- Period
- 2026-07-26 — 2026-09-16
- Method
- Every URL a model returned as a citation is classified by host into one of eight classes; the host is read from the URL as the model gave it, without following redirects. Classes and their rules are printed on the page.
- Method version
- 1.0
- Published
- 2026-09-20
What was measured
22 789 citations extracted from 2 580 model answers to product questions between 2026-07-26 and 2026-09-16. Every answer was produced by asking a real buyer question about a real product in our catalogue; every URL the model returned as a source was recorded with it.
Whose page the citation points at
| Class | Citations | Share | | --- | --- | --- | | the source's own site | 6 816 | 29% | | the source's own site, another domain of the same brand | 1 631 | 7% | | another site in our catalogue | 394 | 1% | | marketplace | 1 319 | 5% | | community and video | 1 506 | 6% | | elsewhere | 10 071 | 44% | | ProductKnowledgeGraph | 250 | 1% | | a search redirect, not a page | 802 | 3% |
The classes are decided in this order: our own domain; the domain of the organization whose product the question was about; the same brand on another domain (decathlon.co.uk and help.nomadgoods.com next to decathlon.com and nomadgoods.com); any other domain in our catalogue; then three explicit lists — marketplaces (amazon, ebay, etsy, walmart, target, aliexpress, temu.), community and video (reddit, youtube, quora, stackexchange, stackoverflow.), and search redirects; everything left over is "elsewhere".
Most cited hosts in "elsewhere": bestbuy.com (152), rei.com (146), eagleseyefinds.com (124), optechusa.com (122), idealo.de (99), ifixit.com (94), alexkwa.com (88), facebook.com (86).
The same question per model
| Provider | Model | Citations | Own site | Marketplace | Elsewhere | Redirect | | --- | --- | --- | --- | --- | --- | --- | | perplexity | sonar | 15 702 | 36% | 6% | 45% | 0% | | openai | gpt-4o-mini | 4 095 | 44% | 1% | 45% | 0% | | anthropic | claude-haiku-4-5-20251001 | 1 157 | 52% | 7% | 37% | 0% | | gemini | gemini-2.5-flash | 1 077 | 3% | 1% | 17% | 74% | | xai | grok-4-latest | 476 | 54% | 5% | 36% | 0% | | mistral | mistral-small-latest | 144 | 0% | 9% | 90% | 0% | | openai | gpt-4o-mini-2024-07-18 | 88 | 3% | 5% | 89% | 0% | | mistral | mistral-medium-latest | 50 | 0% | 0% | 100% | 0% |
The redirect column is why one model looks unlike the rest: gemini-2.5-flash returns its own search-redirect URLs instead of the page, so what it points at cannot be classified without following them — and following them today would date the reading to today rather than to the day of the answer.
Models are named in full, and the differences between them are large enough that a single-model measurement would have been misleading either way.
What this does not say
A citation is not use of a fact. The model linked the page; whether it read the number it then stated is a separate question, measured separately.
The host is read from the URL the model returned, without following redirects. That is deliberate: resolving them today would date the reading to today, not to the day of the answer. Where a model returns its own search-redirect URL, that is counted as a redirect rather than guessed at.
Our own domain appears 250 times. We publish these records, so that number is a measurement of our reach, not a recommendation: it is here because leaving it out of a table about who gets cited would be the one omission a reader could not check.
Why a brand should care about the shape rather than the exact figures
The figures move with the catalogue and the questions. The shape has been stable for the whole period: when a model answers about a product, the page it points at is usually not the maker's. Whatever a buyer reads at that link — last season's specification, a wrong variant, a third-party summary — is what the answer rests on.