AI shopping agents aren't ready to buy on your behalf, study finds
decoder ·
Researchers at the University of Pennsylvania’s Wharton School tested six AI models as shopping assistants choosing fitness watches from a fixed product grid using the ACES simulator. The models’ recommendations changed substantially when they were shown a single external source, with a Wirecutter review causing especially large increases in selections of the Fitbit Inspire 3. Adding multiple sources did not make decisions more balanced, and Wirecutter generally dominated when included. The order in which identical sources were presented also affected results, while delivering sources as a bundle rather than sequentially changed GPT-5.5’s choices. In another experiment, brief user-memory statements such as a preference for hiking pushed several models toward more expensive products even when a cheaper watch was objectively superior across measurable criteria. Gemini 3.5 Flash was comparatively consistent in selecting the objectively best product, while other models showed much larger shifts. The researchers say these inconsistencies make AI shopping agents unreliable for buyers and make it difficult for sellers to optimize products for AI-driven purchasing.
decoder