A new study, entitled “What Is Your AI Agent Buying?” conducted by researchers from Yale University, Columbia University, and the University of Chicago is one of the most important examinations yet of how AI agents evaluate, rank, and recommend products. Published this month, the paper combines academic rigor with real-world relevance, making it essential reading for anyone working in commerce, retail media, or digital transformation.
What makes the methodology so compelling is its precision. The research team used carefully controlled experiments, varying product attributes, similarity, and social influence signals from user-generated content (UGC) to isolate the exact factors influencing AI agent decisions. They tested multiple large language models, ran paired choice and ranking experiments, and examined behavior across both specific product tasks and open-ended recommendation tasks. This level of control gives us one of the clearest, most evidence-based views into how AI agents “shop” and the factors that meaningfully shift their choices.
Here are the main themes and most important lessons.
AI agents depend on structured product information:
- Clean metadata beats creative marketing. Products with complete, well-structured, machine-readable attributes were consistently preferred. AI favors clarity over branding, precision over persuasion. For FMCG teams, this flips the traditional hierarchy as data quality now shapes discoverability.
- AI is extremely sensitive to missing information. The study shows a 20%-40% reduction in selection probability when a key attribute is removed. Agents cannot “assume” or “infer” missing data the way shoppers might. Metadata gaps are now visibility gaps.
- Agents compare products contextually, not in isolation. Small tweaks in product similarity reshaped preference rankings. AI does not evaluate absolute quality; it evaluates relative advantage. This mirrors human cognition but at machine-scale speed, making competitive context even more critical.
Agents follow the crowd:
- UGC rewires how AI makes choices. Ratings, sentiment, and review volume altered model outputs in every test. Social proof is one of the strongest behavioral drivers for AI, just as it is for humans.
- Reviews reshape the recommendations AI delivers. A seemingly small difference in rating (e.g., 4.4 versus 4.1) frequently changed the top-ranked product. Agents interpret rating spreads with more weight than many categories currently appreciate.
- Social proof drives AI picks. More reviews means more trust. High-volume, high-quality UGC signals product safety and reliability to agents.
- User reviews steer AI decisions behind the scenes. When UGC inputs changed, the model’s ranked outputs changed instantly. This is algorithmic gravity as reviews bend the decision space.
Stable, optimizable agent behaviors are emerging:
- Cross-model consistency signals emerging norms. Different models (GPT-based, Claude-type, etc.) showed similar preference structures. This convergence creates predictable rules of engagement for brands.
- Agents assign “trust bonuses” to well-reviewed products. Well-reviewed products were often selected even if they were slightly more expensive. This mirrors human “risk-minimizing” logic and suggests premiumization opportunities in certain categories.
- Agents exhibit stable preference biases that brands can optimize for. Because AI decisions follow consistent patterns, brands can now design product data the same way they design packaging, optimized for how the “buyer” processes information.
The implications are direct and urgent. AI agents are rapidly becoming the first line of product discovery, filtering what shoppers see. Brands and retailers should immediately:
- Elevate metadata to a strategic capability. Fix completeness, structure, and clarity.
- Accelerate UGC and management since reviews significantly shift AI recommendations.
- Build AI-ready product pages, feeds, and attributes. Agents reward precision.
- Share this study internally. Every ecommerce, brand, and retail team needs these insights.
Ultimately, this study shows that AI is becoming not just a shopping influencer but the actual shopper. The brands that master agentic decision logic today will shape the digital shelf tomorrow.
We'll have more on AI in retail at our Commerce Insights Conference coming up on Dec. 10-11. See the agenda and register now.
