Reputation Mechanisms

Reputation Mechanisms
Chrysanthos Dellarocas
R. H. Smith School of Business
University of Maryland
College Park, MD 20742
cdell@rhsmith.umd.edu
June 2005
Revised: August 2005
Abstract
Reputation mechanisms harness the bi-directional communication capabilities of the
Internet in order to engineer large-scale word-of-mouth networks. Best known so far as
a technology for building trust and fostering cooperation in online marketplaces, such
as eBay, these mechanisms are poised to have a much wider impact on organizations.
This paper surveys our progress in understanding the new possibilities and challenges
that these mechanisms represent. It discusses some important dimensions in which
Internet-based reputation mechanisms diﬀer from traditional word-of-mouth networks
and surveys the most important issues related to their design, evaluation, and use. It
provides an overview of relevant work in game theory and economics on the topic of
reputation. It discusses how this body of work is being extended and combined with
insights from computer science, marketing, and psychology in order to take into consid-
eration the special properties of online environments. Finally, it identiﬁes opportunities
that this new area presents for information systems research.
1
Introduction
The Internet oﬀers buyers and sellers unprecedented opportunities to access resources that
were diﬃcult, or impossible, to locate in previous times. Whether these resources are obscure
1

books, highly specialized services, or faraway trading partners, the global connectivity of the
Web has the potential to bring them within everyone’s reach, signiﬁcantly enriching our
economic and cultural lives.
Several of the mechanisms through which cooperation is induced in oﬄine settings, such as
the legal system and stable partnerships, do not work as well on the global, decentralized
Internet (Kollock, 1999). To fully reap the beneﬁts of global connectivity and arms-length
transactions with faraway partners our societies thus need to develop new trust mechanisms
capable of ensuring cooperation and eﬃciency in a universe of strangers.
Reputation networks constitute an ancient solution to the problem of trust-building. The
historical appeal of these networks has been their power to induce cooperation without
the need for costly enforcement institutions. Before the establishment of formal law and
centralized systems of contract enforcement backed by the sovereign power of a state, most
ancient and medieval communities relied on reputation as the primary enabler of economic
and social activity (Benson, 1989; Greif, 1993; Milgrom, North, and Weingast, 1990). Many
aspects of social and economic life still do so today (Klein, 1997).
These most ancient of mechanisms are emerging as one of the most promising solutions to
the problem of building trust on the Internet. Online reputation mechanisms, also known
as reputation systems (Resnick et al., 2000; Dellarocas, 2003a), are using the Internet’s
bi-directional communication capabilities in order to artiﬁcially engineer large-scale word-of-
mouth networks where individuals share opinions and experiences on a wide range of topics,
including companies, products, services, and even world events.
For example, eBay’s feedback mechanism is the primary means through which eBay elic-
its honest behavior and, thus, facilitates transactions among strangers over the Internet
(Resnick and Zeckhauser, 2002). Several other communities also rely on reputation mech-
anisms to promote trust and cooperation. Examples include eLance (online community of
freelance professionals), Slashdot (online discussion forum where reputation scores help pri-
oritize and ﬁlter postings), and Epinions (online consumer report forum where user feedback
helps evaluate the quality of product reviews). Table 1 lists several noteworthy examples of
such mechanisms in use today.
Online reputation mechanisms have a lot in common with their oﬄine counterparts. Their
design and implementation, thus, has a lot to gain from a substantial body of prior work on
reputation formation in economics and psychology. On the other hand, online mechanisms
possess a number of unique properties, whose implications are not yet fully understood.
Speciﬁcally:
2

Web Site
Category
Summary of feedback mechanism
Format of solicited feedback
Format of published
feedback
Citysearch
Entertainment
Users rate restaurants, bars, clubs,
Users rate multiple aspects of
Weighted averages of
guide
hotels and shops.
reviewed items from 1-10 and
ratings per aspect
answer a number of yes/no
reflecting both user
questions; readers rate reviews
and editorial ratings;
as “useful”, “not useful”, etc.
user reviews can be
sorted according to
“usefulness”
eBay
Online
Buyers and sellers rate one another
Positive, negative or neutral
Sums of positive,
auction house
following transactions
rating plus short comment;
negative and neutral
ratee may post a response
ratings received during
past 6 months
(see Section 3)
eLance
Professional
Contractors rate their satisfaction
Numerical rating from 1-5 plus
Average of ratings
services
with subcontractors
comment; ratee may post a
received during past 6
marketplace
response
months
Epinions
Online
Users write reviews about
Users rate multiple aspects of
Averages of item
opinions
products/services; other members
reviewed items from 1-5;
ratings; % of readers
forum
rate the usefulness of reviews
readers rate reviews as
who found a review
“useful”, “not useful”, etc.
“useful”
Google
Search engine
Search results are ordered based on
A Web page is rated based on
No explicit feedback
how many sites contain links that
how many links point to it, how
scores are published;
point to them
many links point to the
ordering acts as an
(Brin and Page 1998)
pointing page, etc.
implicit indicator of
reputation
Slashdot
Online
Postings are prioritized or filtered
Readers rate posted comments
discussion
according to the ratings they receive
board
from readers

Table 1: Examples of commercial reputation mechanisms (in use as of June 2005).
3

Global reach enables new applications. Scale is essential to the eﬀectiveness of reputation
networks. In an online marketplace, for example, sellers care about buyer feedback primarily
to the extent that they believe that it might aﬀect their future proﬁts; this can only happen if
feedback is provided by a suﬃcient number of current customers and communicated to a sig-
niﬁcant portion of future prospects. Theory predicts that a minimum degree of participation
in reputation communities is required before reputation eﬀects can induce any cooperation.
Once this threshold is reached, however, the power of reputation immediately springs to
life and high levels of cooperation emerge in a discontinuous fashion (Bakos and Dellaro-
cas, 2002). Therefore, the vastly increased scale of Internet-based reputation mechanisms is
likely to render them powerful institutions in environments where traditional word-of-mouth
networks were heretofore considered ineﬀective devices1. The social, economic and perhaps
even political consequences of such a trend deserve careful study.
Information technology enables systematic design. In oﬄine settings word-of-mouth emerges
naturally and evolves in ways that are diﬃcult to control or model. The Internet allows this
powerful social force to be precisely measured and controlled through proper engineering
of the information systems that mediate online reputation communities. Such automated
feedback mediators specify who can participate, what type of information is solicited from
participants, how it is aggregated and what type of information is made available to them
about other community members. Through the proper design of these mediators, mechanism
designers can exercise precise control over a number of parameters that are very diﬃcult or
impossible to inﬂuence in brick-and-mortar settings. For example, feedback mediators can
replace detailed feedback histories with a wide variety of summary statistics; they can apply
ﬁltering algorithms to eliminate outlier or suspect ratings; they can weight ratings according
to some measure of the rater’s trustworthiness, etc. Such degree of control can impact
the resulting social outcomes in non-trivial ways. Understanding the full space of design
possibilities and the impacts of speciﬁc design choices on the resulting social outcomes is an
important research challenge introduced by these new systems.
Online interaction introduces new challenges. The disembodied nature of online environ-
ments introduces several challenges related to the interpretation and use of online feedback.
1Three recent incidents illustrate the growing power of online opinion forums to exert inﬂuence on corpo-
rations and other powerful institutions of our society. In December 2002 criticism of controversial remarks
made by U.S. Senator Trent Lott by authors of Web Logs (blogs) eventually led to his resignation from his
post as majority leader. In 2003 Intuit Corporation was forced to remove unpopular copy protection spyware
from its TurboTax software following a wave of very negative reviews posted by customers in online product
forums. In September 2004 scrutiny by blog authors revealed inaccuracies in a story aired by long-serving
and respected CBS anchor Dan Rather. The ensuing events culminated into Dan Rather’s retirement from
the channel.
4

Some of these challenges have their roots in the subjective nature of consumer feedback.
Oﬄine settings usually provide a wealth of contextual cues that assist in the proper inter-
pretation of opinions and gossip (such as familiarity with the person who acts as the source
of that information, the ability to draw inferences from the source’s facial expression or mode
of dress, etc.). Most of these cues are absent from online settings. Readers of online feedback
are thus faced with the task of evaluating the opinions of complete strangers. Other chal-
lenges to feedback interpretation have their root in the ease with which online identities can
be changed. This opens the door to various forms of strategic manipulation. For example,
community members can build a good reputation, milk it by cheating other members and
then disappear and reappear under a new online identity and a clean record (Friedman and
Resnick, 2001). They can use fake online identities to post dishonest feedback and thus try
to inﬂate their reputation or tarnish that of their competitors (Dellarocas, 2004b). Finally,
the mediated nature of online reputation mechanisms raises questions related to the trust-
worthiness of their operators. An important prerequisite for the widespread acceptance of
online reputation mechanisms is, therefore, a better understanding of how such systems can
be compromised, as well as the development of adequate defenses.
This chapter surveys our progress so far in understanding the new possibilities and challenges
that these mechanisms represent. Section 2 introduces a framework for understanding the
role of reputation mechanisms in various settings. Section 3 provides an overview of relevant
past work in game theory and economics. Section 4 then discusses how this stylized body of
work is being extended in order to take into consideration the special properties of online en-
vironments. Sections 5 surveys empirical and experimental work on reputation mechanisms.
Finally, Section 6 summarizes the main points of the paper and discusses the opportunities
that this new area presents for information systems research.
2
Signaling and sanctioning role of reputation mecha-
nisms
The primary objective of reputation mechanisms is to enable eﬃcient transactions in com-
munities where cooperation is compromised by post-contractual opportunism (moral hazard)
or information asymmetries (adverse selection). It is instructive to distinguish the role of
reputation mechanisms with respect to moral hazard from their role with respect to adverse
selection.
5

Moral hazard can be present any time two parties come into agreement with one another.
Each party in a contract may have the opportunity to gain from acting contrary to the
principles laid out by the agreement. For example, on eBay, the buyer typically sends
money to the seller before receiving the goods. The seller then is tempted to keep the money
and not ship the goods, or to ship goods that are inferior to those advertised.
Reputation mechanisms can deter moral hazard by acting as sanctioning devices. If the
community follows a norm that punishes traders with histories of bad behavior (by refusing
to buy from them, or by reducing the price they are willing to pay for their products) and if
the present value of punishment exceeds the gains from cheating, then the threat of public
revelation of a trader’s cheating behavior in the current round provides rational traders with
suﬃcient incentives to cooperate.
Adverse selection is present in situations where sellers have information (about some aspect
of their innate ability, product quality, etc.) that buyers don’t (or vice versa). Such situations
often arise in markets for experience goods. Consider, for example, an online hotel booking
site where hotels of diﬀerent qualities advertise rooms. Consumers cannot be certain about
the true quality oﬀered by each hotel until they have actually stayed there. On the other
hand, hotels do not have an incentive to advertise any of their weak points. Knowing this,
consumers will assume that all hotels are of average quality and will not be willing to pay
more than the average price. Akerlof (1970) shows that such a situation will eventually drive
all, except the lowest quality sellers, out of the market.
Reputation mechanisms alleviate adverse selection issues by acting as signaling devices. For
example, by soliciting and publishing experiences of consumers who have stayed in advertised
hotels, they help the community learn the true quality of each hotel. This, in turn, allows a
better matching of buyers and sellers and a more eﬃcient market.
The most important distinction between (pure) moral hazard and (pure) adverse selection
settings is that, in the former, all sellers are capable of the same type of behavior (e.g.
cooperate, cheat), whereas in the latter case seller behavior is completely constrained by
their innate “type.” The role of reputation mechanisms in pure moral hazard settings is to
constrain behavior whereas the role of such mechanisms in pure adverse selection settings is
to induce learning.
In some real-life settings, moral hazard and adverse selection considerations are simultane-
ously present: Sellers diﬀer in their intrinsic ability levels but, in addition, have a choice of
behavior (which is partially, but not completely, conditioned by their type). For example,
6

certain attributes of the customer experience (location, size of rooms, etc.) can be consid-
ered as part of a hotel’s immutable “type”, whereas other attributes (cleanliness of facilities,
professionalism and politeness of staﬀ, etc.) are the result of the hotel’s level of “eﬀort” and
can be varied strategically on a daily basis. In such settings, reputation mechanisms play
both a sanctioning and a signaling role, revealing the hotel’s true immutable attributes while
providing incentives to the hotel to exert reasonable eﬀort.
In other settings, one of the two roles is dominant. For example, Amazon Reviews primarily
serves a signaling role: it spread information about the (initially privately known, but es-
sentially “immutable") qualities of the products (books, CDs, DVDs, etc.) being reviewed.
eBay, on the other hand, is an example of a mechanism that primarily acts as a sanctioning
device. Under the assumption that all eBay sellers are equally capable of acting in honest
and dishonest ways, eBay’s problem is to deter moral hazard. Accordingly, eBay users do
not rate sellers on the absolute quality of their products but rather on how well they were
able to deliver what was promised on the item description. The role of eBay’s reputation
mechanism is to promote honest trade rather than to distinguish sellers who sell high quality
products from those that sell low quality products.
The distinction between sanctioning and signaling is central in reputation mechanisms.
Throughout this article we shall see that several principles of reputation mechanism design
depend on this distinction. Designers should, therefore, be conscious of their mechanism’s
primary objective and carefully make design choices that maximize the resulting market
eﬃciency, given that objective.
3
Reputation in game theory and economics
Reputation formation has been extensively studied by economists using the tools of game
theory. This body of work is perhaps the most promising foundation for developing an
analytical discipline of online reputation mechanisms. This section surveys past work on
this topic, emphasizing the results that are most relevant to the design of online feedback
mechanisms. Section 4 then discusses how this stylized body of work is being extended to
address the unique properties of online environments.
7

3.1
Basic concepts
According to Wilson (1985), reputation is a concept that arises in repeated game settings
when there is uncertainty about some property (the “type”) of one or more players in the
mind of other players. If “uninformed” players have access to the history of past stage game
outcomes, reputation eﬀects then often allow informed players to improve their long-term
payoﬀs by gradually convincing uninformed players that they belong to the type that best
suits their interests. They do this by repeatedly choosing actions that make them appear
to uninformed players as if they were of the intended type, thus “acquiring a reputation” for
being of that type.
The existence of some initial doubt in the mind of uninformed players regarding the type of
informed players is crucial in order for reputation eﬀects to occur. To see this, consider a
repeated game between a long-run player and a sequence of short-run (one-shot) opponents.
In every stage game, the long-run player can choose one of several actions but cannot credibly
commit to any of those actions in advance. If there is no uncertainty about the long-
run player’s type2, rational short-run players will then always play their stage-game Nash
equilibrium response. Such behavior typically results in ineﬃcient outcomes.
Consider, for example, the following stylized version of a repeated “online auction” game. A
long-lived seller faces an inﬁnite sequence of sets of identical one-time buyers in a marketplace
where there are only two kinds of products:
• low-quality products that cost 0 to the seller and are worth 1 to the buyers, and
• high-quality products that cost 1 to the seller and are worth 3 to the buyers.
Each period the seller moves ﬁrst, announcing the quality of the product he promises to
buyers. Since high quality products are more proﬁtable, the seller will always promise high
quality. Buyers then compete with one another in a Vickrey auction and therefore bid
amounts equal to their expected valuation of the transaction outcome. The winning bidder
sends payment to the seller. The seller then has the choice of either “cooperating” (delivering
a high quality good) or “cheating” (delivering a low quality good). It is easy to see that this
game has a unique subgame perfect equilibrium. In equilibrium the seller always cheats
2In other words, if short-run players are convinced that the long-run player is a rational utility-maximizing
player whose stage-game payoﬀs are known with certainty.
8

(delivers low quality), buyers each bid 1, each buyer’s expected payoﬀ is zero and the seller’s
expected payoﬀ is 1.
The ability to build a reputation allows the long-run player to improve his payoﬀs in such
settings. Intuitively, a long-run player who has a track record of playing a given action (e.g.
cooperate) often enough in the past acquires a reputation for doing so and is “trusted” by
subsequent short-run players to do so in the future. However, why would a proﬁt-maximizing
long-term player be willing to behave in such a way and why would rational short-term players
use past history as an indication of future behavior?
To explain such phenomena, Kreps, Milgrom, Roberts and Wilson (1982), Kreps and Wil-
son (1982), and Milgrom and Roberts (1982) introduced the notion of “commitment” types.
Commitment types are long-run players who are locked into playing the same action3. An
important subclass of commitment types are Stackelberg types: long-run players who are
locked into playing the so-called Stackelberg action. The Stackelberg action is the action to
which the long-run player would credibly commit if he could. In the above “online auction”
example the Stackelberg action would be to cooperate; cooperation is the action that maxi-
mizes the seller’s lifetime payoﬀs if the seller could credibly commit to an action for the entire
duration of the game4. Therefore, the Stackelberg type in this example corresponds to an
“honest” seller who never cheats. In contrast, an “ordinary” or “strategic” type corresponds
to an opportunistic seller who cheats whenever it is advantageous for him to do so.
Reputation models assume that short-run players know that commitment types exist, but
are ignorant of the type of the player they face. An additional assumption is that short-run
players have access to the entire history of past stage game outcomes5. A player’s reputation
at any given time then consists of the conditional posterior probabilities over that player’s
type, given a short-run player’s prior over types and the repeated application of Bayes’ rule
on the history of past stage game outcomes.
In such a setting, when selecting his next move, the informed player must take into account
not only his short-term payoﬀ, but also the long-term consequences of his action based on
3Commitment types are sometimes also referred to as “irrational” types because they follow ﬁxed, “hard-
wired” strategies as opposed to “rational” proﬁt-maximizing strategies. An alternative way to justify such
players is to consider them as players with non-standard payoﬀ structures such that the “commitment” action
is their dominant strategy given their payoﬀs.
4If the seller could commit to cooperation (production of high quality), buyers would then each bid 2 and
the seller’s expected per period payoﬀ would be 2.
5The traditional justiﬁcation for this assumption is that past outcomes are either publicly observable or
explicitly communicated among short-run players. The emergence of online reputation mechanisms provides,
of course, yet another justiﬁcation (but see discussion of complications arising from the private observability
of outcomes in such systems in Section 4.1).
9

what that action reveals about his type to other players. As long as the promised future
gains due to the increased (or sustained) reputation that comes from playing the Stackelberg
action oﬀset whatever short-term incentives he might have to play otherwise, the equilib-
rium strategy for an “ordinary” informed player will be to try to “acquire a reputation” by
masquerading as a Stackelberg type (i.e. repeatedly play the Stackelberg action with high
probability).
In the “online auction” example, if the promised future gains of reputation eﬀects are high
enough6, ordinary sellers are induced to overcome their short-term temptation to cheat and
to try to acquire a reputation for honesty by repeatedly delivering high quality. Expecting
this, buyers then place high bids, thus increasing the seller’s long-term payoﬀs.
In general, reputation eﬀects beneﬁt the most patient player in the game: the player who
has the longest time horizon (discounts future payoﬀs less) is usually the one who is able to
reap the beneﬁts of reputation. Fudenberg and Levine (1992) show that this result holds
even when players can observe only noisy signals of each other’s actions, so that the game
has imperfect public monitoring. They prove that, if short-run players assign positive prior
probability to the long-run player being a Stackelberg type, and if that player is suﬃciently
patient, then an ordinary long-run player achieves an average discounted payoﬀ close to his
commitment payoﬀ (i.e., his payoﬀ if he could credibly commit to the Stackelberg action).
In order to obtain this payoﬀ, the ordinary player spends long periods of time choosing the
Stackelberg action with high probability7.
3.2
Reputation dynamics
In most settings where reputation phenomena arise, equilibrium strategies evolve over time
as information about the types of the various players accumulates. In general, the derivation
of closed-form solutions in repeated games with reputation eﬀects is complicated. Nev-
ertheless, a small number of speciﬁc cases have been studied. The general lesson is that
reputation-based performance incentives are highly dynamic: agents tend to behave diﬀer-
ently in diﬀerent phases of the game.
6In this type of game this requires that the remaining horizon of the seller is long enough and that the
proﬁt margin of a single transaction is high enough relative to the discount factor.
7This result also requires that the stage game is either a simultaneous move game, or, in a sequential-move
game, that the short-run players always observe whether or not the Stackelberg strategy has been played.
10

Initial phase In most cases, reputation eﬀects begin to work immediately and in fact are
strongest during the initial phase, when players must work hard to establish a reputation.
Holmstrom (1999) discusses an interesting model of reputational considerations in the context
of an agent’s “career” concerns. Suppose that wages are a function of an employee’s innate
ability for a task. Employers cannot directly observe an employee’s ability. However, they
can keep track of the average value of past task outputs. Outputs depend both on ability and
labor. The employee’s objective is to maximize lifetime wages while minimizing the labor
put in. At equilibrium, this provides incentives to the employee to work hard right from the
beginning of a career in order to build a reputation for competence. In fact these incentives
are strongest at the very beginning of a career when observations are most informative.
During the initial phase of a repeated game, it is common that some players realize lower or
even negative proﬁts, while the community “learns” their type. In those cases players will
only attempt to build a reputation if the losses from masquerading as a Stackelberg type in
the current round are oﬀset by the present value of the gains from their improved reputation
in the later part of the game. In trading environments, this condition usually translates to
the need for suﬃciently high proﬁt margins for “good quality” products so that the promise
of future gains from sustaining a reputation is persuasive enough to oﬀset the short-term
temptation to cheat. This was ﬁrst pointed out by Klein and Leﬄer (1981) and explored
more formally by Shapiro (1983).
Another case where reputation eﬀects may fail to work is when short-run players are “too
cautious” vis-à-vis the long-run player and therefore update their beliefs too slowly in order
for the long-run player to ﬁnd it proﬁtable to try to build a reputation. Such cases may occur
when, in addition to Stackelberg (“good”) types, the set of commitment types also includes
“bad” or “inept” types: players who always play the action that the short-run players like
least. In the “online auction” example, a “bad” type corresponds to a player who always
cheats (because, for example, he lacks the capabilities that would enable him to deliver high
quality.) If short-run players have a substantial prior belief that the long-run player may be
a “bad” type, then the structure of the game may not allow them to update their beliefs fast
enough to make it worthwhile for the long-run player to try to acquire a reputation.
Diamond’s (1989) analysis of reputation formation in debt markets presents an example of
such a setting. In Diamond’s model there are three types of borrowers: safe borrowers, who
always select safe projects (i.e. projects with zero probability of default); risky borrowers,
who always select risky projects (i.e. projects with higher returns if successful but with
nonzero probability of default); and strategic borrowers who will select the type of project
11

that maximizes their long term expected payoﬀ. The objective of lenders is to maximize their
long term return by oﬀering competitive interest rates, while at the same time being able
to distinguish proﬁtable from unproﬁtable borrowers. Lenders do not observe a borrower’s
choice of projects, but they do have access to her history of defaults. In Diamond’s model,
if lenders believe that the initial fraction of risky borrowers is signiﬁcant, then, despite the
reputation mechanism, at the beginning of the game, interest rates will be so high that
strategic players have an incentive to select risky projects. Some of them will default and
will exit the game. Others will prove lucky and will begin to be considered as safe players.
It is only after lucky strategic players have already acquired some initial reputation (and
therefore begin to receive lower interest rates) that it becomes optimal for them to begin
“masquerading” as safe players by consciously choosing safe projects in order to sustain their
good reputation.
Steady state (or lack thereof) In their simplest form, reputation games are character-
ized by an equilibrium in which the long-run player repeatedly plays the Stackelberg action
with high probability and the player’s reputation converges to the Stackelberg type.
The existence of such steady states crucially depends on the ability to perfectly monitor
the outcomes of individual stage games. For example, consider the “online auction” game
that serves as an example throughout this section with the added assumption that buyers
perfectly and truthfully observe and report the seller’s action. In such a setting, the presence
of even a single negative rating on a seller’s feedback history reveals the fact that the seller
is not honest. From then on, buyers will always choose the low bid in perpetuity. Since
such an outcome is not advantageous for the seller, reputation considerations will induce the
seller to cooperate forever.
The situation changes radically if monitoring of outcomes is imperfect. In the online auction
example, imperfect monitoring means that even when the seller produces high quality, there
is a possibility that a buyer will post a negative rating, and, conversely, even when the seller
produces low quality, the buyer may post a positive rating. A striking result is that in such
“noisy” environments reputations cannot be sustained indeﬁnitely: if a strategic player stays
in the game long enough, short-run players will eventually learn his true type and the game
will inevitably revert to one of the static Nash equilibria (Cripps, Mailath and Samuelson,
2004).
To see the intuition behind this result, note that reputations under perfect monitoring are
typically supported by a trigger strategy. Deviations from the equilibrium strategy reveal
12

the type of the deviator and are punished by a switch to an undesirable equilibrium of the re-
sulting complete-information continuation game. In contrast, when monitoring is imperfect,
individual deviations neither completely reveal the deviator’s type nor trigger punishments.
A single deviation has only a small eﬀect on the beliefs of the short-term players. As a
result, a player of normal type trying to maintain a reputation as a Stackelberg type incurs
only a small cost (in terms of altered beliefs) from indulging in occasional deviations from
Stackelberg play. In fact, it is clear that always playing the Stackelberg action cannot be an
equilibrium strategy, because if the short-term players expect long-term players of normal
type to behave that way, then they can actually deviate at no cost, since any bad outcome
will be interpreted by the short-run players as a result of imperfect monitoring. But the
long-run eﬀect of many such small deviations from the commitment strategy is to drive the
equilibrium to full revelation.
These dynamics have important repercussions for reputation systems in settings with both
moral hazard and adverse selection (for example, eBay, under the assumption that there
exist rational and honest seller types). According to the Cripps, Mailath and Samuelson
result, if eBay makes the entire feedback history of a seller available to buyers and if an eBay
seller stays on the system long enough, once he establishes an initial reputation for honesty
he will be tempted to cheat buyers every now and then. In the long term, this behavior will
lead to an eventual collapse of his reputation and therefore of cooperative behavior. I revisit
the implications of this result for reputation mechanism design in Section 4.2.
Endgame considerations Since reputation relies on a tradeoﬀ between current “restraint”
and the promise of future gains, in ﬁnitely repeated games, incentives to maintain a reputa-
tion diminish and eventually disappear as the end of the game comes close.
One solution to this problem is to assign some post-mortem value to reputation, so that
players ﬁnd it optimal to maintain it throughout the game. For example, reputations can
be viewed as assets that can be bought and sold in a market for reputations. Tadelis (1998)
shows that a market for reputations is indeed sustainable. Furthermore, the existence of
such a market provides “old” agents and “young” agents with equal incentives to exert eﬀort
(Tadelis, 2002). However, the long-run eﬀects of introducing such a market can be quite
complicated since good reputations are then likely to be purchased by “inept” agents for the
purpose of depleting them (Mailath and Samuelson, 2001; Tadelis, 2002). Further research
is needed in order to fully understand the long-term consequences of introducing markets for
reputation as well as for transferring these promising concepts to the online domain.
13

3.3
When is reputation bad?
In traditional reputation theory, publication of a long-term player’s past history of outcomes
is good for the long-term player. One, therefore, is tempted to assume that implementation
of reputation mechanisms is always a “good thing”. Ely, Fudenberg, and Levine (2005),
henceforth referred to as EFL, challenge this assumption and show that there exist settings
where the presence of public histories of past outcomes is unambiguously bad.
EFL generalize an example provided by Ely and Valimaki (2003), henceforth referred to as
EV. EV describe a setting where a mechanic of unknown “character” (rational, dishonest)
is facing a sequence of customers who bring their cars to him for repair. Each car might
need a tune-up or an engine change; only the mechanic is able to determine the correct type
of repair. Rational mechanics are assumed to have a payoﬀ structure that induces them to
perform the correct type of repair in a one-stage game. Dishonest mechanics, on the other
hand, always perform engine changes.
Assume, now, that customers have access to the history of past repairs performed by the
mechanic and will only contract with him if they are suﬃciently conﬁdent that he is going
to perform the correct type of repair. Given the above assumptions, if there is a positive
prior probability that the mechanic is dishonest, after histories with many engine changes,
the short-run players will become suﬃciently convinced they are facing such a bad type and
exit. In order to avoid these histories, a rational mechanic who has had the “bad luck” of
having many customers that need an engine change may then begin to recommend tune-ups
to customers who need engine changes; foreseeing this, the short-run players will chose not
to enter.
Observe that, whereas, in the absence of reputation, rational types play friendly actions, the
presence of a public history of past outcomes induces them to behave in ways that short-run
players like least. Foreseeing this, short-run players choose not to enter; the presence of
reputational information then causes the market to break down.
EFL show that settings where reputation is bad are characterized by the following properties:
(i) a long-run player of privately-known type is facing an inﬁnite sequence of short-run
opponents, (ii) commitment types include “unfriendly” types who play “unfriendly” actions,
that is, actions that short-run players dislike, (iii) short-run players will only enter the game
if they are suﬃciently conﬁdent that the long-run player is going to play one of the friendly
actions, (iv) there exist “bad signals” that are most likely to occur when unfriendly actions
are played but also occur with positive probability when friendly actions are played (in EV
14

the bad signal is “engine change”). Finally, (v) there are some actions that are not friendly,
but reduce the probability of bad signals (such as “always perform tune-up” in EV).
The main result of EFL is that, in a bad reputation game with a suﬃciently patient long-run
player and likely enough unfriendly types, in any Nash equilibrium, the long-run player gets
approximately the payoﬀ that results from non-participation (exit) of short-run players.
EFL show that bad reputation games arise in various “expert advice” settings. This includes
consulting a doctor or stockbroker, or in the macroeconomics context, the decision whether
or not to turn to the IMF for assistance. In EV, the short-run players observe only the
advice, but not the consequences of the advice. EFL consider what happens when the short-
run players observe the consequences as well. They show that the bad reputation eﬀect
persists so long as this additional information is not perfectly accurate.
The EFL/EV result delivers a pessimistic message with respect to the application of rep-
utation mechanisms in expert advice settings, an area where the author believes that such
mechanisms have a lot to oﬀer. On the other hand, the result crucially depends on the as-
sumption that short-term players will exit as soon as the long-term player’s reputation falls
below a threshold. If the long-term player can persuade them to participate (by charging
lower prices or, perhaps, paying them a participation reward) the result breaks down: if a
rational player who has had a stream of “bad luck” has a way to restore his reputation, he will
continue to engage in friendly actions. Nevertheless, reputation mechanism designers must
be aware of the EFL result and its implications. It is important to understand the range
of practical settings in which these results might apply, and therefore to identify classes of
settings for which the development of online reputation mechanisms may not be a good idea8.
3.4
Other extensions to the basic theory
The basic theory assumes that uninformed players are short-term. Facing longer-lived op-
ponents may be worse for the informed player and generally results in less sharp predictions
8A related, but not identical, problem arises when the establishment of a reputation mechanism induces
long-run players to change their behavior in ways that improve their payoﬀ but reduce social welfare. Dranove
et al. (2003) examine the consequences of public disclosure of patient health outcomes at the level of
the individual physician and/or hospital. The intention behind this measure was to address informational
asymmetries in markets for health care. However, it also gave doctors and hospitals incentives to decline
to treat more diﬃcult, severely ill patients. Using national data on Medicare patients at risk for cardiac
surgery, Dranove et al. ﬁnd that cardiac surgery report cards in New York and Pennsylvania led both to
selection behavior by providers and to improved matching of patients with hospitals. On net, this led to
higher levels of resource use and to worse health outcomes, particularly for sicker patients.
15

about reputation eﬀects (Cripps and Thomas, 1995; Cripps, Schmidt and Thomas, 1996).
Quite interestingly, however, in repeated games where a patient player faces one or more
long-lived but less patient opponents, if the more patient player does not observe the less
patient players’ intended actions but only sees an imperfect signal of them, reputation eﬀects
once again become strong and result in lower bounds that are even higher than in the case
where all opponents are myopic (Celentani, Fudenberg, Levine and Pesendorfer, 1996). This
last case is equivalent to a situation where a long-run player faces a sequence of long-run
but “infrequent” players. This is, perhaps, an even more realistic model of relationships in
online communities and therefore an area that deserves further study.
Another assumption underlying most game theoretic models of reputation is that all players
have identical prior beliefs and that behavior is consistent with the concept of Bayesian Nash
equilibrium. These assumptions are probably too stringent and unrealistic in environments
as diverse as large-scale online communities. Fortunately, reputation phenomena arise under
signiﬁcantly weaker assumptions on the knowledge and behavior of players. Watson (1993;
1996) and Battigalli and Watson (1997) demonstrated that reputation eﬀects do not require
equilibrium. They are implied by a weak notion of rationalizability along with two main
conditions on the beliefs of players: First, there must be a strictly positive and uniform lower
bound on the subjective probability that players assign to the Stackelberg type. Second, the
conditional beliefs of short run players must not be “too dispersed”.
4
New opportunities and challenges of online mechanisms
I began this chapter by discussing a number of diﬀerences between online reputation mecha-
nisms and traditional word-of-mouth networks. This section surveys our progress in under-
standing the opportunities and challenges that these special properties imply.
4.1
Eliciting suﬃcient and honest feedback
Most game theoretic models of reputation formation assume that stage game outcomes (or
imperfect signals thereof) are publicly observed. Most online reputation mechanisms, in
contrast, rely on private monitoring of stage game outcomes and voluntary self-reporting9.
9For a comprehensive introduction to games with imperfect private monitoring see Kandori (2002) and
other papers contained in the same issue.
16

This introduces two important new considerations: (a) ensuring that suﬃcient feedback is,
indeed, provided, and (b) inducing truthful reporting.
Economic theory predicts that voluntary feedback will be underprovided. There are two
main reasons for this. First, feedback constitutes a public good: once available, everyone
can costlessly beneﬁt from it. Voluntary provision of feedback leads to suboptimal supply,
since no individual takes account of the beneﬁts that her provision gives to others. Second,
provision of feedback presupposes that the rater will assume the risks of transacting with
the ratee. Such risks are highest for new products. Prospective consumers may, thus, be
tempted to wait until more information is available. However, unless somebody decides to
take the risk of becoming an early evaluator, no feedback will ever be provided.
Avery, Resnick and Zeckhauser (1999) analyze mechanisms whereby early evaluators are paid
to provide information and later evaluators pay in order to balance the budget. They con-
clude that, of the three desirable properties for such a mechanism (voluntary participation,
no price discrimination and budget balance), any two can be achieved, but not all three10.
Since in most reputation mechanisms, monitoring of transaction outcomes is private, an
additional consideration is whether feedback is honest. No generally applicable solution to
this important problem currently exists. Nevertheless, several researchers have proposed
mechanisms that induce truth-telling in restricted settings.
Jurca and Faltings (2004) propose a mechanism that limits false reporting in settings with
pure moral hazard. They consider a bilateral exchange setting where long-run buyers and
long-run sellers transact repeatedly. Each period a buyer is asked to rate a transaction
only if the corresponding seller has claimed to have successfully provided the service. If the
two reports disagree, at least one of the traders must be lying; the center then ﬁnes both
transacting parties diﬀerent (ﬁxed) amounts11. Jurca and Faltings show that, if a buyer
always reports feedback on a particular seller truthfully, the seller also ﬁnds it optimal to
truthfully report transaction outcomes, with the exception of a ﬁnite number of transactions.
10Empirical evidence has, so far, not conﬁrmed the (rather pessimistic) predictions of theory. Consumer
participation in online feedback mechanisms is surprisingly high, even though, in most cases, such mechanisms
oﬀer no concrete participation incentives. Such behavior is consistent with a large body of empirical evidence
(Dichter, 1966; Engel, Blackwell and Miniard, 1993; Sundaram, Mitra and Webster, 1998; Hennig-Thuray
et al., 2004) that has identiﬁed a variety of extra-economic 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.). I return to this point in
Section 5.
11Fines can be implemented by levying refundable listing fees from the two traders at the beginning of
each transaction and conﬁscating these fees in the case of conﬂicting reports.
17

Papaioannou and Stamoulis (2005) propose a similar mechanism that is suitable for Peer-to-
Peer environments. In such environments side-payments are usually not possible and peers
are able to exchange roles. After each transaction both peers submit a rating, with each peer
not knowing the rating submitted by the other. A credibility metric is maintained for each
peer regarding his overall truthfulness record in rating transactions. In case of agreement
between two peers, the credibility metric of each peer is improved. In case of disagreement,
the credibility metric of each peer is deteriorated and both peers are punished. Punishment
amounts to not allowing a peer to transact with others for a period that is exponential to
the peer’s credibility metric. This is enforced by publicly announcing a peer’s punishment
and by punishing other peers when they transact with him. Performance of this mechanism
is analyzed by means of experiments in dynamically evolving peer-to-peer systems with
renewed populations.
Miller, Resnick and Zeckhauser (2005) propose mechanisms for eliciting honest feedback in
environments with pure adverse selection. Their mechanisms are based on the technique of
proper scoring rules (Cooke, 1991). A scoring rule is a method that induces rational agents to
truthfully reveal their beliefs about the distribution of a random variable by rewarding them
on the basis of how a future realization of the random variable relates to the distribution they
announced. A proper scoring rule has the property that the agent maximizes his expected
score when he truthfully announces his beliefs.
Assuming that: (i) users rate a set of products, whose types (e.g. qualities) remain ﬁxed over
time, (ii) each type maps to a distribution of outcomes (e.g. satisfaction levels perceived by
consumers) that can be statistically distinguished from that of every other type, and (iii) all
raters and the center have a common set of prior beliefs over types, or, alternatively, each
rater’s prior beliefs have been communicated to the center, MRZ propose a side-payment
mechanism that asks each rater to announce the outcome a1 she observed, and rewards her by
an amount proportional to log p(a2|a1) where a2 is the outcome reported by a future rater.
MRZ show the existence of a Nash equilibrium where such a side-payment rule induces
truth-telling. They also show that their mechanism can be extended to not only induce
truth-telling, but also ratings of a given precision.
The above mechanisms represent promising ﬁrst steps towards ensuring the credibility of
online feedback. On the other hand, they have several limitations. First, such mechanisms
only work as long as raters are assumed to act independently; all break down if raters
collude. Second, in addition to the desirable truth-telling equilibria., all three mechanisms
induce additional equilibria where agents do not report the truth. Equilibrium selection,
18

thus, becomes an important consideration in practical implementations. The development
of robust mechanisms for eliciting truthful feedback thus remains one of the most important
open areas of research.
4.2
Exploring the design space of feedback mediators
Information technology has added a large degree of ﬂexibility and precision to the design
of reputation mechanisms. Online mechanism designers can control a number of important
parameters that are diﬃcult, or impossible, to inﬂuence in oﬄine settings. Examples of such
parameters include the format of solicited feedback (eBay allows traders to rate a transaction
as "positive", "negative" or "neutral", Amazon Auctions supports integer ratings from 1 to
5, other systems support even higher levels of detail), the amount and type of information
included in a trader’s reputation proﬁle (most systems publish the sum or arithmetic mean
of all posted ratings, some systems highlight recent ratings, other systems provide access to
a trader’s entire ratings history) as well as the frequency with which reputation proﬁles are
updated with new information (most current systems make new ratings publicly available as
soon as they are posted). These parameters impact the consequences of a trader’s current
behavior on the community’s perception of him in the future, and thus, his incentives to
cooperate. This section summarizes our current understanding of how such design choices
aﬀect trader behavior and market eﬃciency.
Granularity of feedback. eBay’s mechanism solicits ternary feedback (positive, neutral, neg-
ative) while Amazon Auctions asks traders to rate transactions on a ﬁner-grained scale of
1-5. What is the impact of the granularity of feedback on the seller’s incentives to cooper-
ate? In the special case of settings with two seller actions (cooperate, cheat) and pure moral
hazard, Dellarocas (2005a) shows that the equilibrium that maximizes cooperation occurs
when buyers divide feedback of arbitrary granularity into two disjoint subsets (the “good
ratings” subset and the “bad ratings” subset) and behave as if feedback was binary, i.e. they
reward the seller by the same amount if he receives any rating that belongs to the “good” set
and they punish him by the same amount if he receives any rating that belongs to the “bad”
set. Maximum eﬃciency is then inversely proportional to the minimum likelihood ratio of
punishment if the seller cooperates vs. if he cheats over all possible ways of dividing ratings
into “good” and “bad” subsets12.
12In more general settings where both adverse selection and moral hazard are present the answer is generally
more complex. Soliciting (and publishing) ﬁner-granularity feedback might increase or decrease a seller’s
incentives to cooperate. See Dewatripont, Jewitt, and Tirole (1999) for a precise statement of the relevant
conditions.
19

Length of published feedback history. How much history should a mechanism publish about
a trader’s past behavior? The answer here crucially depends on the type of setting (moral
hazard, adverse selection, or combined). In pure adverse selection settings the goal of the
mechanism is to promote social learning. More information is, thus, always better. In
contrast, in pure moral hazard settings Dellarocas (2005a) shows that the maximum eﬃciency
that can be induced by a reputation mechanism is independent of the length of published
history: a mechanism that publishes a trader’s entire history performs as well as a mechanism
that only publishes the trader’s single most recent rating. The intuition behind this result
is that, in pure moral hazard settings, the goal of the reputation mechanism is to threaten
players with future punishment if the public outcome of the current transaction indicates
that they cheated. The length of time during which a rating persists in the seller’s reputation
proﬁle only aﬀects the duration of future punishment but not the total amount of punishment.
This makes eﬃciency independent of the length of published history.
Finally, in settings where both moral hazard and adverse selection are present, if the objective
of the mechanism is to induce long-term seller cooperation, Cripps, Mailath and Samuelson’s
result (see Section 3.2) suggests that reputation mechanisms should not publish a seller’s
entire history. (Otherwise, once sellers have established a good reputation, they will be
tempted to occasionally cheat; in the long-run, this will reveal their opportunistic nature
and will drive the system to the undesirable one-shot Nash equilibrium.) On the other hand,
since their role is to promote some degree of learning about the seller’s type, reputation
mechanisms should publish some amount of past history. Fan, Tan and Whinston (2005)
explore the idea of discounting past ratings in such settings. Using simulation, they show
that such policies can sustain a seller’s incentives to cooperate. General design guidelines for
deciding what is the optimal length of history (or, equivalently, what is the optimal discount
factor of past ratings) constitute an interesting open area of research.
Frequency of feedback proﬁle updating. Most systems in use today update a trader’s reputa-
tion proﬁle with new evaluations as soon as these are posted by users. Dellarocas (2005b)
shows that, in pure moral hazard settings, this is not necessarily the optimal architecture.
Speciﬁcally, he shows that, if ratings are noisy and the per-period proﬁt margin of coop-
erating sellers suﬃciently high, a mechanism that does not publish every single rating it
receives but rather, only updates a trader’s public reputation proﬁle every k transactions
with a summary statistic of a trader’s most recent k ratings, can induce higher average levels
of cooperation and market eﬃciency than a mechanism that publishes all ratings as soon
as they are posted. The intuition behind the result is that basing each proﬁle update on
a trader’s most recent k-transaction rating history reduces the impact of spurious negative
20

ratings (because these are amortized over k transactions). On the other hand, it also in-
creases the trader’s temptation to cheat (because he can cheat for k period before news of his
behavior become public). The optimal updating period k is derived from a tradeoﬀ between
these two opposite forces.
4.3
Coping with cheap online identities
In online communities it is usually easy for members to disappear and re-register under a
completely diﬀerent online identity with zero or very low cost. Friedman and Resnick (2001)
refer to this property as “cheap pseudonyms.” This property hinders the eﬀectiveness of
reputation mechanisms. Community members can build a reputation, milk it by cheating
other members and then vanish and re-enter the community with a new identity and a clean
record.
Friedman and Resnick discuss two classes of approaches to this issue: either make it more
diﬃcult to change online identities, or structure the community in such a way that exit
and re-entry with a new identity becomes unproﬁtable. The ﬁrst approach makes use of
cryptographic authentication technologies and is outside the scope of this paper. The second
approach is based on imposing an upfront cost to each new entrant, such that the beneﬁts
of “milking” one’s reputation are exceeded by the cost of subsequent re-entry. This cost can
be an explicit entrance fee or an implicit cost of having to go through a reputation building
(or “dues paying”) stage with low or negative proﬁts. Friedman and Resnick show that,
although dues paying approaches incur eﬃciency losses, such losses constitute an inevitable
consequence of easy name changes.
Dellarocas (2005a) shows how such a “dues paying” approach can be implemented in pure
moral hazard trading environments. He proves that, in the presence of easy name changes,
the design that results in optimal social eﬃciency is one where newcomers pay an entrance fee
and the mechanism only publishes a trader’s single most recent rating. Dellarocas further
demonstrates that, although this design incurs eﬃciency losses relative to the case where
identity change is not an issue, in settings with two possible transaction outcomes and
opportunistic sellers that can freely change identities, its eﬃciency is the highest attainable
by any mechanism.
21

4.4
Understanding the consequences of strategic manipulation
As online reputation mechanisms begin to exercise greater inﬂuence on consumer behavior
the incentive for strategically manipulating them becomes correspondingly stronger. The low
cost of submitting online feedback coupled with the relative anonymity of the rater makes
such manipulation a real problem that needs to be studied and addressed.
Dellarocas (2000; 2004a) has pointed out a number of manipulation scenarios and has pro-
posed a number of immunization mechanisms, based on ideas from clustering and robust
statistics, that reduce the eﬀect of such attacks if the fraction of unfair raters is reason-
ably small (up to 20-30% of the total population). A number of commercial reputation
mechanisms (for example, Amazon Reviews and Epinions) are attempting to address such
problems through the concept of “rate the rater”: members can rate how useful other mem-
bers’ feedback has been to them. Whereas this technique is somewhat eﬀective for separating
high-quality from low-quality postings, it is not eﬀective for reducing strategic manipulation.
Determined manipulators can manipulate the “rate the rater” ratings as much as they can
manipulate the ratings themselves.
The ability to “cheaply” create multiple online identities further complicates attempts to
combat manipulation attacks. For example, an early technique through which eBay traders
attempted to manipulate their reputation proﬁle was for colluding traders to form a ring,
engaging in repeated “fake” transactions with each other for the purpose of artiﬁcially boost-
ing each other’s feedback score. To combat such attacks, eBay is now only counting unique
ratings received from frequent partners towards a trader’s reputation proﬁle. Such defenses
can be defeated, however, if traders can easily create multiple identities. Douceur (2002) has
called this the sybil attack. Cheng and Friedman (2005) provide formal deﬁnitions of sybil-
proofness, i.e. resistance to sybil attacks, and prove the impossibility of developing sybilproof
reputation mechanisms is a broad class of settings.
Since it appears that complete elimination of reputation mechanism manipulation will remain
a formidable task for some time, it is useful to explore the consequences of such activity in
strategic settings. In settings with pure adverse selection, Mayzlin (2003) has analyzed the
impact of strategic manipulation in the context of Usenet groups where consumers discuss
products and services. Mayzlin’s setting involves two rival ﬁrms that wish to inﬂuence
consumer beliefs regarding the quality of their respective products by posting costly “fake”
promotional messages in Usenet groups. Mayzlin’s basic result is that, if the ratio of proﬁts to
manipulation cost is high enough, there exists an equilibrium in which both ﬁrms manipulate
22

but the low quality ﬁrm manipulates more. Promotional chat thus decreases, but does
not completely destroy, the informativeness of online forums. Dellarocas (2004b) extended
Mayzlin’s results in more general settings. His most interesting new result is that there exist
equilibria where strategic manipulation can increase forum informativeness. Such equilibria
arise in settings where ﬁrm revenues are suﬃciently convex functions of consumer perceptions
of their quality. In such settings, the presence of honest consumer opinions can induce
ﬁrms to reveal their own, more precise, knowledge of product qualities by manipulating the
forums at relative intensities that are proportional to their actual qualities. The impact of
manipulation then is to further separate the published ratings of the high quality ﬁrm from
those of the low quality ﬁrm, making it easier for consumers to infer each ﬁrm’s true quality.
4.5
Distributed reputation mechanisms
Traditional reputation theory assumes the existence of a public history of past outcomes.
This implies a centralized architecture where outcomes are either automatically recorded
or explicitly self-reported. It also implies the presence of a trusted mediator who controls
feedback aggregation and distribution. Though the design possibilities of even that simple
architecture are not yet fully understood, centralized reputation mechanisms do not nearly
exhaust the new possibilities oﬀered by information technology.
The growing importance of Peer-to-Peer (P2P) networks (Oram, 2001) is introducing new
challenges to reputation mechanism design. In P2P networks every entity can act both as a
provider and consumer of resources. Entities are assumed to be self-interested and cannot
be trusted to engage in cooperative behavior unless concrete incentives are in place. For
example, in ﬁle-sharing P2P networks, self-interested entities have short-term incentives to
free-ride (consume as much content as they can without contributing any content themselves)
or to contribute low-quality content. Furthermore, there is, usually, no central, universally
trusted entity that can act as a repository of reputational information.
To cope with these challenges, several researchers have proposed decentralized reputation
mechanisms. Two lines of investigation stand out as particularly promising:
Reputation formation based on analysis of “implicit feedback.” Traditional reputation mech-
anisms rely on explicit solicitation of feedback from transaction participants. If reliable
explicit feedback is not available, information about an agent’s type can often be inferred by
analyzing publicly available attributes of the network in which the agent is embedded.
23

Perhaps the most successful application of this approach to date is exempliﬁed by the Google
search engine. Google’s PageRank algorithm assigns a measure of reputation to each web
page that matches the keywords of a search request. It then uses that measure to rank order
search hits. Google’s page reputation measure is based on the number of links that point
to a page, the number of links that point to the pointing page, and so on (Brin and Page,
1998). The underlying assumption is that if enough people consider a page to be important
enough in order to place links to that page from their pages, and if the pointing pages are
“reputable” themselves, then the information contained on the target page is likely to be
valuable. Google’s success in returning relevant results is testimony to the promise of that
approach.
Pujol , Sangüesa and Delgado (2002) apply network ﬂow techniques in order to propose a
generalization of the above algorithm that “extracts” the reputation of nodes in a general
class of social networks. Sabater and Sierra (2002) describe how direct experience, explicit
and implicit feedback can be combined into a single reputation mechanism.
Basing reputation formation on implicit information is a promising solution to problems of
eliciting suﬃcient and truthful feedback. Careful modeling of the beneﬁts and limitations of
this approach is needed in order to determine in what settings it might be a viable substitute
or complement of voluntary feedback provision.
Distributed reputation mechanisms. The majority of decentralized reputation mechanisms
proposed so far are based on variations on the theme of referral networks (Aberer and
Despotovic, 2001; Kamvar et al., 2003; Xiong and Liu, 2004; Yu and Singh, 2002). Since no
centralized reputation repository exists, each agent in a P2P network solicits referrals about
a target agent from a set of neighbors who might, in turn, ask their neighbors, and so on.
Referrals are weighted by the relative amounts of trust that the soliciting agent places on his
neighbors’ advice and combined with personal experiences that the soliciting agent might
have had with the target agent in the past. Based on his subsequent experience with the
target agent, the soliciting agent dynamically adapts both his beliefs regarding the target
agent, as well as the amounts by which she trusts her neighbors’ recommendations13.
Most work in this area is currently based on heuristics and evaluated using simulation. The
development of rigorous results on the eﬃciency bounds and design principles of distributed
reputation mechanisms constitutes an important open area of research.
13See Despotovic and Aberer (2004) for a more comprehensive overview of trust-building mechanisms in
P2P networks.
24

5
Empirical and experimental studies
A growing literature of empirical and experimental studies constitutes an essential com-
plement to game-theoretic and simulation-based analyses of reputation mechanisms. This
section surveys some of the main themes of such work.
5.1
Empirical studies and ﬁeld experiments
Most empirical studies of reputation mechanisms have focused on eBay’s feedback mecha-
nism. The majority of early studies has looked at the buyers’ response to seller reputation
proﬁles. In particular, a large number of studies estimate cross-sectional regressions of sale
prices and probabilities of sale on seller feedback characteristics. Table 2 summarizes the
main results of these studies.
The following points summarize the principal conclusions derived from a collective reading
of these works:
• Feedback proﬁles seem to aﬀect both prices and the probability of sale. However, the
precise eﬀects are ambiguous; diﬀerent studies focus on diﬀerent components of eBay’s
complex feedback proﬁle and often reach diﬀerent conclusions.
• The impact of feedback proﬁles on prices and probability of sale is relatively higher for
riskier transactions and more expensive products.
• Among all diﬀerent pieces of feedback information that eBay publishes for a member,
the components that seem to be most inﬂuential in aﬀecting buyer behavior are the
overall number of positive and negative ratings, followed by the number of recently
(last 7 days, last month) posted negative comments.
Resnick, Zeckhauser, Swanson and Lockwood (2005) point out the potential for signiﬁcant
omitted variable bias in these cross-sectional regressions. They argue that the price diﬀerence
commanded by sellers of higher reputation might be due to several other factors that tend to
exhibit positive correlation with a seller’s eBay score (for example, the clarity and aesthetics
of the item listing, the professionalism of the seller’s email communications, etc.). To better
assess the premium attached to reputation they conduct a controlled ﬁeld experiment in
which a seasoned seller sells identical postcards using his real name and an assumed name.
25

Shorthand
Citation
Items sold
Remarks
BP
Ba and Pavlou 2002
Music, Software,
Positive feedback increased estimated price, but
Electronics
negative feedback did not have an effect
BH
Bajari and Hortacsu 2003
Coins
Both positive and negative feedback affect
probability of modeled buyer entry into the auction,
but only positive feedback had a significant effect
on final price
DH
Dewan and Hsu 2004
Stamps
Higher net score increases price
E
Eaton 2002
Electric guitars
Negative feedback reduces probability of sale, but
not price of sold items
HW
Houser and Wooders 2005
Pentium chips
Positive feedback increases price; negative
feedback reduces it
KM
Kalyanam and McIntyre
Palm Pilot PDAs
Positive feedback increases price; negative
2001
feedback reduces price
KW
Kauffman and Wood 2000
Coins
No significant effects, but negative feedback seems
to increase price (!) in univariate analysis
LIL
Lee, Im and Lee 2000
Computer monitors
Negative feedback reduces price, but only for used
and printers
items
L
Livingston 2002
Golf clubs
Positive feedback increases both likelihood of sale
and price; effect tapers off once a record is
established
LBPD
Lucking-Reiley et al. 2000
Coins
No effect from positive feedback; negative
feedback reduces price
MA
Melnik and Alm 2002
Gold coins
Positive feedback increases price; negative
feedback decreases price
MS
McDonald and Slawson
Dolls
Higher net score (positives -negatives) increases
2002
price
RZ
Resnick and Zeckhauser
MP3 players,
Both forms of feedback affect probability of sale
2002
Beanie babies
but not price contingent on sale
RZSL
Resnick Zeckhauser,
Vintage postcards
Controlled field experiment; established seller
Swanson and Lockwood
commands higher prices than newcomers; among
2005
newcomers, small amounts of negative feedback
have little effect
Table 2: Summary of early empirical studies on eBay (adapted from Resnick et al., 2005).
26

They ﬁnd an 8% premium to having 2000 positive feedbacks and 1 negative over a feedback
proﬁle with 10 positive comments and no negatives.
In another ﬁeld experiment, Jin and Kato (2004) assess whether the reputation mechanism
is able to combat fraud by purchasing un-graded baseball cards with seller-reported grades,
and having them evaluated by the oﬃcial grading agency. They report that, while having a
better seller reputation is a positive indicator of honesty, reputation premia or discounts in
the market do not fully compensate for expected losses due to seller dishonesty.
Using panel data, Cabral and Hortacsu (2005) analyze the impact of reputation on eBay
sellers. They ﬁnd that, when a seller ﬁrst receives negative feedback, his weekly sales rate
drops from a positive 7% to a negative 7%; subsequent negative feedback ratings arrive 25%
more rapidly than the ﬁrst one and don’t have as much impact as the ﬁrst one. They also
ﬁnd that a seller is more likely to exit the lower his reputation is; and that, just before
exiting, sellers receive more negative feedback than their lifetime average.
Cabral and Hortacsu then consider a series of theoretical models (pure moral hazard, pure
adverse selection, combined moral hazard and adverse selection) and measure them against
these empirical results. They are not able to draw deﬁnite conclusions as to which theoretical
model best explains the data; they do, however, conclude that eBay’s reputation system gives
way to noticeable strategic responses from both buyers and sellers.
Another interesting area of empirical work relates to the motivations for participation in
online feedback mechanisms. Economic theory predicts that, being public goods, evaluations
will be underprovided unless evaluators are provided with concrete incentives (Avery, Resnick
and Zeckhauser, 1999). Nevertheless, in systems such as eBay, more than 50% of transactions
receive feedback even though no direct incentives to rate are in place. Dellarocas, Fan and
Wood (2003b) use survival analysis techniques to study the motivations and dynamics of
voluntary feedback submission on eBay. They ﬁnd that reciprocity is an important driver
of participation on eBay’s mechanism: a trader’s propensity to rate a transaction increases
after her partner posts a rating. They also ﬁnd that silence is often an indicator of an
unsatisfactory transaction and propose an empirical method for inferring the true number
of unsatisfactory transactions on eBay from the observed patterns of feedback submission.
Using their method, they estimate that, whereas the percentage of negative feedback on eBay
is less than 1%, the fraction of transactions where, at least one of the two parties remained
unsatisﬁed is closer to 10%.
Product review forums (such as Amazon Reviews) have, so far, received less attention. Nev-
ertheless, their increasing popularity is attracting more empirical research. Hennig-Thurau
27

et al. (2004) conduct an online survey to identify what motivates consumers to rate products
in such forums. Their study identiﬁes four signiﬁcant motives. In order of decreasing signiﬁ-
cance the four motives are: social beneﬁts (genuine fun that results from the social experience
of participating in an online dialog), economic incentives (rewards oﬀered by some sites for
contributing content), concern for other consumers (urge to help others by recommending
good products and warning against bad products) and extraversion/self-enhancement (pos-
itive feelings that result from sharing one’s successes with others; enhancement of one’s
self-image by projecting oneself as intelligent shopper).
Another question of theoretical and practical consequence is the extent to which such forums
inﬂuence consumer behavior. Obtaining reliable answers to this question is challenging. eBay
has the advantage of being a “closed universe”: all transactions and feedback are mediated
through the site and are publicly visible. On the other hand, in the general case consumers
can obtain product ratings from one forum and purchase the product from another site, or
from an oﬄine store. It is, thus, diﬃcult to establish causality between ratings and sales.
The approach of Chevalier and Mayzlin (2003) constitutes a step in the right direction.
Chevalier and Mayzlin examine the eﬀect of consumer reviews on relative sales of the same
set of books on Amazon.com and BarnesandNoble.com, two large online booksellers14. They
ﬁnd that an improvement in a book’s reviews on one site leads to an increase in relative
sales at that site. They also ﬁnd that reviews are overwhelmingly positive and, thus, that
the impact of (less common) negative reviews is greater than the impact of positive reviews.
5.2
Controlled experiments
Controlled experiments are a useful complement to empirical studies, especially in relation
to exploring individual-level perceptions of and reactions to reputation mechanisms.
Keser (2003) reports a number of experiments based on the “trust game.” In this game, one
player (the “buyer”) can choose to send money to a second (the “seller”). This amount is then
substantially increased and the seller can choose to share some of the gain with the buyer.
By removing many of the complexities involved in market transactions, this game provides
a simple context to study the eﬀect of diﬀerent information policies about revealing past
behaviors. Keser ﬁnds that the presence of a reputation mechanism signiﬁcantly increases
both the buyers’ level of investment (trust) as well as the fraction that sellers share with
14Book sales are inferred from the publicly available “sales rank” metric using an approach introduced by
Chevalier and Goolsbee (2003).
28

their buyers (trustworthiness). Furthermore, she ﬁnds that both trust and trustworthiness
are higher when the reputation mechanism publishes the entire history of each player’s past
behavior than when it publishes each player’s behavior in the most recent transaction only.
Questions involving reputation’s eﬀect on market eﬃciency require more complex experi-
mental scenarios. Bolton, Katok and Ockenfels (2002) study trading in a ﬁxed-price market
where buyers can choose whether to send the purchase price to the seller and sellers have
the option of not fulﬁlling their contracts. They compare transaction completion rates in
a setting with random matching of players and public histories of trader fulﬁllment (rep-
utation), to a setting with random matching without reputation, as well as to a market
where the same people interact with each other repeatedly (partners market). They ﬁnd
that, while the presence of the reputation mechanism induces a substantial improvement in
trading eﬃciency, it falls short of the eﬃciency achieved in the partners market.
Chen, Hogg, and Wozny (2004) conduct experiments similar to those performed by BKO but
provide a broader set of endogenous choices for the players. First, players can explicitly decide
who they wish to do business with rather than being paired with a single other player by the
experimenter. Second, both buyers and sellers make fulﬁllment choices and so face a moral
hazard for which reputations are relevant. Third, in addition to settings with automatically
generated reputation, CHW examine games where players self-report transaction outcomes,
allowing them to misreport their experiences as possible punishment for a poor report on
their own reputation. Fourth, prices and trading volumes are determined endogenously. The
ﬁndings of CHW are consistent with the previous experiments: the presence of reputational
information led to a substantial increase of transaction fulﬁllment. Interestingly, CHW
found fulﬁllment rates to be almost as high when traders self-reported transaction outcomes
as when reliable reputational information was automatically generated, indicating that, in
practice, private observability of outcomes might not be as big a problem as theory suggests.
6
Conclusions: Opportunities for IS research
Online reputation mechanisms harness the remarkable ability of the Web to not only dis-
seminate, but also collect and aggregate information from large communities at very low
cost, in order to artiﬁcially construct large-scale word-of-mouth networks. Best known so
far as a technology for building trust and fostering cooperation in online marketplaces, these
mechanisms are poised to have a much wider impact on organizations.
29

For example, in a marketing context, certain types of reputation mechanisms (e.g., product
discussion forums) can serve as a low-cost channel for acquiring and retaining customers,
complementary to advertising (Mayzlin, 2003). In addition, they can assist organizations
better understand consumer reactions to their current product line. At the same time, these
systems aﬀect an industry’s competitive dynamics because they reveal the same information
to all competitors. In a supply chain context, reputation mechanisms can assist organiza-
tions better assess the trustworthiness of prospective ﬁrst-time suppliers, and thus enable
more ﬂexible supply chain management. At the same time they can become a powerful dis-
ciplining mechanism that can ensure fulﬁllment of contractual obligations without recourse
to litigation; they, thus, have the potential of lowering the legal costs of doing business in a
variety of industries (Bakos and Dellarocas, 2002).
The design of online reputation mechanisms can greatly beneﬁt from the insights produced
by more than twenty years of economics and game theory research on the topic of reputation.
These results need to be extended to take into account the unique new properties of online
environments, such as their unprecedented scalability, the ability to precisely design the type
of feedback information that is solicited and distributed, and the relative ease with which
traders can change online identities. The following list contains what the author considers
to be the most important open areas of research in reputation mechanism design:
• Scope and explore the design space and limitations of online reputation mechanisms.
Understand what set of design parameters work best in what settings. Develop models
and prototype implementations of such systems.
• Develop eﬀective solutions to the problems of suﬃcient participation, easy identity
changes and strategic manipulation of online feedback.
• Conduct theory-driven experimental and empirical research that sheds more light on
buyer and seller behavior vis-à-vis such mechanisms.
• Compare the relative eﬃciency of reputation mechanisms to the eﬃciency of more
established mechanisms for dealing with moral hazard and adverse selection (such as
the legal system and advertising);15 develop theory-driven guidelines for deciding which
set of mechanisms to use when.
15See Cooter and Rubinfeld (1989) for an introduction to economic models of legal disputes. See Milgrom
and Roberts (1986) for the role of advertising in revealing a ﬁrm’s unknown quality.
30

• Understand how decision-makers must adapt their strategies to react to the presence
of such mechanisms in areas such as marketing, product development, and customer
service.
The power of online reputation mechanisms has its roots in the strategic side eﬀects brought
about by the increased interdependencies they create among ﬁrms, their customers, their
partners, and their competitors. As the mathematical study of interaction of self-interested
agents, game theory is the natural foundation for the study of online reputation mechanisms.
Other established paradigms of information systems research can play an important role
in translating the conceptual insights of game theoretic models into concrete guidelines for
building (and reacting to) large-scale reputation mechanisms that can inﬂuence the dynamics
of entire industries or societies. For example, computational methods can help analyze
games that may be too complex to solve analytically. Laboratory experiments can inform
about how people will behave when confronted with these mechanisms, both when they
are inexperienced and as they gain experience. Finally, game-theoretic models can often be
approximated by generally more tractable, decision-theoretic approaches16.
There is much work to be done. But it is important that research be conducted now, in the
formative phases of this technology and the social practices surrounding it. There are likely
to be path-dependent eﬀects in the deployment and use of online reputation mechanisms, so
it is important that researchers develop insights into the functioning and impacts of these
systems while they can still have a large impact on practice.
References
Aberer, K. and Despotovic, Z., 2001, Managing trust in a Peer-2-Peer information system,
Proc. 2001 ACM International Conf. on Information and Knowledge Management (CIKM),
310-317.
Akerlof, G., 1970, The market for lemons: Quality uncertainty and the market mechanism,
Quarterly Journal of Economics 84, 488-500.
16See Zacharia et al. (2001) for an example of using simulation modeling to study the eﬀects of reputation
mechanisms on markets with dynamic pricing. See Dellarocas (2001) and Shapiro (1982) for examples of
how the adoption of “reasonable” (common-sense or empirically-driven) assumptions about the behavior of
other players permits the use of decision-theoretic approaches to model reputation phenomena.
31

Avery, C., Resnick, P., R. Zeckhauser, 1999, The Market for Evaluations. American Eco-
nomics Review 89 (3) 564-584.
Ba, S., P. Pavlou, 2002, Evidence of the Eﬀect of Trust Building Technology in Electronic
Markets: Price Premiums and Buyer Behavior. MIS Quarterly 26 (3) 243-268.
Bajari, P., A. Hortacsu, 2003, Winner’s Curse, Reserve prices and Endogenous Entry: Em-
pirical Insights from eBay Auctions. Rand Journal of Economics 34 (2).
Bakos, Y., C. Dellarocas, 2002, Cooperation without Enforcement? A Comparative Analysis
of Litigation and Online Reputation as Quality Assurance Mechanisms, L. Applegate, R.
Galliers, J. I. DeGross, eds. Proc. 23 rd Int. Conf. on Information Systems (ICIS 2002).
Association for Information Systems, Barcelona, Spain, 127-142.
Battigalli, P., and Watson, J., 1997, On Reputation Reﬁnements with Heterogeneous Beliefs,
Econometrica 65, 369-374.
Benson, B., 1989, The Spontaneous Evolution of Commercial Law, Southern Economic Jour-
nal 55 (January) 644-661. Reprinted in D. B. Klein, ed. 1997. Reputation: Studies in the
Voluntary Elicitation of Good Conduct. University of Michigan Press, Ann Arbor, 165-189.
Bolton, G. E., Katok, E., A. Ockenfels, 2004, How Eﬀective are Online Reputation Mecha-
nisms? An Experimental Study, Management Science 50 (11), 1587-1602.
Brin S., L. Page, 1998, The Anatomy of a Large-Scale Hypertextual Web Search Engine.
Computer Networks and ISDN Systems 30 (1-7) 107-117.
Cabral, L., and Hortacsu, A. 2004, The Dynamics of Seller Reputation: Theory and Evidence
from eBay (April 2004), CEPR Discussion Paper No. 4345. http://ssrn.com/abstract=541161
Celentani, M., Fudenberg, D., Levine, D., and Pesendorfer, W., 1996, Maintaining a Repu-
tation against a Long Lived Opponent, Econometrica 64, 691-704.
Chen, K-Y, Hogg, T., and Wozny, N., 2004, Experimental study of market reputation mech-
anisms, Proc. ACM Conference on Electronic Commerce 2004, 234-235.
Cheng, A. and Friedman, E., 2005, Sybilproof Reputation Mechanisms, Proc. of the 3rd
Workshop on the Economics of Peer-2-Peer Systems (P2PECON), Philadephia, PA.
Chevalier, J. A. and Goolsbee, A., 2003, Price Competition Online: Amazon Versus Barnes
And Noble, Quantitative Marketing and Economics 1(2), 203-222.
32

Chevalier, J. A. and Mayzlin, D., 2003, The Eﬀect of Word of Mouth on Sales: Online Book
Reviews, Yale SOM Working Paper No’s. ES-28 & MK-15. http://ssrn.com/abstract=432481
Cooke, R. M., 1991, Experts in Uncertainty: Opinion and Subjective Probability in Science.
Oxford University Press, New York, NY.
Cooter, R.D. and Rubinfeld, D. L., 1989, Economic Analysis of Legal Disputes and Their
Resolution, Journal of Economic Studies 27 (3), 1067-1097.
Cripps, M., Mailath, G., L. Samuelson, 2004, Imperfect Monitoring and Impermanent Rep-
utations, Econometrica 72, 407-432.
Cripps, M., Schmidt, K., and Thomas, J., 1996, Reputation in Perturbed Repeated Games,
Journal of Economic Theory 69, 387-410.
Cripps, M., and Thomas, J., 1995, Reputation and Commitment in Two-Person Repeated
Games without Discounting, Econometrica 63, 1401-1419.
Dellarocas, C., 2000, Immunizing online reputation reporting systems against unfair rat-
ings and discriminatory behavior, Proceedings of the 2nd ACM Conference on Electronic
Commerce. Association for Computing Machinery, Minneapolis, MN, 150-157.
Dellarocas, C., 2001, Analyzing the economic eﬃciency of eBay-like online reputation re-
porting mechanisms, Proceedings of the 3rd ACM Conference on Electronic Commerce. As-
sociation for Computing Machinery, Tampa, FL, 171-179.
Dellarocas, C., 2003a, The Digitization of Word-of-Mouth: Promise and Challenges of Online
Feedback Mechanisms, Management Science 49 (10), 1407-1424.
Dellarocas, C., Fan, M., and Wood, C., 2003b, Self-Interest, Reciprocity, and Participation in
Online Reputation Systems, 2003 Workshop in Information Systems and Economics (WISE),
Seattle, WA.
Dellarocas. C., 2004a, Building Trust On-Line: The Design of Robust Reputation Mecha-
nisms for Online Trading Communities. G. Doukidis, N. Mylonopoulos, N. Pouloudi, eds.
Social and Economic Transformation in the Digital Era. Idea Group Publishing, Hershey,
PA.
Dellarocas C., 2004b, Strategic Manipulation of Internet Opinion Forums: Implications for
Consumers and Firms, MIT Sloan Working Papers No. 4501-04. http://ssrn.com/abstract=585404
33

Dellarocas C., 2005a, Reputation Mechanism Design in Online Trading Environments with
Pure Moral Hazard, Information Systems Research 16 (2), 209-230.
Dellarocas C., 2005b, How Often Should Reputation Mechanisms Update a Trader’s Repu-
tation Proﬁle? Working Paper.
Despotovic, Z. and Aberer, K., 2004, Possibilities for Managing Trust in P2P Networks. Swiss
Federal Institute of Technology (EPFL) Technical Report IC/2004/84, Lausanne, Switzer-
land.
Dewatripont, M. Jewitt, I. and Tirole, J., 1999, The Economics of Career Concerns, Part I:
Comparing Information Structures, Review of Economic Studies 66 (1), 183-198.
Dichter, E., 1966, How Word-of-Mouth Advertising Works, Harvard Business Review 44
(Nov-Dec), 147-166.
Dewan, S. and Hsu, V., 2004, Adverse Selection in Reputations-Based Electronic Markets:
Evidence from Online Stamp Auctions, Journal of Industrial Economics 52 (4), 497-516.
Diamond, D., 1989, Reputation Acquisition in Debt Markets, Journal of Political Economy
97 (4) 828-862.
Douceur, J. 2002, The sybil attack, Proc. 1st International Workshop on Peer-to-Peer Sys-
tems (IPTPS ’02), Cambridge, MA.
Dranove, D., Kessler, D., McClellan, M., and Satterthwaite, M., 2003, Is More Informa-
tion Better? The Eﬀects of “Report Cards” on Health Care Providers, Journal of Political
Economy 111 (3), 555-588.
Eaton, D. H., 2002, Valuing Information: Evidence from Guitar Auctions on eBay, Depart-
ment of Economics Working Paper #0201, Murray State University, Murray, KY.
Ely, J. C., Fudenberg, D. and Levine, D. K., 2005, When is Reputation Bad? Working
Paper.
Ely, J. C. and Valimaki, J., 2003, Bad Reputation, Quarterly Journal of Economics 118 (3),
785-814.
Fan, M., Tan, Y., and Whinston, A. B., 2005, Evaluation and Design of Online Coopera-
tive Feedback Mechanisms for Reputation Management, IEEE Trans. Knowledge and Data
Engineering 17 (3), 244-254.
34

Friedman, E., P. Resnick, 2001, The Social Cost of Cheap Pseudonyms, Journal of Economics
and Management Strategy 10 (1), 173-199.
Fudenberg, D., D. Levine, 1992, Maintaining a Reputation when Strategies are Imperfectly
Observed, Review of Economic Studies 59 (3) 561-579.
Greif, A., 1993, Contract Enforceability and Economic Institutions in Early Trade: The
Maghribi Traders’ Coalition, American Economic Review 83 (June) 525-548.
Hennig-Thurau T., Gwinner K.P., Walsh G., Gremler D.D., 2004, Electronic Word-of-Mouth
via Consumer-Opinion Platforms: What Motivates Consumers to Articulate Themselves on
the Internet? Journal of Interactive Marketing 18 (1), 38-52.
Holmstrom, B., 1999, Managerial Incentive Problems: A Dynamic Perspective, Review of
Economic Studies 66 (1) 169-182.
Houser, D., J. Wooders, 2005, Reputation in Auctions: Theory and Evidence from eBay.
Forthcoming Journal of Economics and Management Strategy.
Jin, G. Z., and Kato, A., 2004, Blind Trust Online: Experimental Evidence from Baseball
Cards, University of Maryland Working Paper.
Jurca, R. and Faltings, B., 2004, CONFESS: Eliciting Honest Feedback without Indepen-
dent Veriﬁcation Authorities. Sixth International Workshop on Agent Mediated Electronic
Commerce (AMEC VI 2004), New York, USA, July 19 2004.
Kalyanam, K., S. McIntyre, 2001, Returns to Reputation in Online Auction markets. Retail
Workbench Working Paper W-RW01-02, Santa Clara University, Santa Clara, CA.
Kamvar, S. D., Schlosser, M. T. and Garcia-Molina, H., 2003, The EigenTrust Algorithm
for Reputation Management in P2P Networks, Proc. of the 12th Infernational World Wide
Web Conference, 640-651.
Kandori, M., 2002, Introduction to Repeated Games with Private Monitoring, Journal of
Economic Theory 102 (1), 1-15.
Kauﬀman, R. J., C. Wood, 2000, Running up the Bid: Modeling Seller Opportunism in
Internet Auctions. Proc. Americas Conference on Information Systems (AMCIS 2000),
Association for Information Systems, Long Beach, CA, 929-935.
35

Keser, C., 2003, Experimental Games for the Design of Reputation Management Systems,
IBM Systems Journal 42 (3), 498-506.
Klein, B., K. Leﬄer, 1981, The Role of Market Forces in Assuring Contractual Performance,
Journal of Political Economy 89 (4) 615-641.
Klein, D., ed. 1997, Reputation: Studies in the Voluntary Elicitation of Good Conduct,
University of Michigan Press, Ann Arbor, MI.
Kollock, P., 1999, The Production of Trust in Online Markets. E.J. Lawler, M. Macy,
S. Thyne, and H.A. Walker, eds. Advances in Group Processes (Vol. 16). JAI Press.,
Greenwich, CT.
Kreps, D., Milgrom, P., Roberts, J., R. Wilson, 1982, Rational Cooperation in the Finitely
Repeated Prisoners’ Dilemma, Journal of Economic Theory 27 (2) 245-252.
Kreps, D., R. Wilson, 1982, Reputation and Imperfect Information, Journal of Economic
Theory 27 (2) 253-279.
Lee, Z., Im, I., S. J. Lee, 2000, The Eﬀect of Negative Buyer Feedback on Prices in Internet
Auction Markets. W. J. Orlikowski, S. Ang, P. Weill, H. C. Krcmar, J. I. DeGross, eds.
Proc. 21st Int. Conf. on Information Systems (ICIS 2000), Association for Information
Systems., Brisbane, Australia, 286-287.
Livingston, J., 2002, How Valuable is a Good Reputation? A Sample Selection Model of
Internet Auctions. Working Paper, University of Maryland, College Park, MD.
Lucking-Reiley, D., D. Bryan, et al., 2000, Pennies from eBay: the Determinants of Price in
Online Auctions, Working Paper, University of Arizona, Tucson, AZ.
Mailath, G. J., L. Samuelson, 2001, Who Wants a Good Reputation? Review of Economic
Studies 68 (2) 415-441.
Mayzlin, D., 2003, Promotional Chat on the Internet, Working Paper #MK-14, Yale School
of Management, New Haven, CT.
McDonald, C. G., V. C. Slawson, Jr., 2002, Reputation in an Internet Auction Market,
Economic Inquiry 40 (3) 633-650.
Melnik, M. I., J. Alm, 2002, Does a Seller’s Reputation Matter? Evidence from eBay Auc-
tions, Journal of Industrial Economics 50 (3) 337-349.
36

Milgrom, P. R., North, D., B. R. Weingast, 1990, The Role of Institutions in the Revival
of Trade: The Law Merchant, Private Judges, and the Champagne Fairs, Economics and
Politics 2 (1) 1-23. Reprinted in D. B. Klein, ed. 1997. Reputation: Studies in the Voluntary
Elicitation of Good Conduct. University of Michigan Press, Ann Arbor, 243-266.
Milgrom, P., J. Roberts, 1982, Predation, Reputation and Entry Deterrence, Journal of
Economic Theory 27 (2) 280–312.
Milgrom, P. and Roberts, J., 1986, Price and Advertising Signals of Product Quality, Journal
of Political Economy 94 (4), 796-821.
Miller, N., Resnick, P., R. Zeckhauser, 2005, Eliciting Honest Feedback: The Peer-Prediction
Method, forthcoming Management Science.
Oram, A. (ed.), 2001, Peer-to-Peer: Harnessing the Power of Disruptive Technologies,
O’Reilly Media, Sebastopol, CA.
Papaioannou, T. and Stamoulis, G. D., 2005, An Incentives’ Mechanism Promoting Truthful
Feedback in Peer-to-Peer Systems, Proc. of IEEE/ACM CCGRID 2005 (Workshop on Global
P2P Computing), May 2005.
Pujol, J. M., Sanguesa, R., J. Delgado, 2002, Extracting reputation in multi agent systems by
means of social network topology, Proc. of the First Int. Joint Conference on Autonomous
Agents and Multiagent Systems, Association for Computing Machinery, Bologna, Italy, 467-
474.
Resnick, P., Zeckhauser, R., Friedman, E., Kuwabara, K., 2000, Reputation Systems, Com-
munications of the ACM 43 (12) 45-48.
Resnick, P., R. Zeckhauser, 2002, Trust Among Strangers in Internet Transactions: Empirical
Analysis of eBay’s Reputation System. Michael R. Baye, ed. The Economics of the Internet
and E-Commerce (Advances in Applied Microeconomics, Vol. 11). JAI Press.
Resnick, P., Zeckhauser, R., Swanson, J., K. Lockwood, 2005, The Value of Reputation on
eBay: A Controlled Experiment, forthcoming Experimental Economics.
Sabater, J., C. Sierra, 2002, Reputation and social network analysis in multi-agent systems,
Proc. of the First Int. Joint Conference on Autonomous Agents and Multiagent Systems,
Association for Computing Machinery, Bologna, Italy, 475-482.
37

Shapiro, C., 1982, Consumer Information, Product Quality, and Seller Reputation, Bell
Journal of Economics 13 (1) 20-35.
Shapiro, C., 1983, Premiums for High Quality Products as Returns to Reputations, Quarterly
Journal of Economics 98 (4) 659-680.
Sundaram, D.S., Mitra, K., and Webster, C., 1998, Word-of-Mouth Communications: A
Motivational Analysis, Advances in Consumer Research 25, 527-531.
Tadelis, S., 1999, What’s in a Name? Reputation as a Tradeable Asset, The American
Economic Review 89 (3) 548-563.
Tadelis, S., 2002, The Market for Reputations as an Incentive Mechanism, Journal of Political
Economy 92 (2) 854-882
Watson, J., 1993, A Reputation Reﬁnement without Equilibrium, Econometrica 61, 199-205.
Watson, J., 1996, Reputation in Repeated Games with No Discounting, Games and Economic
Behavior 15, 82-109.
Wilson, R., 1985, Reputations in Games and Markets. A. Roth, ed. Game-Theoretic Models
of Bargaining, Cambridge University Press, Cambridge, UK, 27-62.
Xiong, L., and L. Liu, 2004, PeerTrust: Supporting Reputation-Based Trust in Peer-to-Peer
Communities, IEEE Transactions on Knowledge and Data Engineering 16 (7), 843-857.
Yu, B. and M. Singh, 2002, An evidential model of distributed reputation management,
Proc. of the First Int. Joint Conference on Autonomous Agents and Multiagent Systems,
Association for Computing Machinery, Bologna, Italy,294-301.
Zacharia, G., Moukas, A., P. Maes, 2000, Collaborative Reputation Mechanisms in Electronic
Marketplaces, Decision Support Systems 29 (4) 371-388.
Zacharia, G., Evgeniou, T., Moukas, A., Boufounos, P., P. Maes, 2001, Economics of Dy-
namic Pricing in a Reputation Brokered Agent Mediated Marketplace, Electronic Commerce
Research 1 (1) 85-100.
38