Link to Success: How Blogs Build an Audience by Promoting Rivals

7/30/2019 Link to Success: How Blogs Build an Audience by Promoting Rivals

1/18

MANAGEMENT SCIENCEArticles in Advance, pp. 118ISSN 0025-1909 (print) ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.1110.1510

2012 INFORMS

Link to Success: How Blogs Build an

Audience by Promoting RivalsDina Mayzlin

Marshall School of Business, University of Southern California, Los Angeles, California 40089,[email protected]

Hema YoganarasimhanGraduate School of Management, University of California, Davis, Davis, California 95616,

[email protected]

Empirically, we find that Web logs (or blogs) often link to other blogs in the same category. We present ananalytical model that explains why a rational blogger may choose to link to another blog. We allow bloggersto differ along two dimensions: (1) the ability to post news-breaking content, and (2) the ability to find news in

other blogs. By linking, a blog signals to the reader that it will be able to direct her to news in other blogs inthe future. The downside of a link is that it is a positive signal on the rivals news-breaking ability. We showthat linking will be in equilibrium when the heterogeneity on the ability to break news is low relative to theheterogeneity on the ability to find news in other blogs. One implication of the linking mechanism is that blogsthat are high on the news-breaking ability are more likely to gain readers. Hence, despite the fact that bloggerslink for purely selfish reasons, the macro effects of this activity is that readers learning is enhanced.

Key words : game theory; social media; linking; signaling; blogsHistory : Received September 30, 2010; accepted November 8, 2011, by J. Miguel Villas-Boas, marketing.

Published online in Articles in Advance.

1. IntroductionIn 1994, a Swarthmore College student, Justin Hall,

created an online personal journal called Links.net,now recognized as the first Web log or blog (Rosen2004). Since then, creating (or blogging) and read-ing blogs have become mainstream online activities.According to the Pew Internet Project, 12% of Internetusers (9% of all adults) say that they blog, and 33%of Internet users (24% of all adults) say that they read

blogs (Smith 2008). The growth of blogs as well astheir perceived influence on purchases has motivatedfirms to engage with bloggers as well. For example,in a Society of Digital Agencies (2010) survey of exec-utives from major global brands, agencies, and othermajor players in the digital space, 18% of respondents

considered blogger outreach top priority and 44%considered it important in 2010.

Blogs are part of the larger set of online socialmedia, which include online forums, bulletin boards,social networking sites, and video sharing sites.Although both blogs and other social media involveuser-generated content, blogs also share some char-acteristics of newspapers. For example, blogs provideinformation to readers, and the mode of transmissionis often one-to-many. David Winer, a blogging pio-neer, gives the following definition: A blog is like a

personal newspaper . Its sort of publishing on asmall scale (Potier 2003).1

For example, consider the blog AVC, at http://www.avc.blogs.com (Musings of a VC [venture cap-italist] in NYC), by Fred Wilson, a partner in UnionSquare Ventures. The blogs posts range from thepersonalIve been in a funk for the past threedays and I dont know why (April 22, 2008)tothe generalSo why is Facebook worth $15 bn andWordpress is worth $200 mm? (April 18, 2008). Someposts break news, such as, Disqus [which UnionSquare Ventures financed] announced a new fea-ture release and an investment today (March 18,2008). Others contain information originally reportedon another blog, for example, Microsoft has appar-

ently agreed to acquire Xobni [included a link to apost on TechCrunch] (April 20, 2008). TechCrunch

broke the story, Two independent sources tell us that

1 The following alternative definition of a blog is provided byWikipedia (http://en.wikipedia.org/wiki/Blog; last modified June27, 2012): A blog is a discussion or information site publishedon the World Wide Web consisting of discrete entries (posts) typ-ically displayed in reverse chronological order so the most recentpost appears first. Until 2009 blogs were usually the work of asingle individual, occasionally of a small group, and often werethemed on a single subject .

1

postedonanyother

website,includingtheauthorssite.Pleasesendanyquestionsre

gardingthispolicytopermissions

@informs.org.

Published online ahead of print July 30, 2012


2/18

Mayzlin and Yoganarasimhan: How Blogs Build an Audience by Promoting Rivals2 Management Science, Articles in Advance, pp. 118, 2012 INFORMS

Figure 1 A Snapshot of Daddytypes.com

Bizarre Childrens Book Contest|DT Childrens Book ReviewContest|about daddytypes|adoption|advice|architecture|art|birth|books|cars|clothing|diy|eBay|education|food|furniture|gay dads|gear notstrollers|health|movies|music|

names|newborn|news|nursery|nyc mens roomchanging tables|parentcompany|pregnancy|safety|strollers|toys|travel|tv|urbanbabywatch|vintage|web|work|

Shop Now

DADADSdadblog network

Click to advertise on dadblogsStardate 63293.1

We have come across Star Trek Onesies at Thinkgeek, which show signs that theGerber Childrenswear Company has apparently abandoned its claim on the Onesiecopyright on this M-Class planet.

The Starfleet Academy cadet Onesie shows the pre-2372 logo with thegramatically corrected latin motto [Ex Astris, Scientia], which locates its designas after2368, when TNG: The First Duty took place, and which featured theinfamous Ex Astra, Scientia logo.

Amazon.com

DaVinci Emily

Convertible BabyCrib

C

E

Sorelle

DaVinci ARCHIVES

Select a Month.

DADBLOGS

baby roadiesbeing daddythe blogfatherscynical daddadcentric

daddy dialecticdaddy dramadaddyzine

New $249.00

New $399.99Best $399.99

DaVinci Kalani

DaVinci

Convertible Crib in

Sorelle Tuscany

Convertible Crib-

Static links = blogroll

Admiral Doctorow noted on BoingBoing-9, the red Onesies are for expendablebabies who wont last an entire episode. And i would add that the blue Onesies arecut very short, for the sexy yeoman/nurse babies.

Dynamic linkStar Trek Uniform Onesies, $16 [thinkgeek.com]

filed under: clothing|tv|geeks|nerds|Onesie|star trek

link|add to del.icio.us|digg this|facebook it, dano|add a comment

Source. Used with permission.

the Microsoft/Xobni deal is moving along and thatMicrosoft signed an acquisition LOI in the last week(April 20, 2008).

Although the nature of news-breaking events dif-fers across domains, links to other blogs are common.(Here by links we mean dynamic links (or perma-

links) between blogs, which are links to specific postsin other blogs, as opposed to static links (or theblog roll) that often appear on the right-hand sideof a site (see Figure 1).) For example, on May 11,2006, Daddytypes.com (The Weblog for New Dadsauthored by Greg Allen) posted an announcementabout a two-day sale at Netto Collection, an upscalechildrens furniture store in Manhattan: Looks likeNetto Collections having a sample sale. I have noidea what is there, but I do know that its already

been going for four hours . The post then pro-vided a link to Daddydrama.com, which originallyhad posted the information on May 9, 2006, two days

before Daddytypes.In a small random sample of blogs, we found

that 61% of blogs2 contained at least one link toanother site in the last 10 posts, with approximately72% of links going to other blogs, 13% to newspa-per sites, and the rest to other sites3 (see A.1 inthe appendix for a description of the data collectionmethod). Hence, we find that bloggers often choose

2 Based on a sample of 258 blogs.3 Based on a random subsample of 438 outgoing links.

to link to another blog. This is surprising on severallevels. First, a reader who follows the outgoing linkmay not return to the original site in the short run.Second, a link implies that the linked blog has inter-esting content, which can improve the readers per-ception of a competing site. For example, after seeing

a link to the furniture sale post, the readers of Daddy-types now realize that Daddydrama can bring themuseful information on sales. This of course may implythat readers will defect to Daddydrama in the future.Note that this is a concern only if sites do not alreadyhave established reputations, which is the case formost blogs to a much bigger extent than for newspa-per sites. Hence, whereas all links may result in theshort-term loss of an eyeball, a link to another blogcreates a stronger competitor, which may be detri-mental in the long run.

One possible explanation for these links is thatbloggers are irrational (or perhaps are not solely con-

cerned about the size of their readership). However,linking may not necessarily be an irrational strat-egy even from an economic perspective. For example,in the same blog survey, we find that the blogs in thetop quartile of subscribers are more likely to link than

blogs in the bottom quartile (see Table 1). Althoughthis anecdotal evidence does not establish causality,it does suggest that linking may not necessarily beassociated with a decrease in readers.

Another possibility is that bloggers link to com-plementary blogs as opposed to direct competitors,

postedonanyother



@informs.org.


3/18

Mayzlin and Yoganarasimhan: How Blogs Build an Audience by Promoting RivalsManagement Science, Articles in Advance, pp. 118, 2012 INFORMS 3

Table 1 Linking in the Worst vs. the Best Blogs

Bottom quartile (no. of subscribers) Top quartile (no. of subscribers)

% Blogs with Avg. no. of % Blogs with Avg. no. of

Category N outgoing links subscribers N outgoing links subscribers

Food 10 10 0 7 86 43

Health 16 50 0 4 75 52

Sports 23 30 0 9 78 10

Movies 13 46 0 4 100 77

Business 21 43 0 9 89 96

Music 8 50 0 4 100 331

Fashion 18 33 0 9 67 25

Politics 14 71 0 11 100 103

Notes. No. of subscribers is the number of people who subscribed to RSS feed of the blog through

Bloglines. Note that this represents a small subset of the blogs total readership, because not all read-

ers subscribe to RSS feeds, and Bloglines is one of many platforms that provides access to RSS feeds.

Outgoing link is a link to another site embedded in 10 most recent posts sampled.

i.e., they link to blogs that provide information in adifferent category. For example, a political blog that

links to a food blog faces a smaller danger of los-ing its readers than one that links to another political

blog. However, in our sample, we find that 73% ofthe outgoing links to blogs are made to blogs in thesame category. That is, more often than not, bloggerslink to direct rivals. In this paper, we provide a the-oretical explanation for this phenomenon. We explainhow linking to rivals may increase a bloggers read-ership and explore the implications of linking on theevolution of the blogosphere.

Why would a link lead to an audience increase? Asour examples demonstrate, one of the primary func-tions of a blog is to provide information to its read-ers. We focus on a particular aspect of information,namely, the ability to deliver timely news.4 To capturethe heterogeneity between blogs, we allow bloggersto differ along two dimensions: (1) the ability to postnews-breaking content and (2) the ability to find newsin other blogs. A blog that is higher on the abilityto find news in other blogs is more likely to gener-ate a link. Hence, a link signals to the consumer thatthe blogger is more likely to deliver timely news bydirecting her to other blogs in cases when the bloggeris unable to break news on his own site. For example,Fred Wilsons (author of AVC) link to the Xobni post

on TechCrunch allows him to signal that he can directreaders to interesting information posted on otherblogs because of his extensive knowledge and interestin the category. Of course, a link signals the bloggersown ability to find news in other blogs, but it also

4 A survey of journalists and editors by Brodeur, a unit of Omni-com, confirms that blogs are an important source of news, even tothe professionals in the media industry: 46% of respondents indi-cate that they find blogs helpful in getting information about break-ing news, and 57% read blogs at least two or three times a week(see Brodeur 2008).

signals a potential rivals news-breaking ability. Forexample, TechCrunchs post about the Xobni deal also

demonstrates its ability to break news because of itswell-placed sources. The relative benefit (positive sig-nal about self) versus the relative cost of linking (pos-itive signal about the other blog) determines whethera link increases a blogs audience and, hence, whetherthe blogger chooses to link in equilibrium. We showthat linking will be in equilibrium when the hetero-geneity on the ability to break news is low relative tothe heterogeneity on the ability to find news in other

blogs. We also show that as information decays at amore rapid rate (information obtained later becomesless valuable), the incentive to link decreases.

As a byproduct of the incentive to link, consumers

can learn more efficiently which blogs deliver news-breaking content. Hence, despite the fact that blog-gers link for purely selfish reasons on the micro level,the macro effect of this activity is that readers learn-ing is enhanced. Thus, through linking, blogs that are

better at breaking news grow their readership morequickly than they would in the absence of linking.This of course also implies that the over-all qualityof the blogosphere improves as well. This effect isfurther accentuated by search engines that commonlyoffer higher placement to sites with more incominglinks.

Although the idea that incoming links containinformation on the quality of a website is not a newone (the most prominent example of a model thatassumes this is the Google search engine algorithm),this is the first paper that shows that linking can beincentive-compatible even in the absence of extrin-sic incentives such as advertising links, for instance.That is, in our model, bloggers link because doing soimproves the readers inference about the blogs qual-ity and ultimately increases the readership to theirsite. Hence, we provide an explanation for why bet-ter sites have more incoming links in equilibrium. In

postedonanyother



@informs.org.


4/18


other words, this paper provides a micro foundationfor models that assume that there is information inlinks.

We organize the remainder of this paper as follows:In 2, we discuss previous literature. We present themodel setup in 3, the main results in 4, followed by

extensions in 5. We conclude in 6, and we discusssome limitations and future work in 7.

2. Previous WorkWe first turn to the question of why blogs may link torivals. Katona and Sarvary (2008) investigate strategiclinking online and propose a market for advertising,such that a website may sell advertising space or buyan incoming link from another site. Some similaritiesmark their article and the current work, in that theyalso find that a site with better content enjoys moreincoming links. However, we differ with regard to theproposed mechanism driving this result; Katona and

Sarvary (2008) focus on an explicit pricing scheme,whereas we address the role of inferences made byreaders when they observe a link.

Because a link is a type of referral, the literature onreferral services is a natural setting to explore the rea-sons behind linking. Garicano and Santos (2004) showthat when an expert diagnoses a problem and decidesto address it or refer it to another expert, differentrevenue-sharing schemes have unique implicationsfor efficiency. Chen et al. (2002) consider infomedi-aries, Internet services that direct visitors to retail-ers that are members of their network. Both of thesepapers deal with an explicit contractual arrangement

between the infomediary and its clients, without anyinferences by clients about the infomediarys abil-ity to refer to others. That is, here experts and sitesrefer to others in exchange for payment. In contrast,in our setting, there is no explicit payment structure

between sites. Finally, Park (2005) examines the refer-ral behavior of experts in repeated relationships withconsumers; an expert may refer the client to anotherexpert who is more qualified to address the clientscurrent problem to maintain a long-term relationshipwith her. This broadly relates to the intuition in ourmodel because in our model linking is also motivated

by the desire to enhance a long-term relationship witha reader. However, the mechanisms in the two papersare very different. In Park (2005), an expert refers hon-estly because he is afraid to be punished by his cus-tomers in the future for dishonesty. In our model, onthe other hand, a blogger links to signal his quality.

The literature on network formation also seeks toexplain why people or firms form links. For exam-ple, Bala and Goyal (2000) and subsequent papers(see Demange and Wooders 2005 and Jackson 2008)study network formation as an equilibrium in a non-cooperative game. In these papers, links are formed

strategically, and the benefit of the link is to typicallyenhance the flow of information. That is, a link yieldsa direct benefit (for example, a customer may learnabout a job opportunity). Despite the extensive litera-ture in this area, to our knowledge, no research stud-ies the issue of a third party (i.e., the reader) who

makes inferences on the basis of the observed patternof links.Finally, link formation may be partially attributed

to reciprocal giving between blogs. Resnick andZeckhauser (2002) and Cabral and Hortasu (2010)find some evidence of reciprocity in buyer-seller feed-

back on eBay. Narayan and Yang (2007) find evidenceof reciprocity in link formation between reviewers onEpinions, and Stephen and Toubia (2010) find evi-dence of reciprocal linking in an online social com-merce marketplace. While reciprocity may explainsome of the linking behavior, we observe linking evenin situations where a reciprocal link is not expected

(as would be the case for a relatively unknown bloglinking to a well-known blog). Here we offer an expla-nation for linking that is above and beyond reciprocalgiving between blogs.

Second, we show that an implication of linking isthat a better quality site receives more incominglinks in equilibrium. The idea that hyperlinks on theInternet contain information on site quality has beenvery influential in search engine design. Kleinberg(1999) proposed that hyperlinks offer valuable infor-mation because they reflect the subjective judgmentsof the author who created them. He further offeredan algorithm, based on incoming links, to uncover

the most authoritative webpages for a given query.Brin and Page (1998) expanded this idea to developPageRank, a more flexible algorithm that calculatesthe authority rank of sites as a function of their incom-ing links, which continues to be the basic framework

behind Googles search engine.The assumption about the informativeness of the

link structure is also analogous to the wisdom of thecrowd hypothesis proposed by Surowiecki (2004).Even in the absence of search engines that amplifythe effect of links, incoming links can increase traf-fic by directing people to the focal site. For exam-ple, Stephen and Toubia (2010) show that additional

incoming links result in a better performance for aretailer in an online social marketplace.

In summary, the idea that sites may link to signal toa third party that they are high quality is novel to theliterature. The result that this in turn leads to bettersites having more incoming links, which implies thatthere is valuable information in links, is commonlyassumed in the literature. Hence, the primary contri-

bution of this paper is in providing a micro founda-tion for why we expect linking to occur in the absenceof an explicit payment scheme.

postedonanyother



@informs.org.


5/18


3. Model

3.1. SetupWe use a finite-period game with an infinite num-

ber of risk-neutral consumers (we refer to a consumeras R, denoting reader) and an infinite number of

blogs.5 To clarify the exposition, we henceforth refer

to the blogger as he and to the reader as she.Bloggers obtain utility from the size of the readership.That is, the per-period utility of blogger A at time t is

VANt (1)

where Nt is the number of As visitors during time t.Here we assume that dV/dN > 0: the blogs utilityis increasing in the number of visitors. The blogger

benefits from an increase in traffic in several ways.First, from a financial perspective, an increase in traf-fic results in an increase in advertising revenue. Moreimportantly, the bloggers social utility is also increas-ing in site traffic because the bloggers social influ-ence is increasing in the number of readers.6 We alsoassume that all bloggers act in a way that maximizestheir expected utility.7

Furthermore, we model bloggers as producers, andreaders as consumers of information. We distinguish

bloggers abilities along the following two dimen-sions: (1) the ability to break news on their ownsite and (2) the ability to find news in other blogs.Although we initially assume that these abilities areindependent, we relax this assumption subsequentlyin an extension. A blog can be either a high (h) type or

a low (l) type with regard to breaking news8

: h typesreceive it with probability v and l types with prob-ability w, where v > w. The prior probability that

5 This technical assumption simplifies the model. Qualitatively, theresults do not change as long as we assume a finite but very largenumber of blogs. The model does not depend on the assumptionthat the number of consumers is infinite.6 According to Lenhart and Fox (2006), 61% of bloggers listed thedesire to motivate others to action as a reason for blogging, and51% listed the desire to influence the way others think as areason.7 Of course in reality bloggers may be partially motivated by behav-ioral phenomena such as altruism. Here we show that linking can

occur even when bloggers are motivated solely by self-interest.8 Alternatively, we could differentiate bloggers according to thecosts of cultivating insider sources or searching. For example,we could add an initial stage where the blogger invests a costlyeffort which determines the probability with which he will find

breaking news in another blog, where the cost would differ acrossblogger types. Hence, a blogger with lower costs of finding insidersources could break news with a higher probability and a bloggerwith lower search costs would be more likely to find news-breakingcontent. Assuming differentiation across costs as opposed to prob-abilities of obtaining the information does not change the resultsqualitatively as long as q > 0 in equilibrium. We thank an anony-mous reviewer for pointing this out.

the blogger is h type on ability to break news is .Thus, the prior probability that a blog breaks news,is 0 = v + 1 w. The high-types superior abilityto break news derives either from its insider sourcesor being in the know through other means, such asones social network. For example, Fred Wilson of the

AVC blog, whose company has a stake in a numberof start-ups, is more likely to break news comparedto a blogger who engages in pure commentary.

Similarly, a blog can be either h type or l type withregard to finding news in other blogs: h types find itwith probability p and l types find it with probabil-ity q (where p > q > 09 and the prior on h type is ).10

The prior probability that a random blog finds newsin other blogs is 0 = p + 1 q. Note that beingan h type here requires the knowledge of the othersites in the category, as opposed to access to special-ized sources. In other words, all bloggers have accessto the information in other sites, but the high-type

blogger has either the ability or the desire to processthe large amount of information scattered across dif-ferent blogs. For example, Greg Allen of Daddytypesappears to be an avid reader of other parenting blogs,which enhances his ability to link to interesting postselsewhere.

Hence, a blogger can be one of four types, hhhllhll where the first letter refers to the abil-ity to break news on his own blog and the secondletter refers to the ability to find news in other blogs.We also consider three benchmark cases where the

bloggers are homogeneous along certain dimensions:(1) the case where there is only heterogeneity on the

ability to find news, (2) the case where there is onlyheterogeneity on the ability to break news, and (3)the case where there is no heterogeneity on eitherdimension.

3.2. The Timeline of the GameThe game consists of two periods of two stages each(see Figure 2). Each period represents one news cycle,where the utility that the reader derives from theinformation depends on the speed with which itreaches her. At the beginning of the game, the blog-gers know their own type, but the readers do not.Moreover, for simplicity, we also assume that blog-

gers do not observe their rivals types: There is noinformational asymmetry between readers and blog-gers on other bloggers quality. After observing the

9 Here we assume that q > 0: The l type can find news in anotherblog with nonzero probability. This rules out a trivial separatingequilibrium in linking (see the discussion following Proposition 1).10 The probability that a blog can find news on another blog isconditional on the event that at least one other blog breaks news.However, the probability of such an event is 1 because the numberof blogs is infinite. Hence, p and q can be treated as independentof v and w.

postedonanyother



@informs.org.


6/18


Figure 2 Timeline of the Game

Rdecides which blog to visit in Period 2a:If A linked to B, Rs choice set is {A, B}If A did not link, Rschoice set is {A, C}

Rviews blog A.A may link to B.If Rgets info through link,

she receives utility uc.

Stage 2a

New info is released.Blog A may break news.As visitor receives u if A

has news.

A links to another blog ifhe finds a blog with news.

If As visitor gets info throughlink, she receives utility uc.

Stage 2b

Stage 1a

Rviews blog A.Info is released.Blog A may break news.R receives u if A has news.

Stage 1b

Period 2

Period 1

posting and linking behavior in the first period, thereader makes inferences on the blogs types, whichin turn will influence her blog choice in the secondperiod. All readers and bloggers face the same game.To simplify the exposition, we outline the timeline ofthe game from the perspective of a random reader Rand the blogs to which she may be exposed.

3.2.1. Period 1. At stage 1a, readers choose to visita random blog and consume its content throughoutPeriod 1.11 That is, a reader (R) visits a random blog(A). Also during this stage a unique piece of verifi-able information is released.12 For example, Microsoft

signs a letter of intent to purchase another companyor Netto announces a furniture sale. Bloggers maygain access to the information depending on theirnews-breaking ability. A blogger j who obtains infor-mation may go on to post it on his blog: aj 0 1,where aj = 1 indicates the action of posting news onjs blog. We assume that, conditional on having accessto breaking news, the act of posting this content iscostless. We also assume that because the news is ver-ifiable, bloggers cannot fabricate news stories. If Ahas posted news (aA = 1), R derives utility u from thepost. Otherwise, she derives zero utility from the post.

At stage 1b, bloggers search other blogs for infor-

mation. Bloggers may gain access to news-breakinginformation on other blogs depending on their abil-ity to find news. A blogger j who finds news inanother blog may go on to post a link to that blog:bj 0 1, where bj= 1 indicates the action of posting

11 We do not need to specify the number of readers that arrive ateach blog as long as that number is finite. For example, we couldassume that the number of arrivals is Poisson-distributed.12 All the results are unchanged if the information is released withprobability < 1.

a link. We assume that, conditional on having accessto breaking news in another blog, the act of linkingis costless. Reader R derives utility u c from theinformation if she sees a link to a news-breaking blog(say B), and the information is novel (which is thecase if A had not posted news in stage 1a: aA = 0).However, a reader who has seen the news at stage 1a(aA = 1) does not derive any direct utility from thelink, though she may learn about Bs ability to breaknews. If the blog does not post news or link to newsin another blog, we assume that R receives utility u,which we normalize to 0. (See the second column of

Table 2 for the summary of Rs utility following Asactions in Period 1.)Note that the value of the information declines

over time; here c is the cost of delay. For example,because Daddytypes linked to the original post on theNetto Collection sales after a time delay, his readersmay have already missed some of the better bargains.In short, our timeline captures the idea that origi-nal posts are more useful to consumers because theycontain fresh information, unlike links, which containrelatively stale news.

Here we abstract away from the possibility that ablogger can plagiarize another blogs content without

attribution. Instead, we assume that news-breakingblogs are credited by blogs who link to them. This

Table 2 Readers Utility and Choice-Set at the End of Period 1

Blog As action in Rs utility in Rs choice set at the end of

Period 1 Period 1 Period 1

aA = 1, bA = 1 u (AB)

aA = 1, bA = 0 u (AC)

aA = 0, bA = 1 u c (AB)

aA = 0, bA = 0 0 (AC)

postedonanyother



@informs.org.


7/18


is realistic for two reasons. First, we observe attri-bution.13 Second, reputational concerns (as shown byPark 2005) may induce truth-telling. Finally, we alsoassume that bloggers cannot fabricate links becausereaders can easily verify the links authenticity byclicking on it.

After stage 1b, R decides which blog to visit inPeriod 2a. If A had linked to B at stage 1b (bA = 1),then R chooses between A and B. IfA hadnt linked toany blog at stage 1b (bA = 0), then R chooses betweenA and a random blog (which we denote by C). Fur-thermore, when making this choice, the reader alsoexperiences a reader-blogspecific random shock toher utility. This is a technical assumption that sim-plifies the analysis.14 Hence, R chooses the blog thatdelivers the greatest total utility, which is the sum ofthe expected utility and the random shock.

Note that the blogger has control in determining hiscompetition. We model the consumers choice as one

between the focal blog (A) and a primary competitor(which may be B or C). If the blogger does not link atstage 1b, his primary competitor in Period 2 is a ran-dom blog (C), which is average in his abilities. If the

blogger links, however, he makes his reader aware ofa news-breaking blog (B), which becomes his com-petitor in the future (see the third column in Table 2).This highlights the downside of linking.15

To summarize, Rs choice at the end of Period 1is affected by her observations of the blogs actionsin Period 1. (See also Table 3 for summary of theinformation structure in different stages of Period 1.)That is, R updates her priors on As abilities based

on his posting and linking behavior. In addition, sheupdates her prior on Bs news-breaking ability if sheobserves a link from A to B. To simplify the analysis,we assume that R does not update her priors about

13 For example, the Smoking Gun website received almost univer-sal credit in the blogosphere for exposing James Freys memoir A

Million Little Pieces as largely fictional (The Smoking Gun 2006).14 There are two reasons to introduce the error term in the model.First, it explains why a blog with a negative outcome for either

breaking news or linking still may attract readers in the next period.Second, the error term allows us to consider how linking affectsthe difference in the expected utilities between blog A and its pri-mary rival, EUA EUj, a continuous incentive, rather than a discrete

incentive, as would be the case in a model without noise. Further-more, the results are independent of the exact distribution of theerror term. Finally, note that we could have added an error term tothe utilities in the first period, too. However, it would be inconse-quential because readers pick blogs randomly in the first period.15 Why is C not part of the choice set in the case when A links toB? We can think of this as an outcome of a more complicated gamewhere R can invest in a (costly) search for another blog follow-ing her observation of As linking behavior. In the online techni-cal appendix (available at http://faculty.gsm.ucdavis.edu/hema/

blogs_tech_appendix.pdf), we show that under certain conditionsR only chooses to search for another blog in the case when A doesnot link.

Table 3 Information Structure During Period 1

As information set at the Rs information set at the

Stage beginningof the stage endof the stage

0 A0 = hhhllhll} R0 = 0 0

1a A1a = A0 , access to breaking news?}

R1a =

R0 aA

1b A1b A = A1a aA access to news

R1b

R = R1a bA

in other blogs?} aB = 1 if bA = 1

Bs ability to find news in other blogs, either becauseshe does not observe Bs links (i.e., information fromB may be consumed from As post) or because shevisits blog B and observes its links only after thenews has become stale and the links have no signalingvalue.16 We further assume that the reader does notlearn about the abilities of any other blog during thistime period, due to time constraints or because infor-mation quickly becomes stale in this environment.

3.2.2. Period 2. At stage 2a, a new unique piece of

verifiable information is released and bloggers maygain access to it depending on their types (h typeswith probability v and l types with probability w).Bloggers who receive the information go on to post it

because there is no strategic reason to do otherwise.Reader R obtains utility u from consuming the infor-mation at this stage.

At stage 2b, blogs link to news-breaking blogsif they can find them. Here all blogs link if theyfind news because again, there are no strategic rea-sons to do otherwise. If As visitor had not seenthe news in stage 2a, she obtains utility u c fromthe news. Signaling in the first period is motivated

by readers desire to learn about bloggers ability tolink in the future, which in this case is the secondperiod. The two-period model represents a simplifi-cation of an infinite-period overlapping generationsmodel, without the added complexities of an infinite-period model.

4. Perfect Bayesian Nash EquilibriumIn our analysis we focus on the decision faced bya random reader R and a random blog A of type hhhllhll. Given the symmetry in the readersdecisions and the bloggers incentives, we can thengeneralize the findings to all blogs and all readers.The pure strategy perfect Bayesian Nash equilibriumin linking consists of the bloggers optimal linkingstrategy at stage 1b as well as the readers beliefson the bloggers abilities following the informationreceived in Period 1.

We first turn to Rs problem after stage 1b. Wesignify by R the information set of R at this

16 The results of our analysis remain qualitatively the same if weassume that R can resolve uncertainty about Bs ability to findnews, but the analysis becomes much more cumbersome.

postedonanyother



@informs.org.


8/18


point, which consists of whether A broke the news(aA) and whether A linked to another blog withnews-breaking content (bA):

R = aA bA aB = 1if bA = 1 (see the second column in Table 3).We denote by R AR jR = Ahh

Ahl

Alh All

jhh

jhl

jlh

jll

R the vector of Rs

beliefs on blog As and blog js type, where j = Bif bA = 1 and j = C otherwise. Hence, the posteriorprobabilities that A will break and find the news areA =

Ahh +

Ahlv +

Alh +

All w and A =

Ahh +

Alhp +

Ahl + All q, respectively. We can similarly define j

and j.Reader Rs utility from blog i (where i Aj after

stage 1b is the sum of the expected utility based onRs updated beliefs about the blogs abilities and arandom shock i R, which we assume to be indepen-dent and identically distributed across readers andacross blogs and distributed on the real line with thecumulative distribution function (CDF) F, where den-

sity is nonzero everywhere,

URi = EUii i AR + i R

= iu + 1 iiu c + i R (2)

Therefore, the probability that R visits blog A atstage 2a is

PrURA > UR

j R

= Prj R A R < EUAA A AR

EUjj j jR

= GEUAAA A

R

EUjjj j

R

(3)

where G is the CDF of the random variable j R A R.Second, we turn to As optimal strategy at stage 1b.

Because by assumption the blogger cannot link if hedoes not find news in another blogs, we focus on thescenario in which A finds news in another blog B.The blogger can condition his linking decision onhis information set at this point, which contains histype ( and whether he broke news earlier (aA (seeTable 3). Blogger A chooses an action (link or no link)that maximizes his expected utility in stage 2a:

b

A = argmaxbA0 1 EV

A

= aA

= argmaxbA01

EuNR +NI+NAR A = aA

where NR are readers who choose the blog randomly,NI are the readers who visit A because of previousincoming links (if A had broken news at stage 1aand other blogs had linked to it at stage 1b), and NA

are As returning readers from Period 1 (each one ofwhom returns with probability given in Equation (3)).Because bA only affects the last term in As utility

function by affecting the beliefs of returning readers,henceforth, we focus on this term. Also, because weassumed that dV/dN > 0, we can show that A links ifdoing so increases the probability that it will be cho-sen over the primary rival in stage 2a. Furthermore,we assume that if the blogger is indifferent between

linking and not linking, he chooses to link. In otherwords, A links if

GEUAA A

AaA bA = 1

EUBB B BaA bA = 1

GEUAA A

AaA bA = 0

EUCC C CaA bA = 0

(4)

Because s density function is assumed to be nonzeroon the real line, the density of j R A R is alsononzero on the real line. This, along with the fact thatG is a CDF, implies that G is a strictly increasing func-

tion. Hence, A links ifEUAA A

AaA bA = 1

EUBB B BaA bA = 1

EUAA A AaA bA = 0

EUCC C CaA bA = 0 (5)

Note that the linking condition in Equation (5) doesnot depend on a specific distribution of : The onlynecessary assumption is that s density function isnonzero on the real line. Intuitively, the blogger willlink if this action makes him look on average more

attractive than his primary rival.Because bloggers observe their own type at the

beginning of the game, in principle there could beseparating equilibria where the bloggers decision tolink depends on his type. However, we can show thatno separating equilibria exist here.

Proposition 1. No fully separating or semiseparatingequilibria in linking exist.

Proof. See the appendix.

The intuition behind the proof of Proposition 1 isthe following. Note that given q > 0, even the lowtype can find news in another blog with nonzeroprobability. A bloggers utility depends solely on thesize of his readership, and linking is costless. There-fore, if a link convinces the reader that the bloggeris likely to be of higher ability, and hence increasesthe probability that the reader will come back in thefuture, all types of bloggers prefer to link if they findinformation in other blogs. Thus, separation is impos-sible. In contrast, if we were to assume that q= 0 thepresence of a link could separate hhlh} from hlll}

because the blogger who is l type on ability to findnews is never able to link.

postedonanyother



@informs.org.


9/18


The discussion above suggests that there are twoways to generate separation: (1) to ensure that thelow type is unable to find breaking news in another

blog, or (2) to introduce a cost of linking that dif-fers by type to enable the single-crossing property.The second assumption is clearly not realistic; con-

ditional on finding news in another blog, posting alink does not require any specialized skills. Similarly,it is difficult to believe that in reality the l type wouldnever find news in other blogs (q= 0), because evenminimal consumption of others content by the blog-ger would result in a nonzero probability of findingnews.

Proposition 1 implies that we can focus on equi-libria where the decision to link depends only onwhether the blogger broke news in stage 1a.17 Thereare four possible pure strategy equilibria: LL,LDL, DLL, and DLDL, where L stands forlink and DL indicates do not link, and the cells refer

to the actions when the blog has or has not brokennews, respectively. For example, in DLL, the blog-ger does not link if he broke news, but does link if hedid not break news.

In equilibrium, how are Rs beliefs at the endof stage 1b affected by the information received inPeriod 1? First, we turn to the inference on the

blogs news-breaking ability. Suppose that the bloggeralways breaks news on his blog if he can (which wewill show to be the case). Because bloggers who are htype on the ability to break news are more likely to doso than bloggers who are l type on this ability (v > w),a news-breaking post is a positive signal on that blog-

gers news-breaking ability. This implies that if Robserves a news-breaking action by blog i (ai = 1she updates i upward: i

R = aA = 1 bA U > 0.

18 On the other hand, if ai = 0 R updatesi downward: i

R = a = 0 bA D < 0.Finally, if R has no information on news-breakingaction, she holds a prior belief on the news-breakingability, i 0.

Second, we examine the inference on the bloggersability to find news in other blogs. Because bloggerswho are h type on the ability to find news are morelikely to do so than bloggers who are l type on thatability (p > q), the presence or absence of a link could

be a signal on the bloggers ability to find news in

17 Of course, it is still the case that linking could send a positive sig-nal on the bloggers type because the h type on ability to find newsis more likely to be able to link than l type. However, this signal isnoisy because both types can find news with nonzero probability.Hence, here the signal is about the information that the bloggerobtained in stage 1b, which in turn has implications on the blog-gers type. Unlike classical signaling models, however, it is not adirect signal on the bloggers type.18 In the proof of Proposition 1 in the appendix, we derive U, 0,D, U, 0, and D as functions of the primitive parameters.

other blogs. However, as we show below, a bloggerwho finds news in other blogs may not always chooseto link to that information. Hence, each equilibriumgenerates a different set of posterior beliefs followingthe observation of links. For example, the absence of alink in LL is a negative signal about the blogs abil-

ity to find news because linking is expected in equi-librium, whereas in DLDL it has no negative effectbecause linking is not expected. That is, if linkingis not expected in equilibrium, i

R = aA bA =0 0. On the other hand, if linking is expected inequilibrium, i

R = aA bA = 1 U > 0 andi

R = aA bA = 0 D < 0.In addition, as is the case for most signaling models,

some actions are not in equilibrium, and in this casewe must specify off-equilibrium beliefs. Specifically,if linking is observed but is not played in equilib-rium, Bayes rule does not apply. There are severalapproaches that have been used to deal with off-

equilibrium beliefs. One approach is not to make anyassumptions on the set of out-of-equilibrium beliefsand to narrow it using a refinement such as the intu-itive criterion (Cho and Kreps 1987) or the D1 crite-rion (Fudenberg and Tirole 1991). Unfortunately, herethe refinements do not constrain the set of beliefs

because the bloggers utility and best response do notdiffer by type. Hence, the approach we use here isto assume a certain set of out-of-equilibrium beliefsfollowing a deviation (McAfee and Schwartz 1994).Because linking is only possible in the case when the

blogger found news in another blog, a link perfectlysignals that the blogger found news in another blog.

For this reason we assume that a link is a positivesignal on the bloggers ability to find news in other

blogs even if linking is not expected in equilibrium,i

R = aA bA = 1 U > 0. Note that this is amodified form of passive beliefs. That is, upon see-ing a deviation, R assumes the prior distribution oftypes, but also takes into account that a link is a cred-ible positive signal on the ability to find news in other

blogs.Finally, As linking behavior impacts the beliefs

about the rivals abilities. That is, if A links to B,R believes that B = U > 0 since the act of linking toB implies that aB = 1. That is, by linking in stage 1b,

A sends a positive signal on the rivals news-breakingability. Note that because R does not have informa-tion on Bs linking behavior in Period 1, her belief onBs ability to find news in other blogs in Period 2 isthe prior: B = 0. On the other hand, if A does notlink to another blog in stage 1b (bA = 0), his primaryrival in Period 2 is a random blog (C): C = 0 andC = 0 because R does not have information on theactions ofC in Period 1. Table 4 summarizes the effectof bloggers actions on Rs beliefs on the two abilitiesfor the focal blog A and the primary rival (which is

postedonanyother



@informs.org.


10/18


Table 4 The Effect of the Blogs Action on Rs Beliefs

Cases A A other other

A breaks news; U U B = U B = 0A links to B

A breaks news; U D if LL or C = 0 C = 0A does not link LDL, else 0

A does not break news; D U B = U B = 0A links to B

A does not break news; D D if LL or C = 0 C = 0A does not link DLL, else 0

B or C depending on As linking action in stage 1c).For example, the first row in Table 4 states that if A

breaks news and links, R updates upward her beliefon As ability to break news and to find news in other

blogs, as well as Bs ability to break news. As westated before, Rs belief on Bs ability to finds news inother blogs does not change.

Let us now consider the following inequalities:

EUAU U EUBU 0

EUAU D EUC0 0 (6)

EUAD U EUBU 0

EUAD D EUC0 0 (7)

EUAD U EUBU 0

< EUAD 0 EUC0 0 (8)

EUAU U EUBU 0

< EUAU 0 EUC0 0 (9)

Combining Rs equilibrium beliefs and As incen-tive to link (see (5)) generates the following four setsof equilibrium conditions:

In (L L), A chooses to link if it breaks news (6)and if it does not break news (7).

In (LDL), A links only if it breaks news: (6)and (8).

In (DLL), A links only if it does not break news:(9) and (7).

In (DLDL), A never chooses to link: (9) and (8).Inequalities (6)(9) can also be rewritten as the dif-

ference between the marginal benefit (increase in own

utility) and marginal cost (increase in rivals utility)from linking; for example, (6) becomes

EUAU U EUAU D

EUBU 0 EUC0 0 0

where the left-hand side can be interpreted as a mea-sure of the incentive to link.

Last, we turn to As optimal strategy in stage 1a.We can see that in this game the blogger alwayschooses to post information since there is no strategic

reason to do otherwise. Hence, the only strategicconsideration that the blogger faces is whether to linkto another blog at stage 1b. Inequalities (6)(9) arewritten from the perspective of blogger A, but it istrivial to show that all bloggers face the same prob-lem. Hence, these equations define linking behav-

ior for all blogs. Similarly, the inference made by Rdescribes the inference made by all readers. Hence,even though we have approached the problem fromthe perspective of a single blog and a single reader,we can obtain the general equilibrium of the game.We will show that the blogger links if the benefit oflinking outweighs the cost. To highlight the benefit aswell as the cost of linking in the main model, and theireffect on the equilibrium outcome, we first presentthree benchmark cases where we assume that blog-gers are homogeneous along certain dimensions.

4.1. Benchmark Cases

Below we present the intuition behind the results.Please refer to the online technical appendix for a for-mal proof.

4.1.1. No Heterogeneity on the Ability to BreakNews. First, consider a modified model where thereis no heterogeneity between bloggers on the ability to

break news, i.e., v = w > 0. Because all bloggers areequally likely to break news, a news-breaking postdoes not provide any new information on the blogstype. Hence, there is no downside to linking: WhenA links to B, R does not draw any positive inferenceabout Bs news-breaking ability. However, the benefitof linking remains: R updates upward her belief onAs ability to find news in other blogs. Because linkingin this case does not send a positive signal on therivals ability, but still sends a positive signal aboutown ability to find news in other blogs, LL is theunique equilibrium. This case highlights the benefitof linking for blogs: In a setting where links do not

benefit the potential rival, a blogger always prefers tolink.

4.1.2. No Heterogeneity on Ability to Find Newsin Other Blogs. Next, consider a modified modelwhere there is no heterogeneity among bloggers onthe ability to find news, i.e., p = q > 0. Here, Rs pos-

terior beliefs on blog As ability to find news remainat 0 regardless of whether A linked. That is, U =D = 0. Because this eliminates the benefit of link-ing, but the cost of linking remains due to the hetero-geneity on the blogs abilities to break news, linkingis not an equilibrium strategy. Hence, DLDL is theunique equilibrium for all parameter values. This casehighlights the cost of linking: In a setting where linksdo not function as signals on bloggers own abilities,there is no strategic reason for the blogger to link tohis rival.

postedonanyother



@informs.org.


11/18


Figure 3 Placement of Equilibria in the (vw)(pq) Space q= 02w= 01 = = 05 u= 10 c= 1

4.1.3. No Heterogeneity on Any Dimension.Finally, consider a scenario where there is no hetero-geneity among bloggers on both the ability to breaknews and the ability to find news in other blogs,i.e., v = w > 0 and p = q > 0. Because all blogs arehomogeneous along both dimensions, there is no up-dating of beliefs, i.e., U = D = 0 and U = D = 0.

This is a knife-edge case where the blogger is exactlyindifferent between linking and not linking, and, byassumption, chooses to link. However, even thoughLL is the unique equilibrium here, R does not gainany information on blog quality through the link.Moreover, any nonzero cost of linking would resultin DLDL, which is not the case for the first bench-mark case.

In sum, these three benchmark cases highlight thefact that linking is motivated by the bloggers desireto signal own ability to find news, while it is tempered

by the realization that links also act as positive signalson the linked blogs ability to break news.

4.2. Existence and Uniqueness Resultsof the Full Model

Next we solve for the equilibria of the full model.In Proposition (2) we summarize our findings onequilibrium existence and uniqueness. We find thatLL, DLL, and DLDL can all be unique, butthere are also regions that contain multiple equilib-ria. Note that, for brevity in the proposition, we pro-vide the conditions for uniqueness only, whereas theremaining details are provided in the appendix (A.3).

We also graphically illustrate the results of the modelin Figure 3, which depicts the relative placementof the different equilibria in the (vw)(pq) space.That is, we fix w, q and vary v, p. The bound-aries of the regions in Figure 3 are defined by theiso-curves derived from Equations (6)(9). For exam-ple, we define 6= as EUAU U EUBU 0 =

EUAU D EUCO 0. In the online technicalappendix, we show that there are regularities in therelative placement of the regions that span differ-ent parameter values.19 The numbering of the regions(Region I, etc.) in Proposition (2) refers to the regionsdepicted in Figure 3.

Proposition 2. (Existence and Uniqueness)Conditions for Uniqueness:

1. Region I: If 1 UU 0u c U 0 10u+0c, LL is unique.

2. Region II: If 1DUDu c < U0 10u+0c, DLDL is unique.

3. Region III: If1UUDu c < U0 10u+0c 1DU0u c, DLL isunique.Regions with Multiple Equilibria:

4. Region IV: The equilibria of the model are DLLand LL.

5. Region V: The equilibria are DLDL and DLL.

19 In particular, we show that 6= and 9= are increasing in v, 8= and7= are either increasing in v or increasing and then decreasing in v.For all 0 q < p 1, 0 w < v 1, 9= lies above 6=, 8=, and 7=; 7=

lies below 6=, 8=, and 9=.

postedonanyother



@informs.org.


12/18


6. Region VI: The equilibria are LL, DLL,LDL, and DLDL.


Proposition 2 and Figure 3 demonstrate the impor-tance of the role of heterogeneity between h and ltypes on the existence of linking as an equilibriumstrategy. By varying the p and v parameters in Fig-ure 3, we alter the heterogeneity between types.In this section we illustrate the intuition behind theresults by focusing on extreme casesminimal ver-sus maximal heterogeneity. For example, if v and ware very close, the heterogeneity across types on theability to break news is minimal. The opposite is thecase if v is much greater than w. Similarly, if p andq are close, there is little distinction across types onthe ability to find news in other blogs, whereas theopposite is true is p is much greater than q.

From Figure 3 we can see that LL is unique

when p is much higher than q and v is close to w.Note that here the heterogeneity on ability to findnews is low, which implies that a link to another blogwould not significantly change the posterior about therival blogs ability to break news: U 0 is small.This of course implies that the marginal cost of link-ing is minimal. In the limit, when v = w or U =0, we obtain the benchmark case of no heterogene-ity on the ability to break news, which we showedresults in the unique LL equilibrium. In contrast,if p is much greater than q, U 0 is large, and themarginal benefit of linking is high. Hence, LL isunique when the marginal benefit of linking is rel-

atively large, whereas the marginal cost is relativelylow. Algebraically, we can also see this result in thecondition on the uniqueness ofLL in Proposition 2:Linking is unique when the benefit of linking (whichis a function of U 0) is significantly greater thanthe cost of linking (which is a function of U 0).

In contrast, we can see from Figure 3 that DLDLis unique when p is close to q (the marginal benefitof linking is low) and v is much higher than w (themarginal cost of linking is high). Again, when p and qare close, there is little heterogeneity on the ability tofind news and little difference in the posterior beliefsfollowing link or no link: U D is small. In the

limit, as U approaches D we replicate the result ofthe benchmark model: DLDL is unique. Similarly,from the second point in Proposition 2 we can see thatDLDL is unique when the cost of linking (which isa function of U 0) is significantly greater than the

benefit of linking (which is a function of U D).The other two equilibria, (DLL) and (LDL), exist

when the marginal benefit and the marginal cost areclose, in which case the outcome depends on the equi-librium beliefs and whether the blog breaks news (seeFigure 3 and Proposition 2). Finally, as we decrease

the heterogeneity along both dimensions, in the limit,only the condition for LL is (trivially) satisfied atequality in Proposition 2. This is the benchmark caseof no heterogeneity along both dimensions where the

blogger is indifferent between linking and no link-ing, and R learns nothing about the blogs types from

a link. Again, because this is a knife-edge case, theresult would be DLDL if we introduce a nonzerocost of linking.

Because content on blogs is substitutable, a blog-ger who is more likely to break news earns a smallermarginal benefit from being able to direct readersto news in other blogs. Hence, a blogger shouldhave less incentive to link when he has brokennews. However, this reasoning does not take intoaccount changes in beliefs across information states:In (LDL), the readers inference following no linkis more punishing if the blog had broken the news.Because LDL is the only equilibrium that defies the

intuition that the incentive to link is greater if the bloghad not broken the news, it is comforting to see thatit is never unique.

In summary, the amount of heterogeneity on thetwo abilities determines whether linking is in equi-librium. That is, as we decrease the heterogeneity onthe ability to break news, the marginal cost of linkingdecreases. On the other hand, a decrease in the het-erogeneity on the ability to find news decreases themarginal benefit of linking. When linking occurs inequilibrium, the blogs that are h type on the ability to

break news are more likely to attract incoming links.Thus, our model provides a micro-foundation for

why incoming links may serve as an enduring mea-sure of quality.20 From a consumers perspective, link-ing increases the attractiveness of the blogosphere,

because links enable readers to locate informationmore efficiently. From bloggers perspective, outgoinglinks enable them to signal their ability to locate infor-mation. It also enables the h type blogs to grow ata faster rate. The desire to signal generates an incen-tive for blogs to promote high-quality rivals, which isa byproduct of selfish behavior and not the result ofaltruism.

4.3. Comparative Statics

Here our objective is to clarify and reinforce the mainintuition of the basic model, i.e., that the decision tolink involves a trade-off between the benefit of link-ing (a positive signal about own ability to find newsin other blogs) and the cost of linking (a positive sig-nal about the primary rivals ability to break news).

20 Note that the equilibrium outcome here (i.e., readers learn aboutthe quality of the blog from the number of incoming links) is similarto that of observational learning where customers learn about thequality of the product from the number of people who purchasedit (or chose not to purchase it) in the past (see Zhang 2010).

postedonanyother



@informs.org.


13/18


We demonstrate how a change in parameters affectsthe incentive to link, where the incentive to link wasdefined as the difference between the marginal benefitand the marginal cost of linking.

In Proposition 3, we focus on two kinds of compar-ative statics. First, we focus on how the changes in

learning associated with posting news-breaking con-tent and linking to another blog impact the incentiveto link. From the previous section, we obtain the intu-ition associated with extreme cases: A large amountof heterogeneity on ability to find news compared tominimal heterogeneity on ability to break news leadsto linking and vice versa. Here we ask whether asmall increase in informativeness of the signal on abil-ity to find news always increases the incentive to linkand whether a small increase in informativeness ofthe signal on ability to break news always decreasesthe incentive to link. We use the following notion ofinformativeness: We consider the signal s on ability

to break news to be more informative than s if s

isthe mean-preserving spread of s: 0 = 0 and U >

U > D > D (see Kim 1995). We can also think of

this increase in informativeness as spreading the pos-teriors on the signal. We can similarly define a moreinformative signal on ability to find news.21 Second,we examine the effect of a change of c (the cost ofdelay) on the incentive to link.

Proposition 3. (Comparative Statics)1. A more informative signal on ability to find news

always increases the incentive to link.2. A more informative signal on ability to break news

decreases the incentive to link if (a) the blog was able tobreak news, or (b) the blog was unable to break news andU D and u c are relatively low.

3. The incentive to link is decreasing in c.


Spreading the posteriors on the ability to find newsin other blogs makes both the positive signal (a link)and the negative signal (no link) more informative.Because this increases the marginal benefit of linkingand has no effect on marginal cost, the overall incen-tive to link increases. In contrast, the effect of spread-ing the posteriors on news-breaking ability has a moreambiguous effect on the incentive to link. A more

informative signal improves the readers assessmentof the rivals ability and hence increases the marginalcost of linking. The effect on the marginal benefitdepends on whether the blog broke news; If it did,the marginal benefit of linking decreases, because thereader infers that she is less likely to rely on a link toobtain news in the next period since she can obtain itdirectly from the blog. However, if the blog did not

21 Note that when we vary v and p in Figure 3, we change both theprior and posterior probabilities.

break news, the marginal benefit of the link increases,because the reader infers that she is less likely toobtain news directly from the blog. Thus, the over-all effect depends on the relative importance of thesetwo factors. In particular, when U D and u c arelow, the marginal benefit of a link is relatively low,

and hence the increase in cost outweighs the increasein the benefit of linking: The blogs incentive to linkdecreases.

Finally, we turn to the effect of a change in delaycost on the incentive to link. Increasing c decreasesthe utility that the reader expects to receive frominformation obtained through a link, which decreasesthe marginal benefit of linking. In addition, increas-ing c increases the benefit of receiving informationdirectly versus through a link, which makes thelinked blog look particularly attractive because of itsnews-breaking ability, which increases the marginalcost of linking. Hence, an increase in c decreases the

incentive to link. Therefore, our model would predictthat in categories where the timeliness of informationis more important, we would observe less linking.

5. Correlated AbilitiesHere we relax the simplifying assumption that theability to break news and the ability to link are inde-pendent.22 In particular, if there were an underlyingexpertise that drives both theses abilities, we wouldobserve a positive correlation. We model this correla-tion across types as follows:

P h type on ability to find news h type on

ability to break news =

P h type on ability to find news l type on

ability to break news = 1

(10)

where 05 1.23

22 In the online technical appendix, we consider two additionalextensions to the main model. In the first extension we endoge-nize the readers choice set by allowing her to choose to searchfor another random blog irrespective of As linking decision. Wefind that this does not qualitatively affect the results of the mainmodel. In the second extension, we examine the extent to whichthe bloggers incentive to link and the readers incentive to learn

about the bloggers abilities are aligned. We find that the incentiveson learning about the rival blog are perfectly misaligned betweenthe blogger and the reader. That is, the blogger links in a way thatminimizes the learning about the rival site, whereas the readersutility is strictly increasing in the precision of learning.23 That is, since 05, we assume that a blog that is high type onthe ability to break news is more likely to be high type on the abilityto find news in other blogs. Note that at = 05, the two abilitiesare independent; an increase in implies an increase in positivecorrelation; and at = 1, the two abilities are perfectly correlated.We can show that the reverse inference holds as well: A blog thatis high type on the ability to find news in other blogs is more likelyto be high type on the ability to break news.

postedonanyother



@informs.org.


14/18


Proposition 4. There exists a > 05, such that1. for all 05 < , the equilibrium existence and

uniqueness results are qualitatively the same as underindependence;

2. for all > three additional regions (each with adifferent set of equilibria) become possible, including the

region in which LDL is unique.Proof. See the online technical appendix.

In the case of correlated abilities, the learning aboutthe two abilities is coupled: A positive signal on abil-ity to break news also results in the inference that the

blogger is more likely to find news in other blogs andvice versa. This significantly complicates the trade-offthat the blogger faces. For example, the blogger whoposts a link also sends a positive signal on own abil-ity to break news, an increase in the marginal benefitof linking. On the other hand, a link to another blognot only signals that a potential rival is more likely

to break news in the future, but also that he is morelikely to find news in another blog, an increase in themarginal cost of linking.

Interestingly, when the amount of correlation islarge ( > ), LDL may be unique. This contra-dicts the earlier intuition that the incentive to linkis lower when the blog can break news. The intu-ition behind this new result is the following: If thetwo abilities are highly correlated, and the blog can

break news, the posterior on his ability to find newsin other blogs may change in a way that increasesthe readers uncertainty on this ability (which would

be the case, for example, if the initial prior were low

and then increased to the intermediate region whereuncertainty is maximized). This in turn implies thatanother positive signal (a link) would result in a sig-nificant jump in the readers belief that the blog ish type on ability to find news in other blogs. Thiseffect may dominate the previously discussed effectthat the ability to find news in other blogs is lessimportant for a blog that is perceived as more likelyto break news in the future. Hence, the incentive tolink may be higher following a news-breaking story.

6. ConclusionEmpirically, linking between blogs is common. Herewe provide a rational explanation for why a blog-ger may link to a potential rival. In the blogosphere,

bloggers are providers, and readers are consumers ofinformation. A blogger links in order to credibly sig-nal to the reader that he is high type on the abilityto find news in other blogs. By doing so, his blog can

become a destination site: A reader can gain access tonews-breaking content even in cases when the blog-ger is unable to break the news on his own. Althoughlinks are referrals to competing sites, recommending

rivals may actually be profitable in certain circum-stances. These individual incentives result in a systemin which high-quality blogs gain prominence, whichenhances consumer learning.

Several interesting marketing implications resultfrom our findings. First, we find that both the num-

ber of incoming and outgoing links may serve as ametric of blog quality. Second, we point out the trade-offs inherent to recommending a rival. Moreover,although we focus in this paper on blogs, we believethat our findings apply to many sites that provideinformation as content. For example, Amazon fea-tures links to other book sellers who often have a newversion of the book at a discounted price. AlthoughAmazon takes a commission from sales generatedthrough such links, it nonetheless chooses to facili-tate a relationship between its customers and another

bookstore. Hence, Amazon Marketplace allows it toact as a destination site. Another example involves the

decision by WashingtonPost.com (among others) tofeature links to related articles and blog posts in otherpublications next to the Post article (Tedeschi 2006).

What can a blogger learn from this paper? Here wesuggest a framework for weighing the relative bene-fits and costs of linking in a context where an impor-tant function of blogs is to provide information tothe readers. That is, a link demonstrates that a blogcan find news in other sites, and this ability is espe-cially valued if there is a large amount of heterogene-ity along this dimension among the blogs within thecategory. The costs of linking arise because of the pos-itive signal sent on the other sites ability to break

news. The comparative statics presented in Proposi-tion 3 also suggest that in a category with a greatneed for up-to-date information (such as sports, forexample), the benefit from linking is lower.

7. Limitations and Future ResearchWe note several limitations to this study, which offeropportunities for further research. First, our modelmakes a number of simplifications. For example,we assume that blogs do not observe each otherstype. Including this type of informational asymmetry(where blogs observe their own type as well as the

type of their rivals) would result in a richer set ofequilibria where bloggers may not post a link in caseswhere the linked blog is in fact of low quality, and notlinking would be less of a negative signal. However,the basic trade-off between linking and not linkingwould remain even in this richer model. In addition,here we take a rather narrow view of informationas breaking news, which implies that informationcontained in two blogs (during the same period) issubstitutable. We could alternatively model infor-mation as any content potentially useful to readers,

postedonanyother



@informs.org.


15/18


which would decrease the substitutability betweenblog content. Note that although our model requiresthe heterogeneity in abilities between blogs, we donot explicitly model the mechanism that gives rise tothis heterogeneity because in our model the abilitiesof bloggers are exogenously determined. Instead we

could allow bloggers to invest in both abilities andthus determine whether bloggers prefer to invest inboth or to specialize.

Another limitation in our model is that we onlyconsider links to news-breaking blogs and do not con-sider links to other types of content such as another

bloggers set of relevant links. Allowing these othertypes of links would allow a link to convey infor-mation on the linked blogs ability to find news inother blogs. Hence, in the future it may be interest-ing to examine the incentives to link to these differ-ent types of content. Note that even in this richermodel the basic trade-offs would remain the same: By

linking a blogger increases the readers perception ofits own quality as well as the quality of a potentialrival.

In addition to modeling the phenomena above,we see several other promising directions for fur-ther research. First, it would be interesting to modelhow search engines (and, in particular, news search)affect the bloggers incentives to link. One concernmay be that as the news search engine becomes veryefficient at directing readers to breaking news sites,the individual bloggers ability to find news in othersites becomes less valuable to the readers, which inturn may decrease the bloggers incentive to link.

Paradoxically, this may also adversely affect the per-formance of search engines because their algorithmsrely on links as inputs. Hence, it may be interestingto investigate the equilibrium of the game betweenthe news search engine, blogs, and the consumer.However, we do not believe that search engines cancompletely replace blogs in breaking news discovery

because the quality of the results that a search engineproduces depends on the users ability to enter news-related search terms.24

Finally, there are many opportunities for empir-ical work in this area. Our model makes severaltestable predictions on when the incentives to link

are greatest. In general, very little is known aboutthe determinants of links between blogs, and between

blogs and other sites, and the implications of these

24 For example, consider a recent college graduate who has an ideafor a high-tech start-up. If she is a regular reader of the AVC blog,she would learn about a 2011 New York summer accelerator pro-gram called SeedStart Media that nurtures start-ups. In contrast,Google would return SeedStart as a first page result only if sheenters a specific search term such as New York tech start-up pro-gram, but would not return it if she enters a more generic termlike tech start-up program.

links. For example, some newspapers have been waryof allowing blogs to link to them without permis-sion.25 Richard Posner, a judge on the United StatesCourt of Appeals, even suggested that such links

be banned (Becker-Posner Blog 2009). Hence, empir-ical work in this area may have important policy

implications.

AcknowledgmentsThe authors thank the department editor, the associateeditor, and two anonymous reviewers for very construc-tive feedback during the review process. The authors alsothank Anthony Dukes, David Godes, Yogesh Joshi, SachinSancheti, Jose Silva, Jiwoong Shin, K. Sudhir, and BirgerWernerfelt for comments that have improved the paper.Finally, the authors thank participants at the Yale Uni-versity School of Management (SOM) Ph.D. seminar, theMarketing Science 2006 Conference, the Summer Institutein Competitive Strategy 2006 Conference, the Yale SOMWednesday lunch seminar, the Carnegie Mellon Univer-

sity Heinz School seminar, the University of Minnesotamarketing seminar, the University of Maryland marketingcamp, the University of Massachusetts at Amherst market-ing seminar, the University of Alberta marketing seminar,the Georgetown University marketing seminar, the Univer-sity of Wisconsin marketing seminar, the Boston Universitymarketing seminar, the Fourth Symposium on StatisticalChallenges in Electronic Commerce Research, the Univer-sity of California, San Diego marketing seminar, and theUniversity of Southern California marketing seminar forhelpful comments. Both authors contributed equally, andtheir names are listed in alphabetical order.

Appendix

A.1. Empirical ExampleTo motivate the research questions posed in the paper,we constructed a data set of blogs in eight categories:politics, health, fashion, food, business, music, sports, andmovies. We used the popular blog search engine Techno-rati to collect a list of blogs for sampling. In particular,we used Technoratis blog directory feature. For each cate-gory, we sorted all blogs according to the recency of theirupdates and selected the most recently updated 150 blogs atthe time of the data collection (April 19, 2007). We then col-lected the last 10 posts (as of June 25, 2007) for each selected

blog using blog feeds available through the Bloglines aggre-

gator. Note that this data also includes outlinks. We alsoused the Bloglines aggregator to collect the RSS feed sub-scriber data. To perform further data cleaning and classifi-cation, we used three independent raters to visit each siteand decide whether it was a blog and whether it belongedin the category to which it had been tagged by Technorati.In case of conflicts, we chose the majority rating. This datacleaning and classification method was used to all the out-links, too.

25 For a discussion of various newspaper policies on linking, see

Online Journalism Blog (2009).

postedonanyother



@informs.org.


16/18


A.2. Proof of Proposition 1Here we present a proof by contradiction. Let us parti-tion the set of all possible blogger types (S= hhhllhll)into two nonempty sets: the set SL and the set SDL Sup-pose that there exists a (semi) separating equilibrium whereall the types in set SL link if they find news and thetypes in the set do not link if they find news in another

blog. In such an equilibrium, let {Link Link} be the poste-riors following a link and {NoLink NoLink} be the posteri-ors following no link. Because the types in set SL chooseto link in equilibrium, it must be the case that for i SL,Vi = EUiLink Link EUB EUiNoLink NoLink EUC, where B is a linked blog (that was able to breaknews), and C is a random blog. Since Vi is independent ofi and is a function of posteriors only, the trade-offs remainthe same for all types. This implies that all types in theset SDL would (weakly) prefer to link if they find news inanother blog, which happens with nonzero probability. Thiscontradicts the assumption that all types in the set SDL donot link in equilibrium. Note that allowing the blogger tostrategically post or withhold breaking news also wouldnot enable a separating equilibrium because the same logicwould hold, albeit with more complicated posteriors, suchas News,Link News,Link, etc. Q.E.D.

A.3. Proof of Proposition 2We show in the main body of the paper that for a certain setof parameters an equilibrium exists if its conditions (a sub-set of Equations (6)(9) in the main body of the paper) aresatisfied for that set of parameters. By substituting particu-lar expressions for expected utilities and simplifying, theseconditions can be rewritten as follows:

1 UU Du c U 01 0u + 0c (6)

1 DU Du c U 01 0u + 0c (7)

1 DU 0u c < U 01 0u + 0c (8)

1 UU 0u c < U 01 0u + 0c (9)

Then we have(1) LL: (6) and (7),(2) LDL: (6) and (8),(3) DLL: (9) and (7),(4) DLDL: (9) and (8),

where U = v2 + 1 w 2/v + 1 w > 0 = v +

1 w > D = v1 v + 1 w1 w/1 v + 1 1 w, and U = p

2 + 1 q2/p + 1 q>0 =p + 1 q > D = p1 p + 1 q1 q/1 p +

1 1 q.For example, LL exists under a given set of parame-

ters if both (6) and (7) are satisfied (or hold) for that setof parameters. More generally, to obtain the existence anduniqueness results, we use the four different conditionsto divide the parameter space into 24 = 16 nonintersectingregions, some of which may be empty. (For example, oneregion is defined where (6), (7), (8), and (9) hold, anotherwhere (6), (7), and (8) hold but (9) does not hold, etc.)Note that each region contains a different set of equilib-ria. (For example, all equilibria exist in the former region,whereas only LL and LDL exist in the latter.)

Even though theoretically 16 distinct regions are possible,some regions may be empty because conditions (6)(9) arenot independent. Note that the right-hand side (RHS) of allthe equations is the same. In addition,

The left-hand side (LHS) of (7) > the LHS of (6)because 1 D > 1 U. Hence, if (6) holds, then (7)holds.

The LHS of (6) > the LHS of (9) because U D >U 0. Hence, if (6) does not hold, then (9) holds.

The LHS of (8) > the LHS of (9) because 1 D >1 U. Hence, if (8) holds, then (9) holds.

The LHS of (7) > the LHS of (8) because U D >U 0. Hence, if (8) does not hold, then (7) holds.

We can show that 10 of the 16 potential regions are emptybecause they violate at least one of the relationshops above.For example, the region where (6) holds and (7) does nothold violates the first relationship. The surviving six regionspartition the parameter space as follows. Figure (3) demon-strates that all of these regions are nonempty.

1. Region I: (6) holds, (8) does not hold, (9) does nothold. As before, if (8) does not hold, then (DLDL) and(LDL) do not exist. Furthermore, because (8) does nothold, (7) holds. Because (6) and (7) hold, LL exists.However, because (9) does not hold, (DLL) does not exist.Note that if (9) does not hold, then (6) holds and (8) doesnot hold. Thus, we can describe this region by stating that(9) does not hold. To summarize, if 1 UU 0 u c U 01 0u + 0c, LL is unique.

2. Region II: (6) does not hold, (8) holds, (7) does nothold. If (6) does not hold, then (L L) and (LDL) do notexist. DLL also does not exist because (7) does not hold.Moreover, when (6) does not hold, then (9) holds. Because(8) and (9) hold, (DLDL) exists. Note that if (7) doesnot hold, then (6) does not hold and (8) holds. Thus, we

can describe this region by stating that (7) does not hold.To summarize, if 1 DU Du c < U 0 1 0u + 0c, DLDL is unique.

3. Region III: (6) does not hold, (8) does not hold. If(6) does not hold, then (L L) and (LDL) do not exist. If(8) does not hold, (DLDL) does not exist. If (6) does nothold, then (9) holds. Also, if (8) does not hold, then (7)holds. Hence, only (DLL) exists. To summarize, if1 U U Du c < U 01 0u + 0c 1 D U 0u c (DLL) is unique.

4. Region IV: (6) holds, (8) does not hold, (9) holds.If (8) does not hold, then (DLDL) and (LDL) donot exist. Moreover, if (8) does not hold, then (7)holds. Because (6) and (7) hold, (L L) exists. Sim-

ilarly, because (9) and (7) hold, (DLL) exists. Tosummarize, if 1 UU 0u c < U 0 1 0u + 0c min1 UU Du c1 D U 0u c, the equilibria are DLL and LL.

5. Region V: (6) does not hold, (8) holds, (7)holds. If (6) does not hold, then (L L) and (LDL)do not exist. Also, when (6) does not hold, then(9) holds. Because (8), (7), and (9) hold, both(DLDL) an d (DLL) equilibria exist. To summarize, ifmax1 UU Du c1 DU 0u c 0 is a small change. This approachconstitutes a mean-preserving spread because the priors arekept constant (0 = 0

0 = 0), and one set of posteri-

ors brackets the other set (U > U > D > D and

U >

U > D > D). Hence, according to Blackwells theorem

(Blackwell 1951, Kim 1995), the signal on ability to find newsin other blogs associated with p q is more informativethan the signal associated with (p q), and the signal on abil-

ity to break news associated with (v

w

) is more informativethan the signal associated with (v w).

We use Bayes rule to derive the posterior probabilitiesunder (v w) and (p q) (the posterior probabilities undervw and pq were derived in A.3) as

U =v + 2 + w /1 2

v + 1 w

D =

v + 1 v + 1

w

1

1 w +

1

1 v 1 w1

U =p + 2 + 1 q /1 2

p + 1 q

D =

p +

1 p

1

+ 1 q

1 q+

1

1 p 1 q1

Because we have a mean-preserving spread, we also knowthat U > U,

D < D,

U > U,

D < D ,

U

0

U 0 = U U > 0 and

U

D U D > 0.

(a) We can easily show that the incentive to link isalways higher under (p q) than under (p q) because

f6pqf6pq=

U D

U D1Uuc > 0,

f7pqf7pq =

U D

U D1Duc > 0,

f8p q f8pq =

U

0U 01Duc > 0,

f9p q f9pq =

U

Documents

Link to Success: How Blogs Build an Audience by Promoting Rivals