Promotional Reviews: An Empirical Investigation of
Online Review Manipulation
By Dina Mayzlin, Yaniv Dover, Judith Chevalier
∗
Firms’ incentives to manufacture biased user reviews impede re-
view usefulness. We examine the differences in reviews for a given
hotel between two sites: Expedia.com (only a customer can post a
review) and TripAdvisor.com (anyone can post). We argue that the
net gains from promotional reviewing are highest for independent
hotels with single-unit owners and lowest for branded chain hotels
with multi-unit owners. We demonstrate that the hotel neighbors
of hotels with a high incentive to fake have more negative reviews
on TripAdvisor relative to Expedia; hotels with a high incentive to
fake have more positive reviews on TripAdvisor relative to Expe-
dia.
User-generated online reviews have become an important resource for con-
sumers making purchase decisions; an extensive and growing literature documents
the influence of online user reviews on the quantity and price of transactions.
1
In theory, online reviews should create producer and consumer surplus by im-
proving the ability of consumers to evaluate unobservable product quality. How-
ever, one important impediment to the usefulness of reviews in revealing product
quality is the possible existence of fake or “promotional” online reviews. Specif-
ically, reviewers with a material interest in consumers’ purchase decisions may
post reviews that are designed to influence consumers and to resemble the re-
views of disinterested consumers. While there is a substantial economic literature
on persuasion and advertising (reviewed below), the specific context of advertising
disguised as user reviews has not been extensively studied.
The presence of undetectable (or difficult to detect) fake reviews may have at
least two deleterious effects on consumer and producer surplus. First, consumers
who are fooled by the promotional reviews may make suboptimal choices. Second,
∗
Mayzlin: Marshall School of Business, University of Southern California, 3670 Trousdale Parkway,
Los Angeles, CA 90089; Dover: Tuck School of Business at Dartmouth, 100 Tuck Hall, Hanover, NH
03755; Chevalier: Yale School of Management, 135 Prospect St, New Haven, CT 06511. The authors
contributed equally, and their names are listed in reverse alphabetical order. We thank the Wharton
Interactive Media Initiative, the Yale Whitebox Center, and the National Science Foundation (Chevalier)
for providing financial support for this project (grant 1128322). We thank Steve Hood, Sr. Vice President
of Research at STR for helping us with data collection. We also thank David Godes and Avi Goldfarb for
detailed comments on the paper. We also thank numerous seminar participants for helpful comments.
All errors remain our own.
1
Much of the earliest work focused on the effect of eBay reputation feedback scores on prices and
quantity sold; for example, Resnick and Zeckhauser (2002), Melnik and Alm (2002), and Resnick et al.
(2006). Later work examined the role of consumer reviews on product purchases online; for example,
Chevalier and Mayzlin (2006), Anderson and Magruder (2012), Berger, Sorensen and Rasmussen (2010),
and Chintagunta, Gopinath and Venkataraman (2010).
1
2 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
the potential presence of biased reviews may lead consumers to mistrust reviews.
This in turn forces consumers to disregard or underweight helpful information
posted by disinterested reviewers. For these reasons, the Federal Trade Commis-
sion in the United States recently updated its guidelines governing endorsements
and testimonials to also include online reviews. According to the guidelines, a
user must disclose the existence of any material connection between himself and
the manufacturer.
2
Relatedly, in February 2012, the UK Advertising Standards
Authority ruled that travel review website TripAdvisor must cease claiming that
it offers “honest, real, or trusted” reviews from “real travelers.” The Advertising
Standards Authority, in its decision, held that TripAdvisor’s claims implied that
“consumers could be assured that all review content on the TripAdvisor site was
genuine, and when we understood that might not be the case, we concluded that
the claims were misleading.”
3
In order to examine the potential importance of these issues, we undertake an
empirical analysis of the extent to which promotional reviewing activity occurs,
and the firm characteristics and market conditions that result in an increase or
decrease in promotional reviewing activity. The first challenge to any such exercise
is that detecting promotional reviews is difficult. After all, promotional reviews
are designed to mimic unbiased reviews. For example, inferring that a review
is fake because it conveys an extreme opinion is flawed; as shown in previous
literature (see Li and Hitt 2008, Dellarocas and Wood 2007), individuals who had
an extremely positive or negative experience with a product may be particularly
inclined to post reviews. In this paper, we do not attempt to classify whether
any particular review is fake, and instead we empirically exploit a key difference
in website business models. In particular, some websites accept reviews from
anyone who chooses to post a review while other websites only allow reviews
to be posted by consumers who have actually purchased a product through the
website (or treat “unverified” reviews differently from those posted by verified
buyers). If posting a review requires making an actual purchase, the cost of
posting disingenuous reviews is greatly increased. We examine differences in the
distribution of reviews for a given product between a website where faking is
difficult and a website where faking is relatively easy.
Specifically, in this paper we examine hotel reviews, exploiting the organiza-
tional differences between Expedia.com and TripAdvisor.com. TripAdvisor is a
popular website that collects and publishes consumer reviews of hotels, restau-
rants, attractions and other travel-related services. Anyone can post a review
on TripAdvisor. Expedia.com is a website through which travel is booked; con-
2
The guidelines provide the following example, “An online message board designated for discussions
of new music download technology is frequented by MP3 player enthusiasts...Unbeknownst to the message
board community, an employee of a leading playback device manufacturer has been posting messages
on the discussion board promoting the manufacturer’s product. Knowledge of this poster’s employment
likely would affect the weight or credibility of her endorsement. Therefore, the poster should clearly
and conspicuously disclose her relationship to the manufacturer to members and readers of the message
board” (http://www.ftc.gov/os/2009/10/091005endorsementguidesfnnotice.pdf)
3
www.asa.org/ASA-action/Adjudications.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 3
sumers are also encouraged to post reviews on the site, but a consumer can only
post a review if she actually booked at least one night at the hotel through the
website in the six months prior to the review post. Thus, the cost of posting a
fake review on Expedia.com is quite high relative to the cost of posting a fake
review on TripAdvisor. Purchasing a hotel night through Expedia requires the
reviewer to undertake a credit card transaction on Expedia.com. Thus, the re-
viewer is not anonymous to the website host, potentially raising the probability
of detection of any fakery.
4
We also explore the robustness of our results using
data from Orbitz.com, where reviews can be either “verified” or “unverified.”
We present a simple analytical model in the Appendix that examines the equi-
librium levels of manipulation of two horizontally-differentiated competitors who
are trying to convince a consumer to purchase their product. The model demon-
strates that the cost of review manipulation (which we relate to reputational risk)
determines the amount of manipulation in equilibrium. We marry the insights
from this model to the literature on organizational form and organizational in-
centive structures. Based on the model as well as on the previous literature we
examine the following hypotheses: 1) hotels with a neighbor are more likely to
receive negative fake reviews than more isolated hotels, 2) small owners are more
likely to engage in review manipulation than hotels owned by companies that
own many hotel units, 3) independent hotels are more likely to engage in review
manipulation (post more fake positive reviews for themselves and more fake neg-
ative reviews for their competitors) than branded chain hotels, and 4) hotels with
a small management company are more likely to engage in review manipulation
than hotels that use a large management company.
Our main empirical analysis is akin to a differences in differences approach
(although, unconventionally, neither of the differences is in the time dimension).
Specifically, we examine differences in the reviews posted at TripAdvisor and Ex-
pedia for different types of hotels. For example, consider calculating for each
hotel at each website the ratio of one- and two-star (the lowest) reviews to total
reviews. We ask whether the difference in this ratio for TripAdvisor vs. Expedia
is higher for hotels with a neighbor within a half kilometer vs. hotels without a
neighbor. Either difference alone would be problematic. TripAdvisor and Expe-
dia reviews could differ due to differing populations at the site. Possibly, hotels
with and without neighbors could have different distributions of true quality.
However, our approach isolates whether the two hotel types’ reviewing patterns
are significantly different across the two sites. Similarly, we examine the ratio of
one- and two-star reviews to total reviews for TripAdvisor vs. Expedia for hotels
that are close geographic neighbors of hotels with small owners vs. large owners,
close neighbors of independent hotels vs. chain-affiliated hotels, and neighbors of
hotels with large management companies versus small management companies.
4
As discussed above, TripAdvisor has been criticized for not managing the fraudulent reviewing
problem. TripAdvisor recently announced the appointment of a new Director of Content Integrity. Even
in the presence of substantial content verification activity on TripAdvisor’s part, our study design takes
as a starting point the higher potential for fraud in TripAdvisor’s business model relative to Expedia.
4 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
That is, we measure whether the neighbor of hotels with small owners fare worse
on TripAdvisor than on Expedia, for example, than the neighbors of hotels owned
by large multi-unit entities. We also measure the ratio of five-star (the highest)
reviews to total reviews for TripAdvisor vs. Expedia for independent vs. chain
hotels, hotels with small owners vs. large owners, and hotels with large man-
agement companies versus small management companies. Thus, our empirical
exercise is a joint test of the hypotheses that promotional reviewing take place
on TripAdvisor and that the incentive to post false reviews is a function of or-
ganizational form. Our identifying assumption is that TripAdvisor and Expedia
users do not differentially value hotel ownership and affiliation characteristics and
the ownership and affiliation characteristics of neighbors. In our specifications,
we control for a large number of hotel observable characteristics that could be
perceived differently by TripAdvisor and Expedia consumers. We discuss robust-
ness to selection on unobservables that may be correlated with ownership and
affiliation characteristics.
The results are largely consistent with our hypotheses. That is, we find that
the presence of a neighbor, neighbor characteristics (such as ownership, affiliation
and management structure), and own hotel characteristics affect the measures of
review manipulation. The mean hotel in our sample has a total of 120 reviews on
TripAdvisor, of which 37 are 5-star. We estimate that an independent hotel owned
by a small owner will generate an incremental 7 more fake positive Tripadvisor
reviews than a chain hotel with a large owner. The mean hotel in our sample has
thirty 1- and 2-star reviews on TripAdvisor. Our estimates suggest that a hotel
that is located next to an independent hotel owned by a small owner will have 6
more fake negative Tripadvisor reviews compared to an isolated hotel.
The paper proceeds as follows. In Section I we discuss the prior literature.
In Section II we describe the data and present summary statistics. In Section
III we discuss the theoretical relationship between ownership structure and the
incentive to manipulate reviews. In Section IV we present our methodology and
results, which includes main results as well as robustness checks. In Section V we
conclude and also discuss limitations of the paper.
I. Prior Literature
Broadly speaking, our paper is informed by the literature on the firm’s strategic
communication, which includes research on advertising and persuasion. In adver-
tising models, the sender is the firm, and the receiver is the consumer who tries
to learn about the product’s quality before making a purchase decision. In these
models the firm signals the quality of its product through the amount of resources
invested into advertising (see Nelson 1974, Milgrom and Roberts 1986, Kihlstrom
and Riordan 1984, Bagwell and Ramey 1994, Horstmann and Moorthy 2003) or
the advertising content (Anand and Shachar 2009, Anderson and Renault 2006,
Mayzlin and Shin 2011). In models of persuasion, an information sender can
influence the receiver’s decision by optimally choosing the information structure
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 5
(Crawford and Sobel 1982, Chakraborty and Harbaugh 2010, and Dziuda 2011
show this in the case where the sender has private information, while Kamenica
and Gentzkow 2011 show this result in the case of symmetric information). One
common thread between all these papers is that the sender’s identity and incen-
tives are common-knowledge. That is, the receiver knows that the message is
coming from a biased party, and hence is able to to take that into account when
making her decision. In contrast, in our paper there is uncertainty surrounding
the sender’s true identity and incentives. That is, the consumer who reads a user
review on TripAdvisor does not know if the review was written by an unbiased
customer or by a biased source.
The models that are most closely related to the current research are Mayzlin
(2006) and Dellarocas (2006). Mayzlin (2006) presents a model of “promotional”
chat where competing firms, as well as unbiased informed consumers, post mes-
sages about product quality online. Consumers are not able to distinguish be-
tween unbiased and biased word of mouth, and try to infer product quality based
on online word of mouth. Mayzlin (2006) derives conditions under which online
reviews are persuasive in equilibrium: online word of mouth influences consumer
choice. She also demonstrates that producers of lower quality products will ex-
pend more resources on promotional reviews. Compared to a system with no
firm manipulation, promotional chat results in welfare loss due to distortions in
consumer choices that arise due to manipulation. The welfare loss from promo-
tional chat is lower the higher the participation by unbiased consumers in online
fora. Dellarocas (2006) also examines the same issue. He finds that there ex-
ists an equilibrium where the high quality product invests more resources into
review manipulation, which implies that promotional chat results in welfare in-
crease for the consumer. Dellarocas (2006) additionally notes that the social cost
of online manipulation can be reduced by developing technologies that increase
the unit cost of manipulation and that encourage higher participation by honest
consumers.
The potential for biased reviews to affect consumer responses to user reviews has
been recognized in the popular press. Perhaps the most intuitive form of biased
review is the situation in which a producer posts positive reviews for its own prod-
uct. In a well-documented incident, in February 2004, an error at Amazon.com’s
Canadian site caused Amazon to mistakenly reveal book reviewer identities. It
was apparent that a number of these reviews were written by the books’ own
publishers and authors (see Harmon 2004).
5
Other forms of biased reviews are
also possible. For example, rival firms may benefit from posting negative reviews
of each other’s products. In assessing the potential reward for such activity, it is
important to assess whether products are indeed sufficient substitutes to benefit
from negative reviewing activity. For example, Chevalier and Mayzlin (2006) ar-
5
Similarly, in 2009 in New York, the cosmetic surgery company Lifestyle Lift agreed to pay $300,000
to settle claims regarding fake online reviews about itself. In addition, a web site called fiverr.com which
hosts posts by users advertising services for $5 (e.g.: “I will drop off your dry-cleaning for $5”) hosts a
number of ads by people offering to write positive or negative hotel reviews for $5.
6 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
gue that two books on the same subject may well be complements, rather than
substitutes, and thus, it is not at all clear that disingenuous negative reviews for
other firm’s products would be helpful in the book market. Consistent with this
argument, Chevalier and Mayzlin (2006) find that consumer purchasing behavior
responds less intensively to positive reviews (which consumers may estimate are
more frequently fake) than to negative reviews (which consumers may assess to be
more frequently unbiased). However, there are certainly other situations in which
two products are strong substitutes; for example, in this paper, we hypothesize
that two hotels in the same location are generally substitutes.
6
A burgeoning computer science literature has attempted to empirically exam-
ine the issue of fakery by creating textual analysis algorithms to detect fakery.
For example, Ott et al. (2011) create an algorithm to identify fake reviews. The
researchers hired individuals on the Amazon Mechanical Turk site to write per-
suasive fake hotel reviews. They then analyzed the differences between the fake
5-star reviews and “truthful” 5-star reviews on TripAdvisor to calibrate their
psycholinguistic analysis. They found a number of reliable differences in the lan-
guage patterns of the fake reviews. One concern with this approach is that it
is possible that the markers of fakery that the researchers identify are not rep-
resentative of differently-authored fake reviews. For example, the authors find
that truthful reviews are more specific about “spatial configurations” than are
the fake reviews. However, the authors specifically hired fakers who had not vis-
ited the hotel. We can not, of course, infer from this finding that fake reviews
on TripAdvisor authored by a hotel employee would in fact be less specific about
“spatial configurations” than true reviews. Since we are concerned with fake re-
viewers with an economic incentive to mimic truthful reviewers, it is an ongoing
challenge for textual analysis methodologies to provide durable mechanisms for
detecting fake reviews.
7
Some other examples of papers that use textual analy-
sis to determine review fakery are Jindal and Liu (2007), Hu et al. (2012), and
Mukherjee, Liu and Glance (2012).
Kornish (2009) uses a different approach to detect review manipulation. She
looks for evidence of “double voting” in user reviews. That is, one strategy for
review manipulation is to post a fake positive review for one’s product and to
vote this review as “helpful.” That is, Kornish (2009) uses a correlation between
review sentiment and usefulness votes as an indicator of manipulation. This
approach isolates one possible type of review manipulation and is vulnerable to
the critique that there may be other (innocent) reasons for a correlation between
review sentiment and usefulness votes: if most people who visit a product’s page
are positively inclined towards the product, the positive reviews may be on average
6
In theory, a similar logic applies to the potential for biased reviews of complementary products
(although this possibility has not, to our knowledge, been discussed in the literature). For example, the
owner of a breakfast restaurant located next door to a hotel might gain from posting a disingenuous
positive review of the hotel.
7
One can think of the issue here as being similar to the familiar “arms race” between spammers and
spam filters.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 7
considered to be more useful.
Previous literature has not examined the extent to which the design of web-
sites that publish consumer reviews can discourage or encourage manipulation.
In this paper, we exploit those differences in design by examining Expedia versus
TripAdvisor. The literature also has not empirically tested whether manipula-
tion is more pronounced in empirical settings where it will be more beneficial to
the producer. Using data on organizational form, quality, and competition, we
examine the relationship between online manipulation and market factors which
may increase or decrease the incentive to engage in online manipulation. We will
detail our methodology below; however, it is important to understand that our
methodology does not rely on identifying any particular review as unbiased (real)
or promotional (fake).
Of course, for review manipulation to make economic sense, online reviews must
play a role in consumer decision-making. Substantial previous research establishes
that online reviews affect consumer purchase behavior (see, for example, Chevalier
and Mayzlin 2006, Luca 2012). There is less evidence specific to the travel context.
Vermeulen and Seegers (2009) measure the impact of online hotel reviews on
consumer decision-making in an experimental setting with 168 subjects. They
show that online reviews increase consumer awareness of lesser-known hotels and
positive reviews improve attitudes towards hotels. Similarly, Ye et al. (2010) use
data from a major online travel agency in China to demonstrate a correlation
between traveler reviews and online sales.
II. Data
User generated Internet content has been particularly important in the travel
sector. In particular, TripAdvisor-branded websites have more than 50 million
unique monthly visitors and contain over 60 million reviews. While our study
uses the US site, TripAdvisor branded sites operate in 30 countries. As Scott and
Orlikowski (2012) point out, by comparison, the travel publisher Frommer’s sells
about 2.5 million travel guidebooks each year. While TripAdvisor is primarily
a review site, transactions-based sites such as Expedia and Orbitz also contain
reviews.
Our data derive from multiple sources. First, we identified the 25th to 75th
largest US cities (by population) to include in our sample. Our goal was to use
cities that were large enough to “fit” many hotels, but not so large and dense
that competition patterns among the hotels would be difficult to determine.
8
In
October of 2011, we “scraped” data on all hotels in these cities from TripAdvisor
and Expedia. TripAdvisor and Expedia were co-owned at the time of our data
collection activities but maintained separate databases of customer reviews at the
8
We dropped Las Vegas, as these hotels tend to have an extremely large number of reviews at both
sites relative to hotels in other cities; these reviews are often focused on the characteristics of the casino
rather than the hotel. Many reviewers may legitimately, then, have views about a characteristic of the
hotel without ever having stayed at the hotel.
8 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
two sites. As of December 2011, TripAdvisor derived 35 percent of its revenues
from click-through advertising sold to Expedia.
9
Thus, 35 percent of TripAd-
visor’s revenue derived from customers who visited Expedia’s site immediately
following their visit to the TripAdvisor site.
Some hotels are not listed on both sites, and some hotels do not have reviews on
one of the sites (typically, Expedia). At each site, we obtained the text and star
values of all user reviews, the identity of the reviewer (as displayed by the site),
and the date of the review. We also obtained data from Smith Travel Research, a
market research firm that provides data to the hotel industry (www.str.com). To
match the data from STR to our Expedia and TripAdvisor data, we use name and
address matching. Our data consist of 2931 hotels matched between TripAdvisor,
Expedia, and STR with reviews on both sites. Our biggest hotel city is Atlanta
with 160 properties, and our smallest is Toledo, with 10 properties.
Table 1 provides summary statistics for review characteristics, using hotels as
the unit of observation, for the set of hotels that have reviews on both sites.
Unsurprisingly, given the lack of posting restrictions, there are more reviews on
TripAdvisor than on Expedia. On average, our hotels have nearly three times the
number of reviews on TripAdvisor as on Expedia. Also, the summary statistics
reveal that on average, TripAdvisor reviewers are more critical than Expedia re-
views. The average TripAdvisor star rating is 3.52 versus 3.95 for Expedia. Based
on these summary statistics, it appears that hotel reviewers are more critical than
reviewers in other previously studied contexts. For example, numerous studies
document that eBay feedback is overwhelmingly positive. Similarly, Chevalier
and Mayzlin (2006) report average reviews of 4.14 out of 5 at Amazon and 4.45
at barnesandnoble.com for a sample of 2387 books.
Review characteristics are similar if we use reviews, rather than hotels as the
unit of observation. Our data set consists of 350,485 TripAdvisor reviews and
123,569 Expedia reviews. Of all reviews, 8.0% of TripAdvisor reviews are 1s,
8.4% are 2s, and 38.1% are 5s. For Expedia, 4.7% of all review are 1s, 6.4%
are 2s, and 48.5% of all reviews are 5s. Note that these numbers differ from the
numbers in Table 1 because hotels with more reviews tend to have better reviews.
Thus, the share of all reviews that are 1s or 2s is lower than the mean share of
1-star reviews or 2-star reviews for hotels. Since the modal review on TripAdvisor
is a 4-star review, in most of our analyses we consider “negative” reviews to be
1- or 2-star reviews.
We use STR to obtain the hotel location; we assign each hotel a latitude and
longitude designator and use these to calculate distances between hotels of var-
ious types. These locations are used to determine whether or not a hotel has a
neighbor.
Importantly, we use STR data to construct the various measures of organiza-
tional form that we use for each hotel in the data set. We consider the ownership,
9
Based on information in S-4 form filed by Tripadvisor and Expedia with SEC on July 27, 2011 (see
http://ir.tripadvisor.com/secfiling.cfm?filingID=1193125-11-199029&CIK=1526520)
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 9
Table 1—User Reviews at TripAdvisor and Expedia
Mean Standard
deviation
Minimum Maximum
Number of TripAdvisor reviews 119.58 172.37 1 1675
Number of Expedia reviews 42.16 63.24 1 906
Average TripAdvisor star rating 3.52 0.75 1 5
Average Expedia star rating 3.95 0.74 1 5
Share of TripAdvisor 1-star reviews 0.14
Share of TripAdvisor 2-star reviews 0.11
Share of Expedia 1-star reviews 0.07
Share of Expedia 2-star reviews 0.08
Share of TripAdvisor 5-star reviews 0.31
Share of Expedia 5-star reviews 0.44
Total number of hotels 2931
Note: The table reports summary statistics for user reviews for 2931 hotels with reviews at both Tri-
pAdvisor and Expedia collected in October of 2011.
affiliation, and management of a hotel. A hotel’s affiliation is the most observable
attribute of a hotel to a consumer. Specifically, a hotel can have no affiliation
(“an independent”) or it can be a unit of a branded chain. In our data, 17%
of hotels do not have an affiliation. The top 5 parent companies of branded
chain hotels in our sample are: Marriott, Hilton, Choice Hotels, Intercontinental,
and Best Western. However, an important feature of hotels is that affiliation is
very distinct from ownership. A chain hotel unit can be a franchised unit or a
company-owned unit. In general, franchising is the primary organizational form
for the largest hotel chains in the US. For example, International Hotel Group
(Holiday Inn) and Choice Hotels are made up of more than 99% franchised units.
Within the broad category of franchised units, there is a wide variety of organi-
zational forms. STR provides us with information about each hotel’s owner. The
hotel owner (franchisee) can be an individual owner-operator or a large company.
For example, Archon Hospitality owns 41 hotels in our focus cities. In Memphis,
Archon owns two Hampton Inns (an economy brand of Hilton), a Hyatt, and
a Fairfield Inn (an economy brand of Marriott). Typically, the individual hotel
owner (franchisee) is the residual claimant for the hotel’s profits, although the
franchise contract generally requires the owner to pay a share of revenues to the
parent brand. Furthermore, while independent hotels do not have a parent brand,
they are in some cases operated by large multi-unit owners. In our sample, 16%
of independent hotels and 34% of branded chain hotels are owned by a multi-unit
owners. Thus affiliation and ownership are distinct attributes of a hotel.
10 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
Owners often, though not always, subcontract day to day management of the
hotel to a management company. Typically, the management company charges 3
to 5 percent of revenue for this service, although agreements which involve some
sharing of gross operating profits have become more common in recent years.
10
In some cases, the parent brand operates a management company. For example,
Marriott provides management services for approximately half of the hotels not
owned by Marriott but operated under the Marriott nameplate. Like owners,
management companies can manage multiple hotels under different nameplates.
For example, Crossroads Hospitality manages 29 properties in our data set. In At-
lanta, they manage a Hyatt, a Residence Inn (Marriott’s longer term stay brand),
a Doubletree, and a Hampton Inn (both Hilton brands). While a consumer can
clearly observe a hotel’s affiliation, the ownership and management structure of
the hotel are more difficult to infer for the consumer.
In constructing variables, we focus both on the characteristics of a hotel and
characteristics of the hotel’s neighbors. The first nine rows in Table 2 provides
summary measures of the hotel’s own characteristics. We construct dummies for
whether a hotel’s affiliation is independent (vs. part of a branded chain). We also
construct a dummy for whether the hotel has a multi-unit owner. For example,
chain-affiliated hotels that are not owned by a franchisee but owned by the parent
chain will be characterized as owned by a multi-unit ownership entity, but so will
hotels that are owned by a large multi-unit franchisee. In our data, the modal
hotel is a chain member, but operated by a small owner. For some specifications,
we will also include a dummy variable that takes the value of one if the hotel is
operated by a large multi-unit management company. This is the case for 35% of
independent hotels and for 55% of branded chain hotels in our data.
We then characterize the neighbors of the hotels in our data. The summary
statistics for these measures are given in the bottom four rows in Table 2. That is,
for each hotel in our data, we first construct a dummy variable that takes the value
of one if that hotel has a neighbor hotel within 0.5km. As the summary statistics
show, 76% of the hotels in our data have a neighbor. We next construct a dummy
that takes the value of one if a hotel has a neighbor hotel that is an independent.
Obviously, this set of ones is a subset of the previous measure; 31% of all of the
hotels in our data have an independent neighbor. We also construct a dummy
for whether the hotel has a neighbor that is owned by a multi-unit owner. In our
data 49% of the hotels have a neighbor owned by a multi-unit owner company.
For some specifications, we also examine the management structure of neighbor
hotels. We construct a variable that takes the value of one if a hotel has a neighbor
hotel operated by a multi-unit management entity, which is the case for 59% of
hotels in our sample.
In our specifications, we will be measuring the difference between a hotel’s re-
views on TripAdvisor and Expedia. The explanatory variables of interest are the
neighbor characteristics, the ownership and affiliation status, and the ownership
10
See O’Fallon and Rutherford (2010).
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 11
Table 2— Own and Neighbor Hotel Affiliation, Ownership and Management and Structure
Hotel Status Share of All
Hotels With
Reviews
Share of
Independent
Hotels
Share of Chain
Affiliated
Hotels
Independent 0.17 1.00 0.00
Marriott Corporation Affiliate 0.14 0.00 0.17
Hilton Worldwide Affiliate 0.12 0.00 0.15
Choice Hotels Int’l Affiliate 0.11 0.00 0.13
Intercontinental Hotels Grp Affiliate 0.08 0.00 0.10
Best Western Company Affiliate 0.04 0.00 0.04
Multi-unit owner 0.31 0.16 0.34
Multi-unit management company 0.52 0.35 0.55
Multi-unit owner AND multi-unit
management company
0.26 0.12 0.29
Hotel has a neighbor 0.76 0.72 0.77
Hotel has an independent neighbor 0.31 0.50 0.27
Hotel has a multi-unit owner neighbor 0.49 0.52 0.49
Hotel has a multi-unit management entity
neighbor
0.59 0.58 0.59
Total Hotels in Sample = 2931
Note: Table shows summary information about brand affiliation, ownership, and management charac-
teristics for 2931 hotelssampled with reviews at TripAdvisor and Expedia and their neighbors within
0.5km.
and affiliation status of the neighbors. However, it is important that our specifi-
cations also include a rich set of observable hotel characteristics to control for the
possibility that TripAdvisor and Expedia users value hotels with different char-
acteristics differently. We obtain a number of characteristics. First, we include
the “official” hotel rating for the hotel. At the time of our study, these official
ratings were reported in common by TripAdvisor and Expedia and are based on
the amenities of the hotel. From STR, we obtain a different hotel classification
system; hotels are categorized as “Economy Class”, “Luxury Class”, “Midscale
Class”, “Upper Midscale Class”, “Upper Upscale Class” and “Upscale Class.”
We use dummy variables to represent these categories in our specifications. We
also obtain the “year built” from STR and use it to construct a hotel age vari-
able (censored at 100 years old). Using STR categorizations, we also construct
dummy variables for “all suites” hotels, “convention” hotels, and a dummy that
takes the value of one if the hotel contains at least one restaurant. Even within
12 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
the same city, hotels have different location types. In all of our specifications, we
include dummies for airport locations, resort locations, and interstate/suburban
locations, leaving urban locations as the excluded type.
III. Theoretical Relationship between Ownership Structure and Review
Manipulation
Previous literature on promotional reviewing (see Mayzlin 2006, Dellarocas
2006) models review generation as a mixture of unbiased reviews and reviews
surreptitiously generated by competing firms. The consumer, upon seeing a re-
view, must discount the information taking into account the equilibrium level of
review manipulation.
In the Appendix we present a simple model that is closely related to the previous
models of promotional reviews but also allows the cost of review manipulation
to differ across firms, a new key element in the current context. In the model
firms engage in an optimal level of review manipulation (which includes both
fake positive reviews for self and fake negative reviews for competitors). The cost
of review manipulation is related to the probability of getting caught, which in
turn increases in each fake review that is posted. This model yields the following
intuitive result: an increase in the firm’s cost of review manipulation decreases
the amount of manipulation in equilibrium. Note that this also implies that if the
firm’s competitor has lower cost of review manipulation, the firm will have more
negative manufactured reviews.
The model reflects the fact that in practice the primary cost of promotional
reviews from the firm’s perspective is the risk that the activity will be publicly
exposed. The penalties that an exposed firm faces range from government fines,
possibility of lawsuits, and penalties imposed by the review-hosting platform. We
use the literature on reputational incentives and organizational form to argue that
this cost is also affected by the size of the entity. In this regard, our analysis is
related to Blair and Lafontaine (2005) and Jin and Leslie (2009) who examine
the incentive effects of reputational spillovers among co-branded entities. Our
analysis is also related to Pierce and Snyder (2008), Bennett et al. (2013), and
Ji and Weil (2009). Bennett et al. (2013) show that competition leads vehicle
inspectors to cheat and pass vehicles that ought to fail emissions testing. Pierce
and Snyder (2008) show that larger chains appear to curb cheating behavior
from their inspectors; inspectors at a large chain are less likely to pass a given
vehicle than are inspectors who work for independent shops. Similarly, Ji and
Weil (2009) show that company-owned units of chains are more likely to adhere
to labor standards laws than are franchisee-owned units. While our analysis is
related to this prior literature, we exploit the rich differences in organizational
form (chain vs. independent, large owner vs. small owner, and large management
company vs. small management company) particular to the hotel industry.
Before we formulate our hypotheses on the effect of entity size on review ma-
nipulation, we note a few important details on the design of travel review sites.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 13
In particular, note that reviews on these sites are hotel-specific, rather than chain
or owner specific. That is, a Hampton Inn in Cambridge, MA has unique reviews,
distinct from the reviews of a Hampton Inn in Atlanta, GA. If one wants to en-
hance the reputation of both hotels positively, one must post positive reviews of
both hotels separately on the site. If one wants to improve the attractiveness
of these hotels relative to their neighbors, one must post negative reviews for
the individual neighbors of each hotel separately on the site. These design fea-
tures make it unlikely that reviews would generate positive reputational spillovers
across hotels - that a fake review by one unit of a multi-unit entity is more pro-
ductive because it creates positive reputational spillovers for other units in the
entity. Note also that while the presence of positive spillovers is conceivable in
the case of chain-affiliated hotel posting positive fake reviews about itself (an im-
proved customer review at one Hampton Inn, for example, could possibly benefit
another Hampton Inn), it seems very unlikely in the case of the ownership variable
since co-ownership is not visible to the customers. Thus, it seems inconceivable
that a positive review for, say, Archon Hospitality’s Memphis Fairfield Inn would
improve the reputation of its Memphis Hampton Inn. Positive spillovers are also
less likely to arise in the case of negative competitor reviews. Posting a negative
review of one hotel will likely only benefit that hotel’s neighbors, not other hotels
throughout the chain.
In contrast to the discussion above, there are sizable negative spillovers associ-
ated with promotional reviews. Each incremental promotional review posted in-
creases the probability of getting caught. A larger entity suffers a greater penalty
from being caught undertaking fraudulent activities due to negative spillovers
across various units of the organization. Specifically, if an employee of a multi-
unit entity gets caught posting or soliciting fake reviews, any resulting government
action, lawsuit, or retribution by the review site will implicate the entire organiza-
tion. Because of this spillover, many larger entities have “social media policies,”
constraining the social media practices of employees or franchisees.
11
To make this concrete: suppose that the owner of Archon Hospitality, which
owns 41 hotels in our sample under various nameplates, were contemplating post-
ing a fake positive review about an Archon Hotel. As discussed above, the benefit
of the fake review would likely only accrue to the one hotel about which the fake
review was posted. To benefit another hotel, another fake review would have
to be posted. However, the probability of getting caught increases in each fake
review that is posted. If the owner of Archon were caught posting a fake review
about one hotel, the publicity and potential TripAdvisor sanctions would spill
over to all Archon hotels. Hence the cost of posting a fake review increases in the
number of hotels in the ownership entity, but the benefit of doing so does not.
This mechanism is also demonstrated in a recent case. The Irish hotel Clare
11
For example, Hyatt’s social media policy instructs Hyatt employees to “Avoid commenting on Hy-
att...only certain authorized individuals may use social media for Hyatt as business purposes...your
conduct may reflect upon the Hyatt brand.” (http://www.constangy.net/nr images/hyatt-hotels-
corporation.pdf, accessed April 10 2013).
14 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
Inn Hotel and Suites, part of the Lynch Hotel Group, was given the “red badge”
by TripAdvisor warning customers that the hotel manipulated reviews after it
was uncovered that a hotel executive solicited positive reviews. TripAdvisor also
removed reviews from other Lynch Hotel Group hotels, and the treatment of
Lynch Hotel Group was covered by news media in Ireland. Although the Lynch
Hotel Group hotels are not co-branded under a common nameplate, TripAdvi-
sor took action against the whole hotel group given the common ownership and
management of the hotels.
12
Thus, the key assumption underlying our owner-
ship/affiliation specifications is that the reputational benefit of posting a fake
review only accrues to one hotel, while the cost of posting the fake review (get-
ting caught) multiplies in the number of hotels in the ownership or affiliation
entity. Hence smaller entities have a bigger incentive to post fake reviews. In
terms of our model, the larger entity bears a higher δ and γ, and hence will fake
fewer reviews in equilibrium based on Proposition 1.
There is an additional incentive issue that applies specifically to ownership and
works in the same direction as the mechanism that we highlight. Drawing on the
literature on the separation of ownership and control, we hypothesize that owner-
operated hotels have a greater incentive to engage in review manipulation (either
positively for themselves or negatively for their neighbors). Owner-operators are
residual claimants of hotel profitability and employee-operators are not. Thus,
owner-operators would have more incentive to post fake reviews because owner-
operators have sharper incentives to generate hotel profitability. An employee of a
large ownership entity would have little to gain in terms of direct profit realization
from posting fake reviews but would risk possible sanctions from the entity for
undertaking fake reviewing activity.
In our paper, we consider the differential incentives of multi-unit entities us-
ing three measures of entity type. First, we consider ownership entities that are
large multi-unit owners versus small owners. For example, this measure captures
the distinction between an owner-operator Hampton Inn versus a Hampton Inn
owned by a large entity such as Archon Hospitality. Our ownership hypotheses
suggest that an owner-operator will have more incentive to post promotional re-
views than will an employee of a large entity. Second, we consider independent
hotels versus hotels operating under a common nameplate. As discussed above,
affiliation is a distinct characteristic from ownership; independent hotels can be
owner-operated but can also be owned by a large ownership entity. We hypoth-
esize that units of branded hotels will have less incentive to post promotional
reviews than will independents. As discussed above, brand organizations actively
discourage promotional reviewing by affiliates (with the threat of sanctions) be-
cause of the chain-wide reputational implications of being caught. Third, we
consider management by a large management company versus management by a
smaller entity. Again in this case, a review posted by the entity will benefit only
12
http://www.independent.ie/national-news/hotel-told-staff-to-fake-reviews-on-TripAdvisor-
2400564.html
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 15
one unit in the entity while the cost of being caught can conceivably spill over to
the entire entity. Unlike owners, hotel management companies are not residual
claimants and unlike franchise operations, do not always engage in profit-sharing.
Thus, while we examine hotel management companies in our analysis, it is less
clear that they have a strong enough stake in the hotel to influence reviewing
behavior.
In summary, we argue that the ownership and affiliation structure of the hotel
affects the costs of the promotional reviewing activity, which in turn affects the
equilibrium level of manufactured reviews. Specifically, based on our simple model
and the discussion above, we make the following three theoretical claims:
1) A firm that is located close to a competitor will have more fake negative
reviews than a firm with no close neighbors.
2) A firm that is part of a smaller entity will have more positive fake reviews.
3) A firm that is located close to a smaller entity competitor will have more
fake negative reviews.
IV. Methodology and Results
As Section II describes, we collect reviews from two sites, TripAdvisor and Ex-
pedia. There is a key difference between these two sites which we utilize in order
to help us identify the presence of review manipulation: while anybody can post a
review on TripAdvisor, only those users who purchased the hotel stay on Expedia
in the past six months can post a review for the hotel.
13
This implies that it
is far less costly for a hotel to post fake reviews on TripAdvisor versus posting
fake reviews on Expedia; we expect that there would be far more review manip-
ulation on TripAdvisor than on Expedia. In other words, a comparison of the
difference in the distribution of reviews for the same hotel could potentially help
us identify the presence of review manipulation. However, we can not infer pro-
motional activity from a straightforward comparison of reviews for hotels overall
on TripAdvisor and Expedia since the population of reviewers using TripAdvisor
and Expedia may differ; the websites differ in characteristics other than reviewer
identity verification.
Here we take a differences in differences approach (although, unconventionally,
neither of our differences are in the time dimension): for each hotel, we examine
the difference in review distribution across Expedia and TripAdvisor and across
different neighbor and ownership/affiliation conditions. We use the claims of Sec-
tion III to argue that the incentives to post fake reviews will differ across different
neighbor and ownership/affiliation conditions. That is, we hypothesize that hotels
13
Before a user posts a review on TripAdvisor, she has to click on a box that certifies that she has
“no personal or business affiliation with this establishment, and have not been offered any incentive or
payment originating the establishment to write this review.” In contrast, before a user posts a review
on Expedia, she must log in to the site, and Expedia verifies that the user actually purchased the hotel
within the required time period.
16 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
with greater incentive to manipulate reviews will post more fake positive reviews
for themselves and more fake negative reviews for their hotel neighbors on Tri-
pAdvisor, and we expect to see these effects in the difference in the distributions
of reviews on TripAdvisor and Expedia.
Consider the estimating equation:
(1)
NStarReviews
T A
ij
T otal Reviews
T A
ij
−
NStarReviews
Exp
ij
T otal Reviews
Exp
ij
= X
ij
B
1
+ OwnAf
ij
B
2
+ Nei
ij
B
3
+
NeiOwnAf
ij
B
4
+
X
γ
j
+ ε
ij
This specification estimates correlates of the difference between the share of
reviews on TA that are N star and the share of reviews on Expedia that are
N star for hotel i in city j. Our primary interest will be in the most extreme
reviews, 1-star/2-star and 5-star. X
ij
contains controls for hotel characteristics;
these hotel characteristics should only matter to the extent that TripAdvisor and
Expedia customers value them differentially. Specifically, as discussed above, we
include the hotel’s “official” star categorization common to TripAdvisor and Expe-
dia, dummies for the six categorizations of hotel type provided by STR (economy,
midscale, luxury, etc.), hotel age, location type dummies (airport, suburban, etc),
and dummies for convention hotels, the presence of a hotel restaurant, and all
suites hotels. N ei
ij
is an indicator variable indicating the presence of a neighbor
within 0.5km. OwnAf
ij
contains the own-hotel ownership and affiliation char-
acteristics. In our primary specifications, these include the indicator variable for
independent and the indicator variable for membership in a large ownership en-
tity. NeiOwnAf
ij
contains the variables measuring the ownership and affiliation
characteristics of other hotels within 0.5km. Specifically, we include an indicator
variable for the presence of an independent neighbor hotel, and an indicator vari-
able for the presence of a neighbor hotel owned by a large ownership entity. The
variables γ
j
are indicator variables for city fixed effects.
Our cleanest specifications examine the effect of Nei
ij
and NeiOwnAf
ij
vari-
ables on review manipulation. Following Claim 1 in Section III, we hypothesize
that a hotel with at least one neighbor will have more fake negative reviews
(have a higher share of 1-star/2-star reviews on TripAdvisor than on Expedia)
than a hotel with no neighbor. In addition, using Claim 3 from Section III, we
hypothesize that the neighbor effect will be exacerbated when the firm has an
independent neighbor, and that the neighbor effect will be mitigated when the
firm has a multi-unit owner or multi-unit management company neighbor.
We then turn to the effects of own-hotel organizational and ownership char-
acteristics (OwnAf
ij
) on the incentive to manipulate reviews. Following the
discussion in Section III, we hypothesize that an entity that is associated with
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 17
more properties has more to lose from being caught manipulating reviews: the
negative reputational spillovers are higher. Hence, we claim that 1) independent
hotels have a higher incentive to post fake positive reviews (have a higher share
of 5-star reviews on TripAdvisor versus Expedia) than branded chain hotels, 2)
small owners have a higher incentive to post fake positive reviews than multi-unit
owner hotels, 3) hotels with a small management company have a higher incen-
tive to post fake positive reviews than hotels that use multi-unit management
company.
Our interpretation of these results relies on our maintained assumption that
TripAdvisor and Expedia users value hotels with different ownership and affilia-
tion characteristics similarly. An important alternative explanation for our results
is that there are important differences in tastes of TripAdvisor and Expedia users
for unobserved characteristics that are correlated with our ownership and neigh-
bor variables. For example, one explanation for a finding that independent hotels
have a higher share of positive reviews on TripAdvisor is that the TripAdvisor
population likes independent hotels more than the Expedia population. We dis-
cuss this alternative hypothesis at length in the robustness section below. Here
we note that this alternative explanation is much more plausible a priori for some
of our results than for others. In particular, we find the alternative hypothesis less
plausible for the specifications for which the neighbor variables are the variables
of interest. For the neighbor specifications, the alternative hypothesis suggests
that, for example, some consumers will systematically dislike a Fairfield Inn whose
neighbor is an owner-operated Days Inn relative to a Fairfield Inn whose neigh-
bor is a Days Inn owned by a large entity like Archon, and that this difference in
preferences is measurably different for TripAdvisor and Expedia users.
Note that our empirical methodology is similar to the approach undertaken in
the economics literature on cheating. The most closely related papers in that
stream are Duggan and Levitt (2002), Jacob and Levitt (2003), and Dellavigna
and La Ferrara (2010). In all three papers the authors do not observe rule-
breaking or cheating (“throwing” sumo wrestling matches, teachers cheating on
student achievement tests, or companies trading arms in embargoed countries)
directly. Instead, the authors infer that rule-breaking occurs indirectly. That is,
Duggan and Levitt (2002) document a consistent pattern of outcomes in matches
that are important for one of the players, Jacob and Levitt (2003) infer cheating
from consistent patterns test answers, and Dellavigna and La Ferrara (2010) infer
arms embargo violations if weapon-making companies’ stocks react to changes in
conflict intensity. In all of these papers we see that cheaters respond to incentives.
Importantly for our paper, Dellavigna and La Ferrara (2010) show that a decrease
in reputation costs of illegal trades results in more illegal trading. Our empirical
methodology is similar to this previous work. First, we also do not observe review
manipulation directly and must infer it from patterns in the data. Second, we
hypothesize and show that the rate of manipulation is affected by differences in
reputation costs for players in different conditions. The innovation in our work is
18 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
that by using two different platforms with dramatically different costs of cheating
we are able to have a benchmark.
A. Main Results
In this Section we present the estimation results of the basic differences in
differences approach to identify review manipulation. Table 3 presents the results
of the estimation of Equation (1). Heteroskedasticity robust standard errors are
used throughout.
We first consider to the specification where the dependent variable is the dif-
ference in the share of 1- and 2-star reviews. Our dependent variable is thus
1 + 2StarReviews
T A
ij
T otal Reviews
T A
ij
−
1 + 2StarReviews
Exp
ij
T otal Reviews
Exp
ij
This is our measure of negative review manipulation. We begin with the simplest
specification: we examine the difference between negative reviews on TripAdvisor
and Expedia for hotels that do and do not have neighbors within 0.5km. This
specification includes all of the controls for hotel characteristics (X
ij
in Equation
1), but does not include the OwnAf
ij
and NeiOwnAf
ij
characteristics. The
results are in Column 1 of Table 3. The results show a strong and statistically
significant effect of the presence of a neighbor on the difference in negative reviews
on TripAdvisor vs. Expedia. The coefficient estimate suggests that hotels with
a neighbor have an increase of 1.9 percentage points in the share of 1-star and
2-star reviews across the two sites. This is a large effect given that the average
share of 1- and 2-star reviews is 25% for a hotel on TripAdvisor.
We continue with our analysis of negative reviews by examining ownership
and affiliation characteristics. We include in the specification all of the own
hotel ownership characteristics and the neighbor owner characteristics (OwnAf
ij
and NeiOwnAf
ij
). For these negative review manipulation results, we do not
expect to see any effects of the hotel’s own organizational structure on its share
of 1- and 2-star reviews since a hotel is not expected to negatively manipulate
its own ratings. Instead, our hypotheses concern the effects of the presence of
neighbor hotels on negative review manipulation. The results are in Column 2
of Table 3. As before, our coefficient estimates suggest that the presence of any
neighbor within 0.5km significantly increases the difference in the 1- and 2-star
share across the two sites. We hypothesize that multi-unit owners bear a higher
cost of review manipulation and thus will engage in less review manipulation. Our
results show that the presence of a multi-unit owner hotel within 0.5km results
in 2.5 percentage point decrease in the difference in the share of 1- and 2-star
reviews across the two sites, relative to having only single-unit owner neighbors.
This negative effect is statistically different from zero at the 1 percent confidence
level. As expected, the hotel’s own ownership and affiliation characteristics do
not have a statistically significant relationship to the presence of 1-star and 2-
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 19
Table 3—Estimation Results of Equation 1
Difference in
share of 1- and
2-star reviews
Difference in
share of 1- and
2-star reviews
Difference in
share of 5- star
reviews
X
ij
Site rating -0.0067
(0.0099)
-0.0052
(0.0099)
-0.0205**
(0.0089)
Hotel age 0.0004***
(0.0002)
0.0003*
(0.0002)
0.0002
(0.0002)
All Suites 0.0146
(0.0092)
0.0162*
(0.0092)
0.0111
(0.0111)
Convention Center 0.0125
(0.0086)
0.0159*
(0.0091)
-0.0385***
(0.0113)
Restaurant 0.0126
(0.0093)
0.0114
(0.0092)
0.0318***
(0.0099)
Hotel tier controls? Yes Yes Yes
Hotel location controls? Yes Yes Yes
OwnAf
ij
Hotel is Independent 0.0139
(0.0110)
0.0240**
(0.0103)
Multi-unit owner -0.0011
(0.0063)
-0.0312***
(0.0083)
Nei
ij
Has a neighbor 0.0192**
(0.0096)
0.0296**
(0.0118)
-0.0124
(0.0119)
NeiOwnAf
ij
Has independent neighbor 0.0173*
(0.0094)
-0.0051
(0.0100)
Has multi-unit owner neighbor -0.0252***
(0.0087)
-0.0040
(0.0097)
γ
j
City-level fixed effects? YES YES YES
Num. of observations 2931 2931 2931
R-squared 0.05 0.06 0.12
Note: *** p<0.01, ** p<0.05, * p<0.10
Regression estimates of Equation (1). The dependent variable in all specifications is the share of reviews
that are N star for a given hotel at TripAdvisor minus the share of reviews for that hotel that are N star
at Expedia. Heteroskedasticity robust standard errors in parentheses. All neighbor effects calculated for
neighbors within a 0.5km radius.
star reviews. The presence of an independent hotel within 0.5km results in an
additional increase of 1.7 percentage point in the difference in the share of 1-star
and 2-star reviews across the two sites. Our point estimates imply that having
an independent neighbor versus having no neighbor results in a 4.7 percentage
point increase in and 1- and 2 star reviews (3.0 percentage points for having any
20 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
neighbor plus 1.7 for the neighbor being independent). These estimated effects
are large given that the average share of 1- and 2-star reviews is 25% for a hotel
on TripAdvisor.
Of course, the neighbor characteristics are the characteristics of interest in the
1- and 2-star review specifications. However, our specifications include the hotel’s
own ownership characteristics as control variables. The estimated coefficients for
the hotel’s own ownership characteristics are small in magnitude and statistically
insignificant. This is consistent with our manipulation hypotheses but seem incon-
sistent with the alternative hypothesis of differences in preferences for ownership
characteristics across TripAdvisor and Expedia users.
We next turn to the specification where the dependent variable is the difference
in the share of 5-star reviews. That is, the dependent variable is
5StarReviews
T A
ij
T otal Reviews
T A
ij
−
5StarReviews
Exp
ij
T otal Reviews
Exp
ij
This is our measure of possible positive review manipulation. Consistent with our
hypothesis that independent hotels optimally post more positive fake reviews, we
see that independent hotels have 2.4 percentage points higher difference in the
share of 5-star reviews across the two sites than branded chain hotels. This effect
is statistically different from zero at the five percent confidence level. Since hotels
on TripAdvisor have on average a 31% share of 5-star reviews, the magnitude of
the effect is reasonably large. However, as we mentioned before, while this result
is consistent with manipulation, we can not rule out the possibility that reviewers
on TripAdvisor tend to prefer independent hotels over branded chain hotels to a
bigger extent than Expedia customers.
We also measure the disparity across sites in preferences for hotels with multi-
unit owners. Consistent with our hypothesis that multi-unit owners will find re-
view manipulation more costly, and therefore engage in less review manipulation,
we find that hotels that are owned by a multi-unit owner have a 3.1 percentage
point smaller difference in the share of 5-star reviews across the two sites. This
translates to about four fewer 5-star reviews on TripAdvisor if we assume that
the share of Expedia reviews stays the same across these two conditions and that
the hotel has a total of 120 reviews on TripAdvisor, the site average. While we
include neighbor effects in this specification, we do not have strong hypotheses
on the effect of neighbor characteristics on the difference in the share of 5-star
reviews across the two sites, since there is no apparent incentive for a neighboring
hotel to practice positive manipulation on the focal hotel. Indeed, in the 5-
star specification, none of the estimated neighbor effects are large or statistically
significant. In interpreting these results, it is important to remember that the
ownership characteristic is virtually unobservable to the consumer; it measures
the difference between, for example, an Archon Hospitality Fairfield Inn and an
owner-operator Fairfield Inn. Nonetheless, it is plausible that TripAdvisor and
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 21
Expedia users differentially value hotel characteristics that are somehow corre-
lated with the presence of an owner-operator (and not included in our regression
specifications). We return to this issue below.
For the 5-star specifications, the hotel’s own ownership characteristics are the
variables of interest, rather than the neighbor variables. Here, we find the es-
timated coefficients of the neighbor characteristics to be small and statistically
insignificant. This finding is consistent with our manipulation hypothesis but
seems inconsistent with the alternative hypothesis that TripAdvisor and Expedia
users have systematically different preferences for hotels with different kinds of
neighbors.
What do our results suggest about the extent of review manipulation on an open
platform such as TripAdvisor overall? Note that we cannot identify the baseline
level of manipulation on TripAdvisor that is uncorrelated with our characteristics.
Thus, we can only provide estimates for the difference between hotels of different
characteristics. However, as an example, let’s consider the difference in positive
manipulation under two extreme cases: a) a branded chain hotel that is owned
by a multi-unit owner (the case with the lowest predicted and estimated amount
of manipulation) and b) an independent hotel that is owned by a small owner
(the case with the greatest predicted and estimated amount of manipulation).
Recall that the average hotel in our sample has 120 reviews, of which 37 on
average are 5-star. Our estimates suggest that we would expect about 7 more
positive TripAdvisor reviews in case b versus case a. Similarly, we can perform a
comparison for the case of negative manipulation by neighbors. Consider case c)
being a completely isolated hotel and case d) being located near an independent
hotel that is owned by a small owner. For the average hotel with 120 reviews,
thirty 1-star and 2-star reviews would be expected as a baseline. Our estimates
suggest that there would be a total of 6 more fake negative reviews on TripAdvisor
in case d versus case c.
Our main results focus on the presence of neighbors and the ownership and affil-
iations of hotels and their neighbors. However, hotels differ structurally not only
in their ownership but also in their management. As explained above, some hotel
units have single unit owners, but these owners outsource day to day management
of the hotels to a management company. In our sample of 2931 hotels, of the 2029
that do not have multi-unit owners, 767 do outsource management to multi-unit
managers. As we explain in Section III, the management company is not residual
claimant to hotel profitability the way that the owner is, but nonetheless, obvi-
ously has a stake in hotel success. As in the case of multi-unit owners, posting
of fake reviews by an employee of a management company could, if detected,
have negative implications for the management company as a whole. Thus, we
expect that a multi-unit management company would have a lower incentive to
post fake reviews than a single-unit manager (which in many cases is the owner).
This implies that hotel neighbors of hotels with multi-unit managers should have
fewer 1- and 2-star reviews on TripAdvisor while hotels with multi-unit managers
22 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
should have fewer 5-star reviews on TripAdvisor, once again if we assume that
the share of Expedia reviews stays the same.
In the first column in Table 4, we use the share difference in 1- and 2-star re-
views as the dependent variable. Here, as before, we have no predictions for the
own hotel characteristics (and none are statistically different from zero). We do
have predictions for neighbor characteristics. As before, we find that having any
neighbor is associated with having more 1- and 2-star reviews, a 3.8 percentage
point increase. As before, an independent hotel neighbor is associated with more
negative reviews on TripAdvisor relative to Expedia and having a large owner
chain neighbor is associated with fewer negative reviews on TripAdvisor. The
presence of a large management company neighbor is associated with fewer neg-
ative reviews on TripAdvisor, although the effect is not statistically significant at
standard confidence level. The presence of a large owner neighbor and the pres-
ence of a large management company neighbor are quite positively correlated. A
test of the joint significance shows that the two variables are jointly significant in
our specification at the 1 percent level.
In the second column of Table 4, we examine 5-star reviews. Here, as before,
the neighbor characteristics are uninformative. As before, independent hotels
have more 5-star reviews on TripAdvisor relative to Expedia and that hotels
with a large owner company have fewer 5-star reviews. In addition, the results
show that a hotel that is managed by a multi-unit management company has a
statistically significant 2.1 percentage point decrease in the difference of the share
of 5-star reviews between the two sites which we interpret as a decrease in positive
manipulation. Notably, the inclusion of this variable does not alter our previous
results; independent hotels continue to have significantly more 5-star reviews on
TripAdvisor relative to Expedia and hotels with multi-unit owners have fewer 5-
star reviews. This result is important because, like a multi-unit owner company,
management by a multi-unit management company is invisible to the consumer.
Thus, altogether, there is suggestive evidence that, like larger owner companies,
larger management companies are associated with less review manipulation.
Unfortunately, it is impossible for us, given these data, to measure the effect
that these ratings’ changes will have on sales. While Chevalier and Mayzlin
(2006) show that 1-star reviews hurt book sales more than 5-star reviews help
book sales, those findings do not necessarily apply to this context. Chevalier and
Mayzlin (2006) note that two competing books on the same subject may indeed
be net complements, rather than net substitutes. Authors and publishers, then,
may gain from posting fake positive reviews of their own books, but will not
necessarily benefit from posting negative reviews of rivals’ books. Thus, in the
context of books, 1-star reviews may be more credible than 5 star reviews. We
have seen that, in the case of hotels, where two hotels proximate to each other are
clearly substitutes, one cannot infer that a 1 or 2 star review should be treated
by customers as more credible than a 5-star review.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 23
Table 4— Management Company Specifications
Difference in share of
1- and 2- star reviews
Difference in share
of 5-star reviews
X
ij
Site rating -0.0047
(0.0100)
-0.0183**
(0.0090)
Hotel age 0.0003**
(0.0002)
0.0002
(0.0002)
All Suites 0.0169*
(0.0091)
0.0144
(0.0112)
Convention Center 0.0163*
(0.0090)
-0.0363***
(0.0113)
Restaurant 0.0110
(0.0092)
0.0323***
(0.0099)
Hotel tier controls? YES YES
Hotel location controls? YES YES
OwnAf
ij
Hotel is Independent 0.0141
(0.0111)
0.213**
(0.0104)
Multi-unit owner -0.0014
(0.0064)
-0.0252***
(.0086)
Multi-unit management company 0.0022
(0.0077)
-0.0211 **
(0.0092)
Nei
ij
Has a neighbor 0.0379***
(0.0142)
-0.0098
(0.0140)
NeiOwnAf
ij
Has independent neighbor 0.0173*
(0.0094)
-0.006
(0.0100)
Has multi-unit owner neighbor -0.0169*
(0.0097)
0.0004
(0.0114)
Has multi-unit management
company neighbor
-0.0183
(0.0125)
-0.0059
(0.0136)
γ
j
City-level fixed effects? YES YES
Num. of observations 2931 2931
R-squared 0.06 0.12
Note: *** p<0.01, ** p<0.05, * p<0.10
Regression estimates of Equation (1). The dependent variable in all specifications is the share of reviews
that are N star for a given hotel at TripAdvisor minus the share of reviews for that hotel that are N star
at Expedia. Heteroskedasticity robust standard errors in parentheses. All neighbor effects calculated for
0.5km radius.
B. Results for One-Time Reviewers
Our preceding analysis is predicated on the hypothesis that promotional review-
ers have an incentive to imitate real reviewers as completely as possible. This is
24 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
in contrast to the computer science literature, described above, that attempts to
find textual markers of fake reviews. Nonetheless, for robustness, we do separately
examine one category of “suspicious” reviews. These are reviews that are posted
by one-time contributors to TripAdvisor. The least expensive way for a hotel to
generate a user review is to create a fictitious profile on TripAdvisor (which only
requires an email address), and following the creation of this profile, to post a
review. This is, of course, not the only way that the hotel can create reviews.
Another option is for a hotel to pay a user with an existing review history to
post a fake review; yet another possibility is to create a review history in order
to camouflage a fake review. Here, we examine “suspicious” reviews: the review
for a hotel is the first and only review that the user ever posted. In our sample,
23.0% of all TripAdvisor reviews are posted by one-time reviewers. These reviews
are more likely to be extreme compared to the entire TripAdvisor sample: 47.6%
of one-time reviewers are 5-star versus 38.1% in the entire TripAdvisor sample.
There are more negative outliers as well: 24.3% of one-time reviews are 1-star and
2-star versus 16.4% in the entire TripAdvisor sample. Of course, the extremeness
of one-time reviews does not in and of itself suggest that one-time reviews are
more likely to be fake; users who otherwise do not make a habit of reviewing may
be moved to do so by an unusual experience with a hotel.
In Table 5 we present the results of the following three specifications. In the
first column, we present the results of a specification where the dependent vari-
able is the share of one-time contributor user reviews on TripAdvisor. Thus, our
dependent variable is
one − time Reviews
T A
ij
/
T otal Reviews
T A
ij
. This cap-
tures the incidence of these suspicious reviews and includes potential positive as
well as negative manipulation. The most striking result is that one-time reviews
are 8.8 percentage points more common for independent hotels. This is consis-
tent with our earlier results, but also could be attributable to legitimate customer
reviewing preferences. Also consistent with our earlier results, we find a negative
impact of multi-unit owner on one-time reviewing activity, and a negative impact
of multi-unit owner neighbors. There is one variable in our specification that does
not have the anticipated sign. The presence of any neighbor is negatively asso-
ciated with “suspicious” reviews (although this effect is insignificant); our model
would predict that this association would be positive.
The other two specifications in Table 5 address the valence of these reviews.
For these specifications, the dependent variable is
one − time NStarReviews
T A
ij
one − time Reviews
T A
ij
−
NStarReviews
Exp
ij
T otal Reviews
Exp
ij
That is, we look at the difference between the share of N-star reviews among
“suspicious” reviews on TripAdvisor and the overall share of N-Star reviews on
Expedia. Ideally, we might want to compare one-time reviews on TripAdvisor
to one-time reviews on Expedia. Unfortunately, Expedia’s reviewer identification
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 25
Table 5—Results for TripAdvisor one-time contributor reviewers
Share of one-time
contributor user
reviews
Difference in share
of 1 and 2 star
reviews
Difference in
share of 5 star
reviews
X
ij
Site rating -0.0176***
(0.0061)
-0.0175
(0.0113)
-0.0083
(0.0102)
Hotel age 0.0003**
(0.0001)
0.00005
(0.0002)
0.0002
(0.0002)
All Suites 0.0086
(0.0065)
-0.0147
(0.0137)
0.0035
(0.0150)
Convention Center -0.0177**
(0.0082)
0.0532***
(0.0147)
-0.0716***
(0.0170)
Restaurant 0.0376***
(0.0068)
0.0079
(0.0126)
0.0329***
(0.0128)
Hotel tier controls? YES YES YES
Hotel location controls? YES YES YES
OwnAf
ij
Hotel is Independent 0.0881***
(0.0079)
-0.0035
(0.0135)
0.0082
(0.0123)
Multi-unit owner -.0135**
(0.0052)
0.0109
(0.0102)
-0.0239**
(0.0117)
Nei
ij
Has a neighbor -0.0091
(0.0080)
0.0285*
(0.0156)
-0.0093
(0.0159)
NeiOwnAf
ij
Has independent
neighbor
0.0002
(0.0066)
0.0203
(0.0133)
0.0027
(0.0130)
Has multi-unit owner
neighbor
-0.0144**
(0.0062)
-0.0150
(0.0125)
-0.0038
(0.0132)
γ
j
City-level fixed effects? YES YES YES
Num. of observations 2874 2874 2874
R-squared 0.35 0.05 0.07
Note: *** p<0.01, ** p<0.05, * p<0.10
Estimation of Equation 1 with the sample restricted to hotels that have at least one review by a one-time
contributor (the reviewer has only submitted one review on TripAdvisor).
The dependent variable in the first column is the share of reviews by one-time contributors among all
TripAdvisor reviews for a given hotel. The dependent variable in the other two columns is the share of
reviews by one-time contributors that are N star for a given hotel at TripAdvisor minus the share of
reviews for that hotel that are N star at Expedia.
Heteroskedasticity robust standard errors in parentheses. All neighbor effects calculated for 0.5km radius.
features render identifying one-time reviewers impossible. Column 2 shows the
case where N = 1 or 2, our specification that focuses on the characteristics of
neighbor hotels. The presence of a neighbor is associated with a 2.9 percent-
age point increase in the share of one-time reviews that are 1 or 2 stars. This
26 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
effect is statistically significant at the ten percent level. The effect of indepen-
dent neighbors and multi-unit neighbors are positive and negative, respectively,
in accordance with our model and previous results. However, these effects are
not significant at standard confidence levels. It is possible that these results are
weak in part because one-time reviews are “suspicious.” TripAdvisor has a policy
whereby hotels can contest suspicious reviews and TripAdvisor may, at its discre-
tion, remove contested “suspicious” reviews from the site. Negative reviews by
one-time reviewers may be more likely to be expunged from the site. In Column
3, we examine 5-star reviews, the specifications in which we focus on own-hotel
characteristics. The effect of hotel independence is positive, as predicted, but not
significantly different from zero. Multi-unit owner has a statistically significant
2.4 percentage point lower difference in the share of 5-star reviews across the two
sites, which is consistent with our hypotheses and earlier results.
Overall, these results confirm our prior results that manipulation of reviews
takes place in a way that is consistent with predicted hotel incentives. However,
our results for “suspicious” reviews are not as compelling as our results for all
“reviews.” Of course, with this analysis we are forced to construct the left hand
side variable using a smaller subset of reviews, which may be noisy. Further, if
fakers are sophisticated in their attempt to avoid detection, they may be avoiding
these suspicious reviewing activities.
C. Robustness Checks
Perhaps the main concern with our results is the potential for selection on un-
observables. That is, TripAdvisor and Expedia users may differ in their taste for
hotel characteristics. We have included many such possible characteristics in our
specifications (hotel age, hotel tier, hotel location type, etc.). Thus, differences
in tastes for these included characteristics are not a problem for our analysis; we
have controlled for these in our specification. However, it is possible that con-
sumer tastes differ across the two websites for unobservable characteristics. This
is a concern if the unobservable characteristics are correlated with the ownership,
affiliation, and neighbor variables of interest. This could in principle lead to sig-
nificant measured impacts of our ownership, affiliation, and neighbor variables
even if ownership, affiliation, and neighbor characteristics are not associated with
review manipulation. A priori, we find this alternative hypothesis less plausible
for any specifications in which the variables of interest are neighbor variables. It
seems unlikely, for example, that TripAdvisor users systematically dislike (rela-
tive to Expedia users) hotels whose hotel neighbors are franchisees that operate
a single hotel. A priori, we find selection on unobservables to be a more plau-
sible concern for specifications in which the variables of interest are own-hotel
ownership and affiliation.
To investigate selection on unobservables, we undertake the following exercise.
Recall that our base specifications include a rich set of control variables. We re-
estimate the base specifications in Table 3, maintaining the neighbor, ownership,
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 27
and affiliation variables but removing all of the control variables. We compare
the result of this no-controls specification to our basic results including all of
the control variables. We examine how much (if at all) inclusion of the control
variables attenuates the coefficients for the variables of interest. If unobservable
characteristics are positively correlated with observable characteristics, one might
expect that the inclusion of additional controls, if they were available, would
further attenuate the coefficients on the variables of interest.
14
The no-control
specifications are shown in Columns 1 through 3 of Table 6. Comparing Table 3 to
Table 6, for the neighbor specification shown in Column 1, reestimation excluding
all control variables actually produces a smaller point estimate of the neighbor
effect. Thus, inclusion of a set of control variables does not attenuate the results
at all. This finding has been interpreted in the literature as assuaging concerns
about selection on unobservables. Similarly, the full neighbor ownership-affiliation
specification in Column 2 of Table 3 can be compared with the specifications with
no control variables in Table 6. The independent neighbor variable has a stronger
measured impact on review differences in the regression with the controls versus
the regression without controls. Again, inclusion of controls does not attenuate
the independence effect. The owner-neighbor effect does attenuate from -0.031
to -0.025 with the inclusion of the control variables. However, our specifications
contain a very rich set of control variables. If we could hypothetically perform
a regression that contained all of the unobservables, and if these unobservables
were as powerful as the observable control variables in attenuating the ownership
effect, the ownership effect would still remain substantial in magnitude. Thus,
for the neighbor specifications in Columns 1 and 2 of Table 3, we conclude that
selection on unobservables is unlikely to be a major explanation for our results.
A priori, selection on unobservables is more plausible for the five star specifica-
tions examining a hotel’s own characteristics. Own hotel ownership and affiliation
are plausibly correlated with characteristics that TripAdvisor and Expedia cus-
tomers could value differently. Again, we examine this issue by comparing the no
controls specifications in Column 3 of Table 6 to the full controls specifications
in Column 3 of Table 3. Here, the alternative hypothesis of selection on unob-
servables is more difficult to reject. Both the multi-unit owner dummy and the
independent hotel dummy are attenuated by approximately 50% when controls
are added to the regression. Thus, our interpretation of these coefficients as ev-
idence for review manipulation relies on the included hotel characteristics being
more powerful than omitted hotel characteristics in explaining the difference in
reviewer behavior on TripAdvisor and Expedia.
Our analysis of Table 6 is one strategy to examine the importance of omitted
hotel characteristics. In Appendix Table 8, we take another approach to exam-
ining omitted hotel characteristics. Here, we reexamine the base specifications
of Table 3, including hotel chain fixed effects for the ten largest hotel brands.
Inclusion of these chain fixed effects allows TripAdvisor and Expedia patrons to
14
See Altonji, Elder and Taber (2005) for a more formal discussion of this test.
28 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
Table 6—Robustness specifications: Specifications with no controls
Difference in
Share of 1- and
2 -star reviews
Difference in
Share of 1- and
2- star reviews
Difference in
Share of 5- star
reviews
X
ij
Site rating – – –
Hotel age – – –
All Suites – – –
Convention Center – – –
Restaurant – – –
Hotel tier controls NO NO NO
Hotel location controls NO NO NO
OwnAf
ij
Hotel is Independent – 0.0093
(0.0092)
0.0429***
(0.0103)
Multi-unit owner – -0.0181**
(0.0075)
-0.0642***
(0.0084)
Nei
ij
Has a neighbor 0.0118
(0.0079)
0.0324***
(0.0098)
-0.0177
(0.0109)
NeiOwnAf
ij
Has independent
neighbor
– 0.0022
(0.0082)
-0.0069
(0.0091)
Has multi-unit owner
neighbor
- -0.0310***
(0.0083)
-0.0211**
(0.0093)
γ
j
City level fixed effects? NO NO NO
Num of observations 2931 2931 2931
R-squared 0.001 0.01 0.04
Note: ***p<0.01, **p<0.05, *p<0.10
Estimation of Equation 1 excluding control variables. The dependent variable in all specifications is the
share of reviews that are N star for a given hotel at TripAdvisor minus the share of reviews that are N
star at Expedia. Heteroskedasticity robust standard errors in parentheses. All neighbor effects calculated
for 0.5 km radius.
Results in this Table can be compared to the “base” specifications in Table 3 to measure whether and if
the inclusion of control variables in Table 3 leads to substantial attenuation of the coefficients of interest.
have a very general form of different preferences. They can have not only differ-
ent preferences for hotel quality tiers and hotel age (all included in the controls
in our base specifications), but also can have different preferences for different
individual hotel brands. These specifications produce results very similar to the
base specifications discussed in 3. Here, the neighbor variables of interest are all
of roughly the same magnitude and significance as in our base specifications. The
only change that inclusion of this variable causes compared to the earlier results
is that the independent own hotel dummy in the 5-star specification is no longer
statistically significant; the ownership variable remains of the expected sign and
statistically significantly different from zero.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 29
Given the importance of our negative review specifications, we next turn to
a few robustness checks that examine the robustness of our neighbor ownership
and affiliation results. Throughout, we have used 1 and 2 star reviews as our
marker of “negative” reviews. We chose this specification in part due to the
summary statistics outlined in Table 1. While 31% of reviews on TripAdvisor are
5s, 1s and 2s together only account for 25% of reviews. Hence, a firm attempting
to denigrate its competitor will often be able to do so effectively with either 1-
or 2-star promotional reviews. Furthermore, a scan of web blogs, etc. suggests
that hoteliers complain to TripAdvisor about fake 2-star negative reviews from
competitors and that TripAdvisor has sometimes deemed such reviews as fake and
removed them.
15
Nonetheless, we provide robustness results where we examine
the basic specification in Equation 1 above, but consider only determinants of 1-
star reviews. This is shown in the first column of Table 7. The results are similar
to the base 1- and 2-star results in Column 2 of Table 3; the own-hotel ownership
and affiliation characteristics have little explanatory power and are insignificant.
The independent neighbor and large company owner neighbor coefficients are
similar in magnitude and significance to the main specification. The “having any
neighbor within 0.5km” indicator variable has a smaller coefficient (although still
of the hypothesized sign) but is not statistically significant at standard confidence
levels.
We also examine the robustness of our results by altering the radius that we use
to define neighbors. In our base specifications in Table 3, we define a neighbor
as a hotel that is very close to the hotel of interest– within 0.5km kilometer of
the hotel of interest. Under this definition, 76% of the hotels in our sample have
a neighbor. In Columns 2 through 4 of Table 7 we re-estimate the specification
of Column 2 of Table 4, but using different radii to define neighbors– 0.3km,
0.7km, and 0.9km. Under the narrower radius definition of 0.3km, 65% of hotels
have a neighbor. Under the wider radius definitions of 0.7km and 0.9km, 82%
and 85% of the hotels in our sample have a neighbor, respectively. These varying
radii specifications are similar to our base specification. As the radius widens, the
dummy for having any neighbor appears to diminish in magnitude and significance
(as nearly every hotel has a neighbor), while the neighbor characteristics maintain
or even increase explanatory power. Thus, we conclude that our results change
with the radius size in a sensible way.
Finally, we examine the robustness of our results to our particular choices of
review site. Specifically, we examine the relationship between our results and the
results that would obtain by replacing the data from Expedia analyzed above
with data from another site, Orbitz.com. Orbitz.com, like Expedia, is primarily
a travel booking site that hosts user reviews. Orbitz is a less popular site than
Expedia; Orbitz had approximately 60 percent fewer page views than Expedia
15
See, for example, “Fake Review Number Two” in http://TripAdvisorwatch.wordpress.com/trip-
advisor-fake-reviews/.
30 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
Table 7—Specifications with negative reviews as dependent variable
Difference
in share of
1-star
reviews
Difference in
share of 1-
and 2-star
reviews
Difference in
share of 1-
and 2-star
reviews
Difference in
share of 1-
and 2-star
reviews
X
ij
Site rating -0.0177**
(0.0076)
-0.0055
(0.0099)
-0.0050
(0.0099)
-0.0048
(0.0099)
Hotel age 0.0005***
(0.0001)
0.0003 **
(0.0002)
0.0003**
(0.0002)
0.0003**
(0.0002)
All Suites 0.0091
(0.0076)
0.0159*
(0.0092)
0.0156*
(0.0091)
0.0158*
(0.0091)
Convention Center 0.0104
(0.0073)
0.0166*
(0.0091)
0.0170*
(0.0091)
0.0163*
(0.0091)
Restaurant 0.0039
(0.0076)
0.0110
(0.0092)
0.0091
(0.0092)
0.0091
(0.0092)
Hotel tier controls? YES YES YES YES
Hotel location
controls?
YES YES YES YES
OwnAf
ij
Hotel is Independent 0.0117
(0.0100)
0.0126
(0.0110)
0.0113
(0.0109)
0.0114
(0.0109)
Multi-unit owner -0.0025
(0.0047)
-0.0012
(0.0063)
-0.0015
(0.0063)
-0.0020
(0.0063)
Nei
ij
Has a neighbor 0.0095
(0.0106)
0.0258**
(0.0102)
0.0125
(0.0132)
0.0137
(0.0147)
NeiOwnAf
ij
Has independent
neighbor
0.0192**
(0.0081)
0.0109
(0.0099)
0.0176*
(0.0095)
0.0207**
(0.0093)
Has multi-unit
owner neighbor
-0.0204***
(0.0075)
-0.0262***
(0.0084)
-0.0252***
(0.0093)
-0.0262***
(0.0096)
γ
j
City-level fixed
effects?
YES YES YES YES
Num. of
observations
2931 2931 2931 2931
Neighbor radius? 0.5km 0.3km 0.7km 0.9km
R-squared 0.09 0.05 0.05 0.05
Note: *** p<0.01, ** p<0.05, * p<0.10
Estimation of Equation (1). The dependent variable in all specifications is the share of reviews that
are N star for a given hotel at TripAdvisor minus the share of reviews that are N star at Expedia.
Heteroskedasticity robust standard errors in parentheses. The radius for which neighbors are calculated
for a given hotel is given in the table.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 31
in 2012.
16
In addition, whereas we expect there to be a large overlap between
TripAdvisor and Expedia audiences due to the companies’ co-marketing efforts
at the time of data collection (see our discussion above), we do not have the
same expectations for TripAdvisor and Orbitz. We provide details of out analysis
in Online Appendix. In summary, while we find that that our Orbitz results are
qualitatively similar to the Expedia results presented in the paper, the magnitude
of some of the Orbitz results is smaller. Overall, we take these results as suggestive
that our findings are robust when examining alternative sites.
V. Conclusion and directions for future work
We propose a novel methodology for empirically detecting review manipulation.
In particular, we examine the difference in review distributions across Expedia
and TripAdvisor, sites with different reviewer identity verification policies, and
across different competitive/ownership conditions. Consistent with our theoret-
ical claims, we find that an increase in hotel incentives to manipulate reviews
results in an increase in our measures of manipulation. Substantively, we find
that hotels with next-door neighbors have more negative reviews on Tripadvi-
sor, and this effect is exacerbated if the neighbor is an independent hotel with
a small owner. That is, we find evidence for negative review manipulation. We
also observe review patterns that are consistent with positive manipualtion: we
find that independent hotels with small owners and small management compa-
nies have more positive reviews on Tripadvisor. While we find evidence for both
negative and positive manipulation, throughout the paper we emphasize the fact
that our results on negative manipulation are more robust to selection issues than
our positive manipulation results. We conclude from our results that promotional
reviewing is sufficiently economically important that actors that are differentially
situated economically will indulge in promotional reviewing to a measurably dif-
ferent extent.
Our paper also contributes to the literature on incentives and organizational
form. Our unusually rich data set allows us to exploit the fact that ownership
patterns in the hotel industry are actually quite complicated. For example, as
discussed previously, a hotel can be franchised to a quite large franchisee company;
we hypothesize that the large franchisee company is less incentivized to engage
in this type of fraudulent activity than a small franchisee. In our paper, we
advance the literature on ownership by utilizing data on these complex ownership
structures. We show that larger organizations appear to be measurably better at
curbing cheating.
While is it not our primary goal, our paper also contributes to the literature on
fake review detection. Previous methodologies in the computer science literature
infer that reviews are more likely to be fake if they contain certain textual markers
16
Data from ComScore, found at http://www.newmediatrendwatch.com/markets-by-country/17-
usa/126-online-travel-market, accessed April 11, 2013.
32 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
of fakery (such as not using spatial language). We have noted that a concern
with these methodologies is that manipulating the textual markers in response to
detection algorithms is relatively inexpensive. In contrast, the organization form
of a hotel and its neighbors are very difficult to alter. Our results suggest that
a detection algorithm could incorporate these factors in assessing the probability
that a given review is fake.
Our paper also has implications for user review system design. Our results sug-
gest that promotional reviews are less common on Expedia than on TripAdvisor.
Thus, the policy of verifying reviews does limit promotional reviews. However,
this limitation comes at a cost: there are far fewer reviews on Expedia than on
TripAdvisor. While the policy used by Orbitz (and now Amazon) of marking
verified and unverified reviews is an interesting compromise, it may discourage
unverified reviews and does not fully solve the review site’s problem of whether
to fully incorporate unverified reviews into summary data.
There are a number of limitations of this work. Perhaps the biggest limitation
is that we do not observe manipulation directly but must infer it. This issue
is of course inherent in doing research in this area. In the paper we deal with
this limitation by building a strong case that the effects that we examine are due
to review manipulation and not due to other unobserved factors. The second
important limitation is that our measure of review manipulation does not include
any content analysis. That is, one could imagine that one way in which a hotel
could increase the impact of a fake review is by making particularly strong claims
in the text of the review. For example, to hurt a competitor, a “traveler” could
claim to have witnessed a bed bug infestation. This is an interesting issue for
future work.
In this work, we are unable to measure the impact that this manipulation has
on consumer purchase behavior. Do consumers somehow detect and discount
fake reviews? Do they discount all reviews to some extent? Do they make poor
choices on the basis of fake reviews? These questions suggest important avenues
for future work.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 33
VI. Appendix
Table 8—Chain Fixed Effects Robustness Check
Difference in share
of 1- and 2-star
reviews
Difference in share
of 1- and 2-star
reviews
Difference in share
of 5- star reviews
X
ij
Site rating -0.0007
(0.0100)
-0.0067
(0.0101)
-0.0193**
(0.0089)
Hotel age 0.0003
(0.0002)
0.0002
(0.0002)
-0.00006
(0.0002)
All Suites 0.0107
(0.0090)
0.0112
(0.0091)
0.0097
(0.0123)
Convention Center 0.0185**
(0.0090)
0.0193**
(0.0092)
-0.0263**
(0.0113)
Restaurant 0.0085
(0.0095)
0.0077
(0.0095)
0.0271***
(0.0100)
Hotel tier controls? Yes Yes Yes
Hotel location
controls?
Yes Yes Yes
Chain-level fixed
effects?
Yes Yes Yes
OwnAf
ij
Hotel is Independent – 0.0053
(0.0135)
0.0079
(0.0119)
Multi-unit owner – 0.0053
(0.0067)
-0.0194**
(0.0086)
Nei
ij
Has a neighbor 0.0205**
(0.0096)
0.0304***
(0.0118)
-0.0121
(0.0119)
NeiOwnAf
ij
Has independent
neighbor
– 0.0162*
(0.0094)
-0.0071
(0.0099)
Has multi-unit owner
neighbor
– -0.0253***
(0.0088)
-0.0018
(0.0097)
γ
j
City-level fixed
effects?
YES YES YES
Num. of observations 2931 2931 2931
R-squared 0.06 0.06 0.13
Note: *** p<0.01, ** p<0.05, * p<0.10
Heteroskedasticity robust standard errors in parentheses.
All neighbor effects calculated for 0.5km radius.
34 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
A. A Simple Model
We propose a very simple and stylized model to fix ideas. The game consists
of two competing firms, A and B, and a continuum of consumers. The time line
of the game is the following:
1) Stage I : Nature draws the true quality of each firm (q
A
and q
B
), where
the two firms’ qualities are i.i.d. random variables with the cumulative
distribution function F and E(q
i
) = q
0
, i ∈ A, B. We assume that the
firms’ true quality is not observable to any of the game’s players.
17
Here,
the two firms a priori are identically distributed, but the model can be easily
generalized to the case where the prior means are not equal. We assume
that all other parameters of the model are common knowledge.
2) Stage II : The firms set prices (p
A
and p
B
), which are observed by all the
players.
3) Stage III : Each firm can surreptitiously (and simultaneously) manufac-
ture positive reviews for itself and negative reviews for its competitor. The
reviews are posted by a third party platform that does not verify the re-
viewers’ identity. We assume that consumers observe all the user ratings,
but they can not differentiate between real and manufactured (or biased)
user reviews. We denote by e
i,i
the effort that firm i invests into positive
self-promotion (manufactured positive reviews), and by e
i,j
the effort that
firm i invests into negative reviews for firm j. The observed firm quality
(ˆq
i
) consists of the firm’s true quality (as conveyed noiselessly by the real
reviews) and the firms’ promotional efforts:
ˆq
A
=q
A
+ e
A,A
− e
B,A
(2)
ˆq
B
=q
B
+ e
B,B
− e
A,B
(3)
That is, firm A’s observed quality (ˆq
A
) consists of the firm’s true quality
(q
A
), the positive self-promotion effort by firm A (e
A,A
) and the negative
effort by its competitor (e
B,A
). Note that while we model the benefit of firm
effort as linear in promotional effort for the sake of simplicity, in reality the
benefit is more likely to be concave in effort. That is, since a rating can’t
be higher than 5-stars, an increase in the number of manufactured positive
reviews is likely to have diminishing marginal returns. Similarly, since a
competitor’s rating can’t be lower than 1 star, an increase in the number
of manufactured negative reviews is likely to have diminishing marginal
returns.
4) We model the manipulation effort as costly to the firm. We can think of this
cost as the cost of writing reviews or as reputation-related risks associated
17
The case where only firms, but not the consumers, observe each other’s true quality yields similar
results, but is considerably more complicated.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 35
with this kind of promotion. That is, if the firm is caught doing this kind
of activity, it will suffer damage to its reputation, which may differ for
different types of firms. We assume that the cost of writing reviews is a
convex function of the effort. That is, compare the cost of writing the first
manufactured review to the cost of writing the 30th review. While the
first review can reflect the owner’s own authentic writing style, the 30th
review must be dissimilar from the reviews that preceded it in order to
avoid detection by the review-hosting platform. Hence we assume that
∂C(e
i,i
, e
i,j
)
∂e
i,i
> 0,
∂C(e
i,i
, e
i,j
)
∂e
i,j
> 0,
∂
2
C(e
i,i
, e
i,j
)
∂
2
e
i,i
> 0,
∂
2
C(e
i,i
, e
i,j
)
∂
2
e
i,j
> 0
The following assumed simple functional form satisfies these conditions:
C(e
i,i
, e
i,j
) =
δ
i
2
(e
i,i
)
2
+
γ
i
2
(e
i,j
)
2
Here δ
i
signifies the damage caused to the firm i if it caught doing self-
promotion, and γ
i
the damage if it is caught posting negative reviews for
its competitor.
5) Stage IV : Finally, the consumer chooses the product that maximizes her
utility. We assume that the products are horizontally differentiated. We use
a simple Hotelling model of differentiation to model consumer choice, where
firm A is located at x = 0, firm B is located at x = 1, and the consumer at
location x chooses A if
E[q
A
|ˆq
A
] − tx − p
A
≥E[q
B
|ˆq
B
] − t(1 − x) − p
B
(4)
We assume that consumers are uniformly distributed on the interval [0, 1].
Since consumers do not observe the true quality directly, their expected
utility from A and B is inferred from the signals generated from user reviews.
The equilibrium concept here is Perfect Bayesian Nash Equilibrium.
We next solve for the firms’ optimal actions by backward induction. We start
with the consumer’s inference in stage IV. After observing the signals ˆq
A
and ˆq
B
,
the consumers’ posterior beliefs on the firms’ qualities are:
E[q
A
|ˆq
A
] = ˆq
A
− ˆe
∗
A,A
+ ˆe
∗
B,A
(5)
E[q
B
|ˆq
B
] = ˆq
B
− ˆe
∗
B,B
+ ˆe
∗
A,B
(6)
where ˆe
∗
A,A
and ˆe
∗
B,A
are the inferred equilibrium effort levels since the consumer
does not observe the firms’ manipulation activity directly.
36 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
Assuming market coverage, the consumer who is indifferent between the two
products is located at point ˆx, where
ˆx =
1
2
+
E[q
A
|ˆq
A
] − E[q
B
|ˆq
B
] + p
B
− p
A
2t
(7)
Hence, the market shares of firms A and B are ˆx and 1 − ˆx, respectively. This
implies the following profit functions for firms A and B, respectively in stage III:
Π
∗
A,Stage 3
= max
e
A,A
,e
A,B
p
A
E
q
A
,q
B
1
2
+
E[q
A
|ˆq
A
] − E[q
B
|ˆq
B
] + p
B
− p
A
2t
− δ
A
e
2
A,A
2
− γ
A
e
2
A,B
2
!
(8)
Π
∗
B,Stage 3
= max
e
B,B
,e
B,A
p
B
E
q
A
,q
B
1
2
+
E[q
B
|ˆq
B
] − E[q
A
|ˆq
A
] + p
A
− p
B
2t
− δ
B
e
2
B,B
2
− γ
B
e
2
B,A
2
!
(9)
Substituting (5) and (6) into (8) and (9), and taking the expectation, we can
re-write the firms’ maximization problem as the following:
Π
∗
A,Stage 3
= max
e
A,A
,e
A,B
p
A
"
1
2
+
e
A,A
+ e
A,B
− ˆe
∗
A,A
− ˆe
∗
A,B
+ c
A
+ p
B
− p
A
2t
#
− δ
A
e
2
A,A
2
− γ
A
e
2
A,B
2
!
(10)
Π
∗
B,Stage 3
= max
e
B,B
,e
B,A
p
B
"
1
2
−
e
B,B
+ e
B,A
− ˆe
∗
B,B
− ˆe
∗
B,A
+ c
B
+ p
A
− p
B
2t
#
− δ
B
e
2
B,B
2
− γ
B
e
2
B,A
2
!
(11)
where c
A
= −e
B,A
−e
B,B
+ ˆe
∗
B,A
+ ˆe
∗
B,B
and c
B
= −e
A,B
− e
A,A
+ ˆe
∗
A,B
+ ˆe
∗
A,A
.
Proposition 1 below presents the optimal manipulation levels for the firms:
PROPOSITION 1: In stage 3 (after the firms have committed to prices p
A
and
p
B
), the optimal promotional levels are the following:
e
∗
A,A
=
p
A
2δ
A
t
; e
∗
A,B
=
p
A
2γ
A
t
(12)
e
∗
B,B
=
p
B
2δ
B
t
; e
∗
B,A
=
p
B
2γ
B
t
(13)
PROOF:
To solve for the optimal promotional levels, we 1) derive the first order con-
ditions of firm A’s profit function by differentiating Equation (10) with respect
to e
A,A
and e
A,B
and by differentiating Equation (11) with respect to e
B,B
and
e
B,A
, and 2) simultaneously solve the system of the four resulting equations. This
yields a unique solution since
∂
2
Π
∗
A,Stage 3
∂
2
e
A,A
< 0,
∂
2
Π
∗
A,Stage 3
∂
2
e
A,B
< 0,
∂
2
Π
∗
B,Stage 3
∂
2
e
B,B
< 0,
∂
2
Π
∗
B,Stage 3
∂
2
e
B,A
< 0
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 37
The Corollary below summarizes several key results that we will use in our em-
pirical analysis:
COROLLARY 1: The following results are implied by Proposition 1:
1) An increase in the reputational costs of manipulation decreases the intensity
of this activity:
∂e
∗
A,A
∂δ
A
< 0,
∂e
∗
A,B
∂γ
A
< 0,
∂e
∗
B,B
∂δ
B
< 0,
∂e
∗
B,A
∂γ
B
< 0
2) Firms engage in negative manipulation of reviews of their competitors: e
∗
A,B
>
0 and e
∗
B,A
> 0, and this activity increases as the costs of manipulation decrease.
Hence, a firm that is located close to a competitor will have more negative reviews
than a firm has no close competitors (which will have no fake negative reviews),
and this difference will be greater if the competitor has lower costs of manipulation.
Finally, we turn to the effect that review manipulation has on consumer choice.
In the basic model consumer can invert the firm’s problem and perfectly discounts
the amount of manipulation. That is, in equilibrium, e
∗
A,A
= be
∗
A,A
, e
∗
A,B
= be
∗
A,B
,
e
∗
B,B
= be
∗
B,B
, and e
∗
B,A
= be
∗
B,A
. Since fake reviews are perfectly discounted, the
consumer would make the same choices in the current setting where fake reviews
are possible and in one where fake reviews are not possible. Despite the fact that
fake reviews do not affect consumer choices in equilibrium, firms prefer to post
reviews. That is, if the firm chooses not to engage in manipulation, the consumer
who expects fake reviews will think that the firm is terrible.
In the Online Appendix we derive the comparative statics under endogenous
prices.
38 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
REFERENCES
Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber. 2005.
“Selection on Observed and Unobserved Variables: Assessing the Effectiveness
of Catholic Schools.” Journal of Political Economy 113(1): 151–184.
Anand, Bharat N., and Ron Shachar. 2009. “(Noisy) Communication.”
Quantitative Marketing and Economics 5(3): 211–237.
Anderson, Michael L., and Jeremy R. Magruder. 2012. “Learning from the
Crowd: Regression Discontinuity Estimates of the Effects of an Online Review
Database.” The Economic Journal 122(563): 957–989.
Anderson, Simon P., and Regis Renault. 2006. “Advertising Content.”
American Economic Review 96(1): 93–113.
Bagwell, Kyle., and Garey Ramey. 1994. “Advertising and Coordination.”
Review of Economic Studies 61(1): 153–171.
Bennett, Victor Manuel, Lamar Pierce, Jason A. Snyder, and
Michael W. Toffel. 2013. “Customer-Driven Misconduct: How Competition
Corrupts Business Practices.” Management Science 59(8): 1725–1742.
Berger, Jonah, Alan T. Sorensen, and Scott J. Rasmussen. 2010. “Pos-
itive Effects of Negative Publicity: When Negative Reviews Increase Sales.”
Marketing Science 29(5): 815–827.
Blair, Roger D., and Francine Lafontaine. 2005. The Economics of Fran-
chising. Cambridge University Press, Cambridge, United Kingdom.
Chakraborty, Archishman, and Rick Harbaugh. 2010. “Persuasion by
Cheap Talk.” American Economic Review 100(5): 2361–2382.
Chevalier, Judith A., and Dina Mayzlin. 2006. “The Effect of Word of
Mouth on Sales: Online Book Reviews.” Journal of Marketing Research 43: 345–
354.
Chintagunta, Pradeep K., Shyam Gopinath, and Sriram Venkatara-
man. 2010. “The Effects of Online User Reviews on Movie Box Office Perfor-
mance: Accounting for Sequential Rollout and Aggregation Across Local Mar-
kets.” Marketing Science 29(5): 944–957.
Crawford, Vincent P., and Joel Sobel. 1982. “Strategic Information Trans-
mission.” Econometrica 50(6): 1431–1451.
Dellarocas, Chrysanthos. 2006. “Strategic Manipulation of Internet Opin-
ion Forums: Implications for Consumers and Firms.” Management Science
52(20): 1577–1593.
Dellarocas, Chrysanthos, and Charles A. Wood. 2007. “The Sound of Si-
lence in Online Feedback: Estimating Trading Risks in the Presence of Reporting
Bias.” Management Science 54(3): 460–476.
Dellavigna, Stefano, and Eliana La Ferrara. 2010. “Detecting Illegal Arms
Trade.” American Economic Journal 2(4): 26–57.
VOL. VOL NO. ISSUE AN EMPIRICAL INVESTIGATION 39
Duggan, Mark, and Steven D. Levitt. 2002. “Winning Isn’t Everything:
Corruption in Sumo Wrestling.” American Economic Review 92(5): 1594–1605.
Dziuda, Wioletta. 2011. “Strategic Argumentation.” Journal of Economic The-
ory 146(4): 1362–1397.
Harmon, Amy. 2004. “Amazon Glitch unmasks War of Reviewers.” New York
Times February 14: A1.
Horstmann, Ignatius J., and Sridhar Moorthy. 2003. “Advertising Spend-
ing and Quality for Services: The Role of Capacity.” Quantitative Marketing
and Economics 1(3): 337–365.
Hu, Nan, Indranil Bose, Noi Sian Koh, and Ling Liu. 2012. “Manipulation
of online reviews: An analysis of ratings, readability, and sentiments.” Decision
Support Systems 52: 674–684.
Jacob, Brian A., and Steven D. Levitt. 2003. “Rotten Apples: An Investiga-
tion of the Prevalence and Predictors of Teacher Cheating.” Quarterly Journal
of Economics 118(3): 843–877.
Ji, MinWoong, and David Weil. 2009. “Does Ownership Structure Influence
Regulatory Behavior? The Impact of Franchising on Labor Standards Compli-
ance?” Boston U. School of Management Research Paper No. 2010-21.
Jindal, Nitin, and Bing Liu. 2007. “Review Spam Detection.” WWW 2007
Proceedings of the 16th International Conference on World Wide Web, 1189 –
1190.
Jin, Ginger Zhe, and Phillip Leslie. 2009. “Reputational Incentives for
Restaurant Hygiene.” American Economic Journal: Microecnomics 1(1): 236–
267.
Kamenica, Emir, and Matthew Gentzkow. 2011. “Bayesian Persuasion.”
American Economic Review 101(October): 2590–2615.
Kihlstrom, Richard E., and Michael H. Riordan. 1984. “Advertising as a
Signal.” Journal of Political Economy 92(3): 427–450.
Kornish, Laura. 2009. “Are User Reviews Systematically Manipulated? Evi-
dence from the Helpfulness Ratings.” Leeds School of Business Working Paper.
Li, Xinxiin, and Lorin H. Hitt. 2008. “Self-Selection and Information Role of
Online Product Reviewsl.” Information Systems Research 19(4): 456–474.
Luca, Michael. 2012. “Reviews, Reputation, and Revenue: The Case of
Yelp.com.” Harvard Business School Working Paper No. 12-016.
Mayzlin, Dina. 2006. “Promotional Chat on the Internet.” Marketing Science
25(2): 155–163.
Mayzlin, Dina, and Jiwoong Shin. 2011. “Uninformative Advertising as an
Invitation to Search.” Marketing Science 30(4): 666–685.
Melnik, Mikhail I., and James Alm. 2002. “Does a seller’s ecommerce repu-
tation matter? Evidence from ebay auctions.” Journal of Industrial Economics
50(3): 337–349.
40 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2013
Milgrom, Paul, and John Roberts. 1986. “Price and Advertising Signals of
Product Quality.” Journal of Political Economy 94(August): 796–821.
Mukherjee, Arjun, Bing Liu, and Natalie Glance. 2012. “Spotting Fake
Review Groups in Consumer Reviews.” International World Wide Web Confer-
ence Committee April 16-20.
Nelson, Phillip. 1974. “Advertising as Information.” Journal of Political Econ-
omy 82(4): 729–754.
O’Fallon, Michael J., and Denney G. Rutherford. 2010. Hotel Management
and Operations. John Wiley and Sons, Hoboken, NJ.
Ott, Myle, Yejin Choi, Claire Cardie, and Jeffrey T. Hancock. 2011.
“Finding Deceptive Opinion Spam by Any Stretch of the Imagination.” Proceed-
ings of the 49th Annual Meeting of the Association for Computational Linguis-
tics, 309–319.
Pierce, Lamar, and Jason Snyder. 2008. “Ethical Spillovers in Firms: Evi-
dence from Vehicle Emissions Testing.” Management Science 54(11): 1891–1903.
Resnick, Paul, and Richard Zeckhauser. 2002. “Trust Among Strangers in
Internet Transactions: Empirical Analysis of eBay’s Reputation System.” In The
Economics of the Internet and E-Commerce, Volume 11. , ed. Michael R. Baye.
Elsevier, Amsterdam, Holland.
Resnick, Paul, Richard Zeckhauser, John Swanson, and Kate Lock-
wood. 2006. “The value of reputation on eBay: A controlled experiment.” Ex-
perimental Economics 9(2): 79–101.
Scott, Susan V., and Wanda J. Orlikowski. 2012. “Reconfiguring relations
of accountability: Materialization of Social Media in the Travel Sector.” Ac-
counting, Organizations and Society 37: 26–40.
Vermeulen, Ivar E., and Daphne Seegers. 2009. “Tried and Tested: The
Impact of Online Hotel Reviews on Consumer Consideration.” Tourism Man-
agement 30(1): 123–127.
Ye, Qiang, Rob Law, Bin Gu, and Wei Chen. 2010. “The influence of user-
generated content on traveler behavior: An empirical investigation on the effects
of e-word-of-mouth to hotel online bookings.” Computers in Human Behavior
27(2): 634–639.
