How Users Drive Value in Two-sided Markets Platform Designs That Matter

Extant research has popularized the perspective that strong network effects produce “winner-take-all” outcomes, which leads platforms to invest in user growth and encourages investors to subsidize these platforms. However, user growth does not necessarily imply strong user stickiness. Without user stickiness, strong network effects in the current period may fade in future periods, thus rendering a user growth strategy ineffective. By adding a time dimension to network effects, we developed a model of cross-period and within-period network effects to explain how different types of network effects drive value. We emphasize that the cross-period same-side network effect contributes to user stickiness, while the within-period cross-side network effect persists conditional on user stickiness. We propose that one reason for platforms having heterogeneous cross-period same-side network effects is because of the “product learning” mechanism: it is expected that products with higher uncertainty have a stronger cross-period same-side network effect. Based on different drivers, we extend the customer lifetime value model (CLV2) to two-sided platform markets, allowing us to measure how different interventions drive platform value. Using Groupon data, we verify our insights and discuss platform design choices that enhance user stickiness when the cross-period same-side network effect is weak.


Introduction
Stronger network effects are commonly believed to drive platforms toward "winner-take-all" outcomes (Dubé et al., 2010;Katz & Shapiro, 1985, 1986;Park, 2004;Schilling, 1999;Shapiro & Varian, 1998).This belief has prompted platforms to invest in user growth with the hope that jump starting a user base will lead to continuous value creation.Venture capitalists stagnated or fallen.Central questions for both platform investors and managers are as follows.When is a user growth strategy effective and when is it not?How do users drive value in a two-sided market?The answers to these questions can inform both investors and managers about platform strategies.
The goal of this research is to understand how users drive platform value.In this study, we propose that user stickiness is an important factor for platforms to consider in terms of their growth strategy.User stickiness refers to the ability of a platform to continually attract users in each period, in particular, how existing users on the platform help attract future users.Fast user growth does not necessarily imply strong user stickiness.Without user stickiness, strong network effects fade and cannot continue to create value.To understand the effects, we built a dynamic model of within-period and cross-period network effects, which decomposes user value into one part related to creating transactions in the current period and the other part related to sustaining user stickiness.Building on value decomposition, we developed a model of customer lifetime value in two-sided markets (CLV2), which maps user value to different network effects.We then identified the precursors that moderate CLV2 and propose corresponding platform strategies.
We consider a "network effect" to mean that a product or service has the property of rising in value as more people use it (Gandal, 1994;Katz & Shapiro, 1985;Liebowitz & Margolis, 1994;Shapiro & Varian, 1998).Network effects can be classified as "same-side" versus "cross-side" depending on whether the interacting users belong to the same group or not (Chu & Manchanda, 2016;Eisenmann et al., 2006). 2 In this study, we added the time dimension to represent within and cross periods, depending on whether the effect happens simultaneously or over time.Thus, a twosided market can exhibit up to four types of network effects: a within-period cross-side network effect (CNE0, with the subscript 0 denoting within period); a cross-period crossside network effect (CNE-1, with the subscript -1 denoting cross-period); a within-period same-side network effect (SNE0); and a cross-period same-side network effect (SNE-1).Adding the different user groups on the two sides, the four network effects manifest eight factors depending on the source and destination of the effect.To make these factors more precise, we let "  " designate network effects from the user group I to the user group J.For example, in a retailing setting,  0 represents a within-period cross-side network effect from consumers to merchants, and  −1 represents a cross-period same-side network effect from consumers to consumers.We summarize the nomenclature in Table 1.
Not all these network effects necessarily exist on every platform.For digital platforms like our focal platform Groupon, SNE-1 and CNE0 are more important than CNE-1 and SNE0 for the following reasons.First, the cross-period cross-side network effect, CNE-1, should be present on platforms where users experience nontrivial switching costs.Thus, users on one side need to form expectations of the current user base on the other side, which can be affected by the user base in previous periods.Most network effect models are based on this expectation-driven view (Farrell & Klemperer, 2007;Katz & Shapiro, 1985, 1986).On many digital platforms, however, consumers can switch platforms with little cost or can make the participation decision after observing the current user base on the other side.In this case, the cross-side network effect is captured through CNE0.Second, the withinperiod same-side network effects SNE0 play only a substantial role when the direct interactions among users create value, for example, on gaming platforms.In contrast, on platforms such as Groupon and Amazon, neither consumers nor merchants directly interact among themselves, leading to ignorable SNE0.Furthermore, because consumer cohorts always join the platform in sequence, SNE0 can in theory be mitigated by shortening the time period.In other words, the cross-period same-side network effect SNE-1 should dominate the withinperiod same-side network effect SNE0 when the time period is reasonably short.Nevertheless, we acknowledge that there may be some platforms where CNE-1 and SNE0 are nonignorable.In these settings, researchers can modify the model based on ours to quantify the specific value of each network effect that matters.
The extant research has shown the importance of strong  0 for platform success (Chu & Manchanda, 2016;Song et al., 2018).We emphasize that the combination of strong user stickiness and strong  0 is needed to sustain a high CLV2 (see Figure 1).Previous studies have found that user stickiness can be affected by factors such as platform quality (Halaburda et al., 2020;Zhu & Iansiti, 2012).In this research, we study how a platform's current users contribute to user stickiness, corresponding to the cross-period same-side network effect.This effect,  −1 , can originate from two mechanisms: observational learning and word of mouth, where the former refers to future users inferring quality information from observing the actions of current users and the latter refers to current users sharing opinions to influence future users.Depending on the nature of product offerings, platforms can differ in terms of observational learning and word of mouth, leading to different  −1 values.When  −1 is high, the platform's user base can continuously attract future users in each period, and  0 persists.When  −1 is low, the effect of  0 fades over time, and a user growth strategy is thus not viable.

Figure 1. CLV2 and Underlying Driving Forces
Through the use of Groupon data, we estimated different network effects and examined the heterogeneous  −1 .We hypothesize that the learning benefit is stronger for experience goods than for search goods because, prior to consumption, the former has higher quality uncertainty than the latter.Thus, we estimated different network effects for search and experience goods.On the consumer side, the cross-period same-side network effect  −1 was estimated to be positive and significant for experience goods but insignificant for search goods.For both experience and search goods, the within-period cross-side network effect of merchants on consumers  0 was estimated to be positive and significant.On the merchant side, the cross-period same-side network effect  −1 and the within-period cross-side network effect  0 were estimated to be positive and significant for both experience and search goods.However,  −1 for experience goods was approximately 70% larger than that for search goods, while  0 was qualitatively similar across the two categories.In sum, as hypothesized,  −1 presents heterogeneity, while  0 is qualitatively similar between experience and search goods on Groupon.
Based on parameter estimates, we calculated CLV2 in the search and experience goods markets, with the results showing that CLV2 for search goods is lower than that for experience goods.Furthermore,  −1 has direct managerial implications for the effectiveness of user growth strategies.We demonstrate that two common marketing strategies that bring users to a platform are less effective in the search goods market than in the experience goods market.
In practical terms, our model allows for an analysis of different platform strategies based on parameter estimates.We specifically discuss two platform design options that enhance user stickiness when  −1 is weak.First, platforms can work on platform design features to introduce or enhance learning so that user stickiness is boosted.For example, Meituan bought Dianping (China's Yelp) to gain access to their user reviews of local businesses.Likewise, Airbnb's host forum helps hosts solve problems and share tips to improve service (Hardy & Dolnicar, 2017).Such user-generated content complements user learning and can help reduce uncertainty even further, which would motivate users to stick with the platform and thus increase CLV2.Second, platforms can consider investing in platform quality to enhance user stickiness.For example, they may introduce information systems and use value-added services (e.g., easier reservation, better account keeping) to retain users.As platform users become stickier due to these services, CLV2 increases, such that a user growth strategy becomes more cost-effective.This research contributes to both academia and practice.First, we developed a model of within-period and cross-period network effects to complement the static and dynamic models of two-sided network effects in the extant literature.We demonstrate that  −1 and  0 drive value through different mechanisms and that a strong  0 does not necessarily imply a strong  −1 .When  −1 is weak,  0 does not persist; therefore, a user growth strategy may be ineffective.Second, given the estimates of different network effects, we propose a model for calculating CLV2.The existing CLV models do not capture the value created through all network effects and thus may underestimate the CLV of platforms such as Amazon and eBay (Gupta et al., 2004).Our model extends traditional CLV research by incorporating various network effects into the CLV calculation.Using our model, platforms can calculate CLV2 based on market-level data and compare the effectiveness of different user acquisition strategies.Third, we propose that one reason for the heterogeneous cross-period same-side network effects is because of the "product learning" mechanism.We verified our hypothesis by estimating heterogeneous  −1 in different product markets on Groupon.For product markets with a weak  −1 , we suggest that platforms enhance user stickiness before taking a user growth strategy.
The remainder of our paper proceeds as follows.The following section reviews the related literature.Next, we introduce our theoretical and empirical models and then describe our dataset and identification strategies.Following this, we present the empirical findings and then analyze the strategic responses when a platform's user stickiness is poor.We conclude the paper with a discussion of our findings, contributions, and future research directions.

Related Literature
Our paper builds on four research streams: two-sided markets, network effect dynamics, platform design, and CLV.Research findings and disagreements for each stream are discussed below.

Static Models of Two-Sided Markets
Static theoretical models have been developed to study the price structure of two-sided markets (Armstrong, 2006;Parker & Van Alstyne, 2005;Rochet & Tirole, 2003), highlighting how one side of a market attracts the other side.Thus, platforms can use a price stimulus for one side to get both sides on board.Empirical works in this context have focused on estimating the network effect size and discussing its impact on market surplus (Rysman, 2004).However, due to the intrinsically dynamic nature of platform growth, static models of two-sided network effects ignore the chain of value creation over time (Armstrong, 2006;Parker & Van Alstyne, 2005;Rochet & Tirole, 2003).Therefore, our paper extends these static models by adding the time dimension and then further clarifies the mechanisms through which different network effects create value.We emphasize that a user growth strategy cannot succeed when network effects do not persist, which is in line with McIntyre and Srinivasan (2017), who argued that little empirical work has investigated the persistence of network effects over time.

Dynamic Models of Network Effects
The literature on network effects exhibits a long-standing interest in market evolution.The decision of platform adoption is often modeled as a function of users' expectations of network size (Cabral, 2011(Cabral, , 2019;;Katz & Shapiro, 1994;Shapiro & Varian, 1998).The assumption that network effects depend on user perceptions of network size is reasonable when user switching costs are high.Platform research has shown that the switching cost is an important determinant of user value (Farrell & Klemperer, 2007).Currently, however, user switching costs on many digital platforms are low.For example, for search, rideshare, group buying, and e-commerce platforms, users can easily observe the offers on competing platforms and switch without being locked in.The possibility of platform switching can dissipate rent and intensify competition (Rysman, 2009).Thus, a large user base in one period does not necessarily guarantee user adoption in the next period.
Empirical contributions in this domain have estimated the magnitude of different network effects.For example, Chu and Manchanda (2016) quantified the extent to which current consumers attract future sellers (and vice versa) in an ecommerce market.Li et al. (2021) further studied user growth across multiple categories of goods and across different time periods on a Daily Deal platform.Insights from these studies show how to optimize investments conditional on the effectiveness of the user growth strategy.
Unlike conventional wisdom, our research focuses on the important role of user stickiness on a platform.User stickiness can be affected by factors such as platform quality (Halaburda et al., 2020;Zhu & Iansiti, 2012): everything else being equal, users are stickier on (i.e., returning to) platforms with high quality.In this study, however, we decomposed user values into different network effects and propose that the cross-period same-side network effect is also critical to user stickiness.Our study explicitly examines the role played by  −1 in terms of CLV2 and a platform's user growth strategy.Parker et al. (2017) showed that firms are transitioning from carrying out production themselves to having their external partners create value, as in the case of Uber and the Amazon marketplace.The design of network effects becomes particularly salient in this context (Schrage, 2012).In this research, we underscore the importance of the persistence of network effects.By investing in platform designs that enhance user stickiness, network effects persist, and each user's value is magnified during the platform's growth.Two types of strategies have been proposed.First, platforms can help current users create more value for future users, leading to stronger network effects and thus a boost in user stickiness.For example, platforms can invest in connecting users to one another, such as enabling file sharing, creating complements, and hacking compatibility (McIntyre & Subramaniam, 2009).Second, platforms can enhance network effects through sustained resource commitment and architectural design (McIntyre & Subramaniam, 2009).For example, firms can give away reference designs and system developer toolkits (Boudreau, 2010) or open application programming interfaces (Benzell et al., 2017).These design choices improve platform quality and thus increase user adoption and stickiness (Halaburda et al., 2020;Tellis et al., 2009;Zhu & Iansiti, 2012).

Customer Lifetime Value
An important stream of research that is closely related to firm value is the concept of CLV in marketing (Gupta et al., 2004;McCarthy & Fader, 2018).This line of research emphasizes user retention and explicitly incorporates user retention into CLV calculations (Gupta, 2005;Gupta et al., 2004;Pfeifer & Farris, 2004).However, extant CLV models are primarily proposed for non-platform businesses and thus omit network effects and ignore the positive externalities that customers create for one another.One primary objective of our research is to explicitly model how network externalities create value across user groups and over time.We therefore extend the CLV framework to two-sided markets, denoted as CLV2.Furthermore, our framework can shed light on the mechanisms through which a specific strategy affects CLV2.In other words, we decompose CLV2 into the value created through within-period network effects and the value created through cross-period network effects.We believe that this framework provides an innovative method to quantify the impact of a strategy on CLV2 so that it is feasible to evaluate and understand the return of different strategies, such as advertising, discounting, and architectural design strategies.Note that user stickiness in this research focuses on platforms' ability to use existing users to attract future users, which is different from user retention.While user retention is only related to existing users' adoption choices, user stickiness through cross-period network effects is related to both existing users and new users.

Modeling Approach
This section presents our model specification.First, we explain how different types of network effects influence the dynamics of two-sided markets and specify an aggregate-data, discrete-choice model to study users' participation choices (Berry et al., 1995).Second, based on the estimates of network effects, we present a model to calculate CLV2.Third, we discuss heterogeneity in network effects across product markets and demonstrate how the strength of  −1 affects the effectiveness of the user growth strategy.

Demand Dynamics in Two-Sided Markets
Building on the static model of Parker and Van Alstyne (2005) and the dynamic model of Chu and Manchanda (2016), we developed a model of within-period and cross-period network effects to study demand dynamics for two-sided markets in settings like retailing.Figure 2 describes the market evolution on the two sides of an online platform, which allows consumers (merchants) in the previous period to influence consumers (merchants) in the current period.To illustrate the market evolution, let us consider a shock to the consumer base at  − 1 and its dynamic effect, as shown in Figure 2. First, this shock affects the consumer base at  via  −1 .Second, an increase in the current consumer base leads to growth in the current merchant base via  0 .The size of each user group and the prices charged to users combined determine the profit during each period in the dynamic system.Here,   ,   ,    , and    correspond to the consumer base, merchant base, price for consumers, and price for merchants, respectively, in period .The same logic applies to a shock to the merchant base.Next, we present how these different network effects are modeled empirically.For exposition, we refer to the two sides as the "consumer side" and the "merchant side," although our modeling framework can be applied to any two-sided market beyond retail.
An important question is why a user growth strategy works for some platforms with strong cross-side network effects  0 but not for others.One explanation is that platforms differ in user stickiness due to heterogeneous same-side network effects  −1 and  −1 .The mechanism underlying heterogeneity could be a learning effect: In the case of CC −1 , future users can "learn" information about product quality from the current users on a platform.Existing literature has identified two primary learning mechanisms: observational learning and word of mouth.The former is based on existing users' choices and actions, and the latter is based on opinions and preferences (Chen et al., 2011;Li et al., 2011).Observational learning based on choice has been documented in papers on social influence (Bikhchandani et al., 1998;Bollinger & Gillingham, 2012;Du & Kamakura, 2011;Zhang, 2010).Informing consumers of other people's choices can reduce the uncertainty of product quality and induce learning, which does not require consumers to share their consumption experience.This effect is also similar to the herding effect, i.e., the fact that a product is popular is informative and leads to more future choices (Bikhchandani et al., 1998;Zhang & Liu, 2012).In contrast, word of mouth is preference based, where existing users share their postconsumption opinions to influence future users' choices.For either user group, which learning mechanism is present or dominant on a platform depends on the nature of the business.Accordingly, platforms with different product offerings differ in their magnitude of observational learning and the word-of-mouth effect (see the subsequent analysis for more details).The same observational learning mechanism can also apply to the merchant side.

Merchants can observe what products other merchants offer via
Groupon.With deal sales information, merchants can also observe what policies tend to yield competitive advantage and mimic success.Thus, consumer choices inform not only future consumers but also future merchants.
Per our previous discussion, the cross-period cross-side network effect and within-period same-side network effect are not part of the system of equations in our research.Take the consumer side as an example.The  −1 related to the lags of the merchant base is excluded because we are modeling a setting with trivial switching costs.For many digital platforms such as retailing platforms, consumers make decisions after observing current offerings on the platform and thus do not need to form expectations about current offerings based on previous offerings.The  0 is excluded because consumers on Groupon do not derive utility from having direct interactions with each other.For retailing platforms in general, consumer cohorts join the platform in a sequence, so the concerns related to within-period same-side network effects in theory can be mitigated by shortening the time dimension.

Consumer Side
In this section, we present the model specification on the consumer side.Several factors jointly determine the utility that a representative consumer, , derives from using the platform in market  in period .First, the utility is related to the platform's attractiveness, which is further decomposed into four components: (1) consumer 's intrinsic preference for the focal platform    (in subsequent analyses, superscript "" denotes the consumer side), which is assumed to vary by market but remain constant over time; (2) the number of consumers in the previous period  ,−1 to capture  −1 ; (3) the number of merchants   to capture  0 that is due to the quantity of supply; and (4) average price   across available supply at  and other product characteristics   .Furthermore, the utility depends on (i) the attractiveness of the rival platform   (Armstrong, 2006;Rochet & Tirole, 2003), which includes the rival's user base in market  at ; (ii) the macroeconomic trends of industry demand and seasonality    that are common to all markets but unobservable to the researchers; (iii) the unobserved market-and time-specific shocks to demand    ; and (iv) the idiosyncratic error    .
There are also market-specific demand shocks    that make the platform more or less attractive across different markets over time.These shocks are observed by platforms and users but unobservable to the researchers, causing an endogeneity concern (more on this in subsequent sections).All other factors are absorbed in the idiosyncratic errors    .
Following convention, we now denote the utility as    =    +    and normalize that of the outside option as zero, which corresponds to consumers choosing rivals or not participating in the market.Assuming that    follows independent and identically distributed (i.i.d.) Type-I extreme value distribution, we derive the market share of the focal platform as follows: (3) Thus, the platform's relative market share is as follows: where  0  is the total market share of the outside option and   is the total market size for consumers in market j at time t.Estimating the parameters of interest based on Equation (4) poses challenges due to endogeneity and market heterogeneity.We discuss our identification strategy in detail in the Estimation and Identification section below.

Merchant Side
The utility specification for merchants is similar to that for consumers.For a representative merchant  in market  at time , the utility of working with the platform is related to the following factors: (1) merchant 's preference for the focal platform    (in subsequent analyses, superscript "m" denotes the merchant side), which is assumed to vary by market but remain constant over time; (2) the number of existing merchants  ,−1 to capture  −1 ; (3) the number of current consumers   to capture  0 ; (4) product price   ; and (5) whether the merchant has had deals on platform   , which captures the merchant's existing experience with the platform.
A merchant's utility further depends on (i) the attractiveness of the rival platform   , (ii) the macroeconomic trends of the industry supply    that are common to all markets and unobservable to the researchers, (iii) the unobserved marketand time-specific shocks to demand    , and (iv) the idiosyncratic error    .Formally, the net indirect utility for a merchant is specified as follows: (5) Following a similar step as that for the consumer side, we denote the utility as    =    +    and normalize that of an outside option as zero.Assuming that    is i.i.d.Type-I extreme value distributed, we derive the relative market share of merchants choosing the platform as follows: where  0  is the market share for merchants choosing the outside option and   is the market size for potential merchants.Furthermore,   and   are the average deal price and characteristics, respectively, in market  at .The coefficients are thus estimated based on Equation (6).Table 2 presents the notation and definition for all our variables and parameters.

CLV2
In their classic work, Gupta et al. (2004) defined customer lifetime value (CLV) as the expected sum of the discounted future earnings generated by a customer.Following their definition, we developed a model to estimate CLV2.Let us take the consumer's CLV2 as an example.We assumed a one-unit increase in the consumer base and quantify the total increase in the future consumer base and merchant base.Then, we calculated CLV2 as the total increase in profit brought about by the increased future consumer base and merchant base.Transform elasticity into a marginal effect: To compute the total increase in the future consumer base and merchant base brought about by a one-unit increase in the consumer base, we first transformed the estimated elasticity of the market share into a marginal effect of the user base using the following standard relationship: where  ‾  is the average market share of  (Berry et al., 1995;Trusov et al., 2009).Thus, the marginal effects of (1 −  ‾  ) for the two sides, respectively.The marginal effects of the deal price on the consumer and merchant bases are Total increase in the user base: We use the formula for the sum of a geometric series to calculate the total increase in future consumer and merchant bases.First, a one-unit increase in the consumer base leads to Total increase in profit: For each consumer and merchant, we used   and   , respectively, to denote the value they generate in one period.We used the discounted cash flow approach to compute value perpetuity (Brealey et al., 2020).Thus, consumer value is represented by the total increase in profit associated with a one-unit increase in the consumer base. 3Formally, we used 2  to denote the consumer's CLV2, which is computed as follows: 3 To simplify this calculation, we assume that there is no discount in profits across periods.
Based on the same logic, we used 2  to denote the merchant's CLV2, which is computed as follows: According to Equations ( 7) and ( 8), 2  and 2  are jointly determined by both  0 and  −1 .The existence of  −1 makes the influence of  0 persist over time and thus amplifies the impact of  0 on 2  and 2  .Therefore, a weak  −1 can explain why platforms with a strong  0 may experience a failed user growth strategy.
In the following sections, we estimate network effects and CLV2 using the empirical setting of Daily Deals.We then develop and test hypotheses about heterogeneous  −1 based on the characteristics of this market and discuss our identification strategies.

Data and Estimation
This section presents the empirical setting and our identification strategy.

Empirical Setting and Data
We estimate our model and discuss the analysis of the user growth strategy using data from the Daily Deal industry.Daily Deal platforms such as Groupon emerged around 2008 as two-sided markets connecting merchants and consumers to discounted deals.Figure 3 shows the growth of Groupon's consumer base, profit, revenue, and market value over the past 10 years.

Figure 3. Groupon Growth from 2010-2019
Groupon experienced rapid growth in its first several years, before reaching a revenue of $2.6 billion in 2013, but then its growth stalled.As a forecast of future growth, its market value experienced a sharp decline in 2012.Analysts observed that the consumer base and revenues exhibited a highly correlated trend.For years, a central question for Groupon senior management was the following: Should Groupon further promote the growth of its user base?This question makes Groupon a valuable setting for our research.
The Daily Deal industry is ideal for answering our research questions for several reasons.Sales of Daily Deals are largely dominated by two leading platforms, Groupon and LivingSocial, making it easier to control for competition effects.During our data collection period, Groupon and LivingSocial made up approximately 59% and 17%, respectively, of the total revenue in the U.S. Daily Deal market. 4In our estimation, we used Groupon as the focal platform to estimate the parameters of interest and used LivingSocial to control for competition.Perhaps more importantly, Groupon offers products and services of various categories (especially including experience and search goods), enabling us to examine how  −1 and  0 vary by product category.By leveraging the panel data structure across categories and markets, we were able to estimate the effects of different network effects in a quasi-multiplatform context while controlling for the effects attributable to platform quality and brand name.An alternative identification strategy 4 Statistica, 2013, http://www.statista.com/statistics/322293/grouponmarket-share-us/would be to analyze different platforms and compare their  −1 and  0 .However, this approach would face tremendous identification challenges because the nature of the products offered by different platforms is indistinguishable from the intrinsic quality of the platforms themselves.
Our sample includes Groupon's largest 108 markets (also known as divisions) from January 2012 to December 2012.Table 3 provides summary statistics by market and week.Note that the platform's consumer base is not directly observed in this study, which is considered a limitation of the data.In the estimation, the consumer base is represented using the transaction volume, which has the advantage of capturing the number of active users rather than that of dormant users who do not make a purchase.Compared with the number of registered accounts, transaction volume is better aligned with the decision-making process underlying Equations ( 2) and ( 5) because consumers and merchants can observe the transaction volume but not the number of inactive consumers.
We would also like to note that we estimated the model using weekly data.In our data collection period, the average deal duration was 4.15 days (median = 4) and 95% of the deals stayed on Groupon for no more than 7 days.Thus, using the week as the time interval better reflects merchants' decision timeline and better captures variation compared to using monthly data.

Estimation and Identification
In this section, we present the estimation and identification strategy for the cross-period and within-period network effects.

Identification Strategy for the Cross-Period Network Effect
We begin by discussing the estimation for  −1 .As shown in Equations ( 4) and ( 6),  −1 is captured by the lag(s) of the consumer and merchant bases.As the dependent variable, i.e., the logarithm of relative market share, is also a function of the user base, our specification is a variant of the dynamic panel linear model.This specification has the advantage of controlling for heterogeneity across panels and providing a distinction between short-and long-run dynamics; however, it requires special attention in the estimation.
It is well known that the ordinary least squares (OLS) estimator for the lagged user base (i.e.,  1 and  2 ) is biased in dynamic panel linear models.By construction, ln  ,−1 in Equation ( 4) and ln  ,−1 in Equation ( 6) are correlated with the market fixed effects.To eliminate unobserved market effects, we applied the first-difference approach proposed by Anderson and Hsiao (1981) and Arellano and Bond (1991).
Furthermore, to identify  1 , the levels of the dependent variable lagged two periods or more are valid instruments in the equation of first-differences (Arellano & Bond, 1991).

Identification Strategy for the Within-Period Network Effect
The identification challenge underlying the parameters for the classic network effect is that the consumer and merchant bases are simultaneously determined in Equations ( 4) and ( 6), causing an endogeneity concern.We address this problem by providing instrumental variables (IVs) for ln   and ln   .
We first present the IVs for the merchant base ln   in Equation (4).A valid instrument should be correlated with the number of merchants but orthogonal to demand shock    .We used two sets of instruments here.First, following the predetermined variable in Arellano and Bond (1991), we used the lag of the merchant base (after logarithm transformation) as the instrument.The argument is that the demand shock to the consumer base in period  should not be correlated with the merchant base in period  − 2 after controlling for all other variables in the model.Second, we also leveraged political advertising as an exogenous variation for identification.Our data collection year, 2012, happened to be a presidential election year in the U.S.During an election year, political candidates and interest groups (including party committees and outside political groups known as political action committees (PACs) and superPACs) invest heavily in television advertising, leading to an increase in ad prices (Moshary et al., 2021).This increased cost of advertising may motivate merchants to look for alternative marketing avenues, such as participating in Daily Deals (Dholakia, 2011;Edelman et al., 2016).Therefore, the amount of political election advertising in the market should be correlated with the number of merchants on deal platforms, providing data variation for identifying the coefficient for ln   .5 We collected data on political advertisements across four types of elections in 2012 (presidential, gubernatorial, House, and Senate elections) (Fowler et al., 2015) and computed the total amount of airtime and ad spending across all candidates and parties in each market  in week . 6In this empirical setting, it often takes some time for the merchant to work with Groupon to determine the terms before the deal shows up on the platform (Li et al., 2018).Thus, we used the political advertising in period  − 1 as the instrument because such a specification captures the time lag associated with the merchant base in .Lagging the advertising variable is also in line with the exogenous assumption that past political advertising should not directly affect current deal demand    .
We followed a similar identification strategy for the consumer base ln   in Equation ( 6).We used two sets of instruments: (1) the lag of the consumer base with a degree of 2 and (2) the precipitation and temperature in each market.
The identification argument for precipitation and temperature is based on the finding that consumers' online shopping behaviors are shown to be influenced by weather conditions.Using data from large-scale field experiments, Li et al. (2017) provides robust evidence that consumers' online shopping behaviors deviate significantly on rainy days compared to sunny days because weather impacts consumers' mood and psychological states.Based on these findings, we argue that the local weekly temperature and precipitation should be correlated with consumers' intention to shop Groupon deals, providing first-stage variation for identification.However, current weather at time  should not be correlated with the current supply-side error    .
Merchant negotiations with Groupon predate Groupon's deal offer dates.When merchants negotiate, the future weather is not known.Therefore, weather during the sales week should not be a determining factor for merchants' decision.The exogenous requirement for using precipitation and temperature as the instrumental variable therefore holds.

Other Variables
Market size: Our specification requires the "size" for each market.On the consumer side, market size is defined as the total number of users who can participate in the Daily Deal industry.Because anyone with internet access can use a deal site, we used the number of internet users as the measure of market size.The data were retrieved from the October 2012 School Enrollment and Internet Use Survey, a supplement to the Current Population Survey (CPS) by the U.S. Census Bureau.
On the merchant side, we measured market size using the number of businesses in each local market in 2012, which was extracted from the County Business Patterns (CBP) database from the U.S. Census Bureau. 7This data source tracks the number of businesses in each market along with their North American Industry Classification System (NAICS) industry code and a short description.For each business, we matched the NAICS with the merchant type on Groupon and LivingSocial and selected the businesses whose categories were available on these two deal platforms.
Price: Finally, the deal price in Equations ( 4) and ( 6) can also be endogenous.A rational platform strategically adjusts the deal price in response to an expected shock in supply and demand.For example, if demand is expected to be low, then the platform is thus incentivized to reduce its price.In the same vein, if supply is expected to be strong, then the platform is also incentivized to adjust its price to increase profits.To address this issue, we used the lagged deal prices as the instrument for   , as proposed and tested in Villas-Boas & Winer (1999).The identification assumption is that the prices from the previous periods correlate with the current price due to common cost factors.However, after controlling for market fixed effects and other variables, they should be exogenous to the supply and demand shock in the current period.This assumption seems reasonable in the Groupon setting because many merchants on the platform provide services to local customers, and the cost of services within the market should thus be correlated over time.Furthermore, Groupon typically features different merchants in consecutive weeks; therefore, lagged prices should be uncorrelated with the current supply or demand shock.
that fell into the same DMA had the same value for the advertising variables.The inclusion of ads from gubernatorial, House, and Senate elections helped address the unequal allocation of campaign resources between battleground and non-battleground states for presidential elections.In our data, only 5 (4.7%)DMAs received zero political advertising in 2012, and the remaining 95.3% had some level of election advertising on TV. 7Retrieved February 2021 from https://www.census.gov/programssurveys/cbp/data/datasets.html.

Experience versus Search Goods
This section develops and tests a specific hypothesis to explain the heterogeneous cross-period same-side network effect  −1 .The need for information can vary by product type and by the extent to which consumers can, in fact, benefit from each other's experiences.Whereas "search" goods like salt or gasoline can be compared on price to know their value, "experience" goods like restaurant meals and haircuts benefit from prior consumption to reveal their value (Huang et al., 2009;Nelson, 1970).In our setting,  −1 largely captures this observational learning effect.
On the consumer side, because the mechanism for observational learning originates from reducing uncertainty, naturally, product categories with higher quality uncertainty benefit more from such learning than do those with lower quality uncertainty, leading to heterogeneous  −1 .
Experience goods and search goods exhibit higher and lower quality uncertainty, respectively.By definition, consumers can search more easily for product attributes of search goods than for those of experience goods.Thus, future users of search goods benefit less from current users, resulting in weaker  −1 .Therefore, we hypothesize that the  −1 of experience goods is stronger than that of search goods.The work of Li and Wu (2018) strengthens this hypothesis as they report that prior sales information on Groupon has a greater impact on consumer choices in the case of experience goods than in that of search goods.
On the merchant side, the difference in  −1 is also caused by observational learning.Since experience goods have higher uncertainty, the strength of  −1 should be stronger for merchants selling experience goods than for those selling search goods.Observing competitors' gains provides an incentive for more merchants to join the platform: merchants who joined the platform in period  − 1 gain a competitive advantage in period  compared to merchants who did not join the platform in period  − 1, so nonparticipating merchants may choose to participate.Note that merchants' switching costs between platforms may not be negligible, contributing to  −1 .But the switching cost is not expected to differ between experience and search goods.Thus, if any differences in  −1 are identified between experience and search goods on the merchant side, they should be because of the heterogeneous observational learning effects.
To test our hypotheses, following Li and Wu (2018), we administered an online survey to classify whether a product or service possesses more experience attributes or search attributes.A total of 818 respondents were asked to imagine that "user reviews are not available" on a platform and then to assess each category based on the following question: "To what extent is it easy (difficult) to evaluate quality without seeing or trying it?"8A score closer to "easy" (coded as 1) corresponded more to a search good and vice versa for experience goods if the score was closer to "difficult" (coded as 7) (Nelson, 1974).To ensure that the questionnaire workload was reasonable, each respondent rated a random selection of 15 categories, with each category, on average, being rated by 136.3 respondents.
Table 4 presents the summary statistics from the survey.Across all 78 categories, the mean score was 4.72 ( = 1.02), and the median was 4.97.We used the median split and coded each category into two groups: experience goods if the average score of that category is greater than the median and vice versa for search goods.Among our categories, the highest scores were experience goods including "hair salons," "facials," and "makeup services," while those with the lowest scores were search goods including "holiday decor," "gifts: candles, phone cases, and stationary," and "office supplies."Note that the mean split yielded very similar results since there were only two categories in the middle switching groups.In subsequent analyses, we used the online survey responses to classify the categories.
To cross-check our experience-versus-search-goods classification, we conducted another survey using a different offline design.Twenty-eight business school graduate students at a major U.S. university participated in the survey.Each student was asked to code the product/service categories on Groupon using a binary response: a "search" type (denoted as   = 0) or an "experience" type (denoted as   = 1).Across all respondents, Fleiss's kappa was 0.603, indicating a moderate level of average intercoder reliability (Fleiss, 1971).
For each category, we used the majority vote: a category was classified as the experience type if more than half of respondents coded it as   = 1.Out of the categories, eight (10.3%) were coded as search goods by MTurk respondents but were then coded as experience goods in the validation survey, 5 (6.4%) switched from the experience to search type, and the remaining 65 (83.3%) were consistent between the two surveys.We estimated the main models based on the validation survey and obtained qualitatively similar results.See the Appendix for detailed model estimates.
check questions and were hence excluded.The detailed survey design and summary statistics are included in the Appendix.

Parameter Estimates
In this section, we present the parameter estimates.First, we estimated the consumer-side model (Equation 4) for search and experience goods, the results of which are presented in Table 5.The estimates for the merchant-side model (Equation 6) for search and experience goods are presented in Table 6.
For each table, we present the fixed effect estimator and the results based on the DGMM estimator after applying firstdifferencing as well as the instruments.
On the consumer side,  −1 is estimated to be heterogeneous between experience and search goods.While  −1 is positive and significant for experience goods (0.096,  < 0.05), it is nonsignificant for search goods (−0.034,  > 0.10).The fixed-effect estimates yield a similar pattern.As hypothesized, experience goods have higher uncertainty prior to consumption; thus, having more existing consumers can shed light on the quality of goods and attract more future consumers, yielding a positive  −1 .However, the quality uncertainty for search goods can be reduced through information searches (e.g., doing online searches or reading industry reports), so the size of the existing consumer base is no longer found to be important for attracting future consumers.
The pattern is different for  0 . 0 is found to be positive and significant for both categories.The  0 for experience goods is estimated to be 1.439 ( < 0.01), and that for search goods is estimated to be 37% higher than that for experience goods (1.970,  < 0.01).The conclusion is qualitatively the same for the fixed-effect estimates.
On the merchant side, as hypothesized, the cross-period sameside merchant stickiness is found to be stronger in the experience-goods market than in the search-goods market.
−1 is estimated to be positive and significant for experience goods (0.417,  < 0.01), 70% higher than  −1 for search goods (0.242,  < 0.01).The comparison is even more substantial based on the fixed-effect estimates, as  −1 is estimated to be positive and significant (0.190,  < 0.01) for experience goods but statistically insignificant (0.024,  > 0.10) for search goods.9 0 is also estimated to be positive and significant for both product markets: the  0 for experience goods is estimated to be 0.423 ( < 0.01), and that for search goods is estimated to be 0.380 ( < 0.01).The fixed-effect estimates reach a similar conclusion.
Next, we present the results for the control variables.Not surprisingly, price is estimated to have a negative effect on the consumer base and a positive effect on the merchant base.For experience goods, consumer price elasticity is estimated to be −0.656( < 0.01), and merchant price elasticity is estimated to be 0.411 ( < 0.01).For search goods, consumer price elasticity is estimated to be −0.448 ( < 0.01), and merchant price elasticity is estimated to be 0.196 ( < 0.01).Furthermore, after controlling for  −1 and  0 , the number of returning merchants is found to be negatively related to the consumer base for experience goods (−0.278,  < 0.05) and search goods (−0.325,  < 0.01), indicating that overall, consumers respond more positively to new merchants.

Robustness Analyses
To verify the validity of our results, it is critical to assess the extent to which the assumptions are met for the DGMM estimator.The first important assumption is that the IVs used for identification should be exogenous to the errors.We present the results from the Hansen test of overidentification restrictions.On the consumer side, the Hansen test statistics are  2 (101) = 101.88 ( = 0.457) and  2 (101) = 94.20 ( = 0.671) for experience and search goods, respectively.On the merchant side, the test statistics are  2 (103) = 106.05( = 0.399) for experience goods and  2 (103) = 97.56( = 0.633) for search goods.These results indicate that the null hypothesis that all IVs are exogenous is not rejected at the 0.05 level.
Robustness Analysis section on the instruments, we acknowledge that the validity of the instruments cannot be fully proven and encourage the readers to keep this limitation in mind when interpreting this parameter.Next, we present the first-stage diagnostic statistics for the IVs (see Table 7).To examine the strength of the instruments, the first-stage model was fit by regressing the endogenous consumer and merchant bases on their respective instruments, controlling for all exogenous variables, including market fixed effects.We began with the consumer-side model.For experience goods, the first-stage -statistic is 7.859 ( < 0.001) for the consumer base and 24.801 ( < 0.001) for the merchant base.For search goods, the first-stage test statistics are also significant ( = 7.378,  < 0.001 for the consumer base and  = 15.189, < 0.001 for the merchant base).
Although the first-stage -statistics are modest for the endogenous consumer base, we want to emphasize that the IV estimates are substantially different across the two types of products: the  −1 is positive and significant (0.096,  < 0.05) for experience goods and nonsignificant (−0.034,  > 0.010) for search goods.Thus, the heterogeneous  −1 between experience and search products does not seem to be driven by the difference in the explanatory power of the IVs.The first-stage statistics are qualitatively similar for the merchant-side model.The -statistics are modest for both endogenous variables, but the p-values are all smaller than 0.001, even after controlling for all exogenous variables, the time effect, and the extensive set of market fixed effects.
Given the modest first-stage F-statistics, we conducted a weak-instrument robust test (Finlay & Magnusson, 2009;Kleibergen, 2002).The test forms a hypothesis on the parameters for the endogenous variables (i.e., consumer base and merchant base in Equations 4 and 6) and is robust to potentially weak instruments, and thus can shed light on the robustness of our estimates.In particular, our test results show that  −1 is statistically different from zero for experience goods but not for search goods, confirming the differences in demand-side learning across product categories.Furthermore, all other estimates for  −1 and  0 on the demand side and merchant side are statistically significant, yielding additional support for our results.The details on this test are presented in the Appendix.
Next, we discuss the results of the Arellano-Bond test, which examines the serial correlation in the first-difference errors.
Because the first-differences of independently distributed errors are autocorrelated by definition, the hypothesis of a zero serial correlation should be evaluated at the order of two.In our case, the Arellano-Bond AR(2) test results do not offer evidence to reject the zero serial correlation assumption (at the 0.05 level), except for the merchant model for search goods, which we thus acknowledge as a limitation in this work.Finally, we conducted two additional robustness analyses to validate our estimates of network effects.First, we wanted to rule out the alternative explanation that the estimated  −1 and  0 are due to the attractiveness of the platform rather than the network effects in the local market.To conduct this analysis, we replaced the local market's  −1 and  0 effects with their nonlocal counterparts.For the new  −1 , we added up the customer base and merchant base (see Equations 4 and 6, respectively) in the previous period across all the markets on Groupon other than the local market.The  0 variables are constructed in a similar fashion.Because these variables reflect the (lagged) customer and merchant bases of the platform in other markets, they capture the network effects due to nonlocal influences (if any).The parameters are presented in the Appendix.From the results, we see that the nonlocal network effects are all estimated to be close to zero and weaker than the local network effects, indicating that consumers and merchants on Groupon are attracted by the  −1 and  0 within the same local market.
This analysis also provides additional validation of the IV estimates of our main models.The nonlocal variables should be uncorrelated with the local market political advertising and weather conditions; therefore, our instruments do not provide the needed variation for identification.As expected, the corresponding coefficient estimates are greatly reduced, providing validation evidence for our IVs in identifying local network effects.
Second, we conducted an analysis that splits merchants into two groups: first-time merchants who offer new deals on Groupon and returning merchants who have previously offered deals on Groupon.According to our learning mechanism, we expected  −1 to exist for returning merchants but not for new merchants.For returning merchants, consumers can observe the sales volume for their existing deals and the uncertainty of the quality is thus reduced through learning.In contrast, no learning 10 This measure is also consistent with the ratio of revenue to gross billing in Groupon's 2012 annual report.Groupon's gross billing and revenue in takes place for first-time merchants because existing customers have not yet seen those merchants on Groupon; therefore, no information has accumulated on them.In other words, previous sales are noninformative in terms of their quality for new merchants.Indeed, our results confirm that the cross-period same-side network effect  −1 is much larger for returning merchants than for first-timers (see the Appendix).The coefficient is estimated to be positive and significant (0.133, p < 0.05) for returning merchants but statistically indistinguishable from zero (−0.030,  > 0.10) for new merchants.This result provides additional evidence to support our proposed learning mechanism.

Analysis of a User Growth Strategy
In this section, we calculate 2  and 2  based on the parameter estimates from the models and examine the effectiveness of the user growth strategy.

CLV2
To calculate 2  and 2  , we first transformed the unit that is free from the elasticity coefficients from Tables 5 and 6 into a unit that is denominated by marginal coefficients, which are needed for marginal analysis in Table 8.We then used the calculation of 2  as an example for demonstration.
Since the marginal cost of serving an additional consumer is almost zero for digital platforms, the   of Groupon is equal to its gross profit generated by a consumer in each period.Note that Groupon's gross profit is not directly observed in the data.Thus, we used revenue multiplied by gross margin to approximate gross profit.First, we used half of the gross billing as a proxy for revenue following Dholakia (2011). 10  North America in 2012 was $2,373.2 million and $1,165.7 million, respectively, yielding a ratio of gross billing to revenue of 49.12%.Gross billing is measured as the total sales on Groupon, which is calculated from the transaction volume and deal price.Note that the transaction volume was used as a proxy for the consumer base.Thus, we used the increased consumer base multiplied by the deal price to calculate the gross billing.Second, in Groupon's 2012 annual report, gross profit accounted for less than 70% of Groupon's revenue.In later years, this percentage decreased to approximately 50%.We use 70% of Groupon's revenue to represent gross profit, which sets an upper bound of 2  .
According to Equation ( 7), we calculated 2  in the experience and search goods markets, with the same logic applying to 2  .Based on our calculation (see Table 9), a consumer is worth $62 and a merchant is worth $31,654 in the experience goods market.In the search goods market, a consumer is worth $54 and a merchant is worth $26,338.
Although the  0 and  0 of experience and search goods are all positive and significant, the 2  and 2  in the experience goods market are larger than those in the search goods market as a result of the differences in  −1 and  −1 .In the experience goods market,  −1 amplifies 2  by 1.1, and  −1 amplifies 2  by 1.7.In the search goods market, the value created on both sides of the market has less opportunity to be amplified over time due to the absence of a significant  −1 and a relatively weaker  −1 .For search goods, the positive shocks received by a product market in one period quickly dissipate and do not result in persistent value creation.

Cost-Benefit Analysis of User Acquisition
Platforms often undertake marketing to increase their consumer base, hoping that through network effects, the incremental growth in their consumer base can lead to future increases in their consumer and merchant bases.In this section, we consider two common marketing strategies: (1) advertising in order to, in essence, buy consumers and (2) implementing price discounts.For each marketing strategy, we computed 2  according to the previous section and compared it with the corresponding consumer acquisition cost.Note that this analysis can be applied to both the consumer and merchant sides.However, platforms' costs to acquire merchants are typically unavailable to researchers, which is the case for Groupon.In its annual report, Groupon reports neither its expenses for acquiring merchants nor the growth of its merchant base.As a result, our cost-benefit analysis focuses on the consumer side.

Intervention-Consumer acquisition through advertising:
We first look at a scenario where the platform uses advertising as the user growth strategy.The consumer-side acquisition cost in this case is calculated as the total marketing expense divided by the total number of acquired consumers (Gupta et al., 2004;McCarthy & Fader, 2018).In its 2012 annual report, Groupon reported a total marketing expense of $336.9 million, with 7.3 million consumers acquired worldwide, yielding an acquisition cost of $46 per new consumer globally.In that same year, in North America, the marketing expense was $105.9 million, but the report did not include the total number of consumers acquired in North America.Based on the proportion of gross billing in North America ($2.37 billion) relative to worldwide sales ($5.38 billion), we estimated the number of new consumers in North America to be 3.22 million, which yields an estimated cost of $33 to acquire each new consumer.Table 10 reports Groupon's net 2  in North America, which is equal to 2  minus the consumer-side acquisition cost.This net 2  is $29 in the experience goods market and $21 in the search goods market, which implies that consumer acquisition through marketing is less effective in the search goods market than in the experience goods market due to the different 2  in these product markets.Note that the fixed and operating costs associated with serving consumers are not included in this calculation.If we amortize these costs to each consumer, then net 2  becomes negative.

Intervention-Price cut:
In a two-sided market, a price cut can stimulate greater participation on the discounted side as well as the other side of the market through various network effects.We analyzed the cost and benefit of a price cut on each side of the market.We used the formula for the marginal effect of the deal price in the CLV2 section above to calculate the increase in the consumer and merchant bases for the promotion on each side, which is then multiplied by 2  and 2  , respectively, to obtain the total increased value.
Table 11 illustrates these results.For experience goods, a $1 cut in the deal price brings 33.2 additional consumers and 0.063 additional merchants, yielding a total increased value of $4,035.
Based on the average transaction volume of each market, the promotion cost is calculated at $3,959.Thus, the net gain from a deal promotion is $76.For search goods, a $1 cut in the deal price brings 4.03 additional consumers and 0.008 additional merchants, yielding a total increased value of $435.Based on the average transaction volume of each market, the promotion cost is calculated as $704.Therefore, the net gain from a deal promotion is $ -269.Due to the difference between 2  and 2  in the experience and search goods markets, a deal promotion for experience goods is cost-effective, but that for search goods is not.

Managerial Implications: Using Platform Designs to Enhance User Stickiness
Our research thus far suggests that for platforms with strong  0 but weak  −1 , platform design choices that enhance user stickiness may offer greater leverage than may common marketing strategies that bring users to a platform.
To further illustrate the managerial implications, we conducted simulations to demonstrate how 2  and 2  increase when user stickiness is enhanced.We used 2  and 2  in the experience goods market as an example to simulate the impact of changing  −1 (i.e., for different levels of  1 ).We simulated three scenarios of  1 (0.1, 0.5, and 0.9), and computed the corresponding updated 2  and 2  .Similarly, we computed the impact of changing  −1 by modifying  2 .The simulation results (see Table 12) show that the changes in  −1 and  −1 have a substantial impact on 2  and 2  .
Boosting  1 from 0.096 (observed in our data) to 0.5 improves 2  by 1.8 times and 2  by 1.4 times.
Boosting  1 from 0.096 to 0.9 improves 2  by 9 times and 2  by 5.2 times.Boosting  2 from 0.417 to 0.5 improves 2  by 1.1 times and 2  by 1.2 times.
Boosting  1 from 0.417 to 0.9 improves 2  by 3.4 times and 2  by 5.8 times.Note that changing  −1 has a greater impact on 2  than on 2  .Similarly, changing  −1 has a greater impact on 2  than on 2  .It is also worth noting that participant acquisition marketing becomes more cost-effective as  −1 or  −1 increases.
The scale of these effects raises the question of whether enhancing user stickiness is feasible.Can managerial decisions concerning platform design have substantial strategic growth implications?Evidence from another Daily Deal platform suggests that these outcomes are indeed possible.Meituan was launched in 2010, two years after Groupon, as a Daily Deal platform using an identical business model that leverages group buying power to offer coupons and volume discounts to price-sensitive consumers.Meituan has adopted platform designs that focus on enhancing user stickiness and sustaining network effects.Below, we present and discuss two major platform designs employed by Meituan to enhance its user stickiness.

Enhancing user stickiness through cross-period network effects:
The first strategy adopted by Meituan is to merge with Dianping (the Yelp equivalent in China) to offer more usergenerated content (UGC) on the platform.As indicated by our research, as new transactions generate more user reviews and ratings, UGC on Meituan creates positive feedback among consumers and reduces product uncertainty.As a result, consumers derive value from the accumulated UGC, which is more difficult to find on competing platforms, leading to increased user stickiness.Prior to this merger, user stickiness on Meituan was sustained primarily through observing the size of the customer base, as is modeled in this search.Because a significant proportion of Meituan's offerings are in the experience goods market, such as restaurant deals, providing UGC from existing customers can add an additional boost to the cross-period same-side network effects by further reducing quality uncertainty.Enhancing user stickiness through factors other than network effects: Furthermore, Meituan has invested heavily in its delivery and reservation systems as well as its merchant information management system.Such design improvements increase both consumers and merchants' dependence on Meituan and reroute transactions through Meituan rather than conducting them on other platforms or off Meituan, effectively enhancing user stickiness.
We note that although both strategies have improved Meituan's user stickiness, their underlying mechanisms are substantively different.Incorporating user reviews on the platform increases user stickiness by enhancing cross-period same-side network effects.To leverage this mechanism, managers need to improve information flow among consumers.By doing so, existing customers become the attracting force for future consumers.In contrast, Meituan's second strategy is built on improving platform quality rather than boosting user network effects.We acknowledge that some platforms may find it challenging to enhance their crossperiod same-side network effects in a cost-effective way.In this case, platform managers can learn from Meituan's second strategy and consider value-added services that can attract users to continue using the platform.

Conclusion
We developed a dynamic model of within-period and crossperiod network effects to study how users drive value in twosided markets.Our model clarifies the mechanisms through which different network effects drive value.For certain platforms, weak  −1 may yield poor user stickiness; thus, strong  0 may not persist.Based on our model coefficients, we developed a model to calculate CLV2, extending customer lifetime value literature to two-sided markets.Then, a platform can compare the effectiveness of different user acquisition strategies.We hypothesized that the mechanism behind heterogeneous  −1 is the "product learning" effect: products with higher uncertainty can expect a higher  −1 .We verified our hypothesis by confirming the different levels of  −1 between search and experience goods.We then discussed the platform's strategic response when  0 is strong but  −1 is weak.In particular, our research suggests that a platform design that enhances user stickiness can increase CLV2.After users become stickier, user acquisition marketing becomes more cost-effective.Overall, our findings remind managers not to overemphasize user growth when user stickiness is poor and, instead, focus on platform design choices that enhance user stickiness.
There are several limitations of this work as well as directions for future research.First, platforms have multiple ways to improve their long-term value.Should a platform invest in user growth, the design of network effects, key assets, or other factors?It is worth gathering real-world data to compare the costs and benefits of these strategies and strengthen our business implications in future research.Second, in our empirical setting, we included the competition effect from the primary competing platform.Although this should capture competition in many markets, we acknowledge that there could be other major platforms that are key players in some markets.Future competition analysis could improve these results.Furthermore, we used the instrumental variable approach; therefore, the conclusions depend on the validity of the instruments.
Although we provide various robustness analyses to support the estimates, we acknowledge that these tests cannot prove the validity of the instruments; thus, our results may not be interpreted as causal.Third, our data does not support a distinction between repeated users and new users when we study user stickiness.Previous research has shown that promoting user retention can be more valuable than attracting new users (Gupta, 2005).Therefore, it would be valuable to examine the impact of network effects on repeated users and new users separately.Finally, the strength of the cross-period and within-period effects could differ substantially across different product markets.We show that they vary by product category on one platform.Although building models and estimations using multiple product categories from the same platform allowed us to better address unobserved heterogeneity in platform quality and brand effect, it would still be interesting to conduct future studies to identify and compare cross-period effects across multiple platforms.

Table 5 . Parameter Estimates for the Consumer-Side Model for Experience and Search Goods (Eq. 4)
FE is the fixed effect estimator, and DGMM is the first-difference GMM estimator.*** p < 0.01, ** p < 0.05, * p < 0.1 Note:

Table 12 .
and   in