffee-relatedmerchandise,andrelateditems(Starbucks,2008).StarbucksisactiveontheWebandiskeenonbuildingupitsonlinecommunity,asevidencedbyaccountsonTwitter,Facebook,andYouTube.ItalsobuiltupitsownonlinecommunitieslikeMyStar-bucksIdea(mystarbucksidea.force.com)forcollectingideastoimprovetheirproductsandservice,andStarbucksV2V(www.v2v.net/starbucks)forgettingpeopletogethertovol-unteerforcommunitywork.Assuch,Starbucksappearstobeacompanyinterestedinthesocialnetworkingcommunities.Starbucksownsthreetwitteraccounts:Starbucks(twitter.com/Starbucks),MyStarbucksIdea(twitter.com/MyStarbucksidea),andStarbucksV2V(twitter.com/StarbucksV2V).SinceMyStarbucksIdeaandStarbucksV2VhadfewertweetsfrompeopletheyfollowedthandidStar-bucks,wechosetostudythecompanyschiefTwitteraccount(i.e.,Starbucks)only.Thecompany“rstopenedthisTwitteraccounton12August2008.Its“rsttweetwasWelcometoStarbucksTwitterland!Ž(twitter.com/Starbucks/statuses/885677980).WecollecteddatafromthedaythecompanystartedusingTwitterto2November2008.Bythen,theyhad7,751followersandfollowed7,779Twit-terusers.Starbuckstwittered322timesandreceived1,585tweets.Weusedamixedmethodapproachandperformedbothquantitativeandqualitativeanalysisof1,907tweetsfromStarbucks,itsfollowers,andpeopleitfollowedtodeter-minethecharacteristicsofhowthecompanyemployedthisaccountasaneWOMmanagementtools.Weusedaction-objectpairapproach(Zhang&Jansen,2008)toqualitativelyanalyzethemicroblogs,anapproachJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092175DOI:10.1002/asi FIG.3.Objectcodes. Wordsinblackarenotcodesandareusedheresolelyforlabelingpurposes.2176JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi TABLE2.Actioncodesandde“nitions. ActionDe“nition AnnouncementDeclaringtheupcomingobjectsAnswerHandlingquestionChitchatCasualconversationCommentExpressingmixedorneutralfeelingsregardingobjectsCon“rmationGivingassuranceorvalidationregardingobjectsConsumingDrinkingoreatingobjectsExpectingLookingforwardtoobjectsfromStarbucksForwardingPointingtopotentialusefulobjectsMaintenanceManagingobjectsMissingFeelingfromthelackofobjectsandexpectingtohavethembackNegativecommentCritiquing,complainingNoti“cationLettingoneknowonobjectsOrderviaTwitterAttemptingtoplaceorderonTwitterPatronizingPhysicallybeinginobjectsorgoingtoobjectsfrequentlyPositivecommentComplimenting,praisingQuestionExpressingconfusionsordoubtstowardobjectsRecommendationProvidingpositiveadviceregardingobjectsRecommendationSeekingadviceregardingobjectsRequestAskingforobjectsResearchExaminingobjectsResponseGivingunnecessaryfeedbackonobjectsSuggestionProvidingideastoimproveobjectsSupplementAddingontoobjects originallydevelopedfortransactionloganalysis.Anactionisaspeci“cexpressiontotheobject,andanobjectisaself-containedinformationobject.Thesetwocomponentstogetherformanaction-objectpairandrepresentoneinter-actionbetweenuserandsystem.Weextendedtheconceptofthismethodandappliedittomicroblogginganalysis.Inourscenario,theobjectismaterialrelevanttoStarbucks,forexample,itscoffee,itsservice,anditspromotion.TheactionisanexpressionaseWOMconcerningtheobjects.Forexample,theactioncanbecriticizing,askingaques-tion,orprovidingasuggestion.Thus,theactionandobjecttogethermakesanaction-objectpair,andathreadofaction-objectpairscantellastoryregardingaspectsoftheStarbucksbrandandcustomersatisfactionwiththatbrand.Forexample,criticizing-coffeepairindicatesthecustomersdissatisfactiontowardStarbuckscoffee.Weusedanopencodingapproachtodevelopcodingschemaforactionandobject.Onetweetcanbecodedwithmultipleaction-objectpairs.Thecom-pleterelationshipofcodesforobjectsisinFigure3,andthecompletelistofcodesforactionsisinTable2.ResearchQuestion#1:WhatAretheOveralleWOMTrendsofBrandMicroblogging?Fromananalysisof149,472tweetscollectedover131-weekperiodsfor50brands,wecalculatedtheresultingsen-timentanalysisbyweekperbrand,asshowninTable3.FromTable3,morethan60%oftheaggregateweeklysentimentsTABLE3.Brandsentimentbyweek. SentimentbyweekOccurrencesPercentage Great19429.8%Swell20030.8%So-so7812.0%Bad10215.7%Wretched426.5%NoTweets345.2%Total650100.0% forthebrandswerepositive(GreatorSwell).Justover22%ofthesentimentbybrandwerenegative(BadorWretched).Asmallpercentage(12%)wasneutral(So-so),andanevensmallerpercentageofthebrands(5%)hadnotweetsinagivenweek.So,althoughthepositivetweetsrepresentedthelargestquantity,therewereasubstantialpercentageofnegativetweets.Priorresearchintheimpressionformulationliteraturehasshownthatnegativecommentshaveagreaterimpactthanpositiveones(Skowronski&Carlston,1989).TheanalysisinTable3isbasedontheSummizealgo-rithmicanalysis.However,wewantedtoevaluatewhetherthisalgorithmicapproachwaseffectivebecausewecouldnotlocateanypublishedworksevaluatingtheautomaticsen-timentanalysisoftweets.Plus,Summizeusesonlythemostrecent125tweetsforthetimeperiodtodeterminetheoverallsentiment.However,thisraisesthequestionofwhetherthisisanaccuratesampletodeterminesentimentforagivenweek.Wewantedtoseeifthisapproachaffectedourresults.Inordertoinvestigatethis,wedownloadedthetop250tweetsof“vebrandsover4randomweeksforatotalof2,610tweetsinordertocomparetheautomaticcodingmethodofSummizetothesametweetsmanuallycodedmethod.Wecodedeachtweetforsentimentusingthecodingschemaoutlinedaboveandcomparedthesentimentresultsfromthe“rst125tweetstothesecond125tweets.Twocodersindependentlycodedeachtweetusingthecodingschemeoutlinedabove.Inter-coderreliabilitywasquitehigh(Cohens0.85).Usingthegenerallinearmodel(Dobson,2002),weidenti“ednodiffer-encebetweenthesentimentdistributionofthe“rst125andsecond125tweetsbybrand(0.45).Therefore,wearereasonablyassuredthatthemostrecent125tweetsarearepresentativesampleoftweetsforgivenweeks.Usingthesentimentofthe“rst125tweetstorepresenttheoverallsentiment,wethencreatedablockdesignandtreatedbrand,sentimentcategory,andtime(i.e.,week)asblockingfactorsbecausetheycouldpotentiallyin”uencethesenti-mentpercentage.Wecomparedmanualcodingwithsystemcodingbyusingthegenerallinearmodel.Therewasnodiffer-encebetweenmanualcodingandsystemcoding(1.00).Therefore,our“ndingsfromthesetwoevaluationsshowthatouralgorithmicapproachisaccurateinclassifyingthetweets.Theimplicationisthatonecanusestandardtextclassi“cationmethodstoanalyzedtweets,andnewmethodsarenotrequired.Inadditiontotheoverallsentiments,weperformedadetailedanalysisnotattheweeklevelbutforthespeci“cJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092177DOI:10.1002/asi TABLE4.Analysisofindividualtweetsforsentiment. Total(all)Percentage(all)(sentimentonly) Great9,4516.3%33.0%Swell5,5583.7%19.4%So-so4,0712.7%14.2%Bad4,5503.0%15.9%Wretched5,0323.4%17.6%Nosentiment120,81080.8%Total149,472100.0%28,662(19.2%) TABLE5.Sentimentchangesbyweek. ChangeOccurrencePercentage Changetonegative18230.3%Changetopositive18430.7%Nochange19532.5%Notweetstonegativetweets81.3%Notweetstopositivetweets132.2%Tweetstonotweets183.0%Total600100.0% tweetsthatmakeupthissentiment.ResultsarereportedinTable4.AsonecanseefromTable4,morethan80%ofthetweetsthatmentionedoneofthesebrandsexpressednosenti-ment.ThisindicatesthatpeopleareusingTwitterforgeneralinformation,askingquestions,otherinformation-seekingand-sharingactivitiesaboutbrandsorproducts,inadditiontoexpressingopinionsaboutbrandsorproducts.Ofthe268,662tweetsexpressingsentiment,morethan52%oftheindividualtweetswereexpressionsofpositivesentiment,whileoftweetswerenegativeexpressionsofopinion.ThisisinlinewithpriorworksuchasthatofAnderson(1998),whoshowedthattherewasaU-shaperelationshipbetweencustomersatis-factionandtheinclinationtoengageinWOMtransfers.Thissuggeststhatextremelypositiveandsatis“edandextremelynegativecustomersaremorelikelytoprovideinformationrelativetoconsumerswithmoremoderateexperiences.However,therearedissimilarmotivationsbehindpositiveWOMandnegativeWOMutterances(Anderson,1998).ThemajorincentiveforpeopletospreadpositiveWOMistogainsocialorself-approval.TheirpositiveWOMutterancesdemonstratetheirsplendidpurchasedecisions.Additionally,altruisticbehaviorofsharingexpertisewithothershasalsobeenshowntomotivatepositiveWOM(Fehr&Falk,2002;Richins,1984).Hostility(Jung,1959;Kimmel,2004)andvengeance(Richins,1983)motivatesdissatis“edconsumerstoengageinnegativeWOM.Infurtherinvestigationofthemicrobloggingtrends,weexaminedhowsentimentchangedfromweektoweek.Table5showsthechangesinsentimentfromweektoweekforeachofthe50brandsoverthe13-weekperiod.Beginningwiththestartingweekforeachbrand,wethencalculatedthechange.Weseethat32%ofthetimetherewasnochangefromoneweektothenext.Morethan64%ofthetime,therewasachangeinsentimentorachangetonotweets.Basedonpriorwork,thispropensityofmicroblogstochangecategorieshasimportantimplicationsforbusi-nesses.Benedicktus&Andrews(2006)reportedthattherewaslimitedlong-termeffectifreputationdidnotdeclinetoalowercategory(e.g.,fromaveragetopoor)andthatmanymoreperiodsofpositivecommentswererequiredtorebuildtrustthanwererequiredtodamageit.However,mostofthechangesweretoadjacentcategories.Sothechangescouldbetheby-productofusingtheLikertscale.Soifabrandwasrightoftheedgeoftwocategories,afewtweetseitherwayinagivenweekcouldmoveabrandsentimentclassi“cationfromonecategorytotheadjacentcategory.Inorderforacompanytoperformbrandmanagement,itisimportanttoknowpeoplesopinionsaboutbrandsandproducts.Itiscriticaltorecognizethecompanyspositioninthemarket,especiallyinitsownindustrysector,andtocomparewithitscompetitors.Therefore,wecomparedthesentimentofbrandswestudiedineachof12industrysectors.Westatisticallycomparedthebrandswithineachofthe13industrysectorstodetermineifthereweredifferencesamongthebrands.Table6presentstheresults.Wefoundstatisticallysigni“cantdifferencesbetweenbrandsinsevenindustriesincludingautomotive,computerhardware,computersoftware,consumerelectronics,food,personalcare,andsportinggoods.Wealsoconductedpost-hocanalysisbyusingaTukeytestata5%family-wiseerrorratetoidentifytheexactdifferencesbetweenbrands.Inauto-motive,MiniClubmanhasadifferentsentimentthanHondaandSmartCar.Incomputerhardware,AveratechasadifferentsentimentthanMacBookAir,iPhone,andLenovo.Incom-putersoftware,WindowsVistaandWindows7havedifferentsentimentclassi“cations.Inconsumerelectronics,Magnavoxhasadifferentsentimentthantherestofthatgroup.Kelloggsandtherestofthebrandsinthefoodcategoryhavediffer-entsentiments.Inpersonalcare,CresthasadifferentbrandsentimentthanAquafreshandOral-BTriumph.Insportinggoods,thesentimentofAdidasOriginalsisdifferentthantherestofthebrands.OverallresultsofthisanalysisareshowninTable6.Thisdifferentiationamongbrandswithinanindustrysec-tionshowsthatmicrobloggingaseWOMisapromisingmeasureforcompaniestouseforcompetitiveintelligence.Companiescanalsousemicrobloggingasapartoftheirmar-ketingcampaignstoattempttodifferentiatethemselvesfromtheircompetitors.Sincethissetof149,472microblogpostingsallcontainedbrandingcommentsidenti“edbySummize,wewantedadatasetasabaseofcomparison.Wedownloaded14,200ran-domtweetsviatheTwitterAPI.Wedownloaded1,092tweetsfromeachweekofthedatasampleperiodusedabove.Weanalyzedeachofthese14,200tweetsforoccurrencesofmentionsofabrandorproduct.Afterimportingour14,200tweetsintoarelationaldatabase,wequeriedthedatabasewithour50brands,identi“ed,andlabeledalltweetsthat2178JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi TABLE6.Brandsentimentcomparisoninindustrysector. Post-hoca IndexIndustrySectorBrand-valueDF-valueBrandLevelALevelB 1ApparelBananaRepublic,H&M,TopShop1.1720.328Nosigni“cantdifferenceamongbrandsentimentacross2AutomotiveHonda,MiniClubman,Prius,4.1640.006*Honda,SmartForTwoASmartForTwo,ToyotaToyota,PriusABMiniClubmanB3ComputerAveratec,Dell,iPhone,Lenovo,5.6140.001*MacBookAir,iPhone,LenovoAHardwareMacBookAirDellABAveratecB4ComputerLeopard,Microsoft,Windows7,5.1930.004*Windows7ASoftwareWindowsVistaMicrosoft,LeopardABWindowsVistaB5ConsumerBRAVIA,Magnavox,Nintendo,12.6150.000*WiiFit,BRAVIA,Nintendo,AElectronicsSony,Toshiba,WiiFitSony,ToshibaMagnavoxB6EnergyExxon,Sunoco0.0011.000Nosigni“cantdifferenceamongbrandsentimentacross7FastFoodArbys,McDonalds,Starbucks,1.7830.168Nosigni“cantdifferenceamongbrandsentimentacrossStarbucksDriveThrough8FoodCheerios,Kelloggs,Malt-O-Meal,5.7230.003*SpecialK,Cheerios,Malt-O-MealASpecialKKelloggsB9InternetServiceAmazon,Facebook,Gmail,0.7850.565Nosigni“cantdifferenceamongbrandsentimentacrossGoogle,KartOO,Yahoo!brands10PersonalCareAquafresh,Crest,Oral-B,6.8730.001*CrestAOral-BTriumphOral-BABAquafresh,Oral-BTriumphB11SportingGoodsAdidas,AdidasOriginals,21.7930.000*Adidas,Reebok,SauconyAReebok,SauconyAdidasOriginalsB13TransportationDHL,FedEx,ForeverStamp0.5220.599Nosigni“cantdifferenceamongbrandsentimentacross Levelsnotconnectedbysameletter(A,B)aresigni“cantlydifferent.mentionedthesebrands.Wethenusedanopencodingtech-niquewherewequalitativelyreviewedindividualtweets.Whenaproductorbrandmentionoccurred,wewouldquerytheentiredatasetforalloccurrenceofthisbrand(i.e.,amod-i“edsnowballtechnique[Patton,1990]).Codersreviewedeachofthesetweetstoverifythattheycontainedabrandmen-tion.Werepeatedthisprocessuntilalltweetsinthedatabasehadbeenexaminedandcoded.Ofthe14,200randomtweets,386tweets(2.7%)con-tainedmentionofoneofthebrandsorproductsfromourlist(Table1).Therewere2,700tweets(19.0%)thatmen-tionedsomebrandorproduct,inclusiveofthebrandsthatweusedinthisstudy.Therefore,microbloggingappearstobearichareaforcompaniesinterestedinbrandandcustomerrelationshipmanagement.Inadditiontodeterminingwhetherthesetweetsmentionedthebrands,weclassi“ed2,700tweetsintogeneralcategories,similartotheworkoutlinedpreviously(Broder,2002;Jansen,Booth,&Spink,2008;Rose&Levinson,2004)classifyingWebqueries.Weclassi“edthesebrandingtweetsintothefollowingfourcategories,againfollowingGlaser&Strausss(1967)codingdevelopmentstrategy.Sentiment:theexpressionofopinionconcerningabrand,includingcompany,product,orservice.Thesentimentcouldbeeitherpositiveornegative.InformationSeeking:theexpressionofadesiretoaddresssomegapindata,information,orknowledgeconcerningsomebrand,includingcompany,product,orservice.InformationProviding:providingdata,information,orknowledgeconcerningsomebrand,includingcompany,prod-uct,orservice.Comment:theuseofabrand,includingcompany,product,orservice,inatweetwherethebrandwasnottheprimaryfocus.Atweetcouldbecodedintomorethanonecategory.Forexample,atweetthatexpressessentimentcouldalsopro-videinformation,oratweetseekinginformationmayalsoprovideinformation.Therefore,thesecategorieswerehierarchical.Thatis,we“rstdeterminedwhetheratweetexpressedsentiment.Ifnot,wethenexaminedittoseewhetheritsoughtinformation.JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092179DOI:10.1002/asi TABLE7.Brandingtweetsbycategory. Classi“cationOccurrencesPercent Comments1,31048.5%Sentiment60222.3%Informationproviding48818.1%Informationseeking30011.1%Total2,700100.0% TABLE8.Linguisticstatisticsfortweets. Tweetlength(words)Tweetlength(characters) Tweet14.2K38.8K2.6K14.2K38.8K2.6Kmeasurestweetstweetstweetstweetstweetstweets Average15.414.315.886.389.1102.66.86.46.636.535.336.4Max333343142155185Min1131116 Ifitdidnotseekinformation,wethendeterminedwhetheritprovidedinformation.Commentswerethecatch-allcategory.TheresultsofthiscodinganalysisareshowninTable7.AscanbeseenfromTable7,mosttweetsthatmentionabranddosoasasecondaryfocus.Thesetweetsaccountforjustunderhalfofthebrandingtweetsinthissample.Usersexpressedbrandsentimentin22%ofthetweets.Interestingly,29%ofthetweetswereprovidingorseekinginformationcon-cerningsomebrand.Thisshowsthatthereisconsiderableuseofmicrobloggingasaninformationsource.Thiswouldindicateseveralavenuesforcompanies,includingmonitoringmicrobloggingsitesforbrandmanagement(i.e.,sentiment),toaddresscustomerquestionsdirectly(i.e.,informationseek-ing),andmonitoringinformationdisseminationconcerningcompanyproducts(i.e.,informationproviding).ResearchQuestion#2:WhatAretheCharacteristicsofBrandMicroblogging?Forthisresearchquestion,wedidalinguisticanalysisofthreesetsoftweets.Fromthe149,472microblogpost-ings,wedownloadedthe“rst100tweets(feweriftherewerefewerthan100tweets)foreachbrandduringeachweekofthedatacollectionperiod.Thisgaveus38,772tweetscontainingbrandingterms.Wealsoseparatelyexam-inedthe2,610tweets,performingqualitativeanalysistoensurethatthissampleadequatelyrepresentedtheover-allpopulation.Finally,weanalyzedthe14,200randomlydownloadedtweets.Webelieveacomparisonofthelinguisticallyanalyzedresultsfromthesethreedatasetswillprovideinsightintothesemanticstructureofbrandingtweets.FromTable8,onecanseethatthestatisticsfortweetsaresimilaracrossallthreesets.Theaveragewords-per-tweetisnearly16.Asacomparison,theaverageWebsearchqueriesisapproximatelythreeterms(Jansen&Spink,2005;Wangetal.,2003).ThelengthoftheaverageEnglishsentenceisabout25.Soattheaggregatetermlevel,tweetshavemoreincommonwithstandardwrittensentencesthanwithrelatedshortexpressions,suchasWebqueries.Oneofthesuc-cessesofthemicrobloggingserviceisthisshortnessofthemicroblog.Itmaybethatthetweetlengthisafamiliarlengthforinformationprocessing.Wethenexaminedtweetsatthespeci“ctermlevel,asshowninTable9.FromTable9weseethattweetscontainmanyoftheskipwords(e.g.,the,a,to,us,for,etc.)thatarecommoninnaturallanguageusage.However,therearesomehighoccurrencesofnonskipwords,whichisuncommoninnat-urallanguageexpressions.Nevertheless,thismaybeduetothefocusonbrandingandthedemographicsoftheTwit-teruserpopulation.Theaveragefrequencyofoccurrencerangedfrom4.3to12.1terms(standarddeviation[SD]rang-ingfrom40…166terms),re”ectingthevarioussizeofourdatasamplethatcausesthewidevariations.Whenexaminingtheaverageprobabilityofoccurrence,weseetheaverageinatightrangefrom0.0003…0.0005(SDrangingfrom0.0030…0.0060),re”ectingthenormalizationofthedatasetsizes.Theclusteringoftheprobabilityofoccurrenceindicatesthatoursamplesarerepresentativeofthesamepopulation.Our“ndingsshowaspreadoftermsthat,inextremelylargesam-plesizes,wouldprobablyfollowapowerlawprobabilitydistributioncommoninmanynaturallanguageandWebser-vices.Moreimportant,thesedatahelpexplainourearlier“ndingthatautomatictextclassi“cationtechniquesworkwithmicroblogging.Itappearsthatthesepostshavemuchincommonwithnaturallanguageutterances.Thisnaturallanguageusageisevenmorepronouncedwhenweexam-inetermpairswiththehighestmutualinformationstatistics(MIS),asshowninTable10.Althoughtherearesomeexceptions,manyofthetermpairsarenaturallanguageinstructureandsemantics.TheaverageMISrangedfrom0.6to1.8(SDrangingfrom4.24…4.96);this“ndingsuggeststhatthereis,again,alargedivergenceoftermsusedintweets.Theselinguistic“nd-ingsindicatethattweetssharesomecharacteristicsofnaturallanguagesentences,whichiswhynaturallanguageclas-si“ersaresuccessfulinautomaticallycategorizingthem.However,therearealsosomedifferences,probablyduetothetechnologyemployedinpostingtweets.ResearchQuestion#3:WhatArethePatternsofMicrobloggingCommunicationsBetweenCompaniesandCustomers?Forcommunicationpatternsbetweenabrandandpoten-tialcustomers(i.e.,followersofthecorporationTwitteraccount),weexploredfouraspectsofthecommunicationpattern:range,frequency,time,andcontent(i.e.,HowvariedwerethetopicsofcommunicationbetweenStarbucksanditscustomers?;HowoftendidStarbucksanditscustomerstwit-tereachother?,Whendidtheytwitter?;andWhatdidtheytwitterabout?).2180JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi TABLE9.Termsintweets. 2.6K38.8K14.2K IndexTermFrequencyProbabilityFrequencyProbabilityFrequencyProbability 1the9460.10164680.3679110.402to7730.08128080.2868250.343a9050.09120500.2656330.284i7090.0796480.2154320.275is4320.0462660.1435970.186of3330.0360180.1335840.187for4070.0471110.1635140.188and4920.0573340.1629080.159in3910.0465870.1427230.1410at1900.0240320.0924920.1211my3630.0463010.1422070.1112you2210.0230960.0721820.1113have23340.0521770.1114be19200.0421340.1115it3270.0342530.0920320.1016on3590.0456230.1219180.1017or17670.0918that2170.0231390.0717050.0919can13530.0720with2230.0240140.0912020.0621just2450.0329870.0711420.0622me1530.0222080.0511410.0623this20240.0410380.0524but10030.0525im9500.05Average(ofentiredataset)4.30.000412.10.000310.90.0005(ofentiredataset)40.00.0030166.20.0040121.40.0060 TABLE10.Termco-occurrence. 2.6Ktweets38.8Ktweets14.2Ktweets TermpairMISTermpairMISTermpairMIS Choosesthem6.90Tingtiding8.51Thankyou4.74Guha6.90Microcompactthey7.54matrix_sbonull4.65Finalizingthe6.90Approachingthe7.46Roomto4.39Fencedthe6.90Newarkto6.44Drandolphthe4.39Paternitythe6.33Tidingtiding6.41Playedwii4.39Importingto6.33Melethe6.38Uploadingthe7.54Chem.i6.33refuseto6.31Mhitoshigoogle7.54Evidenceexceptions6.33stock-indexfutures6.16Servicesgoogle4.35Footprintsin6.33Saddleramsey6.16Haripakorssgoogle4.24Riverwalkthe6.33Decadea6.07Linksgooglegoogle4.19Average(ofentiredataset)1.8(ofentiredataset)4.245.114.96 Concerningrange,Table11showsthatStarbucksreceived1,585tweetsfrom1,038peoplethatitfollowed.Starbucksfollowed7,779twitterusers.PeoplecanonlysendtweetstoStarbucksifitfollowedthem.Therefore,Starbucksreceivedtweetsfrom13.3%ofitsfollowers.Conversely,Starbuckstwittered322times,including77timeswithoutreplyingtoanyoneand245timesreplyingto212followerstweets.SinceStarbuckscanonlysendtweetstopeoplefollowingit,Starbuckstwittered2.7%ofitsfollowersandrepliedto20.4%peopletwitteringit(7,751followers,and1,038peopletwitteringit).Therefore,therangeofcommunicationisrathertight,withasmallnumberofTwittersactiveinthecom-municationroleandalargernumbertakingamorepassivemonitoringrole.Thiscommunicationpatternmirrorsthatinlistservsandwikis,whereasmallnumberofmembersareveryactiveandthemajorityarelurkers(Rafaeli,Ravid,&Soroka,2004).Intermsoffrequency,Table11informsusthatStarbucksreceived45.8%tweetsonlyoncefrom69.9%ofthepeopleitfollowedand43.5%tweetsfromtwotofourtimesfromJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092181DOI:10.1002/asi TABLE11.TwitteringfrequencybetweenStarbucksanditscustomers. Starbuckstwittered77timeswithoutreplyingtoanyone FromStarbucks TwitteringFollowerFollowerTweetTweetfrequency(count)(percentage)(count)(percentage) 118285.9%18274.3%22712.7%5422.0%331.4%93.7%Total212100.0%245100.0% ToStarbucks TwitterusersTwitteringTwitterusersfollowedTweetTweetfrequencyfollowed(count)(percentage)(count)(percentage) 172669.9%72645.8%219118.4%38224.1%3696.7%20713.1%4252.4%1006.3%5121.2%603.8%680.8%483.0%730.3%211.3%800%00%910.1%90.6%1010.1%100.6%1120.2%221.4%Total1,038100.0%1585100.0% FIG.4.TimeseriesoftweetsbetweenStarbucksanditscustomers.27.5%ofthepeopleitfollowed.ItwasrareforpeopletotwitterStarbucksmorethanfourtimes.Starbuckstwitteredonlyonceto74.3%tweetsto85.9%ofthesefollowers.FromthetwitternetworkingbetweenStarbucksandfol-lowersinFigure4,wecanseethatStarbucksandothersusuallytwitteredfewerthanfourtimes.Thus,bothStarbucksanditscustomersdidnotinteractfrequentlywitheachotherduringtheapproximately3-monthdatacollectionperiod.So,atleastforthiscompany,Twitterwasnotanactivemediumforcustomerrelationshipmanagement.TherewerekeyStar-bucksfollowerswhowereactivemembersofthisbrandcommunity,whichisconsistentwithpriorworkinonlinecommunities(Panzarasa,Opsahl,&Carley,2009).However,concerningthetimeaspect,Figure5showsastrongweeklypatternofcommunication.Starbucksanditscustomerstwitteredmostlyduringthemiddleoftheweekandlessduringweekendandthebeginningoftheweek.Therearethreeprominentspikesontherighthand,inpartduetoStarbucksrunningasurveyandaneventonTwitterduringthesetimeperiods.Thisindicatesthatanaturalcycleofcom-municationmayexistforcorporateaccounts.Theincreaseincommunicationforthesurveyandeventalsoshowsthecom-municationreachprovidedbyTwittertointerestedpotentialUsingtheaction-objectapproachtoanalyzecontent,we“rstdropped113tweetsbecausetheywereinforeignlanguagesornotunderstandable.Oftheremaining,we2182JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi FIG.5.TwitteringnetworkbetweenStarbucksanditscustomerbasedonfrequency(fromtoplefttorightbottom,frequencyidenti“ed386uniqueaction-objectpairsand2,490totalaction-objectpairs,asshowninTable12.Themostfrequentlyoccurringpairisapositivecommentoncoffee(9.6%).The“fthmostfrequentlyoccurringpairisrelated,whichisapositivecommentonanoncaffeinebever-age(2.8%).ThesecondandthirdmostfrequentlyoccurringpairsarebothrelatedtotheeventsheldbyStarbucksonTwit-ter.Onewasasurveyoffavoritebeverages,withStarbucksreceiving164tweetsrespondingtothissurvey.StarbucksfollowersactivelyrespondedtoTwittereventssponsoredbyStarbucks,includingHaikuonStarbucksandatriviaques-tioncompetition,withanearlyreleaseStarbucksGoldCardasprize.Starbucksfollowerswerealsokeenonrespondingtopromotions(2.8%). ThereddotinthecenterisStarbucks,theyellowdotsarepeoplewhotwitteredwithitmorethanonce,andthebluelineindicatesthecommunication.ExaminingactionsinTable13,24.8%oftheactionsarepositivecommentsand7.3%arenegativecomments.Thereare17.6%oftheactionsthatareresponses.Also,12.7%arequestions,and11.4%areanswerstoquestions.Therefore,questionsandanswerstogetherare24.06%ofallactionsinthisaccount.These“vecategoriesofactions(i.e.,pos-itivecomments,negativecomments,responses,questions,andanswerstoquestions)altogethermake73.7%.Thus,onecanviewStarbucksTwitteraccountasaplaceforacombinationofcustomertestimony,complaining,feed-back,andQ&A.Thisappearsinlinewithpriorcustomerrelationsresearch(Reinartz,Krafft,&Hoyer,2004),soTwitterappearsaviablecustomerrelationshipmanagementJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092183DOI:10.1002/asi TABLE12.Action-objectpairs. Action-objectpairoccurrence20Action-objectpairoccurrence Action-objectpairObjectActionTotalcountPercentageoccurrenceCountTotalCountPercentage CoffeePositivecomment2389.6%201200.8%FavoritebeverageAnswer1646.6%19000%TwittereventResponse1465.9%183542.2%PromotionResponse702.8%175853.4%NoncaffeinebeveragePositivecomment692.8%162321.3%StorePatronizing552.2%153451.8%TwitterChitchat451.8%142281.1%CoffeeQuestion421.7%134522.1%CoffeebeanPositivecomment331.3%123361.5%PromotionPositivecomment331.3%114441.8%TwitterPositivecomment321.3%109903.6%CoffeebeanQuestion311.2%9121084.3%CardQuestion301.2%85401.6%StarbucksPositivecomment271.1%76421.7%BreakfastPositivecomment251.0%614843.4%CardResponse251.0%513652.6%PromotionQuestion251.0%423923.7%CoffeeOrderviatwitter241.0%3351054.2%BaristaAnswer210.8%2701405.6%CoffeeNegativecomment210.8%11511516.1%TwittereventAnnouncement210.8%Total1,31352.7%Total1,17747.3%386uniqueaction-objectpairs,2,490action-objectpairs TABLE13.Actions. ActionCountPercentage Positivecomment61724.8%Response43917.6%Question31612.7%Answer28311.4%Negativecomment1817.3%Chitchat763.1%Suggestion682.7%Comment622.5%Expecting552.2%Patronizing552.2%Announcement542.2%Request532.1%Forwarding441.8%Noti“cation371.5%OrderviaTwitter281.1%Consuming251.0%Recommendation241.0%Missing230.9%Supplement140.6%Con“rmation130.5%Maintenance110.4%Recommendationrequest100.4%Research20.1%Total2490100.0% DiscussionandImplicationsThisstudyoffersimportantinsightsintomicrobloggingaseWOMcommunications,withimplicationsforbrandingforcorporations,organizations,andindividuals.Therearealsoimplicationsforthesocialeffectsthatsocialcommunicationservices(likeTwitter)arehaving,intermsoffosteringnewrelationshipsinthecommercialsector,speci“callyingaugingmarketplacereactions(i.e.,sentiment),externalcommunica-tion(i.e.,informationproviding),andgatheringmarketplaceinformation(i.e.,informationseeking).Theseimplicationsarethesameforbothcorporationsandindividuals.First,oftheentirepopulationoftweets,19%mentionanorganiza-tionorproductbrandinsomeway.Thisisgoodpercentageandindicatesthatthemicrobloggingmediumisaviableareafororganizationsforviralmarketingcampaigns,customerrelationshipmanagement,andtoin”uencetheireWOMbrandingefforts.Second,about20%ofallmicroblogsthatmentionedabrandexpressedasentimentoropinionconcerningthatcompany,product,orservice.Microbloggingisasocialcom-municationchannelaffectingbrandawarenessandbrandimage,thatmanagingbrandperceptioninthemicrobloggingworldshouldbepartofanoverallproactivemarketingstrat-egy,andmaintainingapresenceonthesechannelsshouldbepartofacorporationsbrandingcampaign.Itisappar-entthatcompaniescanreceivepositivebrandexposureviafollowersandotherswhomicroblogaboutthecompanyandproducts.Twentypercentofthisfast-growingmarketissubstantial.Additionally,with80%oftweetsmention-ingabrandbutexpressingnosentimentsuggestspeoplearealsoseekinginformation,askingquestions,andansweringquestionsaboutbrandsviatheirmicroblogs.Thus,companymicrobloggingaccountsareprobablyasmartideatobothmonitorbrandcommunitydiscussionsandtopushinforma-tiontoconsumers.Thisinformationseekingandbrandandproductcommentingseemstoopenthedoorforsometypeofadvertisingmedium.Similartosearchadvertising,where2184JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi FIG.6.Exampleofcompanyusingmicroblogtoimprovecustomerservice.relevantadsaretriggeredviakeytermsinqueries,itwouldappearthattherecouldbeatweetadvertisingmediumwhererelevantadsaretriggeredbykeywordsintweets.Third,theratioofpositivetonegativebrandingtweetsisabout50%to35%,withtheremainingbeingneutral.Peo-plehaveanopinionandareexpressingthemviamicroblogs.Somepartofthepositivetweetscouldbepartofcommercialviralmarketingbycorporations(i.e.,hiringpersonstopostpositivetweetsaboutcompaniesorproducts);however,giventhelargenumberoftweetsexamined,itwouldseemthatthiswouldbeasmallpartifitexistedatall.Thisimpliesthatcorporationscouldusetweetsascustomerfeedbackaboutproductsandfeaturesthatarewellreceivedbytheconsumers.Duanaetal.(2008)foundthatproductqualityhadapositiveimpactongeneratingpositiveeWOM;therefore,businessescouldgeneratebrandawarenesswithoutlargeoutlaysonadvertisingandmarketing.Asforthe35%ofnegativetweets,thispermitsadirectcustomerexpressionofwhatisnotgoingrightforaproductorservice.Onlineconsumerreviewinfor-mationcanbeusefulforidentifyingconsumerpreferences,“ndingoutproductdefects,andcorrectinginadvertentmis-takes.Duanaetal.(2008)pointedoutthatalargeamountofnegativeeWOMmadeitdif“cultforabusinesstoovercomeitsadverseindustrypositioning.CorporationsshouldthenbeproactiveinevaluatingthesesentimentsinthemicrobloggingeWOMarea.Withmicroblogmonitoringtools,companiescantrackmicroblogpostingsandimmediatelyintervenewithunsatis“edcustomers.Somecorporationsarealreadyget-tinginvolved,asshowninFigure6,whichshowsaChrysleremployeeaddressingamicroblogpostingfromacustomerconcerningherPTCruiser.GetSatisfaction.com(getsatisfac-tion.com)integratesdirectlywithTwitterandautomaticallypullsintweetsofspeci“ccompanies.Fourth,thereisabouta60%swinginsentimentfromweektoweek.Assumingthatthisisnotduetobiasesinclassi“-cation,microbloggingbrandingis”uid,requiringconstantandcontinualmanagement.Giventhatonecaneasilysendamicroblogviaavarietyofmobiledevices,thereisobviouslyanimmediateexpressionorreactiontoanindividualsexpe-riencesofproductsorservices.Thisimmediacyatthepointofpurchaseisacriticalfactorthatseparatesmicrobloggingfromotherformsofproductexpression,suchasblogging,Websites,orproductreviews.Assuch,eWOMrequirescon-tinualandconstantmanaginginthemicrobloggingmediumasithasclosedtheemotionaldistancebetweenthecustomerandbusiness.Fifth,thereisastatisticaldifferenceofbrandswithinindustrysectors,sothemicrobloggingdomainmaybeaJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092185DOI:10.1002/asi goodavenuetoexploretotrackthetrendswithinagivenmarketplace.Wefoundbrandsentimentsweredifferentinsevenindustries.Thosecorporationsthatfellbehindrela-tivetoothersintheindustrysegmentcouldleveragethemicrobloggingtoimprovetheirbrandimagebyananaly-sisofthesecustomersposting.Thosecorporationsaheadofotherscouldlearnfromthosebehindandfurtherenhancetheirbrandimage.Sixth,theterm,co-occurrence,andmutualinformationstatisticsgenerallyconformtotypicalnaturallanguageusage,withsomenotabledifferences.Thereisobviouslysomething(probablythetechnology)thatisalteringthenormalcom-municationpatterns.Thiswouldimply,again,theneedforsomespecializedmarketingeffortandmethodologytoana-lyzethesemicrobloggingposts.However,theresultsofthemanualclassi“cationandautomaticclassi“cationarenotstatisticallydifferent.Thisisparticularlyimportantgiventhatmicrobloggingwillmostlikelygrowandintegrateitselfintotheoveralllandscapeofelectronicexpressionmediums.Inordertomakesenseofthesedataandphenomena,cor-porationswillhavetorelyonautomatedmethods.Basedonthisanalysis,itappearsthatonecanaccuratelyclassifymicrobloggingviaourautomatedmethod.Seventh,weseesomegeneralpatternsinhowcompaniesareleveragingmicrobloggingforeWOMbranding.Theseeffortsareprovidingaplaceforcustomerstoexpressfeelings,providefeedback,askquestions,andgetanswers.Assuch,therearealotofpossibilitiestouseTwitterandsimilarsitesforcustomerrelationsandbrandingefforts.Theseincludehavingmultipleaccountsforvariousareasofthecorporationtoaccommodatetheswingsincustomertraf“canddifferentcustomerexpectationsoftheseservicesasacommunicationsystem(e.g.,oneforsurveysandevents,oneforcommentsandsuggestions,etc.).Consideringtherapidgrowthandpop-ularityofmicroblogging,companiesshouldcomeupwithasystematicwaytohandlecustomersonmicroblogsitestoin”uencebrandimage.Itwouldseemthatmicrobloggingcanbeusedtoprovideinformationanddrawpotentialcus-tomerstootheronlinemedia,suchasWebsitesandblogs.Assuch,monitoringandleveragingmicrobloggingsitescon-cerningonesownbrandandthebrandofcompetitorsisavaluablecompetitiveintelligence.Companiescangetnearreal-timefeedbackbysettingupcorporateaccounts.Com-paniesalsogetvaluablecontentandproductimprovementideasbytrackingmicroblogpostingsandfollowingthosepeoplewhofollowtheircorporateaccounts.Finally,compa-niescanleveragecontactsmadeviamicrobloggingservicestofurthertheirbrandingeffortsbyrespondingtocomments,suggestions,orcommentsaboutthecompanybrand.Naturally,therearelimitationstothisstudy.First,weexaminedmicroblogsfromonlyonemicrobloggingsite.Usersofothermicrobloggingservicesmightdifferintheirusagepatterns.Twitterisbyfarthelargestandmostpopu-larofthesesites,butinvestigationsintootherserviceswillbeanareaforfutureresearch.Second,mostofthebrandsthatweexaminedaremajorbrands,withonlyasmallper-centbeingminorbrandslikeAveratec.Porter&Golan(2006)analyzed501advertisements,including235televisionadver-tisementsand266viraladvertisements.The“ndingsshowedthatFortune500companiescreated62%ofthetelevisionadsanalyzed(146ads),whilenon-Fortune500companiescre-ated38%(89)ofthetelevisionads.However,non-Fortune500companiesproducedthemajorityofviralads,with60%(160ads),comparedto40%byFortune500companies(106ads).Itwouldbeinterestingtocontinuethisinvestigationintosmallorevenlocalbrands.Finally,incomparingman-ualandautomatedcoding,wecomparedthedistributionsofeachcategoryacquiredfromhumancodingandsystemcoding.Onecouldconductamorein-depthanalysisbycom-paringthesentimentofeachtweetcodedbyhumanandbyanautomatedsystem.However,thedistributionofdifferentsentimentcategoriescangiveusafairenoughdescriptionaboutthesentimentoftweets.Thereareseveralstrengthsofthestudy.First,weusedanumberofwell-knownbrandswithmajorimpactfromavari-etyofindustrysections.Thisensuredourresultswouldhavepracticalandin”uentialimplications.Second,weapproachedouranalysisofmicrobloggingfromavarietyofperspec-tivesandparadigms,usingamixedmethodsapproachandemployingbothquantitativeandqualitativemeasures.Thishelpedensurethatour“ndingsarerobust.Third,wefocusonmicrobloggingasanemergingareawithpotentiallysig-ni“cantimpactoneWOMandanchoredthisanalysisinthebrandknowledgeandrelationshipsthatlinkthe“ndingtoconsumerbehavior.Therefore,ourresearchistimelyandhaspracticalimplicationsinthemarketplace.ConclusionInthisresearchweexaminedtheuseofmicrobloggingforeWOMbranding.Examiningseveraldatasetsfromavarietyofangles,ourresearchhasshedlightoncriticalaspectsofthisphenomenon.Theimplicationsofthisresearchincludethatmicrobloggingisapotentiallyrichavenueforcompaniestoexploreaspartoftheiroverallbrandingstrat-egy.Customerbrandperceptionsandpurchasingdecisionsappearincreasinglyin”uencedbyWebcommunicationsandsocialnetworkingservices,asconsumersincreasinglyusethesecommunicationtechnologiesfortrustedsourcesofinformation,insights,andopinions.Thistrendoffersnewopportunitiestobuildbrandrelationshipswithpotentialcus-tomersandeWOMcommunicationplatforms.ItisapparentthatmicrobloggingservicessuchasTwittercouldbecomekeyapplicationsintheattentioneconomy.Giventheeaseofmon-itoringanybrandssentiment,onecanviewmicrobloggingasacompetitiveintelligencesource.TheessenceofeWOMcommunicatingandcustomerrelationshipmanagementisknowingwhatcustomersandpotentialcustomersaresayingaboutthebrand.Microblog-gingprovidesavenueintowhatcustomersreallyfeelaboutthebrandanditscompetitorsinnearrealtime.Additionally,microbloggingsitesprovideaplatformtoconnectdirectly,againinnearrealtime,withcustomers,whichcanbuildandenhancecustomerrelationships.2186JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi Forfurtherresearch,itwouldbeinterestingtoinvesti-gateoccurrencesofbrandhijackingonmicrobloggingservice(e.g.,www.web-strategist.com/blog/2008/08/01/how-janet-fooled-the-twittersphere-shes-the-voice-of-exxon-mobil/).AswithmostinformationservicesontheWeb,microbloggingsitesaresusceptibletoadversarialandspammingmaneu-vers,andbrandhijackingappearstobeanearlyformofadversarialmethods.EarlyresearchinthisareacouldreducethesusceptibilityofmicrobloggingsitestotheseformsofReferencesAllen,J.,&Perrault,C.R.(1986).Analyzingintentioninutterances.InB.J.Grosz,K.Sparck-Jones,&B.L.Webber(Eds.),Readingsinnaturallanguageprocessing(pp.441…458).SanFrancisco:MorganKaufmann.Anderson,E.W.(1998).Customersatisfactionandwordofmouth.JournalofServiceResearch,1(1),5…17.Archak,N.,Ghose,A.,&Ipeirotis,P.G.(2007,12…15August).InR.Caruana&X.Wu(Eds.),Showmethemoney!Derivingthepricingpowerofproductfeaturesbyminingconsumerreviews(pp.56…66).Paperpresentedatthe13thACMSIGKDDInternationalConferenceonKnowledgeDiscoveryandDataMining(KDD07).NewYork:ACMPress.Barton,B.(2006).Ratings,reviews&ROI:Howleadingretailersusecustomerwordofmouthinmarketingandmerchandising.JournalofAdvertisingResearch,7,article87.Bausch,S.,&McGiboney,M.(2008).Nielsenonlineprovidesfastestgrow-ingsocialnetworksforSeptember2008.NewYork:NielsenCompany.Benedicktus,R.L.,&Andrews,M.L.(2006).Buildingtrustwithcon-sensusinformation:Theeffectsofvalenceandsequencedirection.JournalofInteractiveAdvertising,6.RetrievedJune12,2009,fromhttp://www.jiad.org/article77Broder,A.(2002).AtaxonomyofWebsearch.SIGIRForum,36(2),Brogan,C.(2008).FiftywaystouseTwitterforbusiness.RetrievedFebruary21,2009,fromhttp://socialcomputingmagazine.com/viewcolumn.cfm?Brown,T.J.,Barry,T.E.,Dacin,P.A.,&Gunst,R.F.(2005).Spreadingtheword:Investigatingantecedentsofconsumerspositiveword-of-mouthintentionsandbehaviorsinaretailingcontext.JournaloftheAcademyofMarketingScience,33(2),123…138.Cheung,C.M.K.,Lee,M.K.O.,&Rabjohn,N.(2008).Theimpactofelec-tronicword-of-mouth.Theadoptionofonlineopinionsinonlinecustomercommunities.InternetResearch,18(3),229…247.Church,K.W.,&Hanks,P.(1990).Wordassociationnorms,mutualinformation,andlexicography.ComputationalLinguistics,16(1),Conrad,J.G.,Leidner,J.L.,&Schilder,F.(2008,30October).InJ.G.J.Shanahan(Ed.),Professionalcredibility:AuthorityontheWeb.ProceedingsoftheSecondACMWorkshoponInformationCredibilityontheWeb(WICOW08)(pp.85…88).NewYork:ACMPress.Davenport,T.H.,&Beck,J.C.(2002).Theattentioneconomy:Understand-ingthenewcurrencyofbusiness.Cambridge,MA:HarvardBusinessDavis,A.,&Khazanchi,D.(2008).Anempiricalstudyofonlinewordofmouthasapredictorformulti-productcategorye-Commercesales.ElectronicMarkets,18(2),130…141.Dellarocas,C.(2003).Thedigitizationofword-of-mouth:Promiseandchallengesofonlinereputationsystems.ManagementScience,49(10),Dobson,A.J.(2002).Anintroductiontogeneralizedlinearmodels.NewYork:CRCPress.Duana,W.,Gub,B.,&Whinston,A.B.(2008).Doonlinereviewsmatter?„Anempiricalinvestigationofpaneldata.DecisionSupportSystems,45(3),1007…1016.Ebner,M.,&Schiefner,M.(2008).InMicroblogging„morethanfun?InI.ArnedilloSánchez&P.Isaías(Eds.),ProceedingsoftheIADISMobileLearningConference(pp.155…159).Lisbon,Portugal:IADIA.Engel,J.E.,Blackwell,R.D.,&Kegerreis,R.J.(1969).Howinforma-tionisusedtoadoptaninnovation.JournalofAdvertisingResearch,9(December),3…8.Ennew,C.T.,Banerjee,A.K.,&Li,D.(2000).Managingwordofmouthcommunication:EmpiricalevidencefromIndia.InternationalJournalofBankMarketing,18(2),75…83.Esch,F.-R.,Langner,T.,Schmitt,B.H.,&Geus,P.(2006).Arebrandsforever?Howbrandknowledgeandrelationshipsaffectcurrentandfuturepurchases.JournalofProduct&BrandManagement,15(2),Fehr,E.,&Falk,A.(2002).Psychologicalfoundationsofincentives.EuropeanEconomicReview,46(4),687…724.Gelb,B.D.,&Sundaram,S.(2002).Adaptingtowordofmouse.BusinessHorizons,45(4),21…25.Glaser,B.,&Strauss,A.(1967).Thediscoveryofgroundedtheory:Strategiesforqualitativeresearch.Chicago:Aldine.Goldman,E.(2008).Onlinewordofmouthanditsimplicationsfortrademarklaw.InG.B.Dinwoodie&M.D.Janis(Eds.),Trademarklawandtheoryahandbookofcontemporaryresearch(pp.404…429).Cheltenham,UK:EdwardElgar.Goldsmith,R.E.,&Horowitz,D.(2006).Measuringmotivationsforonlineopinionseeking.JournalofInteractiveAdvertising6.RetrievedJune10,2009,fromhttp://www.jiad.org/article76Grice,H.(1969).Utterersmeaningandintentions.PhilosophicalReview,78(2),47…177.Grosseck,G.,&Holotesch,C.(2008,April).CanweuseTwitterforeducationalactivities?PaperpresentedattheFourthInternationalSci-enti“cConferenceeLearningandSoftwareforEducation,Bucharest,Ha,H.-Y.(2006).Theeffectsofconsumerriskperceptiononpre-purchaseinformationinonlineauctions:Brand,word-of-mouth,andcustomizedinformation.JournalofComputer-MediatedCommunication,8,article2.Helps,J.E.,Lewis,R.,Mobilio,L.,Perry,D.,&Raman,N.(2004).Viralmarketingorelectronicword-of-mouthadvertising:Examiningconsumerresponsesandmotivationstopassalongemail.JournalofAdvertisingResearch,44(2),333…348.Hennig-Thurau,T.,Gwinner,K.P.,Walsh,G.,&Gremle,D.D.(2004).Elec-tronicword-of-mouthviaconsumer-opinionplatforms:WhatmotivatesconsumerstoarticulatethemselvesontheInternet?JournalofInteractiveMarketing,18(1),38…52.Huberman,B.A.,Romero,D.M.,&Wu,F.(2009).Socialnetworksthatmatter:Twitterunderthemicroscope.FirstMonday14,(1…5).Jansen,B.J.,Booth,D.,&Spink,A.(2008).Determiningtheinforma-tional,navigational,andtransactionalintentofWebqueries.InformationProcessing&Management,44(3),1251…1266.Jansen,B.J.,&Spink,A.(2005).HowarewesearchingtheWorldWideWeb?Acomparisonofninesearchenginetransactionlogs.InformationProcessing&Management,42(1),248…263.Java,A.,Song,X.,Finin,T.,&Tseng,B.(2007).WhyweTwitter:Understandingmicrobloggingusageandcommunities.InH.Zhang,B.Mobasher,C.L.Giles,A.McCallum,O.Nasraoui,M.Spiliopoulou,J.Srivastava,&J.Yen(Eds.),ProceedingsoftheNinthWebKDDandFirstSNA-KDD2007WorkshoponWebMiningandSocialNetworkAnalysis(WebKDD/SNA-KDD07)(pp.56…65).NewYork:ACMPress.Jung,C.G.(1959).Avisionaryrumor.JournalofAppliedPsychology,4,Katz,E.,&Lazarsfeld,P.F.(1955).Personalin”uence:Thepartplayedbypeopleinthe”owofmasscommunication.NewYork:FreePress.Kiecker,P.,&Cowles,D.L.(2001).Interpersonalcommunicationandpersonalin”uenceontheInternet:Aframeworkforexaminingonlineword-of-mouth.JournalofEuromarketing,11(2),71…88.Kimmel,A.J.(2004).Rumorsandrumorcontrol:Amanagersguidetounderstandingandcombattingrumors.Mahwah,NJ:LawrenceErlbaum.Lee,D.,Jeong,O.-R.,&Lee,S.-G.(2008).Opinionminingofcustomerfeed-backdataontheWeb.InW.Kim&H.J.Choi(Eds.),ProceedingsoftheJOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November20092187DOI:10.1002/asi SecondInternationalConferenceonUbiquitousInformationManagementandCommunication(pp.230…235).NewYork:ACMPress.Liu,B.,Hu,M.,&Cheng,J.(2005).Opinionobserver:Analyzingandcom-paringopinionsontheWeb.InA.Ellis&T.Hagino(Chairs),Proceedingsofthe14thInternationalConferenceonWorldWideWeb(WWW05)(pp.342…351).NewYork:ACMPress.McFedries,P.(2007).Technicallyspeaking:Alla-Twitter.IEEESpectrum,44(10),84.McGrawHillHigherEducation.(n.d.).Starbuckscasestudy.RetrievedFebruary16,2009,fromhttp://www.mhhe.com/business/management/thompson/11e/case/starbucks.htmlMilstein,S.,Chowdhury,A.,Hochmuth,G.,Lorica,B.,&Magoulas,R.(2008).Twitterandthemicro-messagingrevolution:Communication,connections,andimmediacy„140charactersatatime.Sebastopol,CA:OReillyMedia.Montoya-Weiss,M.M.,Voss,G.B.,&Grewal,D.(2003).Determinantsofonlinechanneluseandoverallsatisfactionwitharelational,multichannelserviceprovider.JournaloftheAcademyofMarketingScience,31(4),Muniz,A.M.J.,&OGuinn,T.C.(2001).Brandcommunity.JournalofConsumerResearch,27(4),412…432.Nardi,B.A.,Schiano,D.J.,Gumbrecht,M.,&Swartz,L.(2004).Whyweblog.CommunicationsoftheACM,47(12),41…46.Panzarasa,P.,Opsahl,T.,&Carley,K.M.(2009).Patternsanddynamicsofusersbehaviorandinteraction:Networkanalysisofanonlinecommunity.JournaloftheAmericanSocietyforInformationScienceandTechnology,60(5),911…932.Park,C.,&Lee,T.M.(2009).Informationdirection,websitereputationandeWOMeffect:Amoderatingroleofproducttype.JournalofBusinessResearch,62(1),61…67.Patton,M.Q.(1990).Qualitativeevaluationandresearchmethods(2ded.).NewburyPark,CA:SagePublications.Porter,L.,&Golan,G.J.(2006).Fromsubservientchickenstobrawnymen:Acomparisonofviraladvertisingtotelevisionadvertising.JournalofInteractiveAdvertising,6.RetrievedJune12,2009,fromhttp://www.jiad.org/article78.Postman,J.(2008).SevenrulesforestablishingacorporatepresenceonTwitter.RetrievedFebruary27,2009,fromhttp://www.socialmediatoday.com/SMC/44452Rafaeli,S.,Ravid,G.,&Soroka,V.(2004).De-lurkinginvirtualcommuni-ties:Asocialcommunicationnetworkapproachtomeasuringtheeffectsofsocialandculturalcapital.InR.H.Sprague(Chair),Proceedingsofthe37thAnnualHawaiiInternationalConferenceonSystemSciences(HICSS04)(p.10).LosAlamitos,CA:IEEE.Reinartz,W.,Krafft,M.,&Hoyer,W.D.(2004).Thecustomerrelation-shipmanagementprocess:Itsmeasurementandimpactonperformance.JournalofMarketingResearch,41(3),293…305.Reynolds,G.(2006).AnarmyofDavids:Howmarketsandtechnologyempowerordinarypeopletobeatbigmedia,biggovernment,andotherGoliaths.Nashville,TN:ThomasNelson.Richins,M.L.(1983).Negativeword-ofmouthbydissatis“edconsumers:Apilotstudy.JournalofMarketing,47(Winter),68…78.Richins,M.L.(1984).Word-of-mouthcommunicationasnegativeinforma-tion.AdvancesinConsumerResearch,44,697…702.Richins,M.L.,&Root-Shaffer,T.(1988).Theroleofinvolvementandopinionleadershipinconsumerword-of-mouth:Animplicitmodelmadeexplicit.AdvancesinConsumerResearch,15,32…36.Rose,D.E.,&Levinson,D.(2004).UnderstandingusergoalsinWebsearch.InS.Feldman&M.Uretsky(Eds.),Proceedingsofthe13thInternationalConferenceonWorldWideWeb(WWW04)(pp.13…19).NewYork:ACMPress.Skowronski,J.J.,&Carlston,D.E.(1989).Negativityandextremityinimpressionformation:Areviewofexplanations.PsychologyBulletin,105(January),131…142.Starbucks.(2008).Companyfactsheet.RetrievedFebruary16,2009,fromhttp://www.starbucks.com/aboutus/Company_Factsheet.pdfSundar,S.S.(2008).TheMAINmodel:Aheuristicapproachtounderstand-ingtechnologyeffectsoncredibility.InM.J.Metzger&A.J.Flanagin(Eds.),Digitalmedia,youth,andcredibility(pp.73…100).Cambridge,MA:MITPress.Thompson,C.(2008,5September).Bravenewworldofdigitalintimacy.RetrievedJune12,2009,fromhttp://www.nytimes.com/2008/09/07/magazine/07awareness-t.html?_rThorson,K.S.,&Rodgers,S.(2006).RelationshipsbetweenblogsaseWOMandinteractivity,perceivedinteractivity,andparasocialinterac-tion.JournalofInteractiveAdvertising,6.RetrievedJune12,2009,fromhttp://www.jiad.org/article79Urban,G.(2005).Dontjustrelate„advocate:Ablueprintforpro“tintheeraofcustomerpower.UpperSaddleRiver,NJ:WhartonSchoolPublishing.Wang,P.,Berry,M.,&Yang,Y.(2003).MininglongitudinalWebqueries:Trendsandpatterns.JournaloftheAmericanSocietyforInformationScienceandTechnology,54(8),743…758.Wells,W.,Moriarty,S.,&Burnett,J.(2000).Advertisingprinciplesandpractice.UpperSaddleRiver,NJ:PrenticeHall.Wijaya,D.T.,&Bressan,S.(2008).Arandomwalkontheredcarpet:RatingmovieswithuserreviewsandPageRank.InJ.G.Shanahan(Ed.),Proceedingsofthe17thACMConferenceonInformationandKnowledgeManagement(CIKM08)(pp.951…960).NewYork:ACMPress.Wikipedia.(2009).Starbucks.RetrievedFebruary16,2009,fromhttp://en.wikipedia.org/wiki/StarbucksZhang,M.,&Jansen,B.J.(2008).Usingaction-objectpairsasaconceptualframeworkfortransactionloganalysis.InB.J.Jansen,A.Spink&I.Taksa(Eds.),HandbookofresearchonWebloganalysis(pp.414…433).Hershey,PA:IGI.Zhang,W.,Yu,C.,&Meng,W.(2007).Opinionretrievalfromblogs.InM.J.Silva,A.A.F.Laender,R.Baeza-Yates,D.L.McGuinness,B.Olstad,O.H.Olsen&A.O.Ealcao(Eds.),Proceedingsofthe16thACMConferenceonConferenceonInformationandKnowledgeManagement(CIKM07)(pp.831…841).NewYork:ACMPress.2188JOURNALOFTHEAMERICANSOCIETYFORINFORMATIONSCIENCEANDTECHNOLOGY„November2009DOI:10.1002/asi