Strategic manipulation of Internet opinion forums: implications for consumers and firms.

1. Introduction

The Internet has enabled individuals all over the world to make their personal experiences, thoughts, and opinions easily accessible to the global community "at the click of a mouse." This has led to the creation of a diverse mosaic of online word-of-mouth communities (online forums), where individuals exchange experiences and opinions on a variety of topics ranging from products and services, to politics and world events. Examples of such communities include online product review forums, Internet discussion groups, instant messaging chat rooms, mailing lists, and Web logs (see Schindler and Bickart 2003 for a nice overview).

There is growing evidence that consumers are influenced by opinions posted in online forums before making a variety of purchase decisions (Thompson 2003, Senecal and Nantel 2004, Chevalier and Mayzlin 2006). Similar evidence suggests that forums play an increasingly important role in public opinion formation. Internet forums are thus emerging as an alternative source of information to mainstream mass media, replacing our societies' traditional reliance on the "wisdom of the specialist" by the "knowledge of the many."

Many have argued that the ease with which the Internet can aggregate information from large numbers of, otherwise unrelated, individuals can lead to better informed and, ultimately, more efficient markets and societies. Nevertheless, these same properties (large scale, relative anonymity) make it relatively easy for interested parties to manipulate the information propagated through online forums by anonymously adding their own, strategically biased, messages to the total mix of posted opinions.

Online forum manipulation strategies can take many forms, and firms (or depending on the context of interest, political parties and special interest groups) are getting more sophisticated by the day. The simplest firm strategy is to anonymously post online reviews praising its own products, or bad-mouthing those of its competitors. There is ample evidence that such manipulation occurs. For example, when in February 2004, because of a software error, Amazon.com's Canadian site mistakenly revealed the true identities of some of its book reviewers, it turned out that a sizable proportion of these reviews were written by the books' own publishers, authors, and competitors (Harmon 2004). The music industry is known to hire professional marketers who surf various online chat rooms and fan sites to post positive opinions on behalf of new albums (White 1999, Mayzlin 2006).

Given the potential backlash of such activities, firms are experimenting with less overt methods. Some firms offer rewards to consumers who start favorable conversations about their products on popular online forums. For example, a recent marketing campaign promised prizes to fans who would start conversations in online forums praising singer Lucinda Williams's albums (see http://slate.msn.com, July 26, 2001). Other firms routinely monitor online forums to identify influential community members. They then target them directly and try to persuade them to write favorable reviews by sending them free samples, inviting them to special events, etc. There is at least one professional marketing firm that conducts such campaigns on behalf of its clients (see http://www.electricartists.com). (1)

As more firms, political parties, and special interest groups realize the power of online forums, it is expected that more will engage in direct or indirect word-of-mouth manipulation practices.

It is, therefore, important and timely to understand what the impact of such activity is likely to be on the informativeness of Internet forums and on the payoffs to the various parties who are affected by them. The results of such analyses will be relevant to policy decisions (Is Internet forum manipulation socially harmful?), R & D decisions (Does it pay to invest in technologies that discourage online manipulation? Who should bear the cost of such investments?), and, of course, firm and consumer attitudes toward Internet forums (How much should consumers trust online forums? How much should firms invest in trying to manipulate them?).

This paper contributes to answering some of these questions by analyzing how the strategic manipulation of Internet opinion forums affects the payoffs of consumers and firms in markets of vertically differentiated experience goods. Specifically, I consider settings where one or more firms simultaneously launch products whose true quality is initially unknown to consumers and difficult to verify before purchase and use. I assume that the main source of quality information for consumers is an online product review forum (such as Epinions.com or Amazon Reviews), where past consumers post opinions about their experiences with the goods. New consumers read these opinions and form perceptions about the qualities of the products. Based on these perceptions, they make purchase decisions. Firms can try to manipulate consumer perceptions by posting anonymous reviews that praise their own product, at a cost. All firms are assumed to be strategic; that is, they manipulate opinion forums to maximize their payoff, given their correct anticipation of other firms' strategies and consumer beliefs. Furthermore, consumers are smart; even though they cannot directly distinguish honest opinions from fake opinions, they are aware that manipulation occurs and adjust their interpretation of online opinions accordingly.

My analysis derives three principal results. First, if every firm's manipulation strategy (amount by which the firm inflates its true ratings) is monotonically increasing in that firm's true quality, manipulation increases the informativeness of online forums, in the sense of increasing the ex ante expected payoff of consumers who base their decisions on information published in these forums. In such settings, high-quality firms inflate their, already higher, true ratings more than low-quality firms. Manipulation activity then increases the separation of the probability

distributions of ratings that correspond to adjacent product qualities. This allows consumers to make more accurate inferences about a firm's true quality from its published ratings. The inverse result holds in settings where every firm's manipulation strategy is monotonically decreasing in that firm's true quality. In such settings, low-quality firms manipulate more than high-quality firms; i.e., they are shrinking the gap between their respective ratings and making it more difficult for consumers to infer a firm's true quality (because everybody's ratings will be clustered together). Manipulation then decreases the informativeness of online forums, hurting consumers.

Second, informativeness-enhancing manipulation equilibria exist in settings where every firm's net payoff function, inclusive of the cost of manipulation, is supermodular in firm type and manipulation action. In a broad class of settings, net firm payoffs are super-modular if firm profits (before manipulation costs) are sufficiently steeply increasing convex functions of consumer perceptions of their quality. In such settings, the better a firm is perceived to be, the more it has to gain from being perceived to be even better. This provides high-quality firms with higher incentives to inflate their ratings. In contrast, in settings where firm profits are concave functions of consumer perceptions of their quality, the better a firm is perceived to be, the less it has to gain from being perceived to be even better. It is then low-quality firms that have higher incentives to inflate their ratings, leading to equilibria where manipulation decreases informativeness.

Third, in a broad class of settings, if the cumulative precision of honest ratings is sufficiently high (for example, because a sufficiently large number of consumers post honest opinions online), the cost of manipulation to firms always outweighs its benefits. A high precision baseline signal cannot be substantially affected by firm manipulation. Resources spent on manipulation are then wasted, because they do not change consumer beliefs. Nevertheless, firms have no choice. The fact that anonymous manipulation is possible induces rational consumers to anticipate that firms will engage in it (and thus to appropriately discount the nominal values of online ratings they observe). Firms are then trapped into performing the equilibrium level of manipulation that is expected of them, because as in a rat race, a lower level will bias consumers' perceptions against them.

The overall picture painted by these results has interesting, and somewhat counterintuitive, implications for practice. On the one hand, my analysis shows the existence of settings where forum manipulation is equivalent to a form of quality signaling that benefits consumers. On the other hand, it shows that if consumers come to expect that firms will manipulate, as the volume and quality of user-generated online content increases, there will be a threshold beyond which firms will be trapped into having to engage in profit-reducing online manipulation practices, simply because consumers expect them to. All firms would then be better off if consumers (rationally) expected them to manipulate less. My analysis shows that one way of accomplishing this is by developing filtering technologies that make it costlier for firms to manipulate, and thus lower the equilibrium levels of manipulation. Interestingly, it is firms, and not consumers, that have most to gain from the development of such technologies.

The rest of this paper is organized as follows. Section 2 introduces the main intuitions by analyzing a set of simple monopoly settings. Section 3 shows how the results generalize for a broad class of payoff functions and signal distributions. Section 4 discusses related work. Finally, [section]5 summarizes the strategic implications of our findings for consumers, firms, and forum operators and concludes. An online appendix, available on the Management Science website at http://mansci.pubs. informs.org/ecompanion.html, provides several modeling extensions. Table 1 summarizes the key notation used throughout this paper.

2. The Main Intuitions

This section introduces the main ideas underlying this work by analyzing a series of simple monopoly settings. The emphasis is on conveying the fundamental intuitions using examples that admit closed-form solutions. Section 3 then states the general form of the results in multifirm settings.

Consider a monopolist firm that introduces a product to a new market. The appeal of the product to consumers is the sum of two independent components: (1) a horizontal component (location), representing product attributes whose valuation depends on each individual consumer's taste (e.g., color, shape, look and feel, etc.) and (2) a vertical component (quality), representing attributes whose valuation is identical among all consumers (e.g., ease of use, durability, etc.). I assume that a product's location can be reliably communicated to consumers, whereas a product's true quality q can only become known after the good is bought and consumed. (2)

Consumers are uniformly distributed in the unit interval [0, 1] and have quadratic transportation costs. According to standard theory, the monopolist will locate his product at the interval's center. A consumer's utility from consuming a product of quality q is then given by

[u.sub.i] = f(q) - [16/27](i - [1/2])[.sup.2] - p,

where f(dot) is an arbitrary nonnegative function, p is the product's price, and i [member of] [0, 1] is the consumer's location in the unit interval. The scaling factor 16/27 simply serves to simplify the final expression of equilibrium firm profits. Expected utility maximization and price-taking behavior imply the following demand function:

D = [3[square root of 3]/2][square root of (f(q) - p)].

Variable costs are assumed to be zero or, alternatively, marginal costs are constant and prices are defined net of marginal costs. Firm profits are then simply equal to sales revenues w = Dp. If q is common knowledge, then profit maximization would imply price p = 2f(q)/3, demand D = 3[square root of (f(q))]/2, and sales revenues w = [f(q)][.sup.3/2].

In our setting, q is known to the firm but not to consumers. Consumers share a common prior regarding q. The prior is normally distributed with mean m and precision [tau]. In addition, consumers have access to an exogenously generated, normally distributed signal x of the product's true quality with mean q, precision [[rho].sub.x], and full support. In the context of this paper, this signal can be thought of as the arithmetic mean of online ratings posted by consumers who have already tried the product. (3) The signal's precision then is the sum of precisions of individual ratings. The firm knows the parameters of the consumers' prior and signal. However, it does not know the exact realization of the signal until it has been published by the forum.

Throughout this paper, I am treating word of mouth as an exogenous phenomenon, and do not make an attempt to justify why consumers engage in this costly activity on the basis of economic grounds. This perspective is consistent with a large body of empirical evidence (Dichter 1966, Engel et al. 1993, Sundaram et al. 1998, Hennig-Thurau et al. 2004) that has identified a variety of extraeconomic motivations to explain why consumers engage in (offline and online) word of mouth (desire to achieve social status, utility from engaging in social interaction, altruism, concern for others, easing anger, dissonance reduction, vengeance, etc.). In the context of information systems research, see Dellarocas et al. (2003) and Gu and Jarvenpaa (2003) for empirical studies of extraeconomic drivers of consumer contributions in online forums.

All consumers (and the firm) have access to the same realization of signal x. Let [theta] = E[q | x] denote the mean of consumer posterior beliefs regarding the product's quality after they observe x. I will refer to [theta] as the firm's perceived quality. If the firm sets its price after consumers observe signal x, no signaling of quality through price is possible. Maximization of expected sales revenues then implies price p = 2f([theta])/3 and sales revenues w = [f([theta])][.sup.3/2].

Taking advantage of the anonymity of the online medium, the monopolist can manipulate signal x by posting fake anonymous ratings that praise its product or by providing incentives to past consumers (who would otherwise not have posted online ratings) to do so. This way the monopolist can attempt to increase consumer perceptions of its quality by shifting the mean of the distribution of average ratings from q to q + [eta] at cost c([eta]) = [lambda][[eta].sup.2]. The parameter [lambda] captures the unit cost of manipulation.

Denote by y the signal that results from manipulation of the original signal x. The motivating question of this paper is to understand how…