Journal of Retailing 80(2004)
159–169
The influence of online product recommendations on
consumers’online choices
Sylvain Senecal a ,∗,Jacques Nantel a ,1
a
HEC Montreal,University of Montreal,3000Chemin de la Cote-Sainte-Catherine,Montreal,Que.,Canada H3T 2A7
Abstract
This study investigates consumers’usage of online recommendation sources and their influence on online product choices.A 3(websites)×4(recommendation sources)×2(products)online experiment was conducted with 487subjects.Results indicate that subjects who consulted product recommendations lected recommended products twice as often as subjects who did not consult recommendations.The online recommendation source labeled “recommender system,”typical of the
加拿大滑铁卢大学
personalization possibilities offered by online retailing,was more influential than more traditional recommendation sources such as “human experts”and “other consumers”.The type of product also had a significant influence on the propensity to follow product recommendations.Theoretical and managerial implications of the findings are provided.
©2004by New York University.Published by Elvier.All rights rerved.
Keywords:Online product;Recommendation;Consumers
Introduction
Among all possible advantages offered by electronic com-merce to retailers,the capacity to offer consumers a flexible and personalized relationship is probably one of the most important (Wind &Rangaswamy,2001).Online personal-ization offers retailers two major benefits.It allows them to provide accurate and timely information to customers which,in turn,often generates additional sales (Postma &Brokke,2002).Personalization has also been shown to in-crea the level of loyalty consumers hold toward a retailer (Cyber Dialogue,2001;Srinivasan,Anderson,&Ponnavolu,2002).While there are veral ways to personalize an online relationship,the capacity for an online retailer to make rec-ommendations is certainly amon
g the most promising (The e-tailing Group,2003).Online,recommendation sources range from traditional sources such as other consumers (e.g.,testimonies of customers on retail websites such )to personalized recommendations provided by recommender systems (West et al.,1999).To date,no study has specifically investigated and compared the rela-tive influence of the online recommendation sources on
∗
Corresponding author.Tel.:+14195302422.
E-mail al@hec.ca (S.Senecal),jacques.nantel@hec.ca (J.Nantel).1Tel.:+15143406421.
consumers’product choices.Therefore,the main objective of this study is to investigate the influence of online product recommendations on consumers’online product choices.In addition,we explore the moderating influence of vari-ables related to recommendation sources and the purcha decision.Literature review
Rearch on the u and influence of recommendations on consumers has typically been subsumed
under personal influence or word-of-mouth (WOM)rearch.In addition,as noted by Ron and Olshavsky (1987),rearch on opinion leadership and reference groups also relates to the study of recommendations and to influence in general.
Recommendation sources are considered primarily as information sources.Andrean (1968)propos the fol-lowing typology of information sources:(1)Impersonal Advocate (e.g.,mass media),(2)Impersonal Independent (e.g.,Consumer Reports ),(3)Personal Advocate (e.g.,sales clerks),and (4)Personal Independent (e.g.,friends).Although rearch on personal influence and WOM fo-cus on the latter two information sources,it is notewor-thy that impersonal independent information sources such as Consumer Reports can also rve as recommendation sources.Moreover,the Internet can provide consumers with
0022-4359/$–e front matter ©2004by New York University.Published by Elvier.All rights rerved.doi:10.1016/j.jretai.2004.04.001
160S.Senecal,J.Nantel/Journal of Retailing80(2004)159–169
an additional type of impersonal information source.For instance,electronic decision-making aids such as recom-mender systems are impersonal information sources that provide personalized infor
mation to consumers(Ansari, Esgaier,&Kohli,2000).In an effort to extend Andrean’s (1968)typology to computer-mediated environments,we asrt that information sources can be sorted into one of four groups:(1)Personal source providing personalized ,“My sister says that this product is best for me.”);(2)Personal source providing non-personalized ,“A renowned expert says that this product is the best.”);(3)Impersonal source providing personalized ,“Bad on my profile,the recommender system suggests this product.”);(4)Impersonal source pro-viding non-personalized ,“According to Consumer Reports,this is the best product on the market.”). In consumer rearch,studies on personal influence,so-cial influence,or WOM,can be categorized as studies investigating personal sources providing personalized or non-personalized information.Furthermore,studies dealing with reference groups encompass such sources as well as impersonal sources that provide non-personalized informa-tion.Thus,a new area has emerged in consumer rearch, arising mainly from information technologies such as the Internet:that of impersonal sources that provide person-alized information(Alba et al.,1997;Ansari et al.,2000; Häubl&Trifts,2000;Maes,1999;Urban,Sultan,&Qualls, 1999;West et al.,1999).
Rearch on information sources suggests that personal and impersonal information sources influen
ce consumers’decision-making(Ardnt,1967;Duhan et al.,1997;Gilly et al.,1998;Olshavsky&Granbois,1979;Price&Feick, 1984).For instance,Price and Feick(1984)found that con-sumers planned to u the following information sources for their next durable good purcha:(1)Friends,relatives, and acquaintances,(2)Salespeople,(3)Publications such as Consumer Reports.However,if much is known about the relative likelihood of consumers to consider recommen-dations in the cour of their decision making process,lit-tle is known about how recommendations,especially in a computer-mediated environment,impact consumers’prod-uct choices.
Determinants of recommendation influence
The current study focus on three determinants that could influence the impact of computer-mediated recommenda-tions on consumers’online product choices:the nature of the product recommended,the nature of the website on which the recommendation is propod,and the type of recommen-dation source.
Prior rearch has shown that the type of product affects consumers’u of personal information sources and their influence on consumers’choices(Bearden&Etzel,1982; Childers&Rao,1992;King&B
alasubramanian,1994). Nelson(1970)suggests that goods can be classified as pos-ssing either arch or experience qualities.Search quali-ties are tho that“the consumer can determine by inspec-tion prior to purcha,”and experience qualities are tho that“are not determined prior to purcha”(Nelson,1974, p.730).Since it is difficult or even impossible to evalu-ate experience products before purcha,consumers should rely more on product recommendations for the products than for arch products.In support of this view,King and Balasubramanian(1994)found that consumers asssing a arch ,a35-mm camera)are more likely to u own-bad decision-making process than consumers asssing an experience product,and that consumers evalu-ating an experience ,afilm-processing rvice) rely more on other-bad and hybrid decision-making pro-cess than consumers asssing a arch product.
The nature of the website can also influence the im-pact of a given recommendation.Bad on previous web-site classifications(Hoffman,Novak,&Chatterjee,1995; Spiller&Loh,1998),Senecal and Nantel(2002)sug-gest that recommendation sources can be ud and promoted by three different types of websites:,retailer or manufacturer websites such ),commer-cially linked third ,comparison shopping web-sites such ),and non-commercially linked third ,product or merchant asssment web9月英文
sites such as Consumerreports).More independent websites such as non-commercially linked third parties that facilitate consumers’external arch effort by decreasing arch costs are assumed to be preferred by consumers(Alba et al.,1997; Bakos,1997;Lynch&Ariely,2000).By providing more al-ternatives to choo from and more objective information, independent websites should be perceived as more uful by consumers.In addition,prior rearch on attribution theory suggests that consumers discredit recommendations from endorrs if they suspect that the latter have incentives to recommend a product(for reviews,refer to Folkes,1988; Mizerski,Golden,&Kernan,1979).According to the dis-counting principle of the attribution theory(Kelley,1973), which suggests that a communicator will be perceived as biad if the recipient can infer that the message can be at-tributed to personal or situational caus,consumers would attribute more non-product related ,com-missions on sales)to recommendation sources that are pro-moted by commercially linked third parties and llers than independent third party websites.Conquently consumers would follow product recommendations in a greater propor-tion when shopping on more independent than on less inde-pendent websites.
葡萄牙语国家
In light of rearch on consumers’u of relevant others in their pre-purcha external arch efforts(Olshavsky& Granbois,1979;Price&Feick,1984;Ron&Olshavsky, 1987)and in consideration
of the emergence of online in-formation sources providing personalized recommendations (Ansari et al.,2000),Senecal and Nantel(2002)asrt that online recommendation sources can be sorted into three broad categories:(1)other ,relatives,friends
S.Senecal,J.Nantel/Journal of Retailing80(2004)159–169161
learn chine
and acquaintances),(2)human ,salespersons, independent experts),and(3)expert systems such as recom-mender systems.We posit that the online recommendation sources will have different levels of influence on consumers’online product lection.Brown and Reingen(1987)sug-gest that information received from sources that have some personal knowledge about the consumer have more influ-ence on the latter than sources that have no personal knowl-edge about the consumer.Thus,a recommendation source providing personalized information to ,rec-ommender system)should be more influential than a recom-mendation source providing non-personalized information (e.g.,other consumers).
The fact that both factors,the origin(source)of a recom-mendation as well as the type of website on which it is made, have an impact on the likelihood it has to be followed may find its explanation in Kelman’s(1961)work on source cred-ibility.Kelman(1961)suggests that credibility is a product of expe
rti and trustworthiness.Experti can be viewed as the perceived ability of an information source to know the right answer and trustworthiness as the perceived infor-mation source’s motivation to communicate this experti without bias(McGuire,1969).Although moderated by con-textual factors(for a review,refer to Sternthal,Phillips,& Dholakia,1978),source experti and trustworthiness have been found to be positively correlated with consumers’atti-tude toward the brand,and consumers’behavioral intentions and behaviors(Gilly et al.,1998;Harmon&Coney,1982; Lascu,Bearden,&Ro,1995;Tybout,1978). Hypothes
Bad on the preceding review of the literature we pos-tulate that personal information sources as well as imper-sonal information sources providing product recommenda-tions(Price&Feick,1984)will influence consumers in computer-mediated environments such as the Internet and the World Wide Web.We thus formulate the following gen-eral hypothesis.
H1.Consumers who consult an online information source recommending a given brand will lect that brand in a greater proportion than consumers who do not consult an online recommendation source.
As for the impact that such a recommendation will have on consumers’choice,we formulate three ad
ditional hypothe-s.First,we posit that the nature of the product for which a recommendation is provided will influence the likelihood that it will be followed.Bad on prior rearch on the re-lationship between product type and personal information source influence(Bearden&Etzel,1982;Childers&Rao, 1992;King&Balasubramanian,1994),we put forward the following hypothesis.H2.Online recommendations for experience products will be followed in a greater proportion than online recommen-dations for arch products.弗里特
Second,bad on Alba et al.(1997),Bakos(1997)and Lynch and Ariely(2000),we propo that online product recommendations from more independent websites are more influential than tho from less independent websites.We therefore put forth the following hypothesis.
H3.Online product recommendations consulted on “non-commercially linked third party”websites will be followed in a greater proportion by consumer than if con-sulted on“commercially linked third party”websites,and online product recommendation consulted on the latter type of websites will be followed in a greater proportion than if consulted on“ller”websites.
Finally,we believe,bad on the literature which has dealt with the issue of consumers’u of relevant others in their pre-purcha external arch efforts,that personalized rec-ommendations will
foolishhave a greater influence on consumers than non-personalized ones(Brown&Reingen,1987).Thus follows hypothes four.
H4.Recommendations from information sources offering personalized ,recommender system) will be followed in a greater proportion by consumers than recommendations from information sources providing non-personalized recommendations.
In addition to this t of hypothes,which pertains to the variables that moderate the influence of an online recommendation,we formulate a t of three hypothes which consider potential reasons for which various online recommendation sources may differ in their influence on consumers’choices.First,we expect that the recommen-dation source“other consumers”will be perceived as less expert than“human experts”and“recommender systems”. However,bad on the discounting principle of attribution theory(Kelley,1967),the recommendation source“other consumers”should be perceived as more trustworthy than human experts and recommender systems since the lat-ter two recommendation sources are more susceptible to non-product related attributions.Second,since consumers may also attribute non-product related motivations more easily to recommendation sources promoted by websites that are not clearly independent,we predict that the type of website will have an impact on the perception of the recom-mendation source’s trustwort
hiness.For instance,a human expert who recommends a product on a ller website may be perceived by consumers as less trustworthy than if that person recommended the same product on an independent third party website.Thus,the following hypothes are posited.
162S.Senecal,J.Nantel/Journal of Retailing80(2004)159–169
H5a.The online recommendation sources“human experts”and“recommender system”will be perceived as posss-ing more experti than the online recommendation source “other consumers.”
H5b.The online recommendation sources“human experts”and“recommender system”will be perceived as less trust-worthy than the online recommendation source“other consumers.”
H6.Consumers’trust in the recommendation source will be greater when shopping on a“non-commercially linked third party”type of website than on a“commercially linked third party”type of website,which in turn will be perceived as more trustworthy than the“ller”type of website. Method
Sample
A convenience sample of487subjects was generated on the basis of25742e-mails nt to three popu在线网站翻译
lations of web urs.The initial e-mail stated that two rearchers from a large business school were conducting a study on electronic commerce,and that participants had a chance of winning one of the products about which the experi-ment was designed.Potential participants did not know in advance the types of products that were to be tested. They were told that they would be asked to participate in two ssions in order to complete the study.Overall,630 subjects participated in thefirst ssion and487subjects participated in both ssions.Subjects who participated in both ssions did not significantly differ from subjects who only participated in thefirst ssion with regards to their socio-demographic profile(X2
gender
(2)=0.009, X2age group(4)=2.138,X2profesional situation(6)=7.482, X2income(7)=11.683;all p-values>.05),their product class familiarity(F wine(1,629)=0.316,F calculator(1,630)= 0.951;all p-values>.05),and their product class subjective knowledge(F wine(1,629)=0.758,F calculator(1,630)= 0.000;all p-values>.05).Subjects took an average of6–8 days to complete both ssions of the online experiment, which included a5-day delay between email invitations of ssions1and2.Of the487participants,173were from a specializedfirm list(respon rate:0.6%),59originated from an e-commerce rearch center list(respon rate: 12.0%),203were from an undergraduate student list(re-
spon rate:7.7%)and52could not be traced since they ud a different email address than the ones on the lists. Subjects participated in both ssions of the study from the location where they usually u the Internet.The major-ity of subjects were between the ages of18and29years (84%).Fifty percent were female,one third were working full time(31%);26%of subjects were full-time students and another31%were part-time workers and students.On average,subjects had been using the Internet for4.5years and currently ud it18hours per week.
Procedure
Experiment overview
In thefirst ssion of the experiment,subjects were simply asked to complete an online questionnaire.In the cond s-sion,subjects were asked to perform online shopping tasks on a specific website.During that cond ssion a3×4×2 online experiment was conducted.Thefirst between-subject factor was the website manipulation.Subjects were assigned to one of three types of websites:retailer,third party com-mercially linked to retailers,or non-commercially linked third party.The cond between-subject factor manipulated the source of recommendation.Subjects were assigned to one of the four following conditions:other consumers,hu-man experts,recommender syst
em,or no recommendation source.Finally,the last factor,a within-subject factor,was the product manipulation.During theirfirst online shopping task,subjects were randomly assigned to either a arch or experience product,and they were assigned to the remaining product type for their cond shopping task. Experiment description
To motivate subjects to participate without mentioning the preci goal of the ,the influence of recom-mendation on product choice),a cover story was ud.Sub-jects were told that a two-ssion study was being conducted to asss the commercial potential of various products that a foreign ,Maximo)was interested in introduc-ing to local markets via their website.In addition,partici-pants were informed that they would be asked in the cond ssion of the study to lect three products,and that they had a one in three chance of winning one of the products lected.This procedure was ud to maximize the involve-ment of subjects with their online shopping tasks.Subjects were informed that the average product value was$45.The first ssion questionnaire measured their knowledge and familiarity with the computer mou,calculator and wine product class,2their Internet usage,and some demograph-ics.At the end of the questionnaire,subjects were asked to provide their email address and were told that they would be contacted in the following days for the cond ssion. Five days after thefirst ssion,subjects were
nt an email providing a hyperlink to the cond ssion website. Once on the website,they were asked to logon by entering their email address.Following a brief introduction to the ex-perimental website to remind them of the goal of the study (i.e.,cover story),they were randomly assigned to one of three Maximo websites.Once on the website,subjects were 2In thefirst ssion,subjects were asked to complete a subjective knowledge measurement scale(Flynn&Goldsmith,1999)and a famil-iarity measurement scale(Park,Mothersbaugh,&Feick,1994)for each product category.
S.Senecal,J.Nantel/Journal of Retailing80(2004)159–169163 Table1
Brands ud in the experiment
rvants
Computer Mice Calculators Red Wine
Kensington Mou In A Box Optical Pro Canon P23-DH Ceuso Custera1998
Targus Optical Stroller Mini Mou Casio HR-8L a Les Longeroies1998
Microsoft Intellimou Optical Texas Instruments TI5019Callabriga1995a
Logitech Wheel Mou Optical a Casio HR100LC Coteaux du Languedoc,Les Hauts de Lunes1996 a Product recommended by all recommendation sources.
instructed to read a description of the company to clearly understand which of the three types of websites they were visiting.During the experiment,the information and its ,website layout)were held constant across the different treatments levels.Thus,all three Maximo websites were graphically identical.Subjects were then advid that within the next few minutes they would be asked to shop on Maximo’s website and lect three products from three dif-ferent product class.Although Maximo was prented as a real European company with a professional looking web-site it was in fact afictitious company.However,all products ud in the experiment were actual brands available online (e Table1).
As recommended by Nok,Banaji,and Greenwald (2002),thefirst online shopping task was a warm-up task. Its goal was to familiarize subjects with the structure and functionalities of Maximo’s website.Subjects were shown four computer mice and asked to choo one.They were able to evaluate mice bad on their attributes and they were also randomly assigned to one of the four recommen-dation source treatment levels.Hence,most subjects had the opportunity to consult a recommendation page.Sub-jects were free to consult or not the recommendation page (i.e.,Click or not on the“Our recommendation”button). Remaining subjects were assigned to the control group ,they were randomly assigned to one website treatment level and to the“no recommend
ation”condition of the recommendation source factor.On the recommenda-tion page,the recommendation ,human experts) was described to the subject and it recommended one of four products within the product class.Note that the same product was recommended by all recommendation sources (e Table1).After this initial product asssment,subjects were asked to choo one of the four mice prented.
The warm-up task was followed by a cond online shop-ping task.Subjects were randomly assigned to a product ,calculator or wine3).Product class were ran-domized to control for any order effect.The cond shop-ping task esntially followed the same procedure as the warm-up shopping task.Subjects were assigned to the same recommendation source treatment level and were asked to 3The data collection was performed in Canada where the legal age for drinking is18years old.In addition,no significant relationship was found between subjects’professional ,part time worker,full time worker,full time students,etc.)and their subjective knowledge of wine (F(6,473)=1.322;p>.05).lect one product out of four within the product class(e Table1).The product recommended by all recommendation sources was again the same.Following the cond product choice,subjects not assigned to the control group condi-tion and who had consulted the recommendation , subjects who clicked on the“Our recommendation”button) were asked to complete a recommendation source credibil-ity measurement scale.
Following this cond shopping task,subjects were asked to perform the third andfinal shopping task.As part of this task,subjects were expod to four products of the remaining product ,calculator or wine).The third shopping task esntially followed the same procedure as the cond shopping task.Subjects were assigned to the same recom-mendation source as that of previous shopping tasks.Follow-ing theirfinal product lection,subjects who consulted the recommendation page were again asked to evaluate the rec-ommendation source’s credibility.After having completed all three shopping tasks,subjects were asked to complete a shortfinal questionnaire in which they were prompted to guess the main objective of the experiment.4They then ac-cesd a debriefing page explaining the actual goal of the experiment and were logged out of the cond ssion.The debriefing page explained the real goal of the experiment (i.e.,influence of recommendations on product choices),re-assured subjects about their chance to win one of the prod-ucts they lected,indicated that the collected data would remain confidential,and that all rearchers performing the study had signed a confidentiality agreement.Finally,sub-jects were provided the University Ethics Comittee phone number in order for them to call if they had any questions or comments on the study.
The website treatment
Subjects were assigned to one of the three following web-site treatment levels prented in Table2.We purpoly ud afictitious company in order to control for any past experi-ence.
The recommendation source treatment
Four treatment levels were ud for the recommenda-tion source manipulation:other consumers,human experts, recommender system,and no recommendation.During the
4The following open-ended question was asked to subjects at the end of the cond ssion:“To your knowledge,what is the main goal of this study?”
164S.Senecal,J.Nantel/Journal of Retailing80(2004)159–169
Table2
The website treatment levels
Treatment Level Description Provided to Subjects
Retailer“Maximo is a large European store lling products on the Internet.It is currently asssing the
feasibility of offering new products on its website to consumers in your area.Therefore,it is very
interested in learning your product preferences.In the region,Maximo competes with The Bay,
Staples and Wal-Mart,which also offer their products on the Internet.”
3rd party commercially linked to retailers“Maximo is a large European buying group.It is currently asssing the feasibility of offering
new products on its website to consumers in your area.Therefore,it is very interested in learning
your product preferences.Being an intermediary between consumers and a limited number of
partner-retailers offering their products on the Internet,Maximo offers the best products available
at its partner-retailers.In the region,Maximo has the following partners:The Bay,Staples and
Wal-Mart.”
Non-commercially linked3rd party“Maximo is a large European independent organization offering a product comparison rvice on
the Internet.It is currently asssing the feasibility of offering new products on its website to
consumers in your area.Therefore,it is very interested in learning your product preferences.
Being independent,Maximo lects for you the best products available on all sites offering
products on the Internet.Hence,Maximo offers a rvice similar to that of Consumer Reports.”
experiment,if subjects were assigned to a recommendation source,and if they elected to e the product ,clicked on the“Our recommendation”button), they were expod to a recommended product and a descrip-tion of the recommendation source.Bad on pretest results of consumers’preferences,the cond best preferred prod-uct was always propod by recommendation sources.For the“other consumers”treatment level,the recommendation source for the mou product class was described as follows.“This recommendation is bad on other consumer -lections.In fact,bad on the choices of consumers in your area,we have determined the following preferences: Product Consumers who have
lected the product(%) Kensington’s Mou In A
Box Optical Pro
9
Targus’Optical Scroller
Mini Mou
2
Microsoft’s Intellimou
all at once
Optical
18
Logitech’s Wheel Mou
Optical
71
When subjects were assigned to the“human experts”con-dition,the recommendation source was pre
nted as follows:“This recommendation is bad on an evaluation by our team of experts.Our advisors,experts in this product class, highly recommend this product over the others.”The“rec-ommender system”treatment level was described as follows:“This recommendation results from the analysis of the an-swers to the questionnaire that you completed a few days ago during thefirst pha of the study.Our computer system ana-lyzed your answers and,bad on your personalized profile, the system highly recommends this product over the others.”Thus,subjects were led to believe that the recommendation from the recommender system was personalized bad on their answers to the questions of thefirst ssion.Subjects assigned to the“no recommendation”treatment level did not have the opportunity to consult a recommendation source during their sopping ,no“Our recommendation”button was prent on the Maximo website they visited. Product type treatment
The product type was manipulated by using two different product class.Bad on pretest results,the arch product class ud for the experiment was the calculator,and wine was ud for the experience product class.Since it is the only within factor of the experiment,after their warm-up task,subjects were randomly assigned to either the calculator or the wine product class on theirfirst shopping task and assigned to the other product class for their cond shopping task.
Measures
Thefirst dependent variable was the influence of the rec-ommendation source on consumers’online product choices. The influence was measured by a dichotomous variable. Each product choice was categorized as either a decision to follow or not to follow the product recommendation.The remaining dependent variables were related to the credi-bility of recommendation sources.The measurement scale developed by Ohanian(1990)was ud to asss recom-mendation sources’experti and trustworthiness.Results from a pretest(n=39)and from the experiment(n= 487)both show that the measurement scale is reliable.The Cronbach’s alphas for the experti dimension ranged from 0.88to0.91and from0.84to0.88for the trustworthiness dimension.
Manipulation checks
迷失第三季Following Perdue and Summers(1986),all manipulation checks were performed during pretests.A ries of four pretests were necessary to achieve effective online manipula-tions.After each pretest,necessary iterations were made on
