The Value of Personalized Recommendations (arxiv.org)

Researchers estimate the causal value of personalized recommendations using detailed Netflix viewership data by building a structural discrete-choice model that explicitly includes recommendation-induced utility, low-rank user/item heterogeneity, and flexible state dependence. They exploit idiosyncratic variation introduced by Netflix’s recommendation algorithm to separately identify the value of recommendations versus the intrinsic appeal of titles, and they use model-free diversion ratios to validate the structural fit. The approach combines econometric identification with machine-learning style low-rank representations to recover who benefits from recommendations and why. Counterfactual simulations show that swapping Netflix’s system for a simple matrix-factorization recommender would cut engagement by about 4%, while a popularity-based recommender would cut it by ~12% and reduce consumption diversity. Crucially, the bulk of the engagement uplift comes from effective targeting (matching users to the right titles) rather than mere mechanical exposure; the largest gains accrue to mid-popularity titles rather than broadly popular hits or very niche content. The paper quantifies how personalization reshapes platform consumption and provides a blueprint for measuring recommendation value—insights important for recommender design, platform economics, and policies aimed at discovery and diversity.