How AI Shopping Mediation Changes Retail Assortment and Revenue

Summary

This paper studies retail assortment planning when an AI shopping assistant determines the shortlist that customers see before making a purchase. The retailer first chooses a catalog, an AI intermediary ranks products for customer segments, and customers select from the resulting shortlist under a multinomial logit model. The authors examine both deterministic mediation, where the AI uses a fixed segment-specific ranking, and stochastic mediation, modeled with the Plackett-Luce framework. They build product-ranking data from leading AI models, fit Plackett-Luce ranking models, and test their predictions on held-out assortments. Combining these estimates with synthetic customer preferences, they find that the revenue-maximizing assortment can differ materially when AI mediation is included. Across the tested catalog-capacity settings, optimizing as if no mediation existed produces mean revenue losses of 2.52% to 11.37% compared with the Plackett-Luce optimum. A simpler deterministic plug-in that uses only the ordering of fitted AI utilities performs much better, with mean losses no greater than 0.53%. The paper then analyzes deterministic mediation and shows that, for any fixed number of customer segments, its assortment-optimization problem has a fully polynomial-time approximation scheme.