Rewriting one product title moved an AI agent's choice 80 points

For years we have optimized stores to persuade humans. Now a second kind of buyer is arriving: the AI agent, tasked not just to search but to pick and purchase. A new branch of research is forming around what these agents actually choose, and the early findings are usable today.

In a sandbox study running 1,000 experiments across 8 product categories, researchers isolated the signals that sway agents. When they rewrote an office lamp's title from "SUNMORY Floor Lamps for Living Room" to "SUNMORY Office Floor Lamp," matching how the request was phrased, the product's selection rate jumped 80 percentage points with GPT-5.1, 52 points with Gemini 2.5 Flash, and 41 points with Claude Opus 4.5.

What biases an AI agent

Agents are trained on human decision-making, so they lean on familiar shortcuts:

• Title keywords and their order. A title that mirrors the query ("office floor lamp" for someone asking for an office lamp) wins over a generic one.
• Ratings and review counts. A small rating increase measurably raised the odds of selection in the tests.
• Badges. Positive labels like "Bestseller" help. A "Sponsored" label hurts, agents discount it much as people do.
• Competitive pricing. All else equal, newer models increasingly pick the better-value option.

What to do on your Shopify store

• Write titles that match real intent. Lead with the words a shopper would type or say for that exact product, not internal or collection-style naming. "SUNMORY Office Floor Lamp" over "SUNMORY Floor Lamps for Living Room."
• Protect your ratings and review volume. They are now a ranking signal for agents as well as a trust signal for humans.
• Use honest, positive badges where they are true, and understand that "Sponsored" placement can work against you with agents.
• Keep structured, comparison-friendly detail in your descriptions, so an agent can match your product to specific requirements.
• Re-test after model updates. Different agents rank differently, and the same model can flip behavior between versions. Identify which agents your customers use, then check your key products periodically.

A caveat worth keeping

This is an early-stage working paper, not yet peer-reviewed, and it was run in a controlled sandbox where each factor was isolated. Real shopping journeys mix these signals together, and model behavior is changing fast. Treat this as a reason to get your fundamentals right, clear titles, strong ratings, honest badges, rather than to chase any single model's quirks.

None of this replaces optimizing for humans, who still make the overwhelming majority of purchases. It sits alongside it. If AI referrals are a real channel for you, read our companion insight on how AI-driven traffic actually converts before you invest heavily.

---

Research: Alloua, A., Besbes, O., Figueroa, J.D., Kanoria, Y. & Kumar, A. (2025). Working paper, Columbia University and Yale University. https://dx.doi.org/10.2139/ssrn.5381574