The measured split
Every URL that models returned as a source in answers to product questions, 26.07.2026 to 16.09.2026, classified by whose site it points at:
| Where the citation points | Share | | --- | --- | | the maker's or seller's own site (any of its domains) | 37% | | press, reviews, reference and other sites | 44% | | community and video | 6% | | marketplaces | 5% | | another site in our catalogue | 1% | | search redirects that cannot be classified | 3% |
Classes, rules and per-model figures are in the study.
It depends heavily on the model
The share pointing at the seller's own site was 52% for claude-haiku-4-5-20251001 (Anthropic), 44% for gpt-4o-mini (OpenAI) and 36% for sonar (Perplexity) over the same period. gemini-2.5-flash (Google) returned its own search-redirect URLs in 74% of cases, so what it points at cannot be classified without following them. A measurement taken with one model describes that model, not "AI".
Why third-party pages win
A review page usually does three things a product page often does not: it states the facts a buyer asks about in plain sentences, it compares, and it is reachable without running scripts. Models quote what they can quote.
What a brand can influence, and what it cannot
You cannot make a model ignore a review. You can make your own page a better source than the review for the facts that are yours: identifiers, dimensions, what is in the box, compatibility, what the product does not do. When those facts are stated on your page, dated, and reachable without JavaScript, the model has a primary source to cite instead of a secondary one.
What this does not promise is that it will. The rules we use to separate a citation from the use of a fact, and both from influence, are at /evidence.