Firm-Created Word-of-Mouth Communication: Evidence from a Field Test∗

David Godes and Dina Mayzlin† August 2007

The authors would like to thank participants at the INFORMS 2003 Annual Meeting in Atlanta, marketing

seminars at Boston University, Washington University, Yale University, University of Texas - Dallas and New York University; the CU-Boulder Choice Symposium, and the Marketing Science Conference in Rotterdam for helpful comments on previous drafts of this paper. We would also like to thank Toni Wegner for her thoughtful assistance in experimental design; Dave Balter for generous comments and support as well as Rock Bottom Brewery for the generous donation of time, data and wisdom that made this research possible. Please do not cite or circulate without permission of the authors. Comments welcome. † David Godes is an Associate Professor in the Graduate School of Business Administration, Harvard University, Soldiers Field, Boston MA 02163 email:, phone: 617-495-8040, fax: 617-496-5637. Dina Mayzlin is an Associate Professor in the School of Management, Yale University, New Haven, CT 06520

Firm-Created Word of Mouth Communication: Evidence from a Field Test Abstract In this paper, we investigate the effectiveness of the firm’s proactive management of customer-to-customer communication. We are particularly interested in understanding how, if at all, the firm should go about effecting meaningful word-of-mouth. To tackle this problem, we collect data from two sources: 1) We implemented a large-scale field test in which a national firm created word of mouth through two populations: customers and non-customers, 2) We collected data from two laboratory studies. We break our theoretical problem into two subproblems. First, we ask, “What kind of WOM drives sales?” Motivated by previous research, we hypothesize that for a product with low initial awareness level, the WOM that is most effective at driving sales is created by less loyal customers, not loyals, and occurs between acquaintances (not friends). We find support for this in the field test as well as in a lab setting. Hence we demonstrate the potential usefulness of exogenously-created WOM: conversations are created where none would naturally have occured otherwise. Then, we ask, “Which agents are most effective at creating this kind of WOM?” In particular, we are interested in evaluating the effectiveness of the commonly-used opinion leader designation. We find that while opinion leadership is useful in identifying potentially effective spreaders of WOM among very loyal customers, it is less useful for the sample of less loyal customers. We suggest a measure of one’s social network breadth may help to identify less-loyal customers as potential WOM creators. replicated in a laboratory study. Keywords: Word of Mouth, Promotion, Advertising The results are



“The Beatles said you ‘Can’t buy me love,’ but marketers’ version of that refrain is closer to ‘Can’t buy me buzz.’ Buzz belongs to the people.”’s Jeff Jarvis 1 In September 2005, NBC launched the second season of its reality show about weight loss, “The Biggest Loser.” In preparation for the new season, NBC ran ads in early August asking viewers to fill out a survey at a Web site. Out of all the applicants, 1,000 “biggest” fans were chosen to throw parties during an advanced screening of the show’s premiere. The hope was that this, along with the resulting word of mouth (WOM), would generate interest in the show.2 In 2001, Lee Dungarees wanted to improve its image with teen boys. Their agency identified 200,000 “influentials” from online communities devoted to video games. The firm then emailed each a series of short films from unknown characters who turned out to be protagonists in a video game commissioned by the firm. On average, these films were forwarded to about six people each. To play the game, however, one had to go to a retail store and get a code from a pair of Lee jeans.3 In March of 2006, WD-40 hired Proctor and Gamble to promote their new product extension, the “No-Mess” pen. The product was promoted through P&G’s Vocalpoint, a panel of influential moms who were pre-selected via a survey based on their ability to be “connectors.”4 Hasbro in 2001 launched a new handheld video game called POX. To do so, they ran surveys in Chicago area elementary schools to find the “coolest” kids in each school. Once 1,600 kids were chosen, they were each armed with a backpack filled with samples of the game to be handed out to their friends (Godes and Ofek, 2004). There are several common threads among these examples. “engineer” WOM among their customers. First, the firms in these cases tried to That is, rather than hoping that satisfied customers would

tell people about their products, these firms took actions to increase the number of conversations that were taking place. Second, they each attempted to identify who the “key influencers” would be in their respective situations. NBC used self-reporting, Lee Dungarees used observational methods, Hasbro used a combination of sociometry and self-reporting, and WD-40 found a fit between their product and the Vocalpoint panel. For each firm, the implementation of their WOM campaign played a primary role in their marketing effort during the respective time period. One notable difference is in the approaches undertaken by the marketers of the two mature products, NBC and Lee Dungarees. While NBC recruited the most loyal users for its campaign, Lee Dungarees instead focused its efforts on influentials, regardless of their existing relationship to the product. The past several years have witnessed a marked increase in attention paid to “buzz” in the popular and managerial press. Moreover, marketing managers are increasingly interested in taking actions in order to influence, directly or indirectly, WOM. Hence, we can think of firm-created WOM as a hybrid between
1 Source: 2 Source:

Abbay Klaassen, Adage, Jan 31, 2007, “Marketers Can’t Buy Buzz.” “NBC Goes Viral in Effort To Win With ’Biggest Loser’,” by Gavin O’Malley,, Sep 8,

2005. 3 “Buzz Marketing: Suddenly This Stealth Strategy Is Hot—but It’s Still Fraught with Risk,” By Gerry Khermouch with Jeff Green, Business Week, July 30 , 2001. P. 50. 4 “P&G Provides Product Launchpad, A Buzz Network of Moms,” By Jack Neff,, March 20, 2006.


traditional advertising and consumer word of mouth in that the former is firm-initiated and firm-implemented while the latter is customer-initiated and customer-implemented. WOM marketing, on the other hand, may be characterized as being firm-initiated but customer implemented. In February 2007, a search for "word of mouth marketing" in the title yielded 16 different books at Managers’ motivation for looking toward interpersonal communication as a potential new tool grows out of a sense that traditional media advertising is declining in effectiveness, particularly with respect to younger demographic groups (Keller and Berry, 2003). However, for all of the importance that managers are apparently placing on the creation of these WOM strategies, there has been little academic research looking at WOM from the firm’s perspective. In this paper, we address this issue by first investigating whether a firm is able to orchestrate a WOM campaign that drives sales. As is demonstrated by the quote at the beginning of the paper, there is some skepticism as to whether a firm can in fact “buy” WOM. This is a materially different question from that of whether naturally-occurring WOM may drive sales. should create in order to impact sales. Next, we examine what type of WOM the firm In particular, should a firm adopt NBC’s strategy of focusing on

its loyal customers to spread WOM or should it instead try to adopt Lee Dungarees’ strategy of spreading the message through people who may not have an existing strong relationship to the product? The former strategy is often discussed in the business and popular press.5 Finally, we turn to the firm’s targeting problem in its design of a WOM marketing campaign. What are the characteristics of the disseminators that are associated with higher amounts of impactful WOM? Do these characteristics differ between loyal and less loyal consumers? To what extent are the existing scales — those designed to predict naturally-occurring WOM — effective at measuring the propensity to spread firm-created WOM? To answer these questions, we collect two types of data. First, with the cooperation of two separate firms, we collect field data as part of a WOM marketing campaign. In this campaign, two types of people were engaged in the process of information dissemination: (a) a set of people who had demonstrated behavioral loyalty through their usage of the firm’s product and (b) a panel of people who had no loyalty to the firm. In addition, we run two laboratory studies. The studies allow us to control for variables that are difficult to control in the field and to explore the mechanism behind the effects. We provide three sets of results. Using field data, we empirically demonstrate for the first time, to our knowledge, that the firm can create WOM that drives sales. This is an important result and is distinct from previous work that has demonstrated that naturally-occurring WOM can drive important outcomes (Chevalier and Mayzlin, 2006; Manchanda et al., 2006; Godes and Mayzlin, 2004; Van den Bulte and Lilien, 2003; Foster and Rosenzweig, 1995). This is the first study to our knowledge to demonstrate that discussions created by the firm’s actions rather than simply positive product experiences can have such an effect. From a managerial perspective, this result gives credence to the evolving notion that not only is WOM important but it’s actually something that is under the firm’s control (Biyalogorsky et al., 2001; Mayzlin, 2006).
5 “When

customers are truly thrilled about their experience with your product or service, they become outspoken ‘evangelists’

for your company. Savvy marketing professionals are discovering that this group of satisfied believers can be converted into a potent marketing tool to grow their customer universe.” — Creating Customer Evangelists by Ben McConnell and Jackie Huba, Dearborn Trade Publishing, inside flap, 2003.


Very loyal customers are likely to live in social networks in which either (a) others are also loyal to the firm or (b) others are aware of. The paper concludes with a discussion of the results and suggestions for future research in this area. In this regard. 1993) while (b) follows from the idea that loyal customers are likely to have already engaged in WOM about the firm’s products. a WOM campaign is primarily beneficial to the extent that it results in the spread of information (as opposed to influence or persuasion). we argue — and present evidence — for the idea that it may be more impactful for the firm to target less loyal customers to participate in a WOM campaign. The following Section presents the theoretical background to our hypotheses. (1966). Godes and Mayzlin (2004) demonstrate a positive relationship between online WOM — in particular. Garber et al.v. (2006). the firm’s products. 2 Theoretical Development Beginning with Katz and Lazarsfeld (1955) nearly half a century ago. for a product with initial low awareness level (such as the one studied here). In Section 4. In Section 3. we outline the set-up of the field test. Others have inferred the impact of WOM from the geographical evolution of sales data: Foster and Rosenzweig (1995). but not interested in. Bell and Song (2004). we demonstrate that while “opinion leadership” is associated with higher propensity to spread WOM among loyal consumers. We explore another possible indicator that offers encouraging results: network breadth.Second. The rest of the paper proceeds as follows. While the result is somewhat surprising ex ante. we engage in an exploration of how the firm might find those less-loyal customers most likely to engage in WOM. After testing first the usefulness of the well-known opinion leader construct. Arndt (1967). as one might expect. it actually follows directly from commonly-accepted ideas about social networks. we present results from the field data as well as a follow-up lab experiment on the impact of firm-created WOM on sales. (2004).” The literature has focused primarily on the former which is characterized by conversations that occur naturally among consumers as a 3 . we demonstrate that it is not necessarily the highly-loyal customers that generate the important incremental WOM. shows. (1969) have corroborated the primacy of WOM as a key driver of firm sales. we then engage in an exploratory analysis of another potentially-useful metric: network breadth. On the contrary. Engel et al. Both of these imply that WOM from a less-loyal customer is likely to have a bigger impact. the impact of WOM on consumers’ actions. Context (a) follows from the concept of homophily in social networks (Rogers. 2003). While Van den Bulte and Lilien (2001) question the conclusion reached by Coleman et al. preferences and choices has been of great academic interest. having demonstrated that the firm should look for less-loyal customers to spread WOM. (1966). in fact. the same authors later determined that a more-sophisticated decomposition of the physicians’ adoption decision did. Finally. Researchers such as Coleman et al. yield evidence for the role of interpersonal influence (Van den Bulte and Lilien. its “dispersion” across communities — and ratings for t. We distinguish between “endogenous WOM” and “exogenous WOM. Manchanda et al. Specifically. the same is not true for less-loyal customers. In Section 5. we study who creates impactful WOM again using both the field data and another lab experiment.

it is also possible that inferences can be drawn with respect to firm strategies. That is. Those disseminators that may be most persuasive may not be the ones that will help the firm achieve maximal awareness. The literature on influence has strong predictions on how decision makers are affected by their peers. We begin by developing a theoretical foundation for why some kinds of exogenous WOM are more or less likely to lead to higher sales for the firm. this does not necessarily imply that either creating a substantial amount of exogenous WOM is possible for the firm or that such exogenously-created WOM would have the same beneficial impact on the product’s sales as previous research has demonstrated with respect to endogenous WOM. Reingen et al. consumers discount received messages. the effectiveness of the sender’s WOM communication may be decreased. attractiveness and optimal design of WOM marketing programs by decomposing the problem into two separate sub-problems. (1984) show that consumers’ brand choices within a social group are often congruent. which in turn implies that the program does not work. While the literature has shown the important relationship between endogenous WOM and sales. They show that the firm should offer rewards for customer referrals only if people are somewhat demanding but not too demanding. (2001) who investigate the optimality of customer referral programs. This issue is addressed by Mayzlin (2006) who shows that the firm’s creation of anonymous online WOM may be a profitable equilibrium strategy even when consumers are aware of the possibility that the firm is creating it. a WOM campaign might impact outcomes by affecting either (a) awareness and/or (b) preference. 4 . Is a WOM marketing program an effective way to attract customers? This issue is addressed by Biyalogorsky et al. the latter refers to WOM created as the result of the firm’s actions.function of their experiences with the product.and messagelevel inferences. can firms get people to talk about their products in such a way that it impacts sales? Moreover.1 What Kind of Exogenous WOM Matters? As with other media. On the other hand when the costs of creating anonymous WOM are very low. On one hand. Interestingly. We add to this literature on the feasibility. While this work focused on individual. Assuming an effective WOM marketing program is desired. there may be an inherent tension between achieving these two objectives. exposure to a WOM episode might make someone aware of a product they had not been aware of before or it might persuade them by changing the expected utility they had assigned to that product. We then consider the question of identifying those disseminators who are most likely to create impactful WOM. there’s a question about whether or not the firm should bother building exogenous WOM. 2. how? The answer is not necessarily obvious. one may process the message in such a way that a “change in meaning” may occur. one needs to ask whether it is feasible. (2004) who show that when consumers perceive ulterior motives. given the inferences a recipient of WOM information may draw. when confronted with a persuasion attempt. if so. In contrast. This model has been applied to the WOM domain by Verlegh et al. comparing them to another method of attracting customers: cutting prices. That is. The Persuasion Knowledge Model (Friestad and Wright. 1994) suggests that. For example.

necessarily better than the other. Thus. information transmitted between acquaintances or strangers should ultimately reach more people — i. lead to higher awareness — than if it had been transmitted between friends or relatives. This would seem to indicate that a loyal disseminator — one who is more familiar with the product — is more likely to be persuasive than a less loyal disseminator.There is also an extensive literature documenting that the similarity between sender and recipient may increase the persuasiveness of the communication (Cialdini and Sagarin. One might characterize the choice as being between a “behavioral” measure or an “attitudinal” measure. Swan and Oliver. 5 . 1989. 1982).. Granovetter (1973) showed that it is essential to distinguish between “strong ties” and “weak ties” in understanding the flow of interpersonal information. However. (2001) use cellular automata to investigate A fundamental implication of this research is that information. However. 1993) and hence to have created more WOM than average (Bowman and Narayandas. Following past research. 2001). the definition of loyalty is not without disagreement. We investigate here the role of customer loyalty. 1992. we expect loyal customers to be more satisfied with the product than average (Anderson and Sullivan. Capturing attitudinal loyalty for all but a sample of customers is likely to be a difficult undertaking. 2005. 2001. 1998. such as public luxuries (Bearden and Etzel. we would expect a communication with a friend or a relative to be more persuasive than a conversation with an acquaintance or a stranger. For the purpose of designing a WOM marketing campaign. 1973). Similarity may be especially important when new attitudes and beliefs are formed — such as for a new product (Kardes. Bolton and Drew. the characteristics of WOM that are typically associated with higher persuasiveness may. 1990. the same critical dimension — the extent to which WOM reaches previously-uninformed customers — may also be related to more-easily identifiable customer characteristics as well. we argue that behavioral measures may be more practical since most firms have access to some measure of behavioral loyalty.. be associated ultimately with less breadth of awareness of the message. This is an attractive criterion since it is relatively easy for the firm to select customers for inclusion in a WOM campaign based on loyalty. Kruglanski and Mayseless. on the other hand. it is clear that the measures are quite distinct (Jacoby and Kyner. The disseminator’s expertise with the product has also been found to increase the influence of her advice (Petty et al. Anderson. we implement an attitudinal measure to check the robustness of our results. 2005). In our experimental analysis.e. 1990. the relative macro-level impact of strong and weak ties and find that the latter may have a bigger impact even though the former are activated more frequently. the approach we take in our core empirical analysis is to adopt observed behavior as a measure of loyalty. 2002) — and for certain types of products. Mazen and Leventhal. this is not easily managed by the firm within the context of a WOM marketing campaign. Thus. The former is typically characterized by repeat purchase behavior while the latter “includes a While one is not degree of dispositional commitment” to the brand (Chaudhuri and Holbrook. it is often information communicated via a weak tie that results in a greater increase in the number of new people that are informed. 1972). While the effectiveness of WOM may depend on the strength of the tie across which a message is communicated. An important argument in this work is that weak ties form the bridges between otherwise isolated strong-tie networks. Since those in the same social networks are likely to have similar Goldenberg et al. However. Reichheld and Sasser.

it seems reasonable to expect that this discounting is less likely to occur with respect to the recommendations by loyal customers since they may be able to justify their recommendation with personal experiences. with loyal customers are more likely to already have been informed about the firm and its products before the beginning of the WOM campaign. we expect the persuasiveness of a message to be highest when sent by a loyal customer to friends and relatives. if anything. Here. we expect the ex post breadth of awareness to be higher when the message is sent by a less-loyal customer to acquaintances and. perhaps. 1977). in some cases. Hence. we posit the following two hypotheses: Within the context of an exogenous WOM campaign for a product with low initial awareness levels. which would have important implications on optimal allocation of resources to customer retention. We note that it is likely that persuasion knowledge (Friestad and Wright. H2: The sales impact of incremental WOM to an acquaintance is higher than that to either a friend or a relative. Mela and Vidal-Sanz (2006) find sizable network effects in an online auction site. A traditional calculation of customer lifetime value considers direct purchases only (see Gupta. (2003). As we demonstrate in Section 3. awareness is relatively low for the product under study. More formally. the incremental WOM created by a less-loyal customers would result in greater awareness compared to that created by a loyal customer due to the lower levels of endogenous WOM that were created by the less-loyal customers. if the price of trial is relatively low and if there is not a lot of existing awareness. Hardie. Less-loyal customers are probably less able to do so and may thus risk activating concerns about ulterior motives. If the product is well-known. as was pointed out by Hogan et al. Hence. 1994) will play a role as the recipients of these recommendations.Holmes and Lett. V. We emphasize that the main contribution of the current paper lies in H1. the value of a lost customer does not just include the revenue generated directly by the customer but also the lost value of the social interactions associated with the customer. Specifically. Here. As a result. Lin and Sriram (2006)) and would imply that the loyal customers are the firm’s most valuable asset. Thus. we argue that persuasion 6 . may invoke their “schemer schema” and discount the value of the information. we expect that those with ties. we suggest that the network effects associated with WOM may be especially large for less-loyal customers. However. we expect that a WOM campaign would have a bigger marginal impact by creating conversations where none ordinarily would take place rather than amplifying conversations that are already naturally occurring. it may suggest that less-loyal customers are more valuable than current models give them credit for. On the other hand. Gupta. Whether the firm should maximize awareness or persuasiveness is a function of the status quo prior to the campaign. In summary. weak or strong. Kahn. strangers. On the other hand. we assume that the latter better matches our field study context.Kumar. H1: The sales impact of incremental WOM from a less loyal customer is higher than that from a loyal customer. It is useful to relate this investigation to the CRM literature. then the firm should probably concern itself with persuading consumers of the product’s value. Our test of H2 represents an important field test of the application of the influential results reported by Granovetter (1973). then the firm might concentrate primarily on the spread of information. However.

little research has investigated potential moderators of the opinion leader-WOM relationship. Again.2 Who Creates WOM that Matters? The examples offered in Section 1 suggest that the implementation of a WOM campaign requires that the firm identify effective disseminators of information. While we’d expect a loyal opinion leader to create WOM. With respect to (1).6 Specifically. the latter show that loyal customers of loyalty. we expect intuitively that a consumer who has more ties to others or is. may be particularly likely to engage in negative WOM following a dissatisfying experience. We hypothesize that the same mechanism exists with respect to product familiarity. As a result. it is less clear what a less-loyal opinion leader’s WOM behavior would be. 296). 2.” Opinion leaders risk losing their status when they become “too innovative. The other component of WOM volume is the likelihood of interpersonal interactions. in some sense. With respect to the latter. We consider here two important components that contribute to the probability that a customer creates WOM about a product: interacts with others. she may not be “loyal” because (a) her prior is low and/or (b) the distribution is too diffuse. Clearly. Jacoby and Hoyer (1981). H3 : While opinion leadership is associated with the creation of more WOM for loyal customers. While this has proven to be a reliable scale in general.knowledge is likely to bias against H1. not the main effect Moreover. 7 . and (2) the frequency with which one is important to note that our concern here is the interaction between loyalty and opinion leadership. With respect to (b). While there is less theoretical guidance for this. it may increase the likelihood of negative WOM. this is less true for less-loyal customers. We investigate here the moderating role of loyalty. the customer may be uncertain about the product due to lack of experience with it. In fact. As shown by Anderson (1998) and Bowman and Narayandas (2001). To fulfill this role. H3 represents our view of the likelihood that — conditional on already being in a conversation — one discusses the underlying product.” recommending products that those in their social circle are not ready to adopt. the volume of WOM may follow a U-shaped pattern in satisfaction. we expect an interaction between opinion leadership and loyalty in a model of WOM volume. As noted by Rogers (1993. the literature is less clear. Rogers (1993)). If we conceptualize an opinion leader as a Bayesian with a distribution over the firm’s value. that one discusses a specific product. in the former case. a common approach used in marketing has been to identify “opinion leaders” (see Katz and Lazarsfeld (1955). the opinion leader should demonstrate prudent judgement in decisions about adopting new ideas. conditional on an interaction with another. “If an opinion leader becomes too innovative or adopts a new idea too quickly. p. King and Summers (1970). Bloch and Richins (1983). One role of the opinion leader in a social system is to help reduce the uncertainty about an innovation for his or her followers. being an opinion leader would not increase the likelihood of positive WOM. followers may begin to doubt the opinion leader’s judgement. “more social. opinion leadership should be a less-useful predictor of WOM in a low loyalty context as compared with the high loyalty context. King and Summers (1970) offer a scale for this purpose.” would have more opportunities to disseminate 6 It (1) the probability.

1 Field Test Set-Up The field test included two organizations: an agency (BzzAgent) and a restaurant chain (Rock Bottom Brewery). they are rewarded with prizes including free products. 8 . To do so. She is likely to interact with others more often and is therefore more likely on average to engage in WOM. we collected two types of data: 1) we designed and implemented a field test. While all of the BzzAgents were invited to participate in the campaign. BzzAgent maintains a panel of “agents. 3. BzzAgent trains the agents and manages the process in which the agents create WOM for the client. 99% answered “Yes” when asked whether they “liked” the firm. As part of this agreement. the customers were invited on a rolling basis such that the most loyal customers (those who were the heaviest users of firm’s Category A product) were invited first. To address this. BzzAgent had only run campaigns utilizing their panel of agents and had never worked with its clients’ own customers. Among customers. The field test. The process ultimately yielded 381 customers and 692 non-customers who agreed to participate in the program. BzzAgent is a marketing agency engaged in the business of creating WOM communication for its clients.5% reported that they were indifferent. After a specified number of purchases. and the agency’s panel. Subjects were invited to participate in the field test via an email from either BzzAgent (for non-customers) or Rock Bottom (for customers). involved a comparison of the WOM created by these two populations: the members of the firm’s loyalty program.7 The firm’s product line is split into five categories as shown in Table 1. The firm’s gross sales in the twelve months leading up the program were over $100 million. the agency agrees to lease a specified number of its agents to the client. To our knowledge. we do not reveal detailed or disaggregate financial information. 3 Data Collection To test the hypotheses above. . no other major advertising the request of Rock Bottom. etc.5% answered “No” and . Rock Bottom Brewery did business in 15 markets across the United States. on the other. and 2) as a follow-up to the field test. Prior to this research project. we explore a preliminary approach to measuring the construct we call “network breadth” and assessing its impact on WOM created. We refer to the former as “customers” and the latter as “non-customers” since they had little-to-no information about the retail chain before the study. which lasted 13 weeks. on one hand.” In a standard project. we conducted two laboratory experiments. effort had taken place during the duration of the field test.information to her network. Crucial to our test is the fact that Rock Bottom maintains a loyalty program centered around Category A. coupons and other promotions. then those who were somewhat less loyal. At the time of the study. Several thousand customers hold a card which they present at the time of purchase. The objective was to recruit a total of 1.000 subjects. The email explained what the campaign was about and noted that their participation and performance would qualify them for potential prizes. 7 At The email invitations continued in this fashion until a sufficient number of customers had registered.

the sample populations were mutually exclusive. redeemed. Notably. the more prizes they were able to win. relative. 213 had previously participated in campaigns for the agency’s clients. 86% of non-customers had never heard of the chain prior to the project. The instrument captured important information about the agent’s social networks as well as demographic and attitudinal data that we expected would be useful in building an individual-level model of WOM creation. There appear to be some differences between the customer and non-customer sample. While the broader networks associated with the non-customers may result in more WOM. each subject was directed to a website to fill out an extensive survey. In addition. our field test ran in 15 markets for thirteen weeks. The grading was handled by the agency’s staff as part of their normal contract obligations. on average older. stranger or some other relationship. 17% on three and 7% on four. The WOM creation process officially began in April 2003 and ran through June 2003. the agency reports that most of the points earned by participants are never.Among the 692 non-customers. customers and non-customers do not appear to differ significantly in terms of opinion leadership. are more likely to work full-time. Once she agreed to participate in the campaign. In particular. we simply maintained theirs to allow for data consistency. we would not expect to find any impact of reported WOM on sales. a single report often represented more than a single new person being Each WOM report was graded on its potential to create meaningful WOM. to the extent that actual WOM is significantly lower than that reported. the agents provided information on their relationship with the recipient: whether she was a friend. and eat out more often than noncustomers. However. Participants had an incentive to create meaningful WOM since the higher their scores. Once it began. we wouldn’t expect that this would affect the impact of a given WOM episode. Moreover. Nonetheless. informed. 29% on two. Importantly. Moreover. whether such reports are a useful measure is ultimately an empirical question. After filling out the online survey. customers are. They were directed to a website through which they were to report in detail each time they engaged in a WOM episode. Moreover. is important to note that — since the agency had been in business for a number of years — they had a full battery of questions they asked of the agents as well. we have sales data at the market level for each week of 8 It Hence. We have data on the amount of WOM created by each agent over those weeks. By design. acquaintance. This package also contained specific details on how the campaign would be run and how the agents were to participate. Of these. That is. 47% had worked on one campaign. each participant received a package of information about Rock Bottom and its products as well as specific suggestions for creating WOM for the firm.1). To the extent that the questions we wanted to ask were similar to the ones they asked as part of their routine. none of the non-customers were members of the firm’s loyalty program. 9 .8 Table 2 provides comparisons between groups on the available demographic data. these incentives were extremely low-powered: the average prize was valued at around $15. Note that the self-reported nature of the WOM episodes may give rise to some exaggeration. In summary. the agents were asked to report their WOM creation activity. are less likely to be students. Note that the participants often reported the transmission of information to several people in a single report. on the relationship between the agent and the recipient and on the individual characteristics of each agent. non-customers appear to have somewhat broader networks (we describe this scale in more detail in Section 5. in fact. the bias should be the same for both customers and non-customers.

when the subject categorized her relationship with the recipient as “Other. Note that we focus here on the impact of WOM on category A sales. which is the customer condition. we ask whether the firm can create exogenous WOM that drives sales and. We define W OMijt r targeted. we test them with an aggregate market-level model. 4 The Effect of Exogenous WOM on Sales In this section. we also estimated specifications B that include a year-earlier sales term Si.e.. or N . They were allowed. other} as the total number of reports filed in week t in market i that reflected WOM from someone in condition j (i.t−52 to capture any recurrent shocks that might be seasonal and market specific. customer or non-customer) to a person or people with whom they have a relationship that can be characterized by r. we do not report these results here. for example. Before presenting the results. These ranged from “Internet” to “boyfriend.N X " r∈R X ωr j · r W OMijt # + 15 X i=2 µi + 12 X t=2 τ t + εit (1) A where Sit represents sales of category A in market i in week t. Table 3 shows the summary statistics for the data and Table 4 shows the pairwise correlations.” she was provided the opportunity to describe in more detail what the nature of the relationship was. we only know that a report is made and that the agent felt that r best captured their relationship. however. the non-customer condition. acquaintance. how? Our core analysis is performed on data from the field test. the agent couldn’t report that she told a group of people that were made up of friends and acquaintances. So. We include fixed effects for both the week τ t and the market µi .the campaign and for year-earlier periods. 10 . if so. relative. We also present data from a follow-up lab study that allows us to control some factors that were difficult to control adequately in the field. stranger. We don’t know exactly how many people this WOM episode impacted.10 Since this term is never significant. since this is both where the loyalty program is focused and where the campaign itself was r is the set of possible relationships between the sender and receiver of WOM information. 1 0 As an example in another domain. The main model we estimate is the following A Sit = j∈C. While we expect the combination of these two sets of intercepts to capture most of the local and seasonal shocks.1 Field Test Since H1 and H2 relate to firm sales. it is important to caution against drawing any conclusions from Table 3 regarding 9 Each report allowed only a single designation for r. competition and store location. Our WOM data are captured in W OMijt where r ∈ R = {f riend. Each sender belongs to one of the two possible sample conditions j: C.” We perform some ex-post re-classification of this category below.9 In addition. it might be the case that hotels in New Orleans have a large demand shock during Mardi Gras (February) each year while no other markets would expect to see such a systematic shock. 4. and since the estimates with this term included are qualitatively equivalent. to file two reports if they so chose. We particularly draw the reader’s attention to the µi terms since they are meant to capture all systematic marketlevel factors including market size.

11 . the week effects τ t are included in the estimated model and are also reflected in the reported models’ fit statistics. they may find less-loyal customers to be more helpful. it seems. That is. five of the eleven included effects are significant at the p<.12 analysis at different levels of aggregation of the WOM variables. refer-a-friend programs such as those investigated by Biyalogorsky et al. Most important. we calculate the R2 statistic for the differenced model. we ran a factor analysis on the ten WOM variables ωr for each market ijt i in week t. therefore. The intra-condition correlation is probably due to individual factors such as WOM episodes in which multiple people — friends. Model (3) reflects the completely disaggregated approach shown in Equation 1. acquaintances — were informed at once. 1 3 For reasons of confidentiality and exposition. create exogenous WOM among non-customers that has a significant and measurable effect on sales.10 level Since the coefficients on the lagged variables were not significant with the effect sizes representing systematic weekly swings of around 10% of sales.11 The results from our initial regressions are shown in Table 5. The coefficient on the WOM created by non-customers suggests that each WOM episode yields average incremental category sales of $192. we consider the overall amount of WOM (i. we control for time and market effects. 1973). we do not report the coefficients on either the market or the week effects. Finally. These data are market-level statistics and.e. We first consider the implications of the results in Models (1) .95 for all the models in the table. this model shows that the impact of incremental exogenous WOM created by customers that have no relationship to the firm is significantly higher than the WOM created by customers. (2001) — should not confine themselves to their base of highly-loyal customers. relatives. The average R2 if we include the market fixed effects is over 0. we form a variable W OMit = ¢ P ¡P r j r∈R W OMijt ). Note also that. On the contrary. What it suggests is that firms interested in designing an exogenous WOM program — for example.(3) in Table 5. WOM through acquaintances has significantly more impact than WOM to those 1 1 To check whether this intuition seemed reasonable. we present our In all models. which lagged values of the WOM variables were included. Model (2) decomposes the WOM according to the the sender’s relationship to the P r firm: W OMijt = r∈R W OMijt . The more-detailed analysis in Table 5’s Model (3) demonstrates that. consistent with the theory of weak ties (Granovetter. Model (4) removes removes independent WOM variables that appear to have no explanatory power. and since the results were not qualitatively different. Not surprisingly. and the agent decided to separate the reports across relationship type. This provides initial support for H1. A follow-up factor analysis constrained to two factors using a varimax rotation revealed a pattern of loadings in which all of the customer variables loaded most heavily on one factor and all of the non-customer variables loaded most heavily on the second factor. the R2 we report estimates the percentage of variance explained by the model beyond the market fixed effects. non-customers to create WOM. It is important to note that this does not mean that overall WOM by non-customers is more impactful than that by customers. reflect the different proportions of customers and non-customers in each market. when the objective is to spread information. they are not reported here. though not shown. 13 In Models (1)-(3). variance in the model.the proclivity of customers vs. the results are also clear in reinforcing the message in Godes and Mayzlin (2004): All WOM is not created equal.. This can be seen in Model (2). Details avaialable from the authors. With respect to the former. In Model (1). one sees in Table 4 that there is a fair amount of correlation within the customers and within the non-customers but little across the conditions. Model (5) includes additional data as a follow-up to the main results. they are all quite strong and highly significant. Second. The first is that the firm can. 1 2 In the aggregate model. This analysis yielded only two factors with eigenvalues greater than 1. they explain the bulk of the Note that we also estimated models in With respect to the week effects.

Given the apparent lack of explanatory power of the WOM variables other than Customer-to-Other WOM and Non-Customer-to-Acquaintance. taken to its logical next step. Thus. the effect size and significance level of the Non-Customer-to-Acquaintance coefficient are increased.05 level.16 While the noncustomers’ WOM is not significant at the p<. However. there is likely to be an offsetting force associated with credibility which might render a suggestion from a stranger to be less credible. However. decreased somewhat for the Customer-to-Other WOM. To resolve disagreements (the ex-post classified "friend" and "acquaintance" categories had 50% and 35% agreement 1 4 Another In this model. 12 . 1 5 Our decision to run 10. First. This argument. as the latter can be viewed as an approximation to the bootstrap (Efron and Tibshirani.15 As these results show. we performed a Wald test on the composite hypothesis that these eight other coefficients are all zero. The table contains confidence intervals at both the 90% and 95% levels for each of the coefficients of interest. These results begin to provide compelling evidence for the fundamental idea that focusing exclusively on loyal customers for a WOM marketing campaign may not be optimal. our claim that less-loyal customers may be more valuable to the firm in terms of their ability to create meaningful. there is support for H1. Note that this result is consistent with the results on job search reported by Granovetter (1973). we analyzed in detail the free-text reports submitted with each WOM report. an “online” category or an “indeterminate” category. These results are shown in Table 5’s Model (4).10 level when all possible WOM variables are included (see the re-estimated Model (3)).with stronger ties in the social network. as is standard.14 Table 6 presents the results of this analysis. So. we limit our discussion here to significance levels rather than effect sizes. In an effort to understand better the source and nature of this impact.000 draws per model is due to the fact that we are presenting bootstrap results that are bias-corrected and accelerated. the non-customers’ WOM does achieve significance. we implement only the bootstrap. would seem to suggest that WOM to strangers would be yet more powerful. even under far-less-stringent assumptions. this approach requires at least 1. we estimate a third model in which — besides the market and week fixed effects — we include only Customer-to-Other WOM and Non-Customer-to-Acquaintance. incremental WOM.000 draws to perform well. we re-estimate Table 5’s Models (2)-(4) via a bootstrap approach. as a way of validating the results under a less-stringent set of distributional assumptions. these bootstrap estimates provide further comfort for the robustness of our results. 1 6 Note that the regression-generated coefficients themselves are retained for purposes of performing the bootstrap.41. such as friends or relatives. there are a number of important questions raised in these initial results that we address next. relative. 1993). we hired two raters who were not informed about the goals of the study to use this free text to classify each of these “other” reports into either one of six categories represented by the four existing categories (friend. acquaintance. In the re-estimated Model (2). H2 is also supported. when we exclude those that have no explanatory power. Thus. However. Both of these are approach would have been to use jacknife estimation. the non-customers’ WOM is significant at the p<. According to Efron and Tibshirani (1993). the crucial test of our theory in which customers and non-customers are compared. Thus.91). This test could not reject the null (F=. Another aspect of our results requiring additional analysis is the impact of customers’ WOM to the “Other” category which was somewhat unexpected ex ante. stranger).p=. To do so.

e. we re-estimated Table 5’s 1 7 Indeed. the category — are better at identifying the communities in which this particular category is discussed. and interest in." We re-ran the model and present the results in Table 5. The fact that customers’ WOM has a large effect online but not offline may be viewed as quite consistent with the network explanation we offer above. some of the non-customers had previous experience generating WOM for the agency’s other clients. the points are awarded as much for the originality of their ideas and the richness of Finally.” Among these. Clearly. This. issue. Given this. About 30% were categorized as “online. they were unable to reach agreement on many disputed reports. of the campaign. finding and needs to be interpreted with caution. we decided to separate the “other” category into "other online" and "other not online. the coefficient on non-customers’ WOM to acquaintances is hardly affected by this analysis. indicating the highly subjective nature of the ex-post classification task into these categories. we note that there are good reasons to believe that the impact of the non-customers’ experience may First. the raters discussed the disputed reports in person. the chain’s customers were probably more likely to have spread WOM offline than online prior to the start Hence. we suggest that these agents should not be thought of as “professional WOM Nor are they rewarded directly for the While they receive creators. suggests that the nature of the WOM — not just the size of the impact — generated by loyal customers and non-customers may be different. it may suggest that the nature of the impact of loyal customers’ WOM may be different from that of less-loyal or non-customers. 13 . there seems to be a very large effect of WOM created online by customers. same is not true with respect to non-customers.” In fact. As noted above. This raises potential questions about the source of the results we present here. in this case. their feedback as for any projected impact. impact of their recommendations (i. Further research is needed to explore this interesting issue. the raters had a 90% agreement rate. though the That is. and ex post. Model (5). we would expect that the true expertise associated with identifying opportunities for recommending the firm and with providing a rich description of the firm should have been significantly greater for the firm’s loyal customers. While this is a preliminary. Recall that our claim is that the result is driven by the fact that the social networks of loyal customers are already well informed about the firm while those of less-loyal customers — or. It is possible. creating WOM is hardly a unique and rare technical skill where one would expect to observe significant experience effects. however. non-customers — have not been informed. Clearly. that the source of some or all of the “non-customer” effect Before discussing the results of our empirical analysis of this is their experience with WOM generation. Notably. One hypothesis might be that customers — by virtue of their deeper experience with. they receive little or nothing for their efforts. the number of points awarded had no explanatory power in any of the several specifications of the sales model in which we tested it.17 points for their reports. On the contrary. However. However. though the effect size is slightly lower.. not be substantial.rates respectively). to ensure that our results are not due to the omission of an experience variable. it certainly suggests that all exogenous WOM created by loyal customers is not meaningless. again. the marginal effect of an informative WOM episode from a customer would be higher online than offline. there is nothing like a “commission” plan). this would not explain why the non-customers’ online WOM was not effective while their offline WOM was. After all.

Thus. both observable and unobservable. 14 . in the section that follows. we estimate a similar model to that above including only the firm’s customers and stratifying them according to a behavioral loyalty measure. both AIC and BIC statistics favor the simpler Model (2) over both of these models. the firm’s customers — having had more conversations about the firm’s products already — might have less intrinsic motivation to provide a convincing recommendation.2 will include an attitudinal component as well.Model (2)18 controlling for the number of previous campaigns the agents had worked on. Note as well that our analysis in Section 4. and Neslin et al.1 Loyalty-Based Stratification Then. thank two anonymous referees for suggesting these possibilities. First. Krishnamurthi and Raj (1991) find significant differences in price sensitivities between loyal and nonloyal segments. containing possibly some mixed or negative information. we create quartiles based on each customer’s past visits to the chain: Q1 is the quartile with the most visits and Q4 has the fewest. we split the WOM episodes generated by non-customers into two separate variables: those generated by agents who had previously participated in a campaign and those that hadn’t. Our discretization of the continuous visit data into buckets captures the idea that the impact of customer loyalty often occurs in a non-linear fashion. 4. The results were consistent across these approaches. On the other hand. a Wald test in each case fails to reject the null hypothesis that the coefficients across the different non-customer WOM variables are all equal. the data suggest that this is not driving our reported results. Note that while store visits is a behavioral measure of loyalty. three and four campaigns.1. experiment to more-fully control the context.1. while we can’t rule out completely the possibility that non-customers had some useful experience. in neither case were any of the WOM coefficients significant at a p < . the two populations in this dataset have clear differences. Second. The creation of loyalty buckets or segments 1 8 These 1 9 We results are available from the authors. Finally. the firm’s customers — as a result of having a deeper experience level with the product — might offer recommendations that are more multifaceted. we believe it also captures the way most firms think about their customers. First.2 add further evidence that experience or expertise in WOM-creation is not the primary driving force behind our results. Finally. our lab study that follows in Section 4.10 level (in comparison to Model (2) in which the non-customer WOM coefficient is significant). (1985) find that the promotional acceleration effect was bigger for heavy users than for light users. two.19 In order to test more completely our theory and in an attempt to demonstrate the robustness of our results while controlling for important factors. we further disaggregated the non-customer WOM data by estimating a separate coefficient for the WOM generated by the agents who had participated in zero. While our theory concerns the relative loyalty of those creating WOM for the firm. Second. we implement a lab In order to analyze the impact of loyalty within the firm’s set of customers. Bowman and Narayandas (2001) find differences in WOM behavior between the highly loyal and the rest of the population following a customer-initiated contact with the company. one. In addition to the different experience levels. we provide two additional analyses. that one must attempt to account for to the extent possible. For example. First.1 as well as our lab experiment in Section 4. We did so in several ways.

then WOM to acquaintances should be the target.09. These results also suggest that the results found in Table 5 are not entirely due to unobservable differences between experienced and inexperienced non-customers. Note that.12 and 3450. and loyal to. These results provide some additional strength behind our claim that the firm should consider designing WOM campaigns not just for their highly loyal long-time customers but also for their newer and less-loyal customers. These results also further reinforce the idea that if the firm wants to create really impactful WOM. Since we also include a non-interacted term in the specification (WOM to Anyone). 3480. 4. It is important to stress that their zero visits does not mean that they had never visited the chain. respectively. Most importantly. Rust and Verhoef (2005). less-loyal customers also created impactful WOM primarily via discussions with their acquaintances. We recruited subjects from a university subject pool to participate in a study of web site usage. 3409. This increases the marginal impact of a new recommendation.39. Members in Q4 just joined the loyalty program prior to the beginning of the campaign. These three then formed a “group” 2 0 AIC’s for Table 8’s Models (1)-(4) were: 3428. We include several specifications to test the robustness of the results.00.88 and 3399. The mechanism we propose in Section 2. 3425. collect data on both sides of a WOM episode. One shortcoming of the The laboratory setting allows us to Befield test data is that it does not allow us to directly test this hypothesis since we only have access to the recommender’s loyalty and not the loyalty of others in the network. The results are shown in Table 8. Here in Table 8. it is never the case that the most loyal customers — those in Q1 — deliver the most impactful WOM.20 What these results show is that the impact of WOM from customers in the loyalty program is higher in some cases the less loyal that customer is (see the significant impact of WOM created by Q4 members to acquaintances).24. This also allows us to investigate the mechanism behind the key finding that less-loyal customers spread more-impactful WOM. cause our proposed mechanism requires the existence of past relationships and interactions. the product. we asked each participant to recruit two friends to participate in the study as well. 3530.2 Laboratory Experiment: An Exploration of the Mechanism In this subsection.1 is that less-loyal customers’ friends and acquaintances are less likely to be aware of. In Table 5. Mela and Vidal-Sanz (2006)).76.88. We then interact the quartile number with the WOM variables to create quartile-specific WOM measures. respectively. as outlined above. BIC’s were: 3534. In a common approach in customer relationship management and the customer lifetime value literature (Hartmann and Viard (2006). we interpret the interactions as the difference in the impact of WOM on sales as a function of the relative behavioral loyalty of the sender. the firm involved in this study recruited participants for the campaign by segmenting their customers according to this measure of loyalty. See Table 7 for summary statistics on these constructed quartiles. we see that. we saw that non-customers created impactful WOM to acquaintances. 15 . similarly. Model (4) is preferred based on a Wald test as well as common information criteria. and Gupta. among those models shown in Table 8. we replicate our core results in a lab setting in order to avoid the unobservable differences between populations that commonly exist in field tests.

whether they had visited it (yes/no). she receives a recommendation for Friendster from Dan. If. whether they intend to visit it (7-point scale) and whether they would recommend the site (yes/no). In the second phase of the study.” Each of these was measured on a 7-point Likert scale: ⎡ (V is_regijt + Intentijt )/2 if i visited j prior to time t 0 otherwise ⎤ ⎦ as well as an attitudinal component. In the first phase. which took place 6 days after the first phase.21 The six sites were pre-tested on a different sample to ensure similar average loyalty levels (see Table 9 for the average loyalty levels for the experimental sample). preferred this design to an obvious alternative in which recommendations are for a single site. some subjects received recommendations from senders who were highly-loyal to a site while others the sender’s initial loyalty (phase 1) to the recommended site. 2) the subject 16 . When several sites fit the criteria. and Post Secret. however. 2} is the phase of the study. Moreover. Linda was told. Friendster.for the purpose of the study. We exclude Google News from the set of possible recommendations since in our sample it has a much higher level of awareness than the other sites. here the difference between the two groups is further amplified since we compare very loyal customers to noncustomers. The experimental manipulation was to randomly assign each subject either to a high_loyalty or a non − customer condition based on the initial loyalty of the sender of the recommendation they received. suppose that Dan (the sender) has a loyalty of 6 to Slate and Note that while this dichotomy is similar to the initial analysis of the field test where we compare the effectiveness of members of the firm’s loyalty program and non-customers. In this alternative design.22 The distribution of recommended sites is given in Table 10. We asked each subject whether they had heard of the site (yes/no). The study had two phases. For example. The manipulation was based on That is. After a brief unrelated task. received recommendations from senders who were not users of the site. whether they visit it regularly (7-point scale). 2 2 We discarded responses for three reasons: 1) missing data on questions related to the recommended site. we would still be able to measure loyalty and we would expect there to be variance in loyalty levels across people. Note that this measure of loyalty combines a frequency component Loyaltyijt = ⎣ (2) customer state as Loyaltyij1 = 0.” Hence. Slate. upon logging in. picked a site at random. We define the high_loyalty state as Loyaltyij1 ≥ 6 and the non − had never visited Friendster. the condition into which each receiver is assigned is manipulated randomly making this a true experiment. 2 1 We We obtained 96 usable responses. in this instance. Slashdot. See Appendix B for the recommends Slate. she is in the high_loyalty condition. it is important to notice that while the loyalty levels of the sender are measured. the subjects took an online survey that elicited their current preferences for six news and entertainment websites: Google News. each subject was given a recommendation from a person in their group. “Dan (dan@yahoo. but we would not be able to randomly manipulate to which condition each subject would be assigned. For example. we where t ∈ {1. If his friend Linda (the receiver) is randomly assigned to the non − customer condition. she receives a recommendation for Slate from Dan. Smoking Gun. We define loyalty as the average of two items from the survey: “I regularly visit the site” and “I intend to visit/keep visiting this site. As we can see. we collected the same preference and awareness measures about each of the six sites that we had collected in the first phase. Dan is the sender and Linda is the receiver of a recommendation.

the willingness to recommend is highly correlated with loyalty for each of the websites in the sample. Thus. This is partly due to the fact that a respondent unfamiliar with a site would often leave all of the questions related to that site blank. We next turn to our main results. 2 3 In fact. Combined.the distribution of recommendations is weighted towards Friendster and Smoking Gun. we also need to control for the possibility the there are individual differences in the phase-to-phase changes. Google News has the highest average awareness levels of all of the sites investigated. It may not be surprising ex post that Google News yields only directionally consistent results: the awareness levels are so high that this may be the result of a ceiling effect. offering fewer incremental gains from a WOM campaign. but at t = 2 she indicated that she had not visited the site before). Since there is a time gap between the two phases. the sender) on the change in intention of the receiver following the recommendation. As Table 11 shows. or 3) the responses were inconsistent across phases (e. 17 . We are interested in the effect of a recommendation on the change in the receiver’s intention to visit the recommended site. We use the following dependent variable in the analysis:24 ³ ´ 1X ∆Intenti ≡ Intentib − Intentib − (Intentij2 − Intentij1 ) j2 j1 5 to visit the recommended site while the second term captures her average change in intention (note that the where b is the site about which i receives a recommendation. (ii) the difference in intention allows us to measure the marginal impact of a recommendation. Before presenting our main results on the change in preferences. the sites that enjoyed a relatively higher awareness in our sample. 2 4 We prefer to control for possible individual-level intention changes in this way rather than by including the other-site intention-change measures as independent variables due to concerns about the endogeneity that might arise in such an analysis. We are now able to test for this directly. In addition. we expect that a customer’s social circle is likely to be aware of the same products that she is aware of and loyal to the same products she is loyal to. As argued above.23 Similarly Table 12 demonstrates that the phase 1 loyalty levels of the sender and the receiver are positively correlated in all sites but SlashDot.. receivers’ phase 1 awareness levels of sites is significantly higher when the sender — someone in their social circle — is aware of the site. and (iii) the measure differences out individual and site fixed effects due to differences in awareness or popularity. in our core analysis. As a result. The focus of our analysis is an assessment of the impact of the loyalty of the recommender (i. we regress ∆Intenti on the sender’s loyalty level (our experimental manipulation). we first investigate a crucial assumption behind our theory. for example.. which meant that we could not sample it for a recommendation. for all sites but Google News.e. at t = 1 the respondent indicated that she had visited the recommended site before. these results support our assumption about congruity of awareness and loyalty within a social network. not surprisingly. We choose this particular measure because (i) in this context the number of visits is the relevant sales measure. The first term captures i’s change in intention j j6=b j recommended site is excluded in the second term average). Another important element of our theory is that one talks about the products to which one is loyal. the network of a highly-loyal customer is saturated. As shown in Table 13. we control for the nature of the relationship between the took an unreasonably short amount of time to complete the task.g.

the negative effect of recommender’s loyalty disappears. we control in Table 14’s Model (3) and Model (4) for whether or not the receiver had heard of or been loyal to the site. 2 5 According to our instructions. when this is not the case — when the receiver is aware of or loyal to the product — we should not expect the result to hold. In conclusion. which seems to contradict the earlier result that the word of mouth to acquaintance is especially powerful. consistent with our expectation. however.. In Table 14’s Model (1). 2 7 This finding must be interpreted with caution. the firm would like to target uninformed customers to be receivers of messages. For example. One possible explanation for this is that there are two forces at work here. a change in receiver’s loyalty from 4 to 6 results in decrease in ∆Intent of 0. Thus.17. controlling for the nature of the relationship increases the effect of the recommender’s loyalty (see Table 14’s Model 2). 2 6 The receiver intention change variable has a mean of . 18 . In cases where the subject indicated that she interacted weekly with the recommender. A recommendation by a non-customer results in ∆Intent that is 0. Clearly. given the low value of the F-statistic. in this final specification the recommendation of a friend is more influential than the recommendation of an acquaintance." I. following Model (4).e.e.sender and the receiver (a dichotomous variable equal to 1 for friends and 0 for acquaintances). not friends with) each other. This reinforces the point that the firm should consider targeting less-loyal customers to spread information about its products. As shown in (3) and (4) in Table 14..50. we see that the marginal effect of WOM is lower when it comes from a highly-loyal customer. this is in practice very difficult to do. but our initial designation was "acquaintances. Thus. Notably. respectively. when controlling for receiver’s awareness or loyalty. we also directly asked each subject how often she interacted with the recommender. if possible. this argues that the less-loyal customers as participants in a viral campaign may in some cases deliver a higher marginal benefit to the firm. not that of her friend who may have recruited them both into the study. and (ii) the persuasiveness effect (the recommendation of a friend is inherently more persuasive). This formed our initial classification. From a managerial perspective. a subject was asked to recruit two friends who were acquainted with (i.27 In order to investigate further the underlying mechanism at work. What these results suggest is that employing less-loyal customers to “spread the word” may be a useful proxy for the state of information in their networks. However.63 higher (nearly 400% relative to the mean) than a recommendation by a highly loyal customer.25 Our results are in Table 14. we are controlling for (i). the network effect which — in the field test — may have dominated the second effect. our field test and follow-up lab experiment suggest that recruiting less loyal customers as recommenders allows the firm to reach less-aware and less-loyal consumers whose intentions may be moreeasily influenced. we were concerned that the recruiter may be misinformed about the relationship between the other two members of the group. By including the receiver’s loyalty in the specification in (4). Recall that our theory is based on the idea that less-loyal customers are effective because their network is more likely to consider information about the product to be new. However." we re-classified the relationship to "friends. we usd a respondents’ own characterization of her relationship with another.26 Perhaps not surprisingly. (i) the network effect (the preferences of an acquaintance are less correlated with one’s own preferences which would make a recommendation more powerful).

we use our field average but the same may not be true among less-loyal customers. the firm can identify those customers that will be most likely to create this highly impactful type of WOM. Given the discrete nature of our data for the latter. we estimate: 2 8 Specifically. and how. “How often do you go out No subjects indicated “never. we are interested in the usefulness of selecting for inclusion in a WOM marketing campaign those that are considered to be “opinion leaders. for example — might be predictive of WOM behavior in specific categories. with respect to WOM creation. and 2) within the firm’s customers. In particular.5 Who Creates WOM That Matters? In this section. the number of WOM episodes reported by individual k. Our interest in this section is in determining to what extent.79). A lot. Rarely and Never. For the non-customers we estimate the following model: 15 X i=2 4 X l=1 W OMk = β 1 · AGEk + β 2 · F U LLT IM Ek + β 3 · OLk + µi + ul + εk (3) The dependent variable is W OMk . our data were collected on a five-point semantic scale in response to the question. Occasionally. it is important to control for AGEk . we administered the King and Summers (1970) scale (see the Appendix for a sanitized version of the scale we implemented). Given the group-level differences shown in Table 2.” 5. However. 19 . we found this scale to be of adequate reliability (α = . we report only the discrete model results. OLk is k’s “opinion leadership” score. there exists an interaction between opinion leadership and the customer’s loyalty to the firm: among loyal customers. suggest that WOM to acquaintances may be especially impactful). data to test this hypothesis in two ways: 1) we compare the effect of opinion leadership between customers and non-customers. We again include market level fixed effects µi .1 Field Test H3 proposes that. We note again that this is not necessarily a statement about the overall impact of WOM but only about the creation of marginal or incremental WOM over and above what is being created already. we look for an interaction between the degree of behavioral loyalty and opinion leadership. opinion leaders create more WOM than As in Section 4. we did not have access to any demographic data other than those that we report here." Note to eat?” The scale items were: All the time.1. it is included as a set of dummies ul . Thus. the fit was somewhat inferior. we follow up on the key insight that the firm’s focus in a WOM campaign should be on less-loyal customers or perhaps even non-customers. As with previous studies. While the results were qualitatively similar.28 One would imagine that other demographic variables — gender and income. F U LLT IM Ek (a dichotomous variable equal to 1 when the respondent works full time) as well as how often they eat out. For our analysis of customers stratified by their behavioral loyalty. as well as the extant literature. As part of the survey instrument in the original field test. We consider both the overall number of episodes as well as those WOM episodes involving only acquaintances (recall that our results in Section 4. that we also estimated specifications in which the discrete data were treated as a continuous variable.

To illustrate the magnitude of the interaction effect. this specification suffers from difficult to estimate the parameters simultaneously. and in Table 16’s Model (4). OLk ∗ Q1k . 3 0 Mean levels of total WOM and OL are. k 20 . We also find that the interaction term. We report in this section the R2 statistic that includes market fixed effects. and (c) the effect of opinion leadership is increasing in a customer’s loyalty. First. opinion leaders do not produce more than non-opinion leaders.99. only AGEk is significant. the fit measures slightly though consistently In Table 16. the interaction between The results are shown in Table 15 and Table 16 which allows us to test H3 in a number of ways.4 more WOM episodes per week if he is in the upper quartile on the loyalty scale than if he were in the lower 3 quartiles. (b) in the customer population. We then further investigate whether we find a similar pattern within the customer sample once customers are stratified by their loyalty level: we estimate Equation (4) on the customer sample (see Table 15’s and Table 16’s Model (3)). is positive and WOM than non-opinion leaders. we find that (a) in the non-customer population. In Table 15’s Model (4).1 which deals with panel data.04 and 4. the fit measures are more ambiguous but the core and significant in both tables’ Model (3). Unfortunately. we attempt to combine customers and non-customers into a single sample. Finally. consider the following example based on the estimated coefficients in Table 16’s Model (5): an agent who scores a 5 out of 7 on the opinion leadership scale generates 0. opinion leadership and loyalty. OLk ∗Q1k is positive and significant. a natural way to handle this is to simply include the non-customers in the lowest quartile and re-estimate the model using all of the subjects together. which makes it very favor the more parsimonious Model (3).W OMk = β 1 · AGEk + β 2 · F U LLT IM Ek + β 3 · OLk + β 4 · OLk ∗ Q1k + 15 X i=2 µi + 4 X l=1 ul + εk (4) where Q1 is again a dummy indicating that the customer is in the quartile demonstrating the highest level of behavioral loyalty. respectively: 1. In Table 15.23. but the main effect of OLk is not significant. we compare Model (1) to Model (2) in Table 15. These results are shown in Table 15’s and Table 16’s Model (5) and are again consistent with H3.30 Hence. (That is. these results suggest that selecting only those high on opinion leadership would potentially be a serious error since it might identify only those loyal customers 2 9 Note that in contrast to Section 4. The complication in doing this is that the latter have no “visits” with which to construct quartiles. Note that we still see the main effect of opinion leadership on WOM generation in the customer sample: OLk is still positive significant. we compare estimates of Equation (3) on the non-customer and the firm’s customer samples. opinion leaders do produce more WOM a multicollinearity problem since the correlation between OLk ∗ Q1k and Q1k is 0. Our key variable of interest in this equation is OLk ∗ Q1. the market fixed effects are rarely significant and hence appear to explain little variance. Unlike the models presented in Section 4.29 We investigate the impact of these individual-level characteristics on one’s creation of WOM both overall (Table 15) and specifically to acquaintances (Table 16). and Model (1) to Model (2) in Table 16). the data analyzed here is cross-sectional. Nonetheless.1. In both tables. As an additional robustness check. That is. we also estimate the main effect of loyalty (see Model (4) in Tables 15 and 16). we see that opinion leadership is significant in the customer sample and is not significant in the non-customer sample. result — the interaction — holds in both Model (3) and Model (4).

or less informed about. To the extent that this construct is found to be useful. Thus.” Though not researched directly in the marketing literature. usage behavior and lead users/early adopters. One would expect that. In fact. The rotated results 3 1 These are available from the authors. equivalent. it is less true for less-loyal customers. Note that our use of quartiles seems to pick up well the non-linear impact of loyalty. Thus. though not stellar (α = . the more friends and acquaintances one has and the more “social” one is.10 levels (though the p-value on the interaction effect in the analog to Table 16’s Model (3) is . Rosen (2000). for example. While the marketing literature has focused almost exclusively on opinion leaders. log likelihood.. In addition.70). we retained two factors.” In the academic literature. 21 . we find here that while opinion leadership is associated with higher WOM-creation for very loyal customers. our purpose in this analysis is to investigate the potential usefulness of another metric to serve as an indicator of less-loyal customers’ willingness to create WOM for the firm. Gladwell (2000) contrasts “connectors” with “mavens. one obvious candidate for further inquiry is the breadth of one’s network. the firm’s products. To summarize the results so far. However.31 Thus. the reliability of this scale (i. Importantly. the Cronbach alpha) was acceptable. Note that we also estimated a model in which the number of items purchased. we report only the quartile results. the more WOM one would create. all else equal.” Similarly. has recently focused on this dimension. We first ran an exploratory factor analysis without constraining the number of factors. A linear approach using the number of visits instead of Q1 produces results that are similar though not identical to those in Table 15’s Model (3).willing to create WOM and not the crucial less-loyals. it may not hold at all for the less-loyal customers (see Model (5) in Tables 15 and 16). First.102). future research will be required to refine its measurement to meet more rigorous standards. These people are likely to form the bridges Granovetter (1973) described as being the essential links between social networks. we are interested in exploring other possible metrics to identify the key disseminators of WOM. Since to our knowledge no scale exists for this construct.e. After discarding all factors with eigenvalues below 1 (see scree plot in Figure 1). we checked the construct’s discriminant validity vis-a-vis the opinion leader scale employed in our analysis above. 3 2 It is also important to note that the measures used in this scale were standard items on the agency’s survey. was substituted for the loyalty quartiles. it is relatively clear what network breadth might mean for the diffusion of WOM.32 Before implementing the analysis. Their recommendations are more likely to be received by people who are currently less experienced with. AIC and BIC. draws a distinction between “expert hubs” and “network hubs. Specifically. in fact. the interaction effects in these models are not significant at p<. rather Since the results are qualitatively than simply the number of visits experienced. We make no claim in regard to the optimality of these measures. one finds traces of this construct in Coleman et al. It is essential to note that this aspect of our inquiry is purely exploratory. we assess two important dimensions of the scale. we did not construct the scales for this specific use. Instead. we created our own (see Appendix A) comprised of five items. our analysis in Section 4 suggests that less-loyal customers should be recruited to participate in a WOM campaign. the models do not fit the data as well using traditional fit measures such as R2 . we do not report them here. The popular press. (1966) in their analysis of the physicians’ “interconnectedness.

34 In order to test the effect of an individual’s network breadth on her WOM production.Scree plot of eigenvalues 4 0 0 1 Eigenvalues 2 3 5 Number 10 Figure 1: Scree Plot from Factor Anlaysis using Network Breadth and Opinion Leadership Scale Items of a factor analysis constrained to two factors are presented in Table 17. the lowest loading from the King and Summers items is higher than the highest loading from our network breadth items. 3 4 Note that we are able to replicate the results that follow with a scale that excludes the NET_OUTGOING item which shows extremely low communality. The results are shown in Table 18. (3). as well as an interaction of this measure with the loyalty quartile. 22 . the network breadth measure seems to be particularly associated with higher levels of WOM production for less loyal (non Q1) customers. (3). That is. (5) and (6)) as opposed to the interaction term between opinion leadership and loyalty. the main effect of network breadth is positive and significant in some specifications. The reverse is true for the network breadth items. since a number of data points have missing values for the items in this scale. the interaction term between network breadth and loyalty is consistently negative (see Models (2). it Similarly. indicating that some items — particularly NET_OUTGOING — have large unique variances not explained by the factor. we retain all items and leave for future research the “purification” of the network breadth scale. Interestingly. As we can see from Model (2) and Model (4) in Table 18. network breadth may help a firm identify consumers that are most likely to create WOM for the firm. One possible explanation for this result 3 3 This factor analysis was performed for the firm’s customers only. That is. our sample is smaller. As this is an exploratory analysis. The analogous case holds in the Factor 2 column. While far from perfect. which is again significant and positive. Note that. and (5) in Tables 15 and 16 with the addition of the network breadth measure. The finding that the marginal impact of network breadth for very loyal customers is less than the marginal impact of network breadth for less loyal customers is surprising.33 Factor 1 than Factor 2. The communalities for the network breadth measures are low. going down the is clear that the loadings associated with the King and Summers scale items always are always higher on Factor 1 column. we re-estimate Models (1). The constructs as measured for the non-customers demonstrated significantly less discriminant validity.

That is. To summarize.” (See Appendix C for further details on the scenarios employed). we attempt to shed light on the underlying mechanism at work. However. Moreover. we again find it useful to corroborate these findings in the lab. 5.2 Laboratory Experiment: An Exploration of the Mechanism We present here the results of a lab experiment designed to: (1) replicate the findings in Section 5. Clearly.1 in a more-controlled environment. Scenario B. controls for the network 23 . more research is needed to identify the robustness of — and the mechanism behind — this interesting finding. Each subject was randomly assigned either to a high loyalty condition or to a low loyalty condition for both scenarios. we again capture loyalty using a measure that combines both behavioral and attitudinal aspects: “You have frequently chosen to go there. our results suggest that while opinion leadership is an effective predictor of WOM activity for loyal customers. which is a function of both the network effects and the propensity to talk about the restaurant once in a conversation. The major difference between the two scenarios is that in Scenario A the subject is asked about the total number of people that she would inform about the restaurant. and (2) disentangle the effects of network and opinion leadership in order to learn more about the underlying mechanism. While we expect the customer’s network to have no effect on the latter. Our approach to disentangling network effects from opinion leadership effects is to decompose WOM into two factors: the propensity to engage in a conversation (during the course of which a recommendation may be exchanged) and the propensity to offer a recommendation conditional on having a conversation. Subjects were asked to answer questions about two hypothetical scenarios and to complete identical network breadth and opinion leadership scales as those used in Section 5. We attempt to keep as many of the other factors employed in the field test constant in this environment to allow for attribution of any differences to their proper source. This study was conducted among MBA students at a private university in the Northeast U. it is possible that a customer — perhaps as a function of her opinion leadership or loyalty — may start a conversation in order to recommend a product. a disseminator with a broad network can inform a lot more potential customers than the one with a narrow network as part of the campaign. even though there are other places that are closer to you. our preliminary exploration suggests that network breadth may perform the opposite task: identifying less-loyal customers willing to create WOM. on the other hand. a recommender may be more likely to pass on a message about Rock Bottom to those who are not already informed about the restaurant (due to either altruism or the desire not to annoy). You consider it one of your favorite places to grab a quick bite. a loyal customer with a broad network does not enjoy this advantage over the one with a narrow network since in both cases their networks are already relatively well-informed.S. By separating these two factors in an experimental setting.1.may again be due to the asymmetry of endogenous WOM between very loyal and less loyal customers. this may not be true of less-loyal customers. Given the data limitations and the exploratory nature of the analysis. Here. The exact nature of the effect of opinion leadership and loyalty on both the probability of having a conversation and on the probability of making a recommendation is hard to predict a priori. In principle.

10. However.55 and 3. Once again. A 10% increase in N ET _BREADT H would increase WOM from all subjects by 16%. However. These effects are once again meaningful. it would increase that of loyal customers by 7%.effect by informing the subject that she is already engaged in a conversation with an acquaintance. Each subject was shown both scenarios and the order of presentation of the scenarios was randomized.10. In the analysis below. NET _BREADT H and the dependent variable are: 4. we attempt to keep the perceived quality of the restaurant constant across conditions by stating that the restaurant received a positive recommendation either from a friend or a professional reviewer. 24 . The results of Scenario A are presented in Models (1) . we are not able to do so.v. which differs from our earlier results. network breadth is no longer a factor. the same is not true of less loyal customers. the results change (see Model (4) in Table 19). respectively. 3 6 Mean levels of OL. Moreover. The results presented in Table 19’s Models (1) and (2) replicate the core finding in Section 5. on one hand. OL would have no effect on a non-loyal customer’s WOM creation. we can select the more parsimonious Model is equal to zero in Model (3) at p < 0. Moreover. 2. When we shift our focus to the propensity to recommend conditional on being in a conversation. This seems to suggest that we need to examine separately (i) the motivations to initiate a conversation in order to spread WOM and (ii) the motivation to discuss the product while in a conversation. as we can see in Model (3) (and perhaps due to collinearity issues described above). on 3 5 We select this dependent variable because it accurately maps into the WOM data used in the empirical analysis. Hence. and those that demonstrate network breadth. at mean levels. Moreover.92.41.1: among loyal customers. as the product of two variables is that the variance may be exaggerated.78.(3) of Table 19. both loyalty and opinion leadership now become significant. Thus. opinion leaders create more WOM than non-opinion leaders. this implies that Model (1) would be chosen over Model (1) over Model (2). In particular.35 The dependent variable in Scenario B is the probability that the subject would tell the acquaintance about the restaurant while in a conversation. We once again find that individuals with broader networks are significantly more likely Using the coefficients of Model (1). we have replicated the core findings using as the dependent vairable the square-root of the product. One concern in forming this d. The dependent variable in Scenario A is the expected number of people that the subject would tell about the pizza restaurant: the product of the probability that the subject would tell anyone and the number of people that she would tell conditional on telling someone. The results are qualitatively equivalent so they are not reported here. if we had to choose models. it is difficult to estimate the main effect of loyalty as well as the interaction effects due to high levels of collinearity: the correlation between LOY AL and OL ∗ LOY AL is 0. we can’t reject the null that the coefficient on N ET _BREADT H ∗ LOY AL (3) based on parsimony.94 and the correlation between LOY AL and N ET _BREADT H ∗ LOY AL is 0. a 10% increase in to create WOM: the effect of N ET _BREADT H is positive and significant in Models (1) . We can’t reject the null that the coefficient on LOY AL is zero in Model (2) at p < 0. we control for the order of presentation In both scenarios.36 It was our intention to also replicate the finding with respect to the interaction between network breadth and loyalty that we see in Table 18.(3). it suggests that the management of a WOM campaign should take into account the stark differences between those that score high on measures of opinion leadership. Again.

to explicitly test whether and how the firm should attempt to create exogenous WOM to drive sales. Moreover. on the other hand. networks represent fertile ground for the generation of incremental sales.1 Limitations The paper has many important limitations that should suggest to the reader that these results might best be viewed as starting points for further research. It seems that some measure of network breadth — how many people does one know. loyal customers are not necessarily the cornerstones of a successful WOM campaign. we have suggested a possible path to a solution. If this were a major problem. of course. we have no way to ascertain that any of the WOM episodes have actually taken place. Second. shown that. may be more likely to actively find people to whom the information is relevant. is that while our results suggest that the firm should be using its less-loyal customers to create WOM. because the loyal customers’ networks have probably been informed about the product for some time. however. we have no information on the valence of the recommendations. it is an issue that is encountered by both researchers and practitioners alike in this area. it would likely bias towards a non-significant result. we have shown that for products with low or moderate initial levels of awareness. the examples we cite at the outset of the paper — and the hundreds of other firms following a similar path — may represent a rational and profit-maximizing solution to the promotion problem. while the customer population may give more nuanced information about the firm which may render it less persuasive. control for all the differences between the customer and the non-customer populations. one concern is that the agent population may be more experienced and hence more persuasive on average. In particular. it is very difficult to previously alluded to have to do with the field test portion of the paper. The former may be very useful for spreading information about a firm’s product to those they’re talking to. Those with high network breadth. a large portion of the data that we collect is self-reported and is 25 . we have argued. Third. for example — may be an effective way to find the less loyal key 6. or perhaps nonThe “bad news. Some of the most important limitations that we have First. by the campaign may have little impact. In fact. in some cases. it may not be able to use the tool one has traditionally used for the job: opinion leadership. to our knowledge. 6 Conclusion On one hand. communicator. Nonetheless. In fact. purely exogenous WOM is associated with higher week-to-week sales. However. While we assume that most recommendations are positive due to the nature of the campaign. We have Thus. Finally. This paper represents the first attempt. In fact.” customers’. less-loyal opinion leaders may be no more likely to create WOM for the firm than less-loyals on average.the other. the less-loyal customers’. the incremental WOM created On the other hand. the news seems to be good. we would expect that differences in valence would account for some of the variance in effectiveness. those that are not highly loyal may not be likely to actively seek to establish these conversations.

the simple transfer of information (“Hey. refined and tested in order to identify a meaningful and valid measure to capture this potentially powerful construct. However. this paper does not address. thus. one needs to replicate these results in other categories. we would have liked to have collected information on recommenders’ income and gender. In other categories — particularly those in which there is more risk associated with purchase — the lower brand-level knowledge of non-customers may mitigate the network effects we have demonstrated here. The other clear opportunity for deeper investigation lies in the network breadth construct. It would also be interesting to compare the network breadth scale to other existing scales. Given the clear impact of King and Summers (1970).incomplete. we once again caution that the latter part of the analysis (on who generates WOM that matters) is exploratory in nature. 26 . but raises. Given whom we should be targeting and how we might be able to find them. Once one considers the possibility of a competitor. enlist its customers to “switch” another firm’s loyal customers to its side? Would the same implications The answers to these found here — that non-loyal customers are the key — still hold in such a setting? questions are not at all clear ex ante but would be of great interest both practically and theoretically. the rewards to such an endeavor would appear significant. For example. this problem becomes even more complicated. we did not address the competitive aspects of the WOM creation problem. 6. Finally. such as extraversion. we do not address how extrinsic and intrinsic incentives compare in encouraging WOM production. Clearly one of the key differences between loyal and non-loyal populations is the issue of intrinsic incentives. Finally. An aspect of the WOM creation process with which this paper does not deal is the question of tactics. One could imagine that an experimental approach to this problem — whether lab or field — would yield interesting and useful results. have you ever heard of ___?”) is probably sufficient to generate trial. interesting issues about incentives. The product area is a relatively low risk one and. and foremost.2 Future Research First The paper identifies a number of interesting areas on which future research projects could focus. for example. How can the firm. The measures themselves — as well as the scales used for each — need to be revisited. what should the firm do to encourage them to go out and tell people about the firm? Options range from monetary refer-a-friend programs to recognition programs and beyond. In fact.

44 4.60 0.00 0.00 Note: Bold = significant difference at the p<.Tables Table 1: Split of Sales Across Product Categories Proportion of Sales 21% 60% 14% 2% 3% Category A Category B Category C Category D Category E Table 2: Comparison of Groups Customers Non-Customers Variable Response Distribution (%) Work Status Work full time Student Both How often eat out? All the time A lot Occasionally Rarely Never Sample Mean Age Opinion Leadership Network Breadth 36.11 0.21 0.83 0.18 0.15 0.34 30.24 0.26 13.10 level or lower 27 .54 0.00 0.06 0.33 4.16 0.61 0.23 12.24 0.21 0.00 0.

40 1.02 0. Min 0.22 0.Other Table 3: Summary Statistics WOM Variables Mean N Std.22 0.51 0.08 0.26 0.95 0.11 0.30 28 .Acquaintance Non-Customer .08 0.Friend Customer .44 0.92 1.49 1.53 Table 4: Correlation Table Pairwise Correlations Cust Cust Cust Cust Cust Non-Cust Relative Friend Acq Stranger Other Relative Non-Cust Non-Cust Non-Cust Friend Acq Stranger Customer .33 0.04 -0.19 0.52 0.Acquaintance Customer .03 0 0 0 0 0 0 0 0 0 0 Max 2 10 5 5 6 3 10 6 6 5 180 180 180 180 180 180 180 180 180 180 0.Friend Non-Cust .99 1.14 0.12 -0.Other Non-Cust .95 0.05 -0.21 -0.43 0.11 0.07 0.87 0.82 0.31 0.Other Non-Customer .16 0.Relative Non-Cust .11 0.31 0.10 0.Stranger Non -Customer .Acq Customer .10 0.Friend Non-Customer .Stranger Non-Cust .09 0.44 0.51 0.42 0.41 0.00 0.36 0.61 0.37 0.Variable Customer .61 0.Relative Non-Customer .01 0.Stranger Customer .20 0.07 0. Dev.13 0.Relative Customer .00 0.08 0.13 1.05 0.Stranger Customer .51 0.Other 0.Acq Non-Cust .04 0.57 0.52 0.Friend Customer .24 0.51 0.28 0.08 0.51 0.07 0.

28 -1.760 0.85 -34.42 0.96 -180.007 867.07 -260.60 0.73 1.90 -55.83 366.20 -267.44 -0.63 0.00 0.29 662.305 3.000 (5) N R2 F Test: All Coefficients = 0 Pr > F Notes: (1) t-statistics are presented below the coefficient estimates (2) For all coefficients with p-values below .08 0.1 0.16 -0.90 -1.000 180 0.309 2.17 1.37 -708.23 -91.45 1.23 -0.069 1.56 -0.01 0.032 571.98 68.47 0.99 -0.12 299.96 2039.56 -0.18 180 0.39 -167.00 0.55 192.51 907.75 0.052 2.052 2.48 0.37 -191.Table 5: Aggregate Model Regression Results Aggregate Model .97 0. the p-value is presented in bold next to the estimate 29 .Fixed Effects Regression: SALESi.3989 0.10.01 0.07 130.77 84.50 0.038 1011.t+1 (t-statistics beneath) (1) (2) (3) (4) WOM Overall to Anyone WOM Customer to Anyone WOM Non-Customer to Anyone WOM Customer to Relative WOM Customer to Friend WOM Customer to Acquaintance WOM Customer to Stranger WOM Customer to Other WOM Customer to Other (online) WOM Customer to Other (not online) WOM Non-Customer to Relative WOM Non-Customer to Friend WOM Non-Customer to Acq WOM Non-Customer to Stranger WOM Non-Customer to Other WOM Non-Customer to Other (online) WOM Non-Customer to Other (not online) 64.000 180 0.000 180 0.10 2.16 -0.59 -122.23 0.25 -0.253 3.85 0.81 0.066 1.239 4.289 4.000 180 0.56 0.04 -0.18 153.61 -106.24 -1.83 118.

000 replications **: The confidence intervals are bias-corrected and accelerated and were generated with 1.000 replications.38 -143.74 -336.42 23.93 1813.15 1684.67 95% Confidence Interval** Lower Upper -189.89 -163.30 1638.55 45.29 63. ***: Model (2) .99 -123.75 489.6 117 Quartile 3 Quartile 4 0.64 412.33 1807.91 52.0 84 30 .93 1451.38 1244.75 1659.Loyalty Quartiles Average Visits Count Quartile 1 60.69 45.94 1906.47 19.Model (4) refer to columns of Table 5 Table 7: Summary Statistics .Table 6: Bootstrap Estimates Bootstrapped Confidence Intervals* 90% Confidence Interval** Lower Upper Model (2) *** WOM Customer to Anyone WOM Non-Customer to Anyone Model (3) WOM Customer to Other WOM Non-Customer to Acquaintance Model (4) WOM Customer to Other WOM Non-Customer to Acquaintance -32.83 *: All bootstrap estimates presented here are generated via 10.8 89 Quartile 2 3.14 -538.6 94 10.82 -166.

354 5.13 -367.15 3811.66 1.94 0.Q3 Customer->Friend .10.66 1.89 1.Q4 Customer->Other .67 4345.24 0.08 3914.19 4.46 1.Q3 Customer->Stranger .Table 8: Aggregate Model Regressions — Loyalty Quartiles Aggregate Model .80 1.70 1.33 (4) -125.20 3608.94 5008.t+1 (t.85 1.75 0.19 0.01 1.0000 0.09 2.88 1.88 1.10 314.378 2.0000 Notes: (1) t-statistics are presented below the coefficient estimates (2) For all coefficients with p-values below .62 1071.45 4068.Fixed Effects Regression: SALESi.30 (n = 21) 5.64 3.30 0.52 -0.52 0.48 6941.31 2.69 1.Q1 Customer->Other .055 5568.70 (n = 110) 5.39 6068.12 -1.35 4203.Q4 Customer->Friend .92 1.Q1 Customer->Acquaintance .32 4272.13 -72.013 986.Q2 Customer->Stranger .90 5313.19 180 0.000 4.45 1.63 -1.38 3449.93 0.24 1.72 1.Q4 Customer->Stranger .60 1.88 1.00 1.039 2311.63 144.41 3420.46 4779.0000 -4871.26 3528.41 4085.99 1.009 1232.05 1.74 1.367 3.Q1 Customer->Friend .0000 180 0.18 1.051 1176.42 4295. the p-value is presented in bold next to the estimate Table 9: Initial Loyalty to Site | Visited the Site Smoking Gun Slate Slashdot Post Secret Google News Friendster 4.Q2 Customer->Acquaintance .80 4500.21 0.Q3 Customer->Acquaintance .47 1.07 0.15 (n = 72) 31 .12 3859.81 -1.Q4 Customer->Acquaintance .97 1.074 1370.62 5110.000 2171.13 1.66 1.59 4662.21 180 0.43 5620.51 1851.44 0.27 0.62 1.28 4447.072 1.12 (n = 17) 4.52 137.36 -0.Q2 Customer->Friend .002 2.25 1.59 1.63 -524.22 3639.81 -35.38 3996.02 1.85 0.00 0.Q2 Customer->Relative .22 1.098 4185.60 4675.29 -4191.98 0.55 5082.16 (n = 22) 4.Q3 Customer->Other .62 0.59 0.Q1 Customer->Stranger .42 -182.024 6268.32 1.40 4742.399 2.44 4574.62 4488.42 2.52 0.49 0.Q3 Customer->Relative .30 1.52 4412.43 (n = 90) 4.75 -0.29 1.93 1.28 -1.65 1.56 0.25 180 0.41 4493.Q2 Customer->Other .statistics beneath) (1) (2) (3) WOM Customer to Anyone WOM Non-Customer to Anyone WOM Non-Customer to Acquaintance Customer->Relative .59 -0.Q4 N R2 F Test: All Coefficients = 0 Pr > F 2704.45 3878.

s.532 * Friendster 0. **=p<.01.00% Friendster 86.10.Table 10: Distribution of Recommended Sites Non-Customer High Loyalty Smoking Gun Slate Slashdot Post Secret Friendster 10 7 0 4 12 22 2 2 5 32 Table 11: Comparison of Phase 1 Awareness Levels of Receivers Conditional on the Awareness Levels of Senders Site Pr[aware|sender aware] Pr[aware|sender unaware] Smoking Gun 92.628 ** Slate 0.13% 21.s.44% 17.580 ** Friendster 0.65% Google News 87.099 n.872 * Google News 0.34% 75. ** Table 12: Correlation of Phase 1 Receivers’ and Senders’ Loyalty Levels Site Correlation Smoking Gun 0.314 * Google News 0.0001 Table 13: Correlation of Phase 1 Loyalty and Recommendation Levels Site Correlation Smoking Gun 0.86% 45.12% SlashDot 39.420 ** SlashDot -0.680 * Slate 0. PostSecret 0.843 * PostSecret 0.809 * SlashDot 0.13% PostSecret 44.00% * = p<. **=p<.83% Slate 64.67% 50.001 ** ** * ** n.00% 19.487 ** * = p<.0001 32 .831 * * = p<.

1759 6.53 -1.10.052 -1.000 .0435 0. the p-value is presented in bold next to the estimate -0.692 .15 2.97 0.03 (3) -0.719 0.Table 14: Experimental Results D. observations 96 96 96 R-squared 0.92 (4) -0.630 .426 1.590 1.079 0.55 0.Change(Avg Intention to Visit Other Sites) for Receiver Recommender High Loyalty Friend Receiver Heard about Site at t = 1 Receiver's Loyalty to Site at t = 1 (1) -0.85 96 0.000 Notes: (1) The t-statistics are presented below the coefficient estimates (2) For all coefficients with p-values below 0.11 6.000 -3.127 0.054 No.V: Change(Intention to Visit Recomm Site) .45 0.93 .74 Pr > F 0.0325 0.248 -3.079 -1.1802 F-test: All coefficients = 0 3.78 (2) -0.047 -0.000 33 .760 1.166 -0.12 0.

10.08 3.77 -0.14 -1.010 0.14 -1.003 0.62 Work Full Time -0.26 Eat out occasionally -0.28 Eat our rarely -0.80 0.03 -1.51 -1.37 .024 0.04 .034 -0.060 0.051 -1.27 1.19 1.81 -0.V.93 2.19 0.08 -0.13 0.25 -1.92 .021 0.50 -0.06 Eat out a lot -0.03 -0.19 -1.28 -1.70 Opinion Leadership 0.75 (5) Combined -0.86 1048 .059 0.89 0.48 670 378 378 0.96 378 0.11 0.00 1.42 -0.049 F Test: All Coeffs = 0 3.29 -0.21 -0.46 -0.15 2.: Total WOM (1) (2) (3) (4) Customers SAMPLE Non-Cust Age -0.95 .44 Pr > F 0.047 0.000 (1) t-statistics are presented below the coefficient estimates (2) For all coefficients with p-values below .000 0.059 0.000 N 2 Adjusted R 0.14 -0.59 -2.073 0.003 -3.003 0.10 0.24 -1.05 .Table 15: Individual Model Results: Opinion Leadership.06 -1.03 .22 .59 -1.05 .02 0.39 .88 0.44 -0. the p-value is presented in bold next to the estimate 34 . Total WOM D.25 3.76 -0.23 -0.72 0.64 -2.31 2.12 -2.26 0.43 -1.95 -0.014 2.009 -1.75 -2.01 0.009 -0.06 Q1 Opinion Leadership X Q1 2.

11 .01 .10 0.17 .11 0.83 -0.001 0.046 0.02 -0.44 Eat out occasionally -0.69 0.34 KS_SOURCE 0.040 0.79 0.57 -0.11 -0.13 0.15 -0.20 1.83 Opinion Leadership 0. the p-value is presented in bold next to the estimate Table 17: Discriminant Validity of Network Density for Loyals (Rotated Loadings Shown) Factor 1 Factor 2 Communality KS_TALK 0.049 -1.063 0.38 0.00 0.071 0.00 0.Table 16: Individual Model Results: Opinion Leadership.10 -0.17 NET_INTRODUCE 0.12 -1.035 0.96 -0.02 0.000 (1) t-statistics are presented below the coefficient estimates (2) For all coefficients with p-values below .08 .28 0.21 .69 2.10 -0.20 -0.04 -2.10 -0.24 0.08 .29 NET_OUTGOING 0.15 0.96 1048 670 378 378 0.12 0.30 NET_NUMACQ 0.91 -1.46 0.45 Q1 -2.88 -1.14 0.40 N 2 2.067 0.53 0.06 0.28 -0.81 -0.10.033 0.31 2.53 0.07 0.001 0.02 0.28 35 .13 .78 NET_NUMFRIENDS 0.05 2.049 -3.84 -1.03 0.22 .18 0.27 NET_SOCIAL 0.80 0.11 0.07 -1.00 .55 .64 KS_TOLD 0.48 0.060 -0.03 -0. WOM to Acquaintances D.64 KS_WHO 0.54 KS_ASKED 0.036 0.11 .: WOM to Acq (1) (2) (3) (4) (5) SAMPLE Non-Cust Customers Combined Age 0.97 Work Full Time -0.08 -0.001 0.88 0.24 KS_INFO 0.00 -0.04 -0.95 -1.015 -3.V.10 -0.035 Adjusted R F Test: All Coeffs = 0 2.88 -0.002 -0.56 Eat out a lot -0.70 1.04 0.01 .35 378 4.50 0.72 2.10 0.03 Opinion Leadership X Q1 0.019 0.21 2.000 3.002 0.02 -0.97 Eat our rarely -0.08 0.02 0.01 -2.51 0.74 Pr > F 0.001 0.34 -1.

072 -0.37 .: WOM to Acq (1) (2) (3) (4) (5) (6) SAMPLE Non-Cust Customers Combined Non-Cust Customers Combined Age -.075 .38 -2.320 1.170 2.013 0.01 .13 -0.27 -0.20 -2.013 0.880 1.060 -0.038 (1) t-statistics are presented below the coefficient estimates (2) For all coefficients with p-values below .51 Opinion Leadership .13 0.53 3.61 .314 1.04 -.Table 18: Individual Model: Network Breadth D.67 .031 95 0.084 0.065 0.46 -1.91 -2.V.033 -0.083 1.180 2.86 -2.77 Pr > F 0.081 -1.37 .17 -2.1 0.47 .27 0.14 0.27 0.50 -1. the p-value is presented in bold next to the estimate 36 .001 2.971 .005 -0.02 .020 -1.332 101 0.34 Work Full Time .012 196 0.11 1.82 -1.076 1.038 0.77 0.07 0.98 1.030 1.03 -0.43 .08 -0.147 Adjusted R2 F Test: All Coefficients = 0 1.006 -2.69 0.10.11 .27 0.022 0.33 -0.02 .98 Opinion Leadership X Q1 1.14 0.18 0.088 0.V.51 2.18 0.064 0.018 -0.06 0.51 1.52 1.449 .: Total WOM D.39 1.075 -0.031 -.75 1.29 -1.40 Network Breadth X Q1 -0.012 .04 0.79 N 95 101 0.25 Network Breadth .51 1.03 1.005 196 0.211 2.033 1.23 -1.

04 -0.85 0.16 4.31 -0.30 0.18 181 Scenario B Pr[recommend | conversation] (4) 36.05 (2) -5.Table 19: Scenario A and B Results Scenario A: E[No.27 10. In discussions about Category: (7 .67 2.I Usually Tell My Friends About Category) 6.078 1.21 0.60 -2.Give a Great Deal of Information to 1 .003 3.Very Likely) 3. In general I like to talk to my friends and neighbors about Category (7 .Give Very Little Information) 4.My Friends Usually Tell Me About Category to 1 .093 0.19 0.06 0.0011 0. How many very close friends would you say you have (these are people you would feel comfortable having a very personal conversation with)? 37 . (7 .10 -0.055 1.A Lot of People) 5.21 0.31 Pr > F 0.07 0. Compared with my circle of friends I am ______ to be asked about Category.21 0.Never) 2.71 -1.19 F-test: All coefficients = 0 5. (7 .Very Often to 1 .000 0.Not Very Likely to 1 .000 Control for Order Effect Y Y Y Y t-statistics are presented below the coefficient estimates For all coefficients with p-values below 0.50 0.77 -2.088 0.Often Used As a Source of Advice to 1 . the p-value is presented in bold next to the estimate Appendix A Opinion Leadership Scale Note that all items were measured on a 7-point scale. During the past six months I have told ____ about Category.93 0.74 2.63 0.083 0.29 -0.65 4.22 -0. of people told] D.Not Used As a Source of Advice) Network Breadth Scale 1. (1) Loyal Opinion Leadership Opinion Leadership X Loyal Network Breadth Network Breadth X Loyal -0. When I talk to my friends about Category I: (7 .77 -1.000 0.04 (3) -0. Overall in my discussions with friends and neighbors about Category I am: (7 .10.003 2.007 1.V.36 4.23 0.31 3.63 2.032 2.81 N 2 181 181 231 Adj R 0.No One to 1 . 1.028 2.

I like to be social whenever I can Appendix B 1. I have visited this site: Yes No N/A 3. Which of the following best describes you?: 1: Independent . I have heard of this site: Yes No N/A 2. I would recommend this site: Yes No N/A Appendix C 38 .1: 1-5 2: 6-10 3: 11-15 4: 16-20 5: 21-25 6: 25+ 2.I’m sometimes social sometimes not 3: Outgoing . How often are you responsible for introducing people to each other? (1: Always to 5: Never) 5.I like to do my own thing 2: Middle of the road . How much of your day is spent in a social situation? (1: The majority to 4: Very little) 4. How many acquaintances do you have (if you saw them on the street you?d be comfortable having a long conversation with them)? 1: 1-30 2: 31-50 3: 51-70 4: 71-90 5: 91-110 6: 111+ 3.I’m mainly social but like some time to myself 4: Social Butterfly . I intend to visit/keep visiting this site: (1: Strongly Disagree to 7: Strongly Agree) N/A 5. I regularly visit this site: (1: Strongly Disagree to 7: Strongly Agree) N/A 4.

] (Low Loyalty condition: You have never been there. You consider it one of your favorite places to grab a quick bite. but you heard that it is good. how many people do you think would you tell? ___ Scenario B Imagine you are talking to an incoming MBA student about things to do in town on weekends. even though there are other places that are closer to you.) How likely is it that you will mention this restaurant to somebody (whether it’s a friend. How likely is it that you will mention to her a relatively new local restaurant that received a good review from the local paper and that you had [High Loyalty condition: visited a number of times?] (Low Loyalty condition: not yet visited) _____ % (please fill in % between 0 and 100) 39 .Scenario A Imagine there is a pizza place located about 5 minutes from the business school. [High Loyalty condition: You have frequently chosen to go there. relative. She is a casual acquaintance whom you know from home. acquaintance or stranger) over the course of the next month? _____ % (please fill in % between 0 and 100) Assuming you were to tell anyone about this restaurant over the course of next month.

Etzel (1982). Bowman. An Introduction to the Bootstrap. 2nd Edition’. ‘The effect of word of mouth on sales: Online book reviews’. Eyal. Drew (1992). Peter H. ‘Persuasion: Psychological Insights and Perspectives. Marketing Science 20(1). ‘The chain of effects from brand trust and brand affect to brand performance: The role of brand loyalty’. 82—95. Anderson. Biyalogorsky. Elihu Katz and Herbert Menzel (1966). Arndt. and M. FL. 183—194. W. Journal of Consumer Research 9. Bell. 125—43. Indianapolis. and Sangyoung Song (2004). ‘Customer referral management: Optimal reward programs’. Judith and Dina Mayzlin (2006). Cialdini. pp. 143—169. ‘Customer satisfaction and word of mouth’. and Mary W. 57—70. Efron. Thousand Oaks. Bobbs-Merrill. 291—297. Bearden. and Marsha L. CA. Bloch. and James H. Tibshirani (1993). Medical Innovation: A Diffusion Study.Green. Arjun and Morris B. Bolton. ‘Role of product-related conversations in the diffusion of a new product’. 291—5.Brock and M. 345—54. Coleman. Indiana. Eugene W. Journal of Service Research 1(1). Journal of Marketing 47. Eitan Gerstner and Barak Libai (2001). 81—93. Journal of Marketing Research 38. Chevalier. ‘Reference group influence on product and brand purchase decisions’. ‘Managing customer-initiated contacts with manufacturers: The impact on share of category requirements and word of mouth behavior’.References Anderson. Douglas and Das Narayandas (2001). David R. James S.. Robert and Brad Sagarin (2005). Social contagion and trial on the internet: Evidence from online grocery retailing. 40 . Chapman Hall/CRC. Holbrook (2001). Ruth N. Sage Publications. Journal of Marketing Research 43. in T. Journal of Marketing 65(2). (1998). Principles of interpersonal influence. ‘Mitigating the effect of service encounters’. 5—17. Sullivan (1993). Eugene W. Journal of Marketing Research 4. Marketing Science 12(2). ‘The antecedents and consequences of customer satisfaction for firms’. Boca Raton. 69—81. ‘A theoretical model for the study of product importance perceptions’. Working paper. eds. Marketing Letters 3(1). Richins (1983). Bradley and Robert J. Johan (1967). Chaudhuri.

Gladwell. and Dina Mayzlin (2004). Foster. John H. Garber. The Tipping Point: How Little Things Can Make a Big Difference.Engel. ‘Brand loyalty vs. Marian and Peter Wright (1994).POX. ‘The strength of weak ties’. and John D.Kumar. Lemon and Barak Libai (2003). Mark (1973). ‘Using online consversation to study word of mouth communication’. Journal of Marketing Research 10. 41 . American Journal of Sociology 78(6). ‘From density to destiny: Using spatial dimension of sales data for early prediction of new product success’. Malcolm (2000). 3—8. Journal of Service Research 9(2). Marketing Letters 12(3). Kyner (1973). 139—155. Sunil. New York. Wesley R. 211—223. ‘What is the true value of a lost customer?’. ‘Learning by doing and learning from others: Human capital and technical change in agriculture’. Journal of Service Research 5(3). Bruce Hardie. Tal. Hartmann. Marketing Science 23. Barak Libai and Eitan Muller (2001). The value of a "free" customer. Journal of Advertising Research 17(5).. Boston. 1—9. John E. Nathaniel Lin and Nalini Ravishanker S. ‘Talk of the network: A complex systems look at the underlying process of word-of-mouth’.. Journal of Advertising Research 9. Godes. 196—208. 1176—1210. Gupta. Lett (1977). Case Study: Hasbro Games . David and Elie Ofek (2004). Journal of Political Economy 103(6). 35—40. V. Blackwell and Robert J. Brian Viard (2006). and V. ‘Product sampling and word of mouth’. Working paper. Sunil. Andrew and Mark Rosenzweig (1995). Goldenberg. Granovetter. Kegerreis (1969). ‘The persuasion knowledge model: How people cope with persuasion attempts’. Harvard Business School Publishing. David B. Roger D. 1—31. Friestad. Sriram (2006). Vidal-Sanz (2006). Jacob and David B. Gupta. Carl F. Godes. Little Brown. Jacoby. 1360—80. ‘How information is used to adopt an innovation’. Mela and Jose M. Marketing Science 23(4). Barak Libai and Eitan Muller (2004). ‘What is the true value of a lost customer?’. MA. James E. Do frequency reward programs create switching costs? Working paper. Holmes. 419—429. Hogan. repeat purchasing behavior’. Jacob Goldenberg. Katherine N. New York. William Kahn. The Journal of Consumer Research 21(1). 545—560. Jacob.

and O. Krishnamurthi. CA. Marketing Science 25(2). Ann Arbor. Petty. Raj (1991). ‘Advances in Consumer Research. Elihu and Paul F. 299—303.. 195—208. Free Press.. New York. pp. ‘Optimizing the marketing interventions mix in intermediate-term crm’. Cacioppo. Rosen. Manchanda. Verhoef (2005). Roland T. 43—50.Brock and M. Kardes. Psychological Bulletin 108. 81 — 116. ‘The influence of communicator-recipient similarity upon the beliefs and behavior of pregnant women’. W. Marketing Science 24(3). Harvard Business Review 68. Marketing Science 4(2). Fourth Edition. Association for Consumer Research. John T. Earl Sasser (1990). Mazen. NJ. The role of targeted communication and contagion in new product adoption. Reichheld. 477—489. ‘Classic and current social comparison research: Expanding the perspective’. Journal of Consumer Research 23. Frederick F. Charles W. Richard E. Glencoe. eds. Leventhal (1972). Diffusion of Innovations. and Peter C. 42 . Caroline Henderson and John Quelch (1985). pp. 2nd Edition’. Jacqueline Johnson Brown and Stephen Seidman (1984). Peter. Upper Saddle River. King. ‘Consumer promotions and the acceleration of product purchases’. B. NY. Lazarsfeld (1955). Dina (2006). Puneet. (1993). 155—163. 172—183. Mayzlin. 1—26. The Anatomy of Buzz. Currency. Thousand Oaks.Monroe. NY. Marketing Science 10(2). Lakshman and S. To think or not to think. Kruglanski. Prentice Hall. New York. Brian Foster. Rogers. Neslin. Hoyer (1981). Mayseless (1990). Sage Publications. What if opinion leaders didn’t know more? a question of nomological validity. ‘Zero defections: Quality comes to services’. Emanuel (2000). The Free Press. Katz. Summers (1970). Frank (2002). 289—302. 8’. A. MI. The Influentials.. Scott A. Free Press. Journal of Marketing Research 7. in K. Allen J. ‘Brand congruence in interpersonal relations: A social network analysis’.Jacoby. and W. Personal Influence. Vol.Green. Ying Xie and Nara Youn (2006). Journal of Experimental Social Psychology 8. Ed and Jon Berry (2003). in T. Strathman and Joseph R. ‘Promotional chat on the internet’. Everett M. NY. ed. Priester (2005). Second Edition. Rust. ‘An empirical analysis of the relationship between loyalty and consumer price elasticity’. and John O. Consumer Behavior and Managerial Decision Making. and H.P. Jacob and Wayne D. ‘Overlap of opinion leadership across consumer product categories’. R. 147—165. Keller. Reingen. IL. Working paper. New York. ‘Persuasion: Psychological Insights and Perspectives. 105—11.

J. Van den Bulte. Advances in Consumer Research 31. and Richard L. Verlegh. 516—53. 43 . Christophe and Gary Lilien (2001). Oliver (1989). Christophe and Gary Lilien (2003). John E. Celine Verkerk. Journal of Retailing 65. ‘Customers or sellers? the role of persuasion knowledge in customer referral’. Tuk and Ale Smidts (2004). ‘Medical Innovation revisited: Social contagion versus marketing effort’. 1409—35.Swan. 304—305. Working paper. Mirjam A. Two-stage partial observability models of innovation adoption. Peeter W.. Van den Bulte. American Journal of Sociology 106(5). ‘Postpurchase communications by consumers’.

Sign up to vote on this title
UsefulNot useful