Data, Competition, and Digital Platforms

The paper studies how data and platform design affect competition, pricing, and surplus distribution in digital markets where consumers can buy on-platform or directly from sellers (off-platform). It asks: (i) how does the precision of a platform’s data affect surplus creation and its distribution across consumers, sellers, and the platform, both on and off the platform? (ii) how do these effects depend on the intensity of competition among multiproduct sellers? (iii) how are outcomes shaped by data-collection and data-sharing mechanisms (e.g., personalized vs cohort-based advertising, organic links)? The context is that large digital platforms use data to run managed advertising campaigns that personalize recommendations (product steering) rather than prices, while sellers maintain parallel direct channels. The paper aims to provide a tractable framework to analyze privacy regimes, steering, and showrooming, highlighting the platform’s gatekeeper and competition-manager roles and the regulatory implications for consumer welfare and market power.

The work builds on the literature on information gatekeepers and biased intermediation (e.g., Baye and Morgan 2001; Armstrong and Zhou 2011; de Cornière and Taylor 2019; Inderst and Ottaviani 2012a,b; Rayo and Segal 2010), emphasizing platforms’ steering power via ranking and sponsored links. It connects to models of platforms’ information provision and privacy (de Cornière and de Nijs 2016; Teh and Wright 2022; Zhu et al. 2022) and to nonlinear pricing and market segmentation under information (Mussa and Rosen 1978; Maskin and Riley 1984; Bergemann et al. 2015; Elliott et al. 2020). The paper relates to data externalities and personalization trade-offs (Acemoglu et al. 2022; Ichihashi 2020, 2021; Bergemann et al. 2022), showrooming and multi-channel retail (Wang and Wright 2020; Bar-Isaac and Shelegia 2020; Anderson and Bedre-Defolie 2021), and to partial mechanism design with outside options (Philippon and Skreta 2012; Tirole 2012; Calzolari and Denicolò 2015; Fuchs and Skrzypacz 2015). It also contrasts managed advertising campaigns with data-augmented auctions (Bergemann et al. 2023) and references empirical and experimental work on algorithmic bidding and platform design.

The authors develop a theoretical model with J differentiated multiproduct sellers who can supply any quality q at cost c(q)=q^2/2. A unit mass of consumers has single-unit demand and heterogeneous match values represented by a vector of willingness-to-pay for quality θ=(θ1,…,θJ). Utility from seller j’s product is θj qj. A fraction λ of consumers uses the platform; the platform observes each consumer’s full value profile θ, while sellers know only a prior distribution F of values. Off-platform consumers have only expected values m distributed according to G, with F a mean-preserving spread of G (platform’s information strictly finer). Off-platform, consumers face Diamond-type search: zero cost for the first seller, positive for additional ones, so in equilibrium they visit only the seller with the highest expected value. Sellers post off-platform nonlinear pricing menus (quality-price schedules) to screen consumers with private m, à la Mussa–Rosen/Maskin–Riley. On-platform, the platform runs a managed advertising campaign: it requests a fixed advertising budget from each seller, collects sellers’ pricing/quality schedules conditioned on θ, and selects a single sponsored product per consumer according to a selection rule. The main selection rule studied is efficient steering: the platform sponsors the product that maximizes social surplus θj qj − c(qj). On-platform, the platform reveals to the consumer their value for the sponsored seller’s product (baseline). Consumers may showroom: after seeing the ad and learning θj, they can buy either the sponsored offer or the same seller’s off-platform menu. This induces a showrooming constraint equating on-platform rent to off-platform information rent for that θ. Equilibrium concept is symmetric Perfect Bayesian equilibrium with symmetric off-path beliefs. The analysis characterizes: (i) consumers’ consideration sets and search induced by platform information; (ii) sellers’ optimal off-platform menus given showrooming (solved using control/Hamiltonian methods and highest-order statistics F^{J−1}, G^{J−1}); (iii) on-platform qualities/prices (efficient qualities, rents set by binding showrooming); (iv) platform’s optimal advertising budgets extracting sellers’ surplus above their outside options; (v) comparative statics in λ and J. The framework is extended to: symmetric information for on-platform consumers; organic links (free disclosure of all off-platform menus); cohort-based privacy (platform reveals only preference rankings to sellers; likelihood-ratio order F ≻_lr G ensures tractability); and platform information design (optimal extent of revelation, including pooling regions) with special cases (λ=1; uninformed off-platform consumers).

• Managed campaigns and consideration sets: With even a small informational advantage and off-platform search frictions, the platform can fully steer on-platform consumers’ consideration sets (Proposition 2). Consumers compare only the sponsored seller’s on- vs off-platform offers; nonparticipating sellers lose access to on-platform consumers.
• On-platform offers and showrooming: On-platform, sellers offer efficient qualities (q=θ) and set discounts so the showrooming constraint binds; on-platform rents equal off-platform rents for the same θ. Product steering (not price personalization) extracts surplus under symmetric information for the sponsored seller.
• Off-platform menus distorted: Off-platform quality schedules are further downward distorted relative to standard Mussa–Rosen due to the shadow cost of on-platform rents (Proposition 3). Result: higher qualities at lower unit prices on-platform than off-platform; each variety q is sold at a lower price on-platform (“on-platform only” discounts), even though values receive higher-quality matches on-platform.
• Platform revenue optimality: Efficient steering maximizes joint platform–seller profits; the platform extracts sellers’ surplus above their outside options by charging fixed advertising budgets (Proposition 4). Sellers’ outside option is their off-platform profit with no access to on-platform consumers.
• Symmetric information for consumers: If on-platform consumers know their full θ, platform steering power falls; equilibrium qualities/menus remain as in baseline, but sellers’ outside options increase and equilibrium advertising budgets decrease (Proposition 5; Corollary: any information advantage strictly increases platform revenue).
• Organic links: Displaying all off-platform prices/menus intensifies menu competition, raises equilibrium qualities and consumer rents, lowers sellers’ on-path profits, raises sellers’ outside options, and reduces platform fees (Proposition 6).
• Privacy/cohort targeting: With cohort-based ads (platform reveals rankings to sellers, precise value to consumers), and if F likelihood-ratio dominates G over relevant range, sellers offer the same menu on- and off-platform, equal to the Mussa–Rosen solution for a mixture (λF^{J−1} + (1−λ)G^{J−1}); consumer rents are higher and platform fees lower relative to full disclosure (Proposition 7).
• Information design: For λ=1, efficient matching with full revelation to both sides is revenue-optimal (Proposition 8). With uninformed off-platform consumers, the optimal policy features efficient matching but partial disclosure with pooling around the mean value; thresholds solve moment and slope conditions balancing on-platform surplus and off-platform rents (Proposition 9). As λ→1, pooling shrinks and full revelation emerges.
• Platform size and competition: As platform share λ increases, off-platform qualities and information rents decrease; beyond some λ, low and mid values receive zero off-platform quality (Proposition 10). Increasing the number of sellers J eventually reduces off-platform qualities and rents for all values; for large J, qualities for many types go to zero (Proposition 11). Consumer surplus decreases with λ (and eventually with J), while platform revenue increases with both; as J→∞, the platform captures nearly all created surplus.

The findings show that data-enabled managed campaigns allow platforms to reconfigure competition by steering consumers to a single sponsored seller and compressing consideration sets. The platform’s information advantage and minimal search frictions suffice to grant it bargaining power to extract most of the surplus it creates, constrained by sellers’ outside options in their direct channels. The showrooming constraint links on- and off-platform markets: efforts to reduce off-platform rents permit greater on-platform surplus extraction, generating downward distortions of off-platform quality. These mechanisms address the research questions by tracing how data precision (F vs G), competition intensity (J), and platform size (λ) shape surplus creation and distribution. Policy-relevant insights include that privacy-respecting data governance—such as organic links or cohort-based ads/federated learning—can raise consumer surplus by weakening the platform’s ability to restrain competition and by increasing off-platform quality and rents. Conversely, expanding platform scale increases off-platform distortions and consumer prices, potentially harming consumers despite better matching. Information design matters: full revelation is optimal for large platforms, but with significant off-platform demand and uninformed consumers, partial disclosure that pools mid-values maximizes platform revenue. Overall, managed campaigns differ sharply from transaction-fee retail platforms; fixed budgets substitute for MFNs, and the platform can extract revenue without per-transaction fees, changing incentives for showrooming and competition.

The paper introduces a tractable model of digital intermediation with managed advertising campaigns, product steering, and parallel sales channels. It characterizes equilibrium consideration sets, on- and off-platform menus, and platform revenue extraction, showing that platforms can steer consumers, induce efficient on-platform matching, and appropriate most created surplus, while inducing additional downward distortions off-platform. Privacy-respecting mechanisms (organic links, cohort-based targeting) increase consumer surplus and reduce platform rents. Optimal information design ranges from full revelation for large platforms to partial pooling when off-platform markets are significant. As platform penetration and seller competition increase, consumer surplus declines while platform revenue rises, with off-platform quality and rents falling. Future research avenues include empirically quantifying these effects, allowing heterogeneous seller presence across channels, modeling multi-slot or multi-ad exposures and imperfect steering, endogenizing platform adoption λ via privacy preferences and benefits, and incorporating hybrid platforms’ self-preferencing and variable transaction fees.

• The analysis is purely theoretical with stylized functional forms (quadratic costs, linear utility) and assumes i.i.d. values and expectations with specific order-statistics structures.
• Equilibrium refinement imposes symmetric beliefs off path; outcomes may differ under alternative refinements.
• Single sponsored link and perfect revelation to consumers in the baseline may overstate steering power; multiple ads or noisy revelation could attenuate effects.
• Search frictions off-platform follow the Diamond framework; real-world multi-stage search and heterogeneous costs may produce different consideration dynamics.
• Exogenous platform share λ and fixed competitive structure J; network effects and endogenous adoption are only discussed, not modeled.
• Cohort-based results rely on likelihood-ratio dominance conditions (F ≻_lr G) for tractability.
• No empirical calibration or validation; comparative statics are qualitative.
• Focus on fixed advertising budgets; alternative pricing (auctions with rich signals) is only discussed comparatively.