Burguet et al. (2014) show that the incentive of the search engine to distort content
search and product search, starting from ϕ
O
¼ϕ
S
¼1, depends on the ratio of the follow-
ing terms:
u
σα 1 βðÞm
1
versus
v
1
v
2
m
2
m
1
:
The first expression refers to the costs and benefits of distorting content search, while
the second refers to the costs and benefits of distorting product search. When distorting
content search (with no distortion of product search), consumer surplus falls by a rate u,
but the advertising revenues of the search engine rise at a rate σα(1β)m
1
.
38
Instead,
distorting product search reduces consumer surplus at a rate v
1
> v
2
but increases the
value for the search engine by m
2
m
1
. Burguet et al. (2014) show that, generically,
the search engine will distort at most one type of search, setting the other at the optimal
value. Specifically, if the expression on the left is larger than the one on the right, only
product search is distorted, whereas only content search is distorted if the reverse holds
true. Only if both expressions are the same might both searches be distorted.
Overall, this shows that even without integration of a website with the search engine,
the search engine might have an incentive to distort search due to competition with web-
sites for advertising. The question is, again, whether vertical integration with a website
exacerbates this distortion or reduces it.
To see the incentives of the search engine under integration, suppose, first, that the
search engine is integrated with all websites. Then, the profit of the search engine becomes
π
0
¼F sðÞ 1 μðÞϕ
S
m
1
+1ϕ
S
m
2
+ μm
1
:
The search engine internalizes the externality exerted on websites by distorting ϕ
O
or ϕ
S
because it fully participates in the profits of the websites. This induces the search engine to
improve its reliability, for both content and product search. Thus, the effect of integration
is positive.
To understand the negative effects, consider the more realistic case in which the
search engine is integrated with only a fraction of the websites. Then, it has an incentive
to divert search from non-affiliated websites to affiliated ones. This leads to a different
level of ϕ
O
for affiliated websites than for non-affiliated ones. For example, if ϕ
O
¼ϕ
S
¼
1 without integration, then integration lowers consumer surplus because it may induce
the search engine to reduce ϕ
O
for non-affiliated websites.
To sum up, the literature on search engines shows that even without integration of a
search engine with content providers, the search engine may have an incentive to bias
search results. This bias occurs due to competition for advertisers between the search
38
This is because a reduction of r
S
reduces the probability that a user buys a product through a display link on
a website from σα to σαβ.
490
Handbook of Media Economics
engine and content providers. Integration between the search engine and a content pro-
vider affects the way that competition for advertisers plays out; integration leads to higher
or lower social welfare, depending on the circumstances.
10.4.3 Informatio n Spreading on the Internet
In the previous subsection, we restricted our attention to search engines as the only inter-
mediaries between users and content-providing websites. However, there exist several
other online channels allowing users to find out which website they are potentially inter-
ested in. In what follows, we discuss some of these channels and mechanisms, with a
particular focus on their implications for the spread of information across the Internet.
Specifically, we are interested in whether different users receive the same or differing infor-
mation, according to the channel they use. Because few papers in the literature analyze
these issues, we will confine our discussion to a description of the phenomenon and the
tentative implications for competition and plurality, without presenting a rigorous analysis.
A popular way that users access content apart from using a search engine is to visit a
news website and search for “most-read news” or “most-popular stories.” This device is
offered by most news websites, such as BBC or Bloomberg, the websites of most news-
papers, and also by video-sharing websites, such as YouTube. The standard way in which
websites decide to classify content as most popular or as must-read news is by counting
the absolute number of clicks on this content in the past (correcting for up-to-dateness
and other factors). In this respect, the popularity of stories is similar to a classic network
effect; that is, the more people read a story, the more attractive it will be to others.
39
The
effect of most-popular stories is that users are more likely to obtain the same information.
Even if users are heterogeneous and are interested in different content ex ante, the
pre-selected content of websites is the same, and users access only content within this
pre-selected sample. Therefore, users obtain the same information, which implies that
they become more homogeneous regarding their information. This exerts a negative
effect on plurality.
40
This issue is not (or to a much smaller extent) present in traditional
media, in which the tool of counting the number of clicks and, therefore, a direct
measure of popularity is not feasible.
Additionally, most-popular stories often have a tendency to be self-reinforcing as
most popular. If a story is recommended as highly popular, then more users will read
it, implying that the number of clicks increases, thereby making the story even more pop-
ular. This effect is known as observational learning and is documented by, among others,
Cai et al. (2009), Zhang (2010), and Chen et al. (2011). As a consequence, it is not
39
See Katz and Shapiro (1985, 1994) for seminal papers on network effects.
40
Another effect is that the selection of the content usually depends on the absolute number of clicks but not
on the time users spend on the website. Therefore, it is not clear if websites accurately measure how inter-
esting the respective content is to users.
491
The Economics of Internet Media
obvious whether users read the same stories because they are actually interesting for a
majority of users or if users read them merely because they are recommended.
Contrasting this hypothesis,
Tucker and Zhang (2011) present a mechanism for why
listing “most-read” stories can benefit stories with niche content or narrow appeal. Users
usually have an ex-ante expectation if particular content is of broad versus narrow appeal.
In this respect, a story with content that appeals to a majority of users is more likely to
make it onto the most-read list. Now suppose that a story of broad-appeal content is
ranked fourth on the most-read list, whereas a story of narrow-appeal content is ranked
fifth. Since the broad-appeal story has a higher probability of being part of the most-read
list, users will infer from this ranking that the narrow-appeal story is probably of higher
quality or has particularly interesting insights. Therefore, if both stories are ranked almost
equally, users will be more attracted by the story with narrow-appeal content.
Tucker and
Zhang (2011)
test this hypothesis in a field experiment. A website that lists wedding ser-
vice vendors switched from an alphabetical listing to one in which listings are ranked by
the number of clicks the vendor received. They measure vendors as broad-appeal ones
when located in towns with a large population and as narrow-appeal ones when located
in small towns.
Tucker and Zhang (2011) find strong evidence that narrow-appeal
vendors, indeed, receive more clicks than broad-appeal vendors when ranked equally.
Oestreicher-Singer and Sundararajan (2012) also conduct an analysis to determine if
popular or niche items benefit most from recommendations. In particular, they analyze
the demand effects in recommendation networks by using data about the co-purchase
network of more than 250,000 products sold on Amazon.com. They use the feature
of Amazon.com to provide hyperlinks to connected products. To identify the effect that
the visible presence of hyperlinks brings about, the authors control for unobserved
sources of complementarity by constructing alternative sets of complementary products.
For example, they construct a complementary set using data from the co-purchase net-
work of Barnes & Noble (B&N). The B&N website provides a recommendation net-
work similar to Amazon.com’s, but the product links might be different, and those on
the B&N website are invisible to Amazon.com customers. Therefore, the products linked
on the former website but not on the latter provide an alternative complementary set.
41
Oestreicher-Singer and Sundararajan (2012) find that visibility of the product network
has very large demand effects—i.e., the influence that complementary products can have
on the demand for each other can be up to a threefold average increase. Newer and more
popular products use the attention induced by their network position more efficiently.
The results of
Oestreicher-Singer and Sundararajan (2012) differ from those of Tucker
and Zhang (2011)
. In particular, the former paper finds that popular products benefit more
41
Similarly, products that are linked in the future on the Amazon.com website but not today can be assumed
to be complementary to the focal product today and can be used to construct an alternative complemen-
tary set.
492
Handbook of Media Economics
from recommendation, while the latter find that niche products receive larger benefits.
A potential explanation is that consumers may be more inclined to browse niche websites
when looking for products for a special occasion (such as weddings dresses) than when
looking for more standard products. The contrasting findings could also reflect different
reasons underlying the demand effect—i.e., attention in
Oestreicher-Singer and
Sundararajan (2012)
and observational learning in Tucker and Zhang (2011).
Another way for users to access content is to read what other users recommend. For
example, via the “share” command on Twitter or other social media, users recommend
content to their friends or followers (for some facts on users as curators, see
Section 10.2.1). These friends are highly likely to read what the recommenders
“like,” which is not necessarily what the majority of users are interested in or what friends
of other users like. Therefore, in contrast to the “most-popular” stories, sharing content
leads to different users obtaining different information and therefore does not necessarily
lead to a reduction in plurality. However, users may access only content of a particular
type because they largely ignore or are not aware of recommendations by users who are
not their friends or whom they do not follow. In this respect, sharing content can lead to
narrow or exaggerated views. It is therefore prone to media bias, which is discussed
extensively in Chapters 14 and 15.
It is evident that the flow and diversity of the information depends on the architecture
of the (social) network. For example, the architecture of Twitter is similar to the star net-
work, in which the user in the middle spreads information to all its followers. However,
two followers may not necessarily exchange information directly with each other but
only through the user they follow. By contrast, on Facebook, mostly groups of users
interact, implying that there are more direct links and direct information sharing among
these users.
42
Most-read news and individual users sharing news are two extreme forms of spreading
information on the Internet. Whereas the former depends only on the absolute number
of clicks, the latter depends on a user’s subjective evaluation.
43
In between these two
forms are recommendations provided by websites. These recommendations are based
partly on content (as in the case of most-popular stories) and partly on the specific user
(as in the case of sharing information by users).
Regarding content, a website has many different forms of selecting recommendations
to users. An extreme one is based purely on an algorithm, such as the absolute number of
clicks in the past, and does not involve any editorial selection.
44
The other extreme is a
42
For detailed analyses of network formation, see, e.g., Jackson and Wolinsky (1996) or Jackson (2010).
Banerjee et al. (2014) provide a recent analysis of how gossip spreads within a network.
43
Thus, to formalize the former, standard models with aggregate network effects can be used, whereas for
the latter, the link structure of the social network has to be taken into account.
44
For example, this is the case with Google News. For a more detailed discussion on news aggregators, see
Section 10.3.3.
493
The Economics of Internet Media
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.