AI Shopping Agents Show 90-99% Bias in Picks

Ask an AI shopping assistant to pick you a fitness watch, and the answer may hinge on nothing more than which article happened to load onto the screen first. That's the finding from a new study out of the Wharton School at the University of Pennsylvania, which put six current AI models through a test designed to mimic how agentic shopping tools actually work: reading a product page, pulling in outside sources, then making a call. The models included both mini and frontier-level versions, and the setup used something called the ACES simulator, short for Agentic e-Commerce Simulator, which feeds the AI a screenshot of a product grid and lets it reason its way to a pick.

The task itself was simple. Choose one fitness watch from a fixed set of options, acting as a personal shopping assistant would for a real customer. Researchers first checked what each model preferred with no outside input at all, then started introducing single sources, like a Reddit thread or a review site, to see how much that shifted the outcome. What they found suggests these agents are far more persuadable than a shopper might expect, and not evenly so across sources or models.