Do Online Reviews Affect Product Sales? The Role of Reviewer Characteristics and Temporal
School of Information Systems
Singapore Management University
Department of Accounting & Information Management
School of Management
University of Texas at Dallas
Department of Information Systems and Operation Management
College of Business Administration
University of Texas at Arlington
* Corresponding Author
+ ACKNOWLEDGMENTS: We would like to thank guest editors Paul Tallon, Indranil Bardhan, and Alok Gupta
and the anonymous reviewers for their valuable feedback on earlier versions of this manuscript. All remaining errors
and omissions are our responsibility.
+ Information Technology and Management, Vol 9, Issue 3, 2008
Do Online Reviews Affect Product Sales? The Role of Reviewer Characteristics and Temporal
Online product reviews provided by consumers who previously purchased products have become
a major information source for consumers and marketers regarding product quality. This study extends
previous research by conducting a more compelling test of the effect of online reviews on sales. In
particular, we consider both quantitative and qualitative aspects of online reviews, such as reviewer
quality, reviewer exposure, product coverage, and temporal effects. Using transaction cost economics and
uncertainty reduction theories, this study adopts a portfolio approach to assess the effectiveness of the
online review market. We show that consumers understand the value difference between favorable news
and unfavorable news and respond accordingly. Furthermore, when consumers read online reviews, they
pay attention not only to review scores but to other contextual information such as a reviewer’s reputation
and reviewer exposure. The market responds more favorably to reviews written by reviewers with better
reputation and higher exposure. Finally, we demonstrate that the impact of online reviews on sales
diminishes over time. This suggests that firms need not provide incentives for customers to write reviews
beyond a certain time period after products have been released.
Keywords: Word-of-Mouth, Online Product Reviews, Transaction Cost Economics, Uncertainty
Reduction, Efficient Market, Portfolio Analysis
Word-of-mouth (WOM) communication is considered a valuable marketing resource for consumers and
marketers and a reliable and effective metric for measuring customer loyalty with critical implications for
a product’s success. WOM communication includes all forms of information exchange among consumers
regarding the characteristics and usage of particular products, services, or vendors. It is widely considered
to be a major driver for the diffusion of new products and services [3, 6, 12].
Online product reviews have become a major informational source for consumers due to the fast
spread of WOM1 communication through the Internet. Reichheld  claims that a customer’s propensity
to recommend a product to others – termed “referral value” – is the most important success measure in
business today. Reichheld  argues that referral value may predict firm performance even better than
traditional measures such as customer satisfaction. Hence, online product reviews have fundamental
implications for management activities such as reputation building and customer acquisition.
Previous research has studied the impact of online product reviews on product sales with a variety
of regression models [14, 15, 26]. This stream of literature provides useful insights by linking online
reviews with sales; most of the studies show a positive correlation between the average review score and
product sales. However, one implicit assumption in these studies is that consumers consider only the
scales of review scores when they make a purchase decision. To the best of our knowledge, none of the
previous studies considered other informational aspects of online reviews such as the quality reputation of
a reviewer and his or her exposure to the online community (the number of times a reviewer’s name is
exposed to the public), the information environment or the age of a product. The latter variables are more
compelling measurements of the information content of online reviews because they are directly related to
the intrinsic quality of the reviews in terms of reliability and trustworthiness. Further, in trying to
1 Although the phrase “word-of-mouth” generally refers to oral communication, in this paper we are using this term
to refer to person-to-person virtual communication.
understand the effectiveness of online reviews, it remains unclear what role time periods play in affecting
sales. That is, should firms like Amazon.com encourage buyers to provide reviews for all items or only
for newly released items?
This paper extends previous research by linking changes in online review scores to changes in
sales while considering other important dimensions of online reviews such as the quality and exposure of
a reviewer, the information environment, and the age of a product. We use a “market reaction” lens to
assess the effectiveness of online reviews. Treating the online review environment as a “market,” we
argue that online reviews are like market signals that contain information about the quality of an item.
This analogy helps us to use a portfolio methodology, typically used in the finance literature, to assess the
effectiveness of the online reviews. We show that consumers use the information embedded in online
reviews to reduce the uncertainty involved in purchase decisions, thereby enabling them to choose the
item with the lowest transaction cost.
The paper makes three primary contributions to the research literature. First, we show that the
online review market behaves as an “efficient market” that understands the value difference between
favorable news and unfavorable news and responds accordingly. By demonstrating the effectiveness of
the online review market, this research sensitizes managers to the importance of improving the underlying
quality of items for sale and investing the necessary effort in managing customer expectations and
reactions to products. Second, we show that consumers pay attention to elements other than review
scores, such as reviewer quality, reviewer exposure, and product coverage. The market is more responsive
to a review written by someone with a better reputation and more exposure, while it responds less to an
item that is more extensively covered by reviewers; prior research that ignored these dimensions may
potentially overestimate the impact of review scores. Our paper explicitly considers these factors and
demonstrates the roles they play. Consequently, it provides a more complete understanding of how online
reviews influence the sale of an item. Third, over time, online reviews do not affect sales equally. The
relationship between online reviews and sales depends on the “age” of a product; the longer an item has
been on the market, the smaller the impact online reviews will have on its sales.
The paper proceeds as follows. Section 2 summarizes the related literature. Section 3 describes
the theoretical framework and research hypotheses. Section 4 describes the research setting and
methodology. Section 5 provides data analysis and results. Section 6 contains a discussion of the findings,
their implications, and some concluding remarks.
2. Literature Review
While word-of-mouth has been studied extensively in the marketing literature, it is only recently that
online product reviews have begun to draw the attention of marketing and information systems
researchers. We summarize a cross-section of research in Table 1. As stated above, previous studies focus
on the quantitative aspects of online reviews by linking the level of online reviews to the level of sales. In
this study, we link the change in reviews to the change in sales by considering both quantitative and
qualitative aspects of online reviews.
Author(s) Data sources Findings
and Ravid (2003)
200 films released
between late 1991
and early 1993 from
Baseline Services in
- Both positive and negative reviews are correlated with
weekly box office revenues over an eight-week period.
However, the impact of negative reviews (but not that of
positive reviews) diminishes over time.
- Negative reviews hurt more than positive reviews help
box office performance, but only in the first week of a
website - WOM information offers significant explanatory
power for both aggregate and weekly box office
revenue, especially in the early weeks after opening.
- However, as measured by the percentages of positive
and negative messages, most of this explanatory power
comes from the volume of WOM, not its valence.
Godes and Mayzlin
Viewership data from
Nielsen ratings and
observed in Usenet
The dispersion of conversations about weekly TV shows
across Internet communities is positively correlated with
the evolution of viewership for these shows.
Box office sales data
from Baseline, Inc.
Critical reviews correlate with late and cumulative box
office receipts but do not have a significant correlation
with early box office receipts.
and user review data
collected from the
public web sites of
- Reviews are overwhelmingly positive at both sites.
- An improvement in a book’s reviews leads to an
increase in relative sales at that site.
- The impact of 1-star reviews is greater than the impact
of 5-star reviews.
Clemons, Gao, and
Sales data from the
craft beer industry
and review data from
The variance of ratings and the strength of the most
positive quartile of reviews play a significant role in
determining which new products grow fastest in the
and Zhang (2004)
User reviews posted
on Yahoo! Movies
A newly-derived revenue forecasting model that
incorporates the impact of both publicity and word-of-
mouth on a movie’s revenue trajectory predicts the
movie’s total revenues accurately.
Duan, Gu, and
- Box office sales are significantly influenced by the
number of online postings.
- Ratings of online user reviews have no significant
impact on box office sales.
Survey Consumers who are more familiar with a specific
retailer are less likely to be affected by negative reviews
of that retailer.
Chen, Wu, and Yoon
Review and sales
More recommendations are associated with higher sales,
while consumer ratings are not found to be related to
Chen, Fay, and
and J.D. Power &
Controlling for price and quality, number of online
postings is positively related to automobile sales.
Hu, Pavlou, and
A field study and data
The most satisfied and the most disgruntled consumers
are the most likely to post reviews. Therefore, the
average rating may not be a fair evaluation of the
Table 1. A Summary of the Related Literature.
3. Theoretical Background and Development of Hypotheses
Building on earlier work by Nobel prize winner Ronald Coase, Williamson  developed the theory of
Transaction Cost Economics (TCE). TCE specifies variables (asset specificity, uncertainty, and
transaction frequency) that determine why a certain transaction is conducted in a particular form (market,
hierarchy, or hybrid) and whether the market or the hierarchy has a lower transaction cost under the two
main assumptions of human behavior (bounded rationality and opportunism). Williamson argues that
firms will choose a channel that minimizes their total cost, which is comprised of both production and
transaction costs. Transaction costs occur because decision makers have limited cognitive processing
power and cannot consider all possible scenarios (bounded rationality). Also, people may not be truthful
about their intentions all the time and may act in a self-interested manner to take advantage of unforeseen
circumstances (opportunism attributable to information asymmetry).
TCE has been successfully used to analyze issues such as internal organization, vertical
integration and contracting, resource allocation, outsourcing decisions, etc. In the area of E-commerce,
researchers have adopted TCE to explain both firm-level and individual-level issues. For example, Liang
and Huang  proposed that consumers will choose a channel that has a lower transaction cost in
deciding whether to buy from online stores or traditional stores. TCE is also a viable theory for explaining
online consumer behavior.
When consumers decide which items to purchase on a given E-commerce website, they must go
through a transaction process. It starts with searching for relevant products, followed by comparing prices,
evaluating product quality, ordering, delivering, and post-sales services such as customer service and
support. Online transactions of experience goods can involve product, process, and psychological
uncertainties because the product descriptions might not provide sufficient information and the quality of
the product can only be evaluated after trying or inspecting it. Product uncertainty refers to the situation in
which consumers find after consuming a product that what they bought is different from what they
perceived it to be at the shopping stage. Process uncertainty refers to the case in which consumers
purchase products from undesired vendors, while psychological uncertainty refers to all of the emotional
costs associated with the uncertainty. Overall, uncertainty refers to the costs associated with unexpected
outcomes tied to information asymmetry. Therefore, a higher level of uncertainty implies a higher
transaction cost, which will result in lower sales.
The goal of an E-commerce participant is to identify the intrinsic quality of a product based on all
available information, and then to purchase the product with the lowest transaction cost or with the lowest
uncertainty. To begin, a consumer may or may not possess any prior quality information about the
product, and may or may not have previously conducted business with the online vendors involved. In
such a scenario, there are both financial and psychological uncertainties associated with the product and
the online vendors. According to Uncertainty Reduction Theory , whenever consumers lack knowledge
of a product or of the outcomes of consuming that product, they will engage in uncertainty reduction
efforts to mitigate and eliminate the risk associated with the uncertainty and to maximize the outcome
value. Consumers can reduce the quality uncertainty by drilling down to obtain more details about the
product’s author, publishers, and subject. Consumers can then try to understand the returns policy and
product warranty to further reduce uncertainty. For search goods, consumers may stop here because they
are already informed about the value of the products. However, for experience goods, product uncertainty
may still be high. To reduce this uncertainty, consumers will actively seek other information, such as
online reviews written by previous customers. Overall uncertainty reduction theory provides a framework
through which we can understand how individuals use different online information, such as online
reviews, to: 1) infer product quality; 2) reduce product uncertainty; and 3) make a final purchase decision.
In this study, we investigate how consumers utilize online reviews to reduce the uncertainties
associated with online purchases. Figure 1 provides an overview of the key conceptual constructs that we
examine in this study. As Figure 1 illustrates, we focus on three sources that may influence a consumer’s
interpretation of online reviews and subsequent purchase decisions. We discuss each one in detail in the
Insert Figure 1 about here
3.1 The Information Content of Online Reviews
Extending the market metaphor to online reviews, we suggest that when consumers purchase experience
goods such as music CDs through the Internet, they first form a quality evaluation based on the
combination of product information, their own personal tastes, and recommendations from friends or
relatives. Due to the nature of experience goods, they will read reviews written by previous customers to
help determine the value and quality of a product and to reduce the uncertainty associated with consuming
that product. Reduced uncertainty should result in decreased transaction costs. Out of all the products that
meet a consumer’s requirements, the consumer will then select the one with lowest transaction cost.
Online reviews written by previous customers provide information about an item’s perceived
value. These reviews are helpful for making purchase decisions because they provide new customers with
indirect experiences and help prospective customers reduce the uncertainties involved in inferring product
quality. Product quality, which is the aggregate of all consumers’ perceived values, reflects a product’s
In this paper, we use the term favorable news if a newly released review for any single item is
better than its prior average (prior consensus reviews). On the other hand, if the newly released review is
worse than the prior average for that product, we call it unfavorable news. Both types of news can change
consumers’ expectations about product quality. Favorable news may convert a consumer from “not
buying” to “buying” because it reduces the quality uncertainty and, hence, the total transaction cost for a
new customer. On the other hand, unfavorable news may convert a consumer from a potential “buying”
consumer to a “not buying” one. In other words, favorable news and unfavorable news contain different
information about product quality.
Hypothesis 1: Products with favorable reviews enjoy better sales than products with unfavorable
In this study, we investigate whether the marginal change in sales associated with the favorable
news group exceeds that of the unfavorable news group. This is a more conservative test than testing the
impact of each group of news on sales separately, in that the former method controls for the potential
differences in risk-return relations for items included in our study. By controlling for the unobserved
heterogeneity of items, we can ensure that our results are less likely to be distorted by sample selection
3.2 The Role of Reviewer Quality
When consumers read online reviews, they will not limit themselves to the numerical scores alone.
Consumers are likely to pay attention to reviewer credibility as well. To some degree, online reviews are
not verifiable and may not be objective and credible to potential customers. Consumer reviews are user-
generated and they measure product quality and valuation from a user’s perspective . Review scores
are based more on reviewers’ own experiences rather than on underlying characteristics of the product. In
such cases, the reviews should have limited influence on other consumers’ evaluations because consumers
might think the reviewers have not provided unbiased quality assessments for the product. In other words,
not all reviews have the same influence on consumers and consumers might selectively pay attention to
the reviews written by reviewers with better quality reputations because such reviews are more
trustworthy and reliable.
Trust can be defined as the expectation that an engaging partner will forgo short-term outcomes
obtained through opportunistic behavior even when there is uncertainty about long-term benefits .
Chiles and McMackin  examined ways to incorporate trust and reputation into TCE. The honoring of
moral obligations generates trust, and trust leads to the constraining of opportunistic behavior by way of
reputation . An entity builds its reputation by consistently engaging in trustworthy behavior. Trust
reflects all of the historical trustworthy behaviors exerted by the entity and is a strong signal of reliability
to third parties, no matter whether they have or have not conducted transactions with the entity before
Without trust, information-exchanging parties need to constantly monitor the information being
provided to guard against opportunistic behavior. Trust alleviates the monitoring and safeguarding costs
associated with a contract because each party believes that the other party will act in a proper way to
generate long-term benefits. Reputation about such trustworthiness decreases the cost of finding a
contract partner . Trust and reputation will thus lead to reduced behavioral uncertainty and decreased
transaction costs because trust in a contractual relationship can result in more accurate and timely
exchange of information and greater influence on the information receiver.
In an online review environment, there is enormous information asymmetry between online
reviewers and new customers. Consumers may be inclined to give more weight to reviews written by
reviewers with higher quality reputations because they perceive these reviews to be more credible and
trustworthy. Reviewers with better reputations will help decrease a product’s quality uncertainty because:
1) the market has previously found that these reviewers have the necessary expertise to assess product
quality; and 2) they are less likely to engage in opportunistic behavior such as accepting payment from
vendors for writing fake reviews that simply promote product sales. Thus, consumers might ignore the
reviews written by lower quality reviewers because consumers perceive that the background and
motivation of these reviewers prevents them from writing high quality reviews.
Hypothesis 2: The difference in sales between favorable news and unfavorable news is significantly
different from zero for reviews written by higher quality reviewers, but not for reviews written by
lower quality reviewers.
3.3 The Role of Reviewer Exposure
To some degree, reviews written by consumers on E-commerce websites are similar to reports written by
analysts about market securities. The former expresses a reviewer’s evaluation of product quality, the
latter reveals an analyst’s assessment of a company’s valuation. In the finance literature, analyst quality is
measured in two ways: analyst reputation [7, 20, 25, 34] and analyst exposure to the community .
Analysts’ reputation and exposure affect both the information content of the signals they send to the
market and the efficiency of the price discovery process for a market security.
Prior studies show that superior analysts elicit stronger market responses for their forecast
revisions because their reputations affect the way that market participants perceive those forecast
revisions . In the analyst forecast literature, Bonner, Hugon, and Walther  documented that market
participants react more strongly to forecast revisions issued by celebrity analysts (i.e., analysts with
greater media exposure). Following Boorstin , they defined a celebrity as a famous person who is
known for his name recall instead of performance-related qualities.
Conceptually, exposure is different from quality reputation. Exposure here refers to media
exposure of a reviewer in the online review community. It can be measured by how many times a
reviewer writes reviews on an online community website. In addition to being influenced by higher
quality reviewers, consumers may pay more attention to higher exposure reviewers for reasons similar to
those outlined above. Because consumers might ignore the reviews issued by reviewers with lower
exposure, favorable (unfavorable) news written by such reviewers might not change the uncertainties
associated with the consumption of a product or consumers’ transaction costs for buying such products.
Thus, favorable news might not solicit a different market response from unfavorable news if written by
Hypothesis 3: The difference in sales between favorable news and unfavorable news is significantly
different from zero for reviews written by higher exposure reviewers, but not for reviews written by
lower exposure reviewers.
3.4 The Role of Product Coverage and Age of an Item
Empirical research shows that security price reactions to unanticipated information conveyed to the
market by actual earnings and earning forecasts are more substantial for smaller firms because the amount
of private pre-disclosure information is an increasing function of firm size [1, 25]. For items with lower
product coverage, that is, items with a smaller number of reviewers, there is an often limited amount of
quality information about that item other than online reviews written by these reviewers. Therefore,
reviewers play a very important role in terms of informing consumers of product quality and reducing
uncertainty for such products. Each new reviewer might reveal additional product quality information to a
new customer. The incremental impact of the reviews issued by a reviewer will be bigger when an item
has fewer pre-existing reviewers that covered it before. After a product receives a critical mass of
reviewers, new reviewers generally disseminate only a limited amount of new information. Thus, a new
reviewer can not significantly reduce the uncertainty and has little or no impact on the transaction cost
associated with buying that product.
One way of characterizing the product information environment is by counting the total number of
reviewers that have commented on that product, which is similar to the number of analysts covering a
firm. We classify products into two categories: high-coverage products and low-coverage products. High-
coverage products are products whose total number of reviewers is above the median of our sample in a
given batch; low-coverage products are ones whose total number of reviewers is below the median of our
sample in a given batch.
Hypothesis 4: The difference in sales between favorable news and unfavorable news is significantly
different from zero for reviews issued about low-coverage products but not for reviews issued about
Besides product coverage, another factor that may affect the impact of online reviews on sales is
the age of the item, that is, how long an item has been selling on the market. Although product coverage
is likely to co-vary with the age of an item (a product is likely to have more reviewers if it has been on the
market longer), conceptually these are distinct notions. While age refers to the time period an item has
been in existence, product coverage refers to the number of reviewers for that item. In the initial phase of
a product’s introduction, there is a limited number of sources of product quality information. Hence, the
market is likely to rely heavily on online reviews for purchase decisions during this time. However, as the
market gains experience with the product with the passage of time, consumers can obtain product
information from other sources such as recommendations from friends, newspapers, and magazine
comments. Therefore, we posit that online reviews will have greater impact in the initial phase of a
product lifecycle than in later phases. In other words, the impact of online reviews on sales will decrease
Hypothesis 5: The market’s reaction to favorable news and unfavorable news is significantly
different from zero ONLY for a newly released product; as the age of a product increases, the difference
will fall to zero.
4. Method and Measurement Development
We collected our data from Amazon.com’s Web Service (AWS). These data allow us to examine the
effectiveness of online reviews and how consumers react to contextual elements such as reviewer quality,
reviewer exposure, product coverage, and the impact of online reviews on sales over time. A panel of
books, DVDs, and videos was randomly chosen in July 2005. We used panel data because compared with
cross-sectional data, panel data are more suitable for studying the dynamics of adjustments because they
control for unobserved heterogeneity [2, 10]. For each item, we collected its price, sales, and review
information for several months at approximately three-day intervals. We identified each session by a
unique sequence number. Because of some technical glitches in AWS, we had to exclude certain
sequences in which only partial data were collected. For example, during several sessions, AWS did not
respond to our queries or was offline and we were therefore only able to process partial or no data during
these sessions. In total, we obtained 26 batches of review and item-level data.
Table 2 provides summary statistics for our panel data. The data include some very popular books,
such as The World Is Flat: A Brief History of the Twenty-first Century by Thomas L. Friedman (sales rank
fluctuates between No. 1 to No. 7), Freakonomics: A Rogue Economist Explores the Hidden Side of
Everything by Steven D. Levitt (sales rank fluctuates between No. 3 to No. 11); popular DVDs, such as
The Simpsons (sales rank fluctuates between 26 to 236) and Star Wars (sales rank fluctuates between 27 to
141); and videos, such as Shall We Dance, Cinderella and John Wayne: American Hero: The John Wayne
Story. On Amazon.com, consumers can only report an integer product review on a 1-star to 5-star scale,
where 1-star = least satisfied and 5-star = most satisfied. The average review scores for books, DVDs, and
videos are 3.87, 4.07, and 4.02, respectively. This observation is consistent with Chevalier and Mayzlin’s
 finding that for books in both Amazon.com and BarnesandNoble.com websites, product reviews are
Insert Table 2 about here
4.2 The Portfolio Approach
We adopt a portfolio approach to our investigation of whether customers of Amazon.com understand the
difference between favorable news and unfavorable news and respond accordingly. The meaning of a
portfolio in this context is different from a traditional finance context, where a portfolio represents a
basket of securities, typically designed to reduce risk. Here our portfolio comprises products and events
(favorable and unfavorable) that share similar characteristics. Our favorable (unfavorable) news group
includes events where a newly released review for a product has a higher (lower) score than its previous
average review score. Conceptually this is similar to Sloan’s  study where he tested whether market
valuations incorporated fully the information provided by different earnings components. Our method can
also be viewed as a variation of matching sample techniques where variables of interest across treatment
(in our case favorable news group) and control groups (in our case unfavorable news group) are
compared. Matching sample techniques have been widely used across different disciplines including
psychology, economics, and management science [23, 27, 29, 32].
We next define how we measure the change in sales, reviewer quality, reviewer exposure,
product coverage, and the age of an item.
4.3. Sales Change
There are two types of events of interest in this study: favorable news events and unfavorable news
events. A favorable (unfavorable) news event occurs when a newly released review for an item has a
higher (lower) score than the previous average review score for that item. We are interested in knowing
whether customers at Amazon.com understand the information value difference between favorable news
and unfavorable news and respond accordingly by either buying or refusing to buy the product.
Because we can precisely pinpoint the review date of each item, we limit our event window to a
starting point (day 0) to estimate the sales change associated with favorable (unfavorable) news. We start
at day 0 because it is unlikely that a consumer discloses a review to the Amazon community before it is
actually posted online.
In the Amazon market, even when there is no newly released review for a product, the sales of
that product still fluctuate. In order to compute the actual marginal change in sales associated with
favorable news or unfavorable news, we need to adjust the actual change in sales by a market
performance factor and associated risk factors, as is common in portfolio approaches in the finance
literature. We, therefore, adopt a variation of the Fama and French  model to adjust for the overall
performance of the Amazon.com marketplace and for risk factors that might affect the sales of individual
items. Fama and French  use the average return from a benchmark portfolio, using size (market
equity) and book-to-market (the ratio of book equity to market equity) to adjust actual firm returns and
produce a measure of abnormal return. In the context of Amazon.com, we believe that each item has some
“normal” changes that are driven by the product sub-category and its list price. These are the factors that
can explain the cross-sectional variance of expected “normal” sales change. Any extra changes over and
above the “normal” changes are classified as “abnormal.”
We estimate the average normal change in sales for a benchmark portfolio of products comprising
all products within the same product sub-category and with similar Amazon list prices. For product sub-
categories, we use the classification scheme provided by Amazon.com. For example, within the book
category, sub-categories include history, children, diet, etc. The difference between the change in actual
product sales and the change in sales of the benchmark portfolio signifies the abnormal sales related to
We describe below how we estimate the abnormal sales associated with each review event:
1) In step one, we estimate the change in sales by subtracting sales at time t-1 from sales at time
t (i.e., actual Sales Change t = Sales t – Sales t-1 ).
2) In step two, for every data collection batch, we estimate the average change in normal sales
for each benchmark portfolio formed based on product sub-categories and the Amazon.com
3) Finally, we compute the “abnormal” sales at time t by subtracting the figure obtained in step
two from that in step one (i.e., the Abnormal Salest = Actual Sales Changet – Average
Change in Sales for the Benchmark Portfolio t ).
Instead of providing the actual sales number, Amazon.com provides the sales rank information of
the item. Product sales rank is shown in descending order where 1 represents the best selling product.
Consequently, there is a negative correlation between product sales and sales rank. We use SalesRank as a
proxy for product sales (with the opposite sign). Henceforth, unless stated differently, whenever we refer
to change in sales, it represents an “abnormal” change in sales rank.
4.4 Reviewer Quality
Reviewer quality is measured by the overall quality of the reviews written by reviewers. For each review
posted online, Amazon.com also reveals how many customers read it and how many consider it “useful.”
To assess the quality of a reviewer, we retrieve all the reviews ever written by that reviewer using AWS.
Then, we estimate the mean of the number of “useful” votes divided by the number of total votes of all
the reviews ever written by that reviewer.
Based on this measure, we classify a reviewer into a high-quality or low-quality group. High-
quality reviewers are those whose average up-to-date quality score is above the median, while the low-
quality reviewers are those whose average up-to-date quality score is below the median.
4.5 Reviewer Exposure
To test the conjecture that consumers indeed pay more attention to high-exposure reviewers, one can
classify reviewers into two categories – higher exposure and lower exposure. Higher exposure reviewers
are those whose up-to-date total number of reviews is above the median for our sample, while the lower
exposure category includes those whose total number of reviews is below the median.
Due to limitations in AWS, we could not get reviewer quality and reviewer exposure information
for all items. Hence, from the panel data, we select a sample of items for which we are able to get the
necessary reviewer quality information. To make sure these items are representative of the original panel
data, we compare the means of the average rating, reviewer quality, reviewer exposure, and product
coverage for this group to those of the panel data. Our analysis shows that there is no significant
difference between these two groups.2
4.6 Product Coverage
Product coverage measures the total number of consumers that have reviewed a product. A high-coverage
(low-coverage) product is a product whose total number of reviewers is above (below) the median of our
4.7 Age of an item
2 We also evaluate hypothesis 1 using all of the panel data as well as the sub-sample. These analyses yield consistent
results, as we discuss later.
To examine the dynamics of review scores, we first define a concept of “age” for an item, which is the
number of days between the publishing date of an item and our data collection date. Then for each item,
we divide all the available reviews into three stages of equal duration based on the age of the item. Stage
1 is the earliest stage starting right after an item is released to the Amazon market; while stage 3 is the
most recent period. We choose the relative age instead of absolute age because each item sold on Amazon
has its own release date and, therefore, its own absolute age. This absolute age varies from several months
to several years with a very large variance. Thus, it is difficult to compare the review scores based on an
absolute age of an item. Since we are interested in the temporal properties of online reviews, using the
relative age of the item allows us to pool items with different absolute ages together under the assumption
that reviews of different items have similar trends over each stage.
5.1 Results: Are Changes in Online Reviews Associated with Changes in Sales?
We find support for hypothesis 1, which argues that sales respond differently to “favorable” versus
“unfavorable” reviews. Panel A of Table 3 presents the results of our analysis of differences in abnormal
changes in sales rank when we pool books, DVDs, and videos together. The change in the mean sales
rank for unfavorable news exceeds that of favorable news by 1196.4 (p-value = 0.06).3
Insert Table 3 about here
Panel B of Table 3 presents the results for different product categories. For books, the difference
in mean abnormal sales for favorable news events and unfavorable news events is insignificant
(difference in sales rank = 3810, p-value = n.s.). However, for DVDs, unfavorable news increases the
sales rank of an item by 165.37, while favorable news decreases the sales rank by 195.70. The difference
in sales rank associated with unfavorable news and favorable news is 361 (p-value = 0.035). For videos,
3 Recall that we use sales rank to approximate sales. Sales rank is a function of actual sales so that an increase in
sales rank is associated with decreasing sales.
unfavorable news increases the sales rank by 301.05, while favorable news decreases the sales rank by
236.60. Similarly, for videos, the difference in sales rank of 537 between those items that have received
unfavorable versus favorable news is statistically significant (p-value = 0.015).
As a way to assess the robustness of our results, we also estimate the difference between
unfavorable and favorable news using Wilcoxon Z-statistics4. The results based on splitting our sample at
the median are consistent with results obtained using means. Overall, except for the book category, the
sales of the unfavorable news group decreased, while sales of the favorable news group increased. In
addition, the difference in sales between the favorable and unfavorable news groups is significantly
different from zero, consistent with hypothesis 1.
5.2 Results: Role of Reviewer Quality in Sales Change
Because reviewer quality information is not available for the entire panel, we use a sub-sample of the
panel data to evaluate hypotheses 2, 3, and 4. We first validate whether the sales change difference
between favorable and unfavorable news holds in this reduced sub-sample. Panel A of Table 4 presents
the results of our analysis of the differences in sales change between favorable and unfavorable news for
the sub-sample panel. As before, the market reacts to favorable and unfavorable news differently; on
average, the market responds more favorably to favorable news. The difference in the average change in
sales for unfavorable and favorable news events is 6312.7 (p-value < 0.01).
Insert Table 4 about here
Panel B of Table 4 presents the results disaggregated by reviewer quality. We find support for
hypothesis 2, which predicts that the impact of reviews on sales between favorable and unfavorable news
is different when the reviews were written by higher quality reviewers. We find that for books, DVDs,
and videos, the difference in mean sales for the unfavorable and favorable news events is 8950 for the
4 The skewness of the change in sales is approximately -0.60; the kurtosis is approximately 7.38, which might also
indicate that the data are slightly skewed.
higher quality reviewer group (p < 0.001), while the difference is not significantly different from zero (p-
value = 0.129) for the lower quality group.5 This shows that consumers react to favorable and unfavorable
news differently when the review is written by a higher quality reviewer, but consumers feel indifferent
between favorable and unfavorable news when the review is provided by a reviewer of lower quality.
5.3 Results: The Role of Reviewer Exposure in Changes in Sales
We find support for hypothesis 3, which predicts that consumers will react differently to favorable and
unfavorable news depending on whether the review is written by a higher or lower exposure reviewer.
Panel C of Table 4 shows that consumers value unfavorable and favorable news differently when the
review is written by a higher exposure reviewer (difference = 7712.2, p-value = 0.003), but that
difference is only moderate for a lower exposure reviewer (difference = 4876, p-value = 0.075).
5.4 Results: The Role of Product Coverage in Changes in Sales
We present the results for the effect of product coverage in Panel D of Table 4. We find support for
hypothesis 4 that predicts that consumers value favorable and unfavorable news differently when the
items have lower product coverage. Under a low-coverage scenario, a newly created review is more likely
to reveal additional information and change consumers’ quality expectations regarding a product.
However, when there are many pre-existing online reviewers, a new review, regardless of its favorable or
unfavorable content, is unlikely to introduce enough additional information to change consumers’
Because tabular classifications do not control for other item characteristics and potential
interaction effects that can affect expected sales, we specify the following model and estimate the results
using multiple regression.
5 The median Z-test results are very similar to the mean difference results.
In this model, the dependent variable is the change in sales. To capture the effect of the level of
reviewer quality and the types of reviews (favorable or unfavorable), we define a categorical variable
called Signal with the following values: +1 represents favorable news written by a high quality reviewer,
0 is a review written by a low-quality reviewer, and -1 represents unfavorable news written by a high
quality reviewer. The signal variable represents a qualitative feature of the signal sent to the Amazon.com
market by the reviews. To assess the impact of a reviewer’s exposure on Amazon.com, we define an
indicator variable called Exposure. This variable equals 1 if a review is written by a reviewer with more
than the median number of exposures, and 0 otherwise. We further include a dummy variable to capture
the level of reviewer coverage an item receives (Coverage). It equals 1 if an item is followed by more
than the median number of reviewers, and 0 otherwise. We also allow for interactions of the signal
variable with the exposure and coverage variables by including interaction terms: Signal x Exposure, and
Signal x Coverage. Product category dummies are also included to represent fixed effects due to item-
Table 5 presents the results. The coefficient for Signal in this table is negative and statistically
significant (coeff. = -3106, p-value = 0.006). This confirms the importance of reviewer quality and the
types of reviews as a determinant of abnormal product sales. The interaction between Signal and
Reviewer Exposure is negative and significant, indicating that a review has greater information content if
it comes from a high exposure reviewer. In contrast, the coefficient of the interaction between Signal and
Product Coverage is positive and significant, suggesting that the reaction to Signal is higher for items
with low product coverage. Overall, these results are consistent with the ones reported in Table 4 and lend
further credence to those findings. Furthermore, Coverage has a direct and indirect (through Signal)
impact on product sales, while Exposure only influences sales through Signal.
Insert Table 5 about here
To summarize, the evidence thus far supports the view that reviews written by high quality
reviewers, high exposure reviewers, and for products with less reviewer coverage, have a greater impact
on sales than reviews written by lower quality reviewers, lower exposure reviewers, or for products with
already significant reviewer coverage.
5.5 Results: The Role of Time
In this section, we examine the temporal effects of reviews on sales when a review contains favorable
news or unfavorable news (hypothesis 5). Based on the age of an item, for each product category (Books,
DVDs, and Videos), we classify favorable and unfavorable news into three sub-groups: early stage,
medium stage, and later stage. Then for each sub-group, we compare the mean (median) abnormal
difference in sales between the favorable and unfavorable news portfolio. The results are presented in
Insert Table 6 about here
Overall, except for books, the abnormal difference in sales between favorable and unfavorable
news reviews is significant only for items in the early stages of the product lifecycle. For books, the
difference in sales for the unfavorable and favorable news event is significantly different from zero for the
medium stage (difference = 8679.8, p-value = 0.045), but not significant for the early stage (difference =
3175.7, p-value = 0.441) and later stage (difference = -1428, p-value = 0.792). For DVDs, the difference
in sales between the unfavorable and favorable news portfolio is significantly different from zero for the
early stage only (difference = 621.93, p-value = 0.013). Videos behave similarly to DVDs; the difference
between the unfavorable and favorable portfolio is significant for early stage only (difference = 839.11, p-
value = 0.027). 6
6 We also estimate the difference between favorable and unfavorable news using Wilcoxon Z-statistics. The median testing
results are consistent with the mean testing results.
These results are consistent with hypothesis 5. The impact of a review on sales is a decreasing
function of age. As time elapses, the difference between the information provided by favorable and
unfavorable reviews declines to zero. Consequently, hypothesis 5 is supported.
Our goal in this study is to assess the quantitative and qualitative impact of online reviews on product
sales. We want to assess the effectiveness of online reviews and the extent to which sales react to
contextual information regarding reviewer quality, reviewer exposure, and product coverage. We use data
from a popular online retailer, Amazon.com, to test our hypotheses.
Consistent with our arguments, we find that changes in online reviews are associated with
changes in sales. We also find that, besides the quantitative measurement of online reviews (i.e., review
scores), consumers pay attention to other qualitative aspects of online reviews such as reviewer quality
and reviewer exposure. Furthermore, we find that a consumer’s reaction to online reviews is stronger for
the items that have less product coverage; that is, new online reviews are more informative when items
have fewer pre-existing reviewers. Finally, we find that the review signal moderates the impact of
reviewer exposure and product coverage on product sales. Consumers are fully able to appreciate the
differential impacts of high-quality signals vs. low-quality signals. Lastly, the impact of online reviews on
sales is a decreasing function of the age of the product.
Taken together, our study integrates econometric data with insights from a portfolio approach to
reveal how consumers use online review information. Unlike previous studies that focus on linking levels
of review scores with levels of product sales, we study how consumers use quantitative and qualitative
aspects of online reviews to make purchase decisions. We show that online reviews reduce the uncertainty
and decrease the transaction costs of online transactions. In essence, consumers respond through their
purchase behavior to quality information embedded in online reviews.
6.2 Implications for Research
The paper has implications for online WOM communication and online consumer behavior, as described
6.2.1 Implications for Online WOM Communication
Online WOM communication is becoming a popular informational source for consumers and marketers.
As researchers focus on the impact of average online review ratings on consumer relationship
management and product success, there is a need to understand how consumers use online reviews,
whether they understand the information embedded in reviews, whether they rely on online reviews to
make purchase decisions, and under what circumstances a review is likely to impact sales. This paper
contributes to this emerging literature by addressing these fundamental but largely neglected questions.
6.2.2 Implications for Online Consumer Behavior and Practice
The econometric results in our study suggest that over time, the impact of online reviews on sales
diminishes as consumers begin to receive quality related information from other channels. For the
medium and later stages in the life of a product, online reviews may still influence sales, but there is no
abnormal sales difference between favorable and unfavorable news except in the case of books. Online
retailers, product manufacturers, and companies that specialize in collecting and disseminating product
quality information may need to pay more attention to early stage reviews and to find a way to promote
favorable reviews at that stage, when consumers pay more attention to online reviews. After this stage,
the impact of online reviews on sales begins to decline. Also, online retailers and product manufacturers
should encourage and nurture high quality and high exposure reviewers, since the actions of these
reviewers have a direct impact on product sales.
6.3 Managerial Implications
Our results suggest that the market for reviews is efficient and that consumers are rational. Over the long
run, a strategy of recruiting reviewers to write good reviews of a vendor’s own products and bad reviews
of competitors’ products is unlikely to succeed. Consumers are able to tell the authenticity of a review and
differentiate a good reviewer from a bad reviewer. Firms should identify reviewers with better reputations
and higher exposure and try to promote new products to them in the hopes that they will respond with
favorable reviews. Those reviewers usually act as early adopters and opinion leaders in the consumer
community. Their tastes and judgments will determine which items other consumers are more likely to
adopt in the future.
6.4 Limitations and Suggestions for Future Research
This study has several limitations that create interesting opportunities for future research. First, even
though our results hold for our sample of DVDs and videos, we did not observe a similar set of results for
books. Future research could examine this result by considering heterogeneous properties among different
product categories. Second, this paper does not consider the textual content or length of the reviews,
factors that may also indicate review quality. Future research could take these factors into consideration in
an attempt to document how consumers respond to newly released reviews.
To conclude, online WOM communication in the form of online product reviews has become a
major informational source for consumers and marketers. In large part, by linking the average rating of an
item to its sales, the literature has assumed that consumers use only quantitative information aspects of
online reviews to make purchase decisions. To overcome this problem, this study proposes a portfolio
approach to demonstrate that consumers understand and use both the quantitative and qualitative
information embedded in online reviews. This study encourages further research in this area as a way to
derive deeper insights into the broader implications of online WOM communication.
1. Atiase, R.K. Predisclosure information, firm capitalization, and security price behavior around
earnings announcements. Journal of Accounting Research, 23, 1 (1985), 21-36.
2. Baltagi, B.H. Econometric Analysis of Panel Data (Second ed.). Chichester: John Wiley & Sons,
3. Bass, F.M. A new product growth for model consumer durables. Management Science, 15, 5 (1969),
4. Basuroy S., Chatterjee, S., and Ravid, S. A. How critical are critical reviews? The box office effects of
film critics, star-power, and budgets. Journal of Marketing, 67, 4 (2003), 105-117.
5. Berger, C.R., and Calabrese, R.J. Some exploration in initial interaction and beyond: Toward a
developmental theory of communication. Human Communication Research, 1 (1975), 99-112.
6. Biyalogorsky, E., Gerstner, E., and Libai, B. Customer referral management: Optimal rewards
programs. Marketing Science, 20, 1 (2001), 82-95.
7. Bonner, S., Walther, B., and Young, S. Sophistication-related differences in investors’ models of the
relative accuracy of analysts’ forecast revisions. The Accounting Review,78 (2003), 679-706.
8. Bonner , S.E., Hugon, J.A., and Walther, B.R. Investor reaction to celebrity analysts: The case of
earnings forecast revisions. Working paper (2005).
9. Boorstin, D. The Image: A Guide to Pseudo-Events in America. New York: Random House, 1987.
10. Boulding, W. Unobservable effects and business performance: Do fixed effects matter? Marketing
Science, 9, 1 (1990), 88-91.
11. Bradach, J. L., and Eccles, R. G. Price, authority, and trust: From ideal types to plural forms. Annual
Review of Sociology, 15 (1989), 97-118.
12. Brown, J.J., and Reingen, P.H. Social ties and word-of-mouth referral behavior. Journal of Consumer
Research, 14,3 (1987), 350-362.
13. Chatterjee, P. Online reviews: Do consumers use them? Advances in Consumer Research, 28, 1
14. Chevalier, J., and Goolsbee, A. Measuring prices and price competition online: Amazon and Barnes
and Noble. Quantitative Marketing and Economics, 1, 2 (2003), 203-222.
15. Chevalier, J., and Mayzlin, D. The effect of word of mouth on sales: Online book reviews. Journal of
Marketing Research, 43, 3 (2006), 345-354.
16. Chen, P.Y., Wu, S.Y., and Yoon, J. The impact of online recommendations and consumer feedback on
sales. In Proceedings of the International Conference on Information Systems. Washington, D.C., 2004,
17. Chen, Y., Fay, S. and Wang, Q. Marketing implications of online consumer product reviews.
Working paper, Department of Marketing, University of Florida (2003).
18. Chiles, T.H, and McMackin, J.F. Integrating variable risk preferences, trust, and transaction cost
economics. The Academy of Management Review, 21, 1 (1996), 73-99.
19. Clemons, E.K., Gao, G., and Hitt, L.M. When online reviews meet hyper differentiation: A study of
craft beer industry. Journal of Management Information Systems, 23, 2 (2006), 149-171.
20. Clement, M., and Tse, S. Do investors respond to analysts’ forecast revisions as if forecast accuracy is
all that matters? The Accounting Review, 78 (2003), 227-249.
21. Dellarocas, C., Awad, N., and Zhang, X. Exploring the value of online reviews to organizations:
Implications for revenue forecasting and planning. In Proceedings of the 24th International Conference
on Information Systems. Washington D.C, 2004.
22. Duan, W., Gu, B., and Whinston, A. Do online reviews matter? An empirical investigation of panel
data. Working paper, University of Texas at Austin (2005).
23. Eberhart, A.C., Altman, E.I., and Aggarwal, R. The equity performance of firms emerging from
bankruptcy. The Journal of Finance, 54, 5 (1999), 1855-1868.
24. Fama, E., and French, K. The cross-section of expected stock returns. Journal of Finance, 47, 2
25. Gleason, C.A., and Lee, C.M.C. Analyst forecast revisions and market price discovery. The
Accounting Review, 78, 1 (Jan. 2003), 193-226.
26. Godes, D., and Mayzlin, D. Using online conversations to study word of mouth communication.
Marketing Science, 23, 4 (2004), 545-560.
27. Hendricks, K.B., and Singhal, V. R. The long-run stock price performance of firms with effective
TQM programs. Management Science, 47, 3 (2001), 359-368.
28. Hu, N., Pavlou, P. and Zhang, J. Can online reviews reveal a product’s true quality? In Proceeding of
the ACM Electronic Commerce Conference. Ann Arbor, MI: ACMEC, 2006.
29. Huston, T.L., Ruggiero, M., Conner, R., and Geis, G. Bystander intervention into crime: A study
based on naturally-occurring episodes. Social Psychology Quarterly, 44, 1 (1981), 14-23
30. Eliashberg, J., and Shugan, S.M. Film critics: Influencers or predictors? Journal of Marketing, 61
(April 1997), 68-78.
31. Liang, T.P. and Huang, J.S. An empirical study on consumer acceptance of products in electronic
markets: A transaction cost model. Decision Support Systems, 24 (1998), 29-43.
32. Krishnan, J., and Krishnan, J. Litigation risk and auditor resignations. The Accounting Review, 72, 4
33. Liu, Y. Word of mouth for movies: Its dynamics and impact on box office revenue. Journal of
Marketing, 70, 3 (2006), 74-89.
34. Park, C.W., and Stice, E.K. Analyst forecasting ability and the stock price reaction to forecast
revisions. Review of Accounting Studies, 5 (2000), 259-272.
35. Reichheld, F.F. The one number you need to grow. Harvard Business Review, 81, 12 (2003), 46-54.
36. Sloan, R. Do stock prices fully reflect information in accruals and cash flows about future earnings?
The Accounting Review, 71, 3 (1996), 289-315.
37. Stickel, S.E. Reputation and performance among security analysts. Journal of Finance, 47 (Dec.
38. Williamson, O.E. Transaction-cost economics: The governance of contractual relations. Journal of
Law and Economics, 22, 2 (1979), 233-261.
Figure 1: Factors Contributing to Consumers’ Reactions to Online Reviews
Age of the Product
Note: Variable definitions are in Sections 4.4, 4.5, 4.6, and 4.7
Online_WOMMarket_Nov2007_Revised - final edit suggestions Page 30 of 33
Table 2: Summary Statistics
Amazon Longitudinal Data (July 2005 – Jan 2006)
Category #Reviews #Amazon Items #Distinct Items Avg_Rating
Book 6,759,764 261,187 10,052 3.87
DVD 4,056,340 258,736 9,988 4.07
Video 4,371,833 259,736 10,000 4.02
Table 3: Abnormal Sales Change Difference between Favorable News and Unfavorable News
with favorable news
- favorable news)
(1) - (2)
Mean 1286 89.65 1196.4 -2.940
Video N 162561 16226 (0.059) (0.003)
Mean 4828.4 1017.8 3810.6 1.0173
Book N 3742 4000 (0.146) (0.3090)
Mean 165.37 -195.7 361.02 -1.9353
DVD N 6857 6727 (0.035) (0.053)
Mean 301.05 -236.6 537.62 -2.543
Video N 5657 5499 (0.015) (0.011)
All p-values are based on 2-tailed tests.
1 Note that the number of items in this table is fewer than reported in Table 2 for longitudinal data
because the items that do not have any ratings change associated with them are not relevant for our
Table 4: Abnormal Sales Change Difference between Favorable and Unfavorable News
Portfolios: The Role of Reviewer Quality, Reviewer Exposure, and Product Coverage
(1) - (2)
Mean 2524.8 -3788 6312.7 -2.3883
and Video N 1713 1446 (0.001) (0.017)
Mean 5675.1 -3276 8950.9 -2.3182
Quality N 796 784 (0.001) (0.020)
Mean -209.9 -4394 4184.5 -1.1946
Quality N 917 662 (0.129) (0.232)
Panel C Mean 2555.2 -5157 7712.2 -2.2412 High
Exposure N 886 732 (0.006) (0.0250)
Mean 2429 -2384 4876 -1.1474
Exposure N 827 714 (0.075) (0.251)
Mean -307.9 -3889 3581.5 0.7563
Coverage N 900 741 (0.120) (0.449)
Mean 5660.6 -3681 9341.7 -3.8836
Coverage N 813 705 (0.002) (0.000)
Table 5: Regression Results of Impact of Review on Abnormal Sales
Intercept α0 772.42
SIGNAL α1 -3106.16***
COVERAGE α2 -2026.98**
EXPOSURE α3 203.76
SIGNAL X COVERAGE α4 3696.72***
SIGNAL X EXPOSURE α5 -1711.95**
Adj R-Sq = 0.0070
1) The variable definitions are in Section 3.4.
2) All of the p-values are based on two-tailed test. * indicates significance at 10%; ** indicates
significance at 5%; and *** indicates significance at 1%.
3) The model also includes book and DVD dummies (not shown).
4) The small adjusted R-square is consistent with the abnormal returns in accounting and
information systems literature.
Table 6: The Role of Item Age in Abnormal Sales Change Difference between Favorable and
Unfavorable News Portfolio
(1) - (2)
Mean 7371.5 4195.8 3175.7 -0.2602
Stage N 1633 1733 (0.441)
Mean 7955.5 -724.4 8679.8 1.1083
Stage N 1183 1228 (0.045)
Mean -3651 -2224 -1428 1.1085
Stage N 926 1039 (0.792)
Mean 373.74 -248.2 621.93 -2.6407
Stage N 3376 3294 (0.013) (0.008)
Mean 40.93 -170.4 211.32 -0.2262
Stage N 2097 2035 (0.501) (0.821)
Mean -154.4 -108.7 -45.72 -0.284
Stage N 1384 1398 (0.896) (0.775)
Mean 497.53 -341.6 839.11 -3.9013
Stage N 2685 2347 (0.027) (0.000)
Mean 102.8 -272.6 375.38 -0.6715
Stage N 1515 1708 (0.323) (0.502)
Mean 145.1 -23.32 168.42 1.033
Stage N 1457 1444 (0.619) (0.302)
* For Table 6, we use the full sample as in Table 3.