Assessing the Causal Impact of Recommendation Algorithms on Product Discovery, Demand Allocation, and Concentration in Platform-Mediated Retail

Abstract

Platform-mediated retail increasingly relies on recommendation algorithms to organize vast catalogs, reduce search frictions, and personalize product discovery. At the same time, these systems can reshape how demand is allocated across sellers and products, potentially altering market concentration and the distribution of economic surplus. This paper develops a causal framework for assessing the impact of recommendation algorithms on product discovery, demand allocation, and concentration outcomes within a large online retail platform. The core challenge is that recommendations are endogenous to user behavior, product performance, and platform objectives, so naive comparisons conflate algorithmic effects with demand shocks and quality differences. We formalize exposure as a treatment delivered through ranked recommendation surfaces and define discovery as the transition from unobserved to considered and then to purchased states under limited attention. The empirical design leverages plausibly exogenous variation from algorithmic experiments, policy-driven parameter shifts, and discontinuities created by rank-threshold rules, combined with panel data on impressions, clicks, add-to-carts, purchases, prices, fulfillment, and inventory. We estimate both reduced-form causal effects and a structural mapping from exposure to consideration sets and choice probabilities. We then connect micro-level effects to macro outcomes by decomposing changes in concentration indices into components attributable to exposure reallocation versus preference shifts. Results are interpreted through mechanisms including attention allocation, substitution patterns, and feedback loops in learning-to-rank systems. The analysis provides a unified approach to quantifying when recommendations broaden discovery versus concentrate demand, and how these effects vary across product maturity, category complexity, and consumer heterogeneity.