AI Shopping Agents Show Alarming Concentration Bias, Funnel Demand to 2-3 Products While Ignoring Rest

Summary

A groundbreaking research paper from Columbia Business School researchers reveals troubling patterns in how AI shopping agents make purchasing decisions in e-commerce environments. Using ACES, a provider-agnostic auditing framework, the study found that autonomous AI agents exhibit severe "choice homogeneity," consistently concentrating consumer demand on just a few "modal" products while completely ignoring viable alternatives. This behavior persists across multiple AI providers and represents a fundamental departure from human shopping patterns.

The research, led by Amine Allouah, Omar Besbes, Josué D Figueroa, Yash Kanoria, and Akshit Kumar, uncovered that these AI preferences are highly unstable—model updates can dramatically reshuffle market shares overnight. The agents demonstrate strong position biases that vary by provider and model version, even in text-only "headless" interfaces where visual ranking shouldn't matter. Agents consistently penalize products marked as "sponsored" while rewarding platform endorsements, and their sensitivity to price, ratings, and reviews varies sharply across different AI models.

Perhaps most concerning, the researchers demonstrated that seller-side AI agents can exploit these biases through simple, query-conditional product description tweaks to capture significant market share gains. The findings suggest that agentic e-commerce markets will be fundamentally volatile and manipulable, raising urgent questions about platform design, competitive fairness, and the need for regulatory frameworks. The research highlights that as AI agents increasingly mediate online commerce, continuous auditing will be essential to prevent market concentration and ensure fair competition.

• Seller-side AI agents can exploit these biases through simple description modifications to capture significant market share
• Findings raise urgent questions about platform design, competitive fairness, and the need for continuous auditing and regulatory oversight

Editorial Opinion

This research should be a wake-up call for both e-commerce platforms and regulators. If AI agents funnel the majority of demand to just 2-3 products based on opaque, unstable preferences rather than genuine product superiority, we're looking at a future where algorithmic bias could determine winners and losers in online retail. The fact that these preferences can be gamed by seller-side AI agents is particularly troubling—it suggests an arms race where success depends less on product quality and more on optimizing for AI decision-making quirks. The volatility introduced by model updates could create chaos for small businesses unable to continuously adapt to shifting AI preferences.