- PayPal's tests found that enriching thin product data helped shopping agents, but over-enriching made agents hallucinate more in some cases.
- Dinh's advice was to add structured, relevant content and skip filler, and to assess each catalog first, because no single recipe fits every business.
Adding more text to product listings is not a safe way to help AI shopping agents. Nixon Dinh, a PayPal director of product working on agentic commerce, made this point on the AI Engineer show. PayPal tested ways to enrich merchant product data. Enrichment helped overall, but in some cases too much of it made agents hallucinate, meaning they produced answers the data did not support.
What was said
Dinh explained that old keyword search matches literal words, so a search for jogging footwear can miss running shoes. Semantic search matches meaning instead. It turns a query into vectors, numeric patterns that capture intent. Dinh warned that adding too much meaning can blur focus and lead to hallucination.
PayPal's test started from a plain listing: a blue men's running shoe with a short description. The enriched version added a fuller title, a richer description, attributes such as fit and materials, and customer reviews.
Enriched data beat the bare baseline head to head. Merchants with the thinnest catalogs gained the most. But unstructured filler did not help. Dinh gave the example of store boilerplate such as a family-owned welcome message, which he said is often irrelevant to an agent. His summary of the over-enrichment result was blunt.
"when we over-enriched, the agent tended to hallucinate more, and it underperformed." His closing advice was short: "more text isn't better." He added that well-organized, high-quality content is what comes out ahead.
Why it matters
Our reading: this turns a familiar habit into a risk. Teams that wrote more copy to win at keyword SEO may assume more content is always safer for agents. Dinh's account suggests volume can work against you when the reader is a machine.
For anyone who runs or buys a catalog, the practical steps follow from what he said. Check what data you already have. Fill real gaps in attributes first. Cut filler that tells an agent nothing. Test changes against a baseline, as PayPal did, rather than assuming extra text is harmless.
The risk also scales. An agent that misreads a listing does so for every shopper who delegates a search to it.
The other side
Dinh was careful about the limits. He called this one set of experiments, and said the over-enrichment problem showed up in some cases, not all. He gave no concrete recommendation for every business. He also gave no measure of how much enrichment tips into too much.
He conceded that, unlike the old SEO era, strategies for this new one remain largely unclear. The results are PayPal's own, shared by a company that sells agentic commerce services, and the excerpts do not show an independent check. What kind of hallucination appeared, and what it cost merchants or shoppers, was not described.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Keyword Search Is Dying. Is Your Catalog Ready for AI Agents? — PayPal (2026-10-04)
