The Market Structure for Internet Search Engines

Transcription

1 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 137 The Market Structure for Internet Search Engines RAHUL TELANG, UDAY RAJAN, AND TRIDAS MUKHOPADHYAY RAHUL TELANG is an Assistant Professor of Inforation Systes at H. John Heinz III School of Public Policy and Manageent, at Carnegie Mellon University. He received his Ph.D. in Inforation Systes fro GSIA, Carnegie Mellon University in 00. His research interests include consuers use of new technologies like search engines and peer-to-peer networks, and epirical and analytical odels of software security and vulnerabilities. His current work on e-business exaines loyalty on the Internet, copetition and ipact of used-goods arkets on retailers and suppliers. His recent papers have studied the ipact of patching on software quality, arkets for software vulnerabilities, and so on. His research has appeared in Journal of Marketing Research, Journal of Inforation Theory and Application, aong others journals. UDAY RAJAN is an Associate Professor of Finance at the University of Michigan Business School. He received his Ph.D. in Econoics fro Stanford University in His priary areas of interests are gae theory and applied gae theory. He is broadly interested in strategic interaction and arket structures in different contexts. Current IS-related research projects include copetition aong Internet search engines, the effects of personalized pricing, and internet referral services. His research has appeared in journals such as Aerican Econoic Review, Journal of Econoic Theory, and Journal of Finance. TRIDAS MUKHOPADHYAY is the Deloitte Consulting Professor of e-business at Carnegie Mellon University. His research interests include strategic use of IT, business-to-business coerce, business value of inforation technology, econoics of cyber security, and software developent productivity. His research appears in Inforation Systes Research, Journal of Manageent Inforation Systes, Counications of the ACM, Journal of Manufacturing and Operations Manageent, MIS Quarterly, Oega, IEEE Transactions on Software Engineering, Journal of Operations Manageent, Accounting Review, Manageent Science, Decision Support Systes, Journal of Experiental and Theoretical Artificial Intelligence, Journal of Organizational Coputing, International Journal of Electronic Coerce, Aerican Psychologist, and other publications. He has been on the Editorial Boards of Inforation Systes Frontier, Journal of Manageent Inforation Systes, Journal of Organizational Coputing and Electronic Coerce, Manageent Inforation Systes Quarterly, Journal of the Association for Inforation Systes, International Journal of Electronic Coerce, and Manageent Science and Inforation Systes Research. ABSTRACT: The Internet search engine arket has seen a proliferation of entrants over the past few years. Whereas Yahoo was the early arket leader, there has been Journal of Manageent Inforation Systes / Fall 004, Vol. 1, No., pp M.E. Sharpe, Inc / 004 $

2 138 TELANG, RAJAN, AND MUKHOPADHYAY entry by both lower-quality engines and higher-quality ones (such as Google). Prior work on quality differentiation requires that low-quality products have low prices in order to survive in a arket with high-quality products. However, the price charged to users of search engines is typically zero. Therefore, consuers do not face a tradeoff between quality and price. Why do lower-quality products survive in such a arket? We develop a vertical differentiation odel that explains this phenoenon. The quality of the results provided by a search engine is inherently stochastic, and there is no charge for using an engine. Therefore, users who try out one engine ay consult a lower-quality engine in the sae session. This residual deand allows lower-quality products to survive in equilibriu. We then extend our odel to incorporate horizontal differentiation as well and show that residual deand leads to higher quality and less differentiation in this arket. Engines want to attract copetitors custoers and therefore have a strong incentive to be siilar to each other. KEY WORDS AND PHRASES: e-coerce, arket structure, product differentiation, residual deand, search engines. THE MARKET FOR INTERNET SEARCH ENGINES has witnessed rapid growth since its inception. Search engines and search engine based portals consistently rank as soe of the heavily visited sites in the arket. 1 On the consuer side, surveys indicate that search engines are the ost iportant prootional ethod used by e-coerce sites and they represent the ost coon way the new sites are discovered by users. Users spend a significant aount of tie on search engines looking for relevant inforation [8]. This arket has also seen any changes on the side of the firs. Yahoo was an early entrant and arket leader. Subsequently, the arket has witnessed the entry of different engines with both lower-quality engines and higher-quality ones (such as Google, now the preeinent engine). Hundreds of engines are in the arket, despite the presence of a few well-established sites such as Google and Yahoo. The industrial organization literature has long established that different products can coexist in a arket at the sae tie (see, e.g., [, 5, 11, 14, 17]). Typically, this is deonstrated in odels of vertical (quality) or horizontal (taste) differentiation. A key feature of vertical differentiation odels is that a low-quality good ust have a lower price than a high-quality good. Otherwise, all consuers would buy the highquality product. However, the price charged to users of search engines is typically zero, with revenues being earned fro advertisers. Therefore, consuers do not face a tradeoff between quality and price. Why do lower-quality products survive in such a arket? Siilarly, a key result in horizontal differentiation odels is that axiu differentiation aong products is optial when users incur quadratic transportation costs [6]. This enables firs to charge higher prices, and leads to higher profits. Since the price is not a strategic variable in the search engine arket, what outcoe should we expect?

3 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 139 In this paper, we first consider just vertical differentiation, and show that it is exactly this zero-price feature that allows products of differing qualities to survive in this arket. Since the quality of the results fro a particular search is inherently stochastic, soe users ay not be fully satisfied with their search. As the price of sapling ultiple products in a single session is close to zero, these users now have an incentive to visit a lower-quality engine. The ability to saple any products quickly and cheaply is a unique feature of inforation goods on the Internet, and represents a departure fro the previous literature on quality differentiation. This phenoenon, which we call residual deand for the low-quality product, allows the low-quality product to survive. Then, we consider a odel that incorporates horizontal differentiation as well, and show that neither axiu nor iniu differentiation is optial. There is a tradeoff on optial location engines would like to be siilar to get residual deand, but not too siilar, since it ay lead to intense quality copetition. Although our odel is specific to search engines, our insights generalize to inforation goods, including news sites such as CNET or the New York Ties, and shopbots such as Mysion ( and Shopping.co ( Our odel is based on two key properties of the arket for Internet-based inforation goods. First, the explicit price paid by the consuers for the use of these products is zero; the priary source of revenue for these products is advertiseents. Despite a soewhat depressed online ad arket in the past two years, online advertising is still a substantial part of the search engines revenue strea. For exaple, Yahoo still earns close to 65 percent of its revenues (about $1 billion) through advertiseents. With the success of Google ( and Overture ( in generating ad revenues via better targeting and technology, the online ad industry grew alost 11 percent and grossed about $7.3 billion in 003 [7]. With the use of rich edia (audio and video) and creative ads (contextual ads), alost all online publishers showed growth in their online revenues in 003 [9, 15]. In fact, search engines are considered the best online ad vehicle. Moreover, due to difficulties involved in icropayents for such content, free digital content exclusively based on advertising revenues will continue to be a large and substantial arket. Second, soeties users ay saple ore than one product or service during a single session. As Lawrence and Giles [13] and Bradlow et al. [3] ention, search engines aintain databases that contain only a fraction of the inforation in the universe. Moreover, they often run different algoriths on their database, so that engine results are inherently stochastic. This often induces a visit to another search engine within the sae session, which restarts the search process. Based on actual usage data, Telang and Mukhopadhyay [19] report that users switched to a second engine during percent of search sessions. For exaple, the rate of switching was 15 percent for those who first went to Yahoo, while it was 31 percent for Infoseek. The statistics fro Nielsen/NetRatings (see Figure 1) also suggest that users saple ore than one engine when looking for inforation. These nubers are only for searchrelated traffic, and the report notes that Because a web surfer ay visit ore than one service, the cobined totals exceed 100 percent [18]. For January 004, the totals for the top 5 engines alone add up to approxiately 13 percent.

4 140 TELANG, RAJAN, AND MUKHOPADHYAY 0% 10% 0% 30% 40% GG 39.4% YH 30.4% MSN 9.6% AOL 15.5% AJ 8.5% Figure 1. Hoe and Work Users, Search Engine Reach in January 004 Source: Nielsen/NetRatings for SearchEngineWatch.co. Notes: GG Google; YH Yahoo; AOL Aerica Online; AJ AskJeeves. Suary of Results and Key Contribution We start with a duopoly odel of vertical differentiation, in which all users prefer a higher-quality engine to a lower one. We consider sequential entry. In the first period, there is only one fir in the arket. An entrant enters in the second period and copetes with the incubent. Since users incur no cost for using an engine, all users saple the higher-quality product first. In previous vertical differentiation odels, such as Moorthy [14] and Shaked and Sutton [17], the lower-quality product has a lower price and hence captures soe of the arket. If both products had the sae price, no lower-quality good could exist in this circustance. In our odel, soe proportion of consuers go on to saple a lower-quality product because quality is stochastic. In other words, on average, the high-quality engine yields good results. However, it ay not yield the site a user needs for a given search ter. A different engine, even one with lower average quality, ay find a different site. We show that this residual deand enables a lower-quality product to survive. To focus on rapid iproveent in technology, we allow the entrant to have an advantage in technology, odeled as a cost advantage. A late entrant can enter with the newest technology, whereas an incubent is locked into an earlier technology. Interestingly, we show that if the entrant s cost advantage is low, the incubent overinvests in quality in the first period, and reains the high-quality provider in the

5 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 141 second period, with the entrant offering a low quality. For exaple, any engines (e.g., Lycos) that entered after Yahoo were considered lower-quality engines. 3 Conversely, if the entrant has a large advantage in technology, the incubent ay choose to not copete head-on and underinvests in quality in the first period. The entrant then takes over as the high-quality provider, uch as Google did over Yahoo. In practice, we expect product positioning (that is, horizontal differentiation) to atter in the search engine arket as well. Therefore, we consider a odel with both vertical and horizontal differentiation. Since sequential entry odels ay be intractable in this context (see [16]), we consider a structure in which engines first choose their location (or differentiation strategy) and then choose their quality. We show that, with quadratic transportation costs (a standard assuption in the literature), engines want to be neither axially nor inially differentiated. Moreover, residual deand reduces this differentiation even ore. Therefore, engines have incentives to be siilar, but not the sae. We suarize the key contributions of our paper as: We introduce the concept of residual deand and show that the presence of such deand changes the search engine vendor fir s strategy in iportant ways. Significantly, residual deand leads to higher overall deand for engines and hence ore opportunities for product proliferation. Therefore, a lower-quality engine ay still be profitable even without any horizontal differentiation fro the doinant player. We provide a odel of both vertical and horizontal differentiation in the search engine arket, where pricing is not a strategic variable. To our knowledge, there is no other odel of copetition between firs where the price is set to zero in this anner. Although television prograing has a siilar zero-price feature, uch of the work in this area focuses on policy issues rather than copetition (see, e.g., [1]). Model of Quality Differentiation TWO FIRMS COMPETE IN QUALITY IN THE MARKET. We interpret quality as the ability of the engine to provide results or inforation that satisfies a user. For exaple, for search engines, this definition of quality includes the two coon attributes of quality: (1) the quantity of inforation retrieved, and () its relevance to the user. Our definition of quality ay be interpreted as a reduced-for notion that encopasses both of these attributes. A higher quality is ore valued by the user because it iplies either a higher quantity of inforation, or inforation of greater relevance, or both. We assue that the qualities of the firs are coon knowledge. Using the product requires both tie and effort on the part of a consuer. Therefore, the lower the quality of a product, the less likely it is that a consuer will be willing to invest his or her tie and effort. Hence, we odel deand as increasing in quality. If a fir is a onopolist and offers quality q, the deand in a given period is D(q). We assue that the deand curve is linear, with D(q) = a + bq, where a 0 (so that a fir offering a zero quality has no custoers) and b > 0. 4

6 14 TELANG, RAJAN, AND MUKHOPADHYAY Our odel encopasses two periods, 0 and 1. The deand curve is assued to be the sae in each period. In Period 0, Fir 1 (the incubent) is the only fir in the arket, and chooses a quality level q 0. In Period 1, Fir (the entrant) enters, and offers quality q. The incubent ay increase its quality fro q 0 to q 1. However, we assue that quality cannot decrease, so that q 1 q 0. Most digital products require an up-front investent in hardware and software developent [1]. The firs need to set up the infrastructure of Web servers, content gathering, and provision. For exaple, in the case of search engines, Web robots crawl the World Wide Web and index the inforation in a local database. When a user inputs a search ter, queries are run on the local database and results are returned. Once the up-front investents have been ade, the arginal costs are relatively low. Therefore, we assue that Fir 1 can at least reain at its original quality q 0 by incurring sall arginal costs. Hence, we restrict q 1 to be at least as high as q 0. Fir 1 has a cost function C(q) assued to be strictly increasing and strictly convex. Fir has a cost function λc(q), where λ ranges between zero and one. When λ = 1, neither fir has a cost advantage. When λ < 1, the entrant has a cost advantage. λ captures the idea that the costs associated with technology are constantly falling (or, alternatively, that the technology itself is iproving). The incubent, since it established its infrastructure in Period 0, is locked into the old technology. These are onetie setup costs that enable repeated use of the product. The arginal cost of additional users is low in practice. We assue it to be zero. On any given visit to a particular engine, a user ay not be fully satisfied with the outcoe. The higher the quality q of the engine, the greater the chance that the user will be satisfied. We assue that for a given visit, a user is satisfied with the results with probability p(q). Suppose a user in Period 1 tries out Fir 1 first. Then, with probability p(q 1 ), the user finds the inforation she needs and her quest for inforation ends. With probability (1 p(q 1 )), the user is dissatisfied with the inforation received. Then, if Fir s product is of sufficiently high quality, she proceeds to use it. We say that Fir has a residual deand in this case; there are soe users who used Product 1 who still want to saple Product. To facilitate analytical solutions, we assue that p(q) = q, and C(q) = kq. Paraeter restrictions are iposed to ensure that chosen quality levels satisfy 0 < q < 1. As shown below, interestingly, firs deand and profit functions are discontinuous, depending on whether the fir has a higher or lower quality than its copetitor. Therefore, though equilibria with siilar qualitative features exist with ore general functions for deand, cost and probability, characterizing these analytically is quite difficult. Revenues of each fir are directly linked to its deand; we assue that a fir earns advertising revenue of r in each period per consuer that visits it. As noted by Hoffan and Novak [10] and Dewan et al. [6], advertising revenues depend on the nuber of hits a Web site obtains. 5 Therefore, the profit of an engine can be written as π = D(q)r kq, where D is the deand, and r is the revenue per user. For convenience, we define c = k/r, and renoralize the profit function by dividing by r and write it as π = D(q) cq. This renoralization affects the level of profit earned by a

7 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 143 Figure. Consuer Deand When q > q 1 > q 0 fir, but not its profit-axiizing quality level (i.e., its strategy). In the analysis below, our ain concern is with the strategies of the firs, and whether profits are greater or less than zero. Neither of these is affected by the renoralization. In Period 0, D(q 0 ) = a + bq 0 consuers use Fir 1, as it is the only product in the arket. Entrant enters in Period 1. Any user ay be dissatisfied with the results offered by her first-choice product. As entioned before, all dissatisfied users will switch to the other product if its quality is high enough. To illustrate the deand of both firs, consider the following two cases. Case 1 Let q > q 1 > q 0. Then, Fir has the high-quality product and Fir 1 the low-quality one in Period 1. Figure shows the deand curve. Consider the deand faced by Fir in Period 1. The users D D 0 = b(q q 0 ) are all new users in this period (that is, they did not use any product at Period 0). All new users will use the higher-quality product first. In this case, since q > q 1, all these users visit Fir as their first choice in Period 1. In addition, all of the users D 0 who visited Fir 1 in Period 0 will also switch to the higher-quality entrant in Period 1. Hence, the total deand for Fir is siply a + bq. With probability (1 q ), each of these users will be dissatisfied with the output of Fir. Users in the region [D 1, D ] end their search, since q 1 is too low for the. The reainder will switch to Fir 1. Hence, the deand for Fir 1 is D 1 = (1 q )(a + bq 1 ). Now, the profit of Fir is ( ) ( ) ( ) π q, q, q = D q, q, q C q = a+ bq λ cq, (1)

8 144 TELANG, RAJAN, AND MUKHOPADHYAY where we use the noralization c = k/r. The profit of Fir 1 is the su of its profits in Periods 0 and 1, with no discounting. The cost of Fir 1 is cq 0 in Period 0, and c(q 0 q 0 ) in Period 1, for a total cost of cq 1. Its deand in Period 0 is a + bq, and deand in Period 1 is given by D 1. Hence, ( ) ( )( ) f q q q a bq q a bq cq π,, = () Case Now Fir 1 is the high-quality one in Period 1, and Fir the low-quality one. In particular, suppose that q 1 > q and q 1 q 0, where q 0 can be greater or less than q. Then, the first- and second-choice deands for the products change in an appropriate anner. The profit function of Fir 1 is now deterined to be ( ) ( ) ( ) t q q q a bq a bq cq π,, = (3) Notice that, in this situation, all users who used Fir 1 s product in Period 0 coe back to it as the first choice in Period 1, since it is the higher-quality fir. Hence, there is no residual deand for Fir 1: any user dissatisfied with Fir has already tried Fir 1 s product and found it unsatisfactory. The profit function of Fir is ( ) ( )( ) f q q q q a bq cq π,, = 1 + λ. (4) If q = q 1, then firs are assued to share new custoers equally (that is, those consuers in the region b(q 1 q 0 )). Notice the discontinuity in the profit functions associated with being a high-quality instead of a low-quality provider. 6 This discontinuity has the sae flavor as the discontinuity in the Bertrand odel of duopoly. The difference is that, in a standard Bertrand odel, if one fir has a higher price than the other, its deand falls to zero. In our odel, if one fir offers lower quality than the other, its deand does fall, but can reain strictly positive because no fir can perfectly satisfy its consuers. Dissatisfied consuers will visit other products or services. This feature of this arket, we argue, ay allow any firs of differing qualities to reain in the arket, even though the price discriination is not possible. The solution concept we use is pure strategy subgae-perfect equilibriu. Subgaeperfection is natural here, since it is a sequential gae. Even in the second period, pure strategy equilibriu exists, so we do not consider ixed strategies. First, we show that, in equilibriu, Fir 1 sets q 1 = q 0. Suppose, instead, that q 1 > q 0. Then, Fir 1 incurs a total cost of cq 1 over the two periods. Increasing first-period quality to q 1 leads to additional deand (hence additional revenue) in Period 0, at no extra cost (since the aount c(q 1 q 0 ) is incurred anyway in Period 1). Hence, a strategy that has q 0 < q 1 is doinated by one that has q 0 = q 1. Note that the proposition below holds for all values of λ. All proofs are in the Appendix.

9 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 145 Proposition 1 (Incubent Quality Choice): In equilibriu, the incubent chooses the sae quality in both periods, so that q 1 = q 0. For the rest of the paper, we set q 1 = q 0. Therefore, in equilibriu, there will be new users in Period 1 (that is, users who did not attept a search in Period 0) only if q > q 1. Next, we define the quality levels the incubent and entrant would offer if each were a onopolist. Let q 1 be Fir 1 s onopoly level, and q that of Fir. Then, q 1 solves the proble ax q π 1 (q) = (a + bq) cq. The first-order condition is bq cq = 0, fro which we have q 1 = b/c. The onopoly level of the entrant, q, is found as the solution to ax q π (q) = a + bq λcq. The first-order condition is bq λcq = 0, fro which we have q = b/λc < q 1. We aintain the following assuptions on the paraeters: Assuption 1: (i) λc > (a+b). (ii) a + (b (1 b/c))/4c > 0. Part (i) of the assuption ensures that the cost stays within a range such that the entrant will not be profitable at a quality of 1. Therefore, it will choose a quality strictly less than 1. Since λ 1 in this odel, this also ensures that the incubent will not choose a quality of 1. Part (ii) ensures that the entrant akes a strictly positive profit when λ = 1, the incubent offers its onopoly quality q 1 = b/c, and the entrant plays a best response. If λ is strictly less than 1, then this condition is ore likely to be satisfied, so it is sufficient to assue at λ = 1. As explained below, this condition is required in Proposition. No Cost Advantage for Entrant FIRST, CONSIDER THE CASE OF λ = 1; that is, the entrant has no cost advantage over the incubent. Under the assuptions ade, we show that the incubent, Fir 1, will offer its onopoly quality level, and the entrant will be the follower at Period 1. In this situation, the presence of the entrant does not affect Fir 1 at all. It continues to offer a onopoly quality and earn a onopoly profit. However, Fir still finds it profitable to enter the arket because of deand fro the custoers of Fir 1 that are dissatisfied with the results of their search query. For exaple, Lycos and Infoseek, which entered the arket after Yahoo, are widely perceived as lower-quality engines. Proposition (Low-Quality Entrant): Suppose the entrant does not have a cost advantage. In equilibriu, the incubent engine offers the onopoly quality (q 1 = b/c) in both periods and the entrant chooses a lower quality q = (b(1 q 1 ))/c < q 1. That is, despite there being no price difference across the two products, the lowerquality product also survives in equilibriu. This is because of the residual deand. If there is no residual deand, then (absent horizontal differentiation, which we consider in the fourth section) a lower-quality engine cannot survive in the arket. For the consuer, both products have the sae price (zero). Hence, the arket cannot sustain a lower-quality product unless there is residual deand fro the higher-quality fir.

10 146 TELANG, RAJAN, AND MUKHOPADHYAY If there is one very high-quality engine in the arket, residual deand will be negligible, and the duopoly is unsustainable. The high-quality engine becoes a onopoly. Part (ii) of Assuption 1 ensures that, when the incubent offers a onopoly quality, the entrant can enter the arket at a low-quality level and still ake a positive profit. To deonstrate the nature of the strategic interaction between the firs, we consider a nueric exaple and depict the best response of the entrant to every possible q [0, 1] that the incubent ight offer. This exaple represents a base case, and all other exaples in the paper consider variants of the sae paraeters. The paraeters for this exaple are a = 0.05, b = 1, c =, λ = 1. Figure 3 deonstrates the reaction function of Fir, given a quality level q 1 chosen by Fir 1. When q 1 is low, close to zero, Fir has the high-quality product, and chooses its onopoly level, q = 0.5. As q 1 rises to just above 0.5, Fir prefers to copete with Fir 1, and chooses q = q 1 + ε. However, at q 1 = q[ (0.43 in this exaple), Fir is indifferent between being a high-quality and low-quality provider. For q 1 q[, Fir has the low quality, and its quality falls as q 1 increases. Finally, when q 1 rises above 0.6, Fir exits the arket altogether (i.e., sets q = 0). The reaction function, therefore, has two points of discontinuity. The first is at q[, as entioned before. The second occurs when Fir 1 chooses a quality against which Fir can no longer ake a positive profit. Since Fir 1 sets q 0 = q 1, we can odel the best response of Fir 1 as choosing an optial point on Fir s reaction function. That is, Fir 1 coputes its profit for each point on the reaction function, and chooses the quality q 1 at which its profits are axiized. In the exaple, when Fir 1 chooses q 1 = q 1 = b/c = 1/, Fir chooses q = 0.15 < q 1, which is the equilibriu described in Proposition. Notice that when Fir 1 chooses q 1 = 0.43, Fir is exactly indifferent between being a low-quality (q = 0.14) and high-quality provider (q = ε for soe sall ε > 0). Define the critical quality threshold to be the quality level of the incubent such that the entrant is indifferent between offering a lower-quality and a higher-quality product. This threshold is q[ in Figure 3. Entrant Cost Advantage As shown in Proposition, when there is no cost advantage to the entrant, the incubent is the high-quality fir in equilibriu. However, this effect can be doinated by a high-cost advantage for the entrant. That is, under soe conditions, the entrant can also be the high-quality fir in equilibriu. Note that if the incubent and entrant both have access to the sae technology in the second period, the results of the previous section will continue to hold. In this section, we analyze the effect of a cost advantage to the entrant (λ). The incubent in our odel has a first-over advantage. By choosing q 1 > q[ (where q[ is the quality level at which the entrant is indifferent between being high quality or low quality, as defined previously), it can always force the entrant to offer low quality in equilibriu. Hence, the effects of the paraeter λ can be analyzed through its effect

11 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 147 Figure 3. Reaction Function of the Entrant on q[. If q[ increases, this is costly for the incubent: it has to offer higher quality than otherwise, to force the entrant to offer a low-quality product. If q[ falls, the incubent is better off. Consider the effect of increasing cost advantage to the entrant. As λ falls below 1, q[ begins to rise. Lea 1: As technology costs λ for the entrant fall, the critical quality threshold q[ of the incubent increases. Note that since q[ < q 1 at λ = 1, there will be a value of λ, λ, at which q[ = q 1. Until this value of λ is reached, it is clear that the equilibriu of Proposition will continue to hold. The incubent offers higher quality, and the entrant offers lower quality. The entrant will offer a real copetitive threat to the incubent only for λ < λ. When q[ > q 1, it becoes costly for the incubent to force the entrant to offer low quality. The incubent now copares its own profit as high quality (as a leader, it ust offer q[ now) and as low quality (where it plays soe best-response strategy, with the entrant offering higher quality). Note that the incubent offers q 1 = q[ > q 1. In other words, the incubent has to offer higher than onopoly quality when the entrant enters with lower costs. This also outlines the strategic advantage the incubent has. It can offer higher quality and force the entrant to becoe a low-quality fir even though the entrant has a cost advantage. Also, note that since q 0 = q 1, the incubent overinvests in quality in the

12 148 TELANG, RAJAN, AND MUKHOPADHYAY first period itself to be a quality leader. Tyagi [1] shows that the fir with lower costs occupies the ost desirable location. Our results extend this result to a two-period setting with differential qualities. In the two-period setting, as long as the entrant does not have too high a cost advantage, the incubent can overinvest in quality and reain the quality leader. Eventually, when the entrant s cost advantage becoes very high, the incubent has no option but to offer lower quality. For exaple, with the eergence of Google, Yahoo is a lower-quality player and generates less traffic than Google. We show first that, when the entrant offers high quality, it offers its own onopoly quality level, q = b/λc. Proposition 3 (High-Quality Entrant): For a high cost advantage (λ < λ l ), the entrant becoes a high-quality fir in Period 1 and offers its own onopoly quality level. The incubent is now a low-quality fir and offers q 1 = (b/c) (b /4λc ) in both periods. Therefore, when the entrant is a high-quality fir in equilibriu, the quality of the incubent is increasing in λ. As λ falls, q rises. The deand for the incubent is now the residual deand created by dissatisfied users of Fir. As this deand shrinks, the quality offered by Fir 1 falls as well. Therefore, the prospect of falling technology costs in Period 1 affects the quality offered by the incubent in Period 0. If the incubent believes that the entrant will copete with a uch superior technology in the next period, then, despite being a first-over, it prefers to offer a lower quality in Period 0 itself. The incubent, therefore, strategically underinvests in quality. It anticipates copetition fro a low-cost entrant and chooses a low quality in the first period itself. Therefore, with sequential entry and quality differentiation, an incubent can exhibit both high and low investent, depending on the cost structure of the entrant. An incubent who expects to be a doinant high-quality player in the future tends to overinvest in infrastructure. On the other hand, an incubent who does not expect to be the leading player in the arket tends to underinvest in infrastructure. If λ is sufficiently low, then q[ can be substantially higher than q 1. At this point, exercising the first-over advantage is costly to the incubent, and it ay prefer to be a low-quality fir in equilibriu. Horizontal and Vertical Differentiation IN THIS SECTION, WE CONSIDER THE EFFECTS of residual deand when there is both horizontal and vertical differentiation. The previous section outlined the iportance of residual deand in allowing the existence of a low-quality engine. Here, we highlight how firs horizontal differentiation strategy changes in the search engine arket as a result of residual deand. With horizontal differentiation, we look at firs choosing location siultaneously rather than sequentially. As Prescott and Visscher [16] note, the sequential entry odel

13 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 149 with firs choosing locations and then prices does not adit an analytic solution. Interestingly, their nueric solution in the duopoly case also yields axial differentiation, which is the standard result in the siultaneous odel as well. First we odel firs choices assuing there is no residual deand, and then we show how residual deand changes their strategies. We now assue that users are heterogeneous in ters of their preferences. For each user, the utility fro a given engine depends on how close it is to his particular taste. 7 In particular, a user ay care about the interface, nature of results, presentation of results, or other such attributes. We assue the taste paraeter is between 0 and 1. Users are uniforly located on the line segent [0, 1], and a user s location deterines his or her taste. A user faces a quadratic transportation cost (i.e., a disutility incurred if an engine does not exactly atch his or her tastes) for using an engine at any other location. 8 As before, we assue that the price for using an engine is zero, and the utility of a user is linear in quality. The net utility to a user fro an engine of quality q located at a distance d fro the user is u(q, d) = q d. Note that, analogous to the previous section, we could set this to a + bq d. Since the scale of the transportation cost is arbitrary, without loss of generality we can set b = 1. Setting a = 0 does not affect the qualitative nature of the results; a value a < 0 would erely iply a iniu quality an engine ust have to obtain a nonzero deand. The engines first choose their locations (i.e., how differentiated they are) and then their respective qualities. The previous literature (see [5]) has established that, in the presence of quadratic transportation costs, products are axially differentiated in a duopolistic arket structure. To siplify the analysis, we assue that engines are syetric in ters of their cost structure. Suppose Engine 1 locates at a distance x fro the left and offers quality q 1 and Engine locates at a distance y fro the right and offers quality q. Figure 4 clarifies this. We also ake the usual assuption that the arket is fully covered in equilibriu, so that each consuer visits an engine. First, suppose there is no residual deand. In equilibriu, the arket of each engine consists of the users close to it. There will be soe consuer (denoted by the solid dot in Figure 4) between the locations of the two engines, who is indifferent between the two. Suppose this indifferent custoer is located at a distance t fro Engine 1. Then, his or her distance fro Engine is 1 (x + y) t. Hence, his or her utility fro Engine 1 is q 1 t, and fro Engine q (1 (x + y) t) = q (1 (x + y)) t + (1 x y)t. We find t by equating these utilities, which yields 1 ( x+ y) q1 q t = +. (5) ( 1 x y) Then, the deand for Engine 1 is D 1 = x + t and for Engine D = 1 (x + t). As before, we assue the noralized (relative to advertising revenue) cost of quality for both engines is cq. Therefore, the profit functions are π = x+ t cq and π = 1 ( x+ t) cq. 1 1

14 150 TELANG, RAJAN, AND MUKHOPADHYAY Figure 4. Horizontal Differentiation Substitute t into the profit equations and take the first-order conditions. This yields the optial qualities q 1 = q = 1/4c(1 x y). If x = y = 0, we have axial differentiation. In this case, the arket would be covered only if q 1 = q 1/4. This requires that c 1. Further, the restriction q 1, q < 1 requires that c > 1/4. Hence, we assue c (1/4, 1]. Now, substitute q 1 and q into the profit equations, and solve for optial locations x and y. This yields x + y = (1 1/( /3 c 1/3 )). Since the firs are syetric, we consider only the syetric solution x = y = 1/(1 1/( /3 c 1/3 )). Now, axial differentiation obtains if x = y = 0. This requires c = 1/4. Hence, for all c > 1/4, axial differentiation fails to obtain. Further, it is clear that x, y < 1/, so that inial differentiation also fails to obtain. Therefore, we have a result that is in contrast to standard pricing odels. In the case of the search engine arket, we get neither axial nor inial differentiation. Note that this result is without residual deand. Differentiation with Residual Deand We next consider the effects of residual deand on firs strategies. Now, users of an engine ay switch to the other engine if they are dissatisfied with the results provided by the first engine. For exaple, users fro Engine will be willing to use Engine 1 as long as their distance fro Engine 1 is no higher than q 1. Users at a distance further than q 1 fro Engine 1 will obtain a negative utility fro the engine and not use it. Thus, there are two cases to consider. Suppose q 1 < (1 x). Then, the user located at 1 (extree right on the line in Figure 4) obtains a negative utility fro Engine 1, so will not use it. All users within a distance q 1 of Engine 1 are willing to use it. Of these, users to the left of Engine 1, and those within a distance t on the right, use it as a first choice. This coprises a total ass x + t. The reaining users of ass q 1 t use Engine as the first choice, but will switch to Engine 1 if dissatisfied. Since the probability that they will be dissatisfied with Engine is still 1 q, the residual deand generated for Engine 1 is (1 q )( q 1 t). Therefore, the total deand for Engine 1 in this case is D 1 = x + t + (1 q )( q 1 t) = x + (1 q ) q 1 + tq. If, on the other hand, q 1 > (1 x), then even the user located at 1 obtains a positive utility fro Engine 1. In this case, all dissatisfied users of Engine will switch to Engine 1. Since a ass 1 (x + t) of users use Engine as their first choice, the residual deand for Engine 1 is (1 q )(1 (x + t)). Hence, the total deand for Engine 1 in this case is D 1 = x + t + (1 q )(1 (x + t)) = 1 (1 (x + t))q. Siilarly, the total deand for Engine can be expressed as

15 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 151 D ( ) + ( -( + + )) ( - ) Ï Ôy+ 1-q1 q 1 x y t q1 if q 1 y = Ì ÓÔ 1- ( x+ t) q1 if q 1-y ( ) (6) Given the deands, the profits for each engine i can be written as π i (q 1, q, x, y) = D i cq i. As before, consider the two-stage gae: at Stage 1, engines siultaneously choose their locations (i.e., their differentiation strategies), and at Stage, they siultaneously choose quality levels. In principle, this gae can be solved in a siilar anner to the one without residual deand. First, consider Stage, with locations already fixed. The first-order conditions for the profit functions yield the optial quality choices. These can then be substituted back into the profit functions, and the optial locations deterined fro that. However, with residual deand, even the qualities cannot be obtained in closed for. 9 Instead, we isolate the effect of residual deand in the following anner. Suppose both engines locate at their optial points when there is no residual deand. We show that the presence of residual deand drives the closer together if costs are sufficiently high, and further apart if costs are sufficiently low. Proposition 4 (Residual Deand and Horizontal Differentiation): There exists a c[ between 0.5 and 1 such that, if c > c[, residual deand leads to less horizontal differentiation, and if c < c[, it leads to greater horizontal differentiation between search engines. Although we cannot deterine the equilibriu of the gae in closed for, we can copute a nueric solution for any value of c. Nuerically, we can also deterine that c[ in Proposition 4. To obtain soe insight into the solution, we plot the equilibriu values of location x and quality q with respect to cost c both with and without residual deand. We show these values in Figure 5, which has c on the X-axis and qualities and locations on the Y-axis. The thick lines indicate the residual deand case, and the dotted lines indicate the case without residual deand. The top two lines in Figure 5 denote quality (q) and the botto two lines denote location (x). Note that once c > 0.344, the differentiation decreases with residual deand (i.e., x increases) and quality increases with residual deand (i.e., q also increases). As one expects, the equilibriu quality decreases with higher cost, but at a uch lower rate when there is residual deand. What is the intuition for such a result? A higher cost for quality leads to lower qualities in equilibriu. But when residual deand exists, this has the effect of increasing the residual deand for each fir (since a greater proportion of users will be dissatisfied with their first search). This in turn provides an incentive for the engines to ove to the center to grab as any dissatisfied users as possible. Also, since the residual deand consists of users who are at a greater distance than an engine s priary users, an engine has an incentive to increase its quality a little (copared to the

16 15 TELANG, RAJAN, AND MUKHOPADHYAY Figure 5. Cost Versus Quality and Location With and Without Residual Deand case of no residual deand) to capture ore of the residual arket. Therefore, with residual deand and horizontal differentiation, we observe a higher level of quality than otherwise, and less differentiation. What is the iplication of this result? Although it is difficult to epirically verify the horizontal differentiation strategy of engines, cursory observation supports this hypothesis. In ters of look and feel, and presentation of results, and even type of results, all ajor engines are soewhat siilar. Moreover, the ajor engines do not try to differentiate across different search categories, although they ay be better at retrieving results fro one category than another [3]. Since pricing is not an option, the desire to attract a copetitor s users restrains the engines fro differentiating axially. Therefore, we should not expect strong horizontal differentiation aong ajor engines. Finally, we briefly coent on asyetric cost structures in this setting. For exaple, suppose Engine 1 has cost c 1 and Engine has cost c > c 1. All else being equal, a lower cost for Engine 1 iplies that, in equilibriu, Engine 1 will have higher quality. This, in turn, suggests that Engine has less incentive to locate close to Engine 1. Hence, in equilibriu, an inferior quality engine (Engine ) would like to differentiate ore fro Engine 1, and the engines will locate further apart.

17 THE MARKET STRUCTURE FOR INTERNET SEARCH ENGINES 153 Conclusion WE SHOW THAT QUALITY DIFFERENCES CAN EXIST in a arket even when the consuer is paying the sae price (zero, for Internet inforation goods) for all goods. New entrants ay enter with low quality, and be viable in the arket. This is because, in any given search session, soe proportion of users will saple ore than one product. This leads to a residual deand for lower-quality products that creates roo for ore firs to enter the arket. Vertical differentiation as a result of this residual deand is different fro that considered in the previous literature, where lower prices are used to sustain low-quality products. The phenoenon of ultiple product sapling extends to any inforation goods on the Internet, including search engines, shopbots, and news sites. Soe users regularly visit ore than one news site, even for the sae news ite. Our results extend ore generally to all such goods, where the quality of content or service provided is probabilistic in nature and the consuer can engage in product sapling at little cost. Of course, a fir in such a arket is aware that soe visitors to its site will soon be consulting a copetitor. Hence, by either placing a direct link to a copetitor (perhaps in the for of an advertiseent), the fir can accrue additional revenue. This phenoenon is evidenced aong search engines, for exaple, with Yahoo and Netscape, which offer links to any other search engines. Firs ay even benefit fro partnering with other firs so that when a user switches, the switch is to a partner in the sae faily (e.g., Yahoo and Google earlier, or Lycos and Hotbot). 10 We also deonstrate that the desire to attract a copetitor s users provides firs with an incentive to locate closer to each other. In other words, ajor engines have fewer incentives to differentiate, and this differentiation decreases with residual deand. Clearly, anagers should take into account the phenoenon of residual deand when devising quality and differentiation strategy. By odeling horizontal and vertical differentiation jointly, we provide a coprehensive analysis of the search engine arket structure. Without residual deand and horizontal differentiation, we can expect a onopolistic arket where one engine will doinate. Finally, our odel also considers the role of new technology. For exaple, Google is considered a top engine because of the new technology it uses in indexing pages ( We also show that an incubent ay underinvest or overinvest in quality, depending on the entrant s technology level. Eventually, the arket equilibriu depends on a cobination of all factors entioned above. Hence, today we see a nuber of firs in the content provider arket that capture soe traffic without being either a first-over or having a higher quality than copetitors. Due to residual deand, our odel points to the fact that ore firs can survive in a arket. One liitation of our analysis is that we consider only a duopolistic arket structure, whereas there are several search engines in the arket at any given tie. While odels with ultiple firs are less tractable analytically, the intuition obtained fro duopoly odels does go through. We expect any search engines to have soe loyal consuers (so that there is an aspect of horizontal differentiation), and also to benefit

18 154 TELANG, RAJAN, AND MUKHOPADHYAY fro residual deand (so that the overall reach of search engines exceeds 100 percent). A future extension should incorporate both these aspects into a odel with ultiple firs. Acknowledgents: The authors thank the guest editors and two anonyous reviewers for their valuable suggestions. NOTES 1. For exaple, see the coscore Media Metrix top 50 sites for January 004, available at The issue of new technology and copetitive advantage has been studied in the IS literature by, aong others, Cleons [4]. 3. Quality in this context ay be ultidiensional. Based on popularity, coverage, precision, and relevance, Lycos ranks lower than Yahoo in ost surveys. See tactics/reviews.php3. 4. Forally, this condition iplies that consuers have a different willingness-to-pay for quality. For a fixed quality, consuers who obtain a positive utility fro the engine will use it. Suppose, for exaple, the expected utility to a user of type θ fro using an engine of quality q is u(θ, q) = θ + α + βq, where α 0, β > 0, and θ is uniforly distributed over [0, 1]. Then, users in the range [ α βq, 1] will use the engine, leading to a deand curve linear in q. 5. Recently, soe engines are using auction to sell their advertising space. Moreover, soe engines allow inforation providers to pay ore oney to be ranked higher in their search results []. But even the coplex contracts can be siplified to per-user or per-ipression rate [6]. As long as engine revenue is soe function of its deand, all our results will continue to hold. 6. This discontinuity iplies, aong other things, that in equilibriu it cannot be that q 1 = q. If q 1 = q, each fir has an incentive to increase its quality by ε > In this section, we odel user s utility, and derive the deand fro this. 8. As entioned by Tirole, the proble is that the linear-cost odel is not very tractable [0, p. 80]. Hence, we assue a quadratic cost. 9. For exaple, if q 1 < (1 x), we have (dπ 1 )/(dq 1 ) = (1 q )/( q 1 ) + (q )/((1 x y)) cq 1, with a siilar expression for dπ /dq. Setting both derivatives equal to zero does not lead to a closed-for solution for q 1 and q. 10. The page contains a list of current partnerships. REFERENCES 1. Beebe, J. Institutional structure and progra choices in television arkets. Quarterly Journal of Econoics, 91, 1 (1977), Bhargava, H., and Choudhary, V. Inforation goods and vertical differentiation. Journal of Manageent Inforation Systes, 18, (Fall 001), Bradlow, E., and Schittlein, D. The little engines that could: Modeling the perforance of World Wide Web search engines. Marketing Science, 19, 1 (1999), Cleons, E. Market doinance as a precursor of a fir s failure: Eerging technologies and copetitive advantage of new entrant. Journal of Manageent Inforation Systes, 13, (Fall 1996), d Aspreont, C.; Gabszewicz, J.; and Thisse, J.F. On Hotelling s stability in copetition. Econoetrica, 47, 5 (Septeber 1979), Dewan, R.; Freier, M.; and Zhang, J. Managing Web sites for profitability: Balancing content and advertising. In R.H. Sprague Jr. (ed.), Proceedings of the Thirty-Sixth Hawaii