- PayPal's Nixon Dinh said merchant catalogs syndicated to ad platforms were built for human search, so agents may struggle to find and read products.
- In PayPal's tests, enrichment helped thin catalogs most, but in some cases too much made the agent more prone to hallucinate.
PayPal says product catalogs built for ads and human search are not ready for AI shopping agents. Nixon Dinh, a PayPal director who runs product for its agentic commerce work, made the case on the AI Engineer podcast. A merchant that looks fine to people may be hard for an agent to read.
What was said
Dinh described a move from the search era, where shoppers type keywords, to an intent era, where they describe a need in their own words. Many merchants syndicate one catalog to Google, Meta or other ad platforms. Dinh said that format was built for human search, and "the spec wasn't built for agents."
The gap comes from how the searches work. Dinh said "keyword search is quite precise. It's very literal. Semantic search is just smart, but it's very fuzzy." A keyword search for blue running shoes returns only listings with those words. A search for footwear for jogging may miss the same shoes. Semantic search matches on meaning instead, and he said it needs very different data structures.
PayPal tested three kinds of enrichment on merchant data: attribute filling, buyer context and review trust signals. The excerpts do not define them. Enriched data beat an unenriched baseline, and the thinnest catalogs gained most. But Dinh said "more text isn't better. Structured quality content is the one thing that wins out." He called filler like a family-owned greeting irrelevant to an agent at times. In some cases, over-enriching made the agent more prone to hallucinate, meaning state untrue things, and it did worse.
Why it matters
Our reading: the catalog feed is now an interface for machines, yet many merchants run it as an ad-platform chore. Budget signers could ask three questions. Who owns the feed, and do they know agents read it? Which products have the thinnest data, since thin catalogs gained most in PayPal's tests? Does each product carry clear facts, such as attributes and who it suits, rather than slogans? Buyers of commerce software can ask vendors how their feeds serve agents.
The other side
Dinh said this was a single set of experiments and that PayPal has no concrete recommendation for every business. He advised assessing your own catalog. The excerpts give no figures for the gains, and they do not say where helpful enrichment ends and excess begins. Dinh also speaks for a company that sells agentic commerce services. He said agent models are largely a black box and rank differently, so the right target is unclear.
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)
