Across the stores running AI Visibility, the app has now generated descriptions for roughly two million products. That is a large enough sample to say something about which catalogs assistants can actually use, and the answer surprised us in its plainness.
Structure beats volume
The stores that get cited in AI answers are not the ones with the most products. They are the ones whose products are described in a way a model can resolve into a recommendation: a clear product type, a price, a material or specification, and a policy that says whether it can be returned.
A catalog of 80 well-described products outperforms a catalog of 4,000 whose titles carry SKU fragments and whose descriptions were written for a search engine rather than a reader.
A model does not reward a long catalog. It rewards a catalog it can answer a question from.
Three patterns that showed up repeatedly
Product type is doing most of the work. Stores that set a real product type — "wool overcoat" rather than "apparel" — appear in far more specific queries. It is one field, it is usually already populated with something vague, and improving it costs nothing.
Variants are where catalogs go wrong. A product with eleven near-identical variants tends to produce eleven near-identical entries, which crowds out everything else in the file. Consolidating variant descriptions was the single largest improvement we made to the generator this year.
Policies get quoted more than expected. Shipping thresholds and return windows appear in AI answers about as often as prices do. Shoppers ask whether something can be returned, and a store whose policy is machine-readable gets named while one whose policy lives in a page of prose does not.
What this means for the app
- Variant consolidation is now on by default for every store
- Product type is weighted more heavily than collection membership
- Store policies are described explicitly rather than linked
- Thin descriptions are flagged in the dashboard rather than silently indexed
What we still do not know
We can see what we generate and we can see what changes when we generate it differently. We cannot see inside any model, and we are wary of the confident claims being made in this space by people in the same position. Treat anyone offering you a guaranteed ranking in an AI answer with the same suspicion you would apply to the same promise about search.
The honest version is narrower: a catalog a model can read has a chance of being recommended, and one it cannot read does not. That is the part we can act on.