- Period
- 2026-07-25 — 2026-09-23
- Method
- Every request carrying an AI crawler user agent is checked against the provider's published IP ranges or by forward-confirmed reverse DNS at the moment it arrives, and the result is stored with the request.
- Method version
- 1.0
- Published
- 2026-09-23
What was measured
16 959 crawler requests to ProductKnowledgeGraph between 2026-07-25 and 2026-09-23. Each one carried a user agent naming an AI provider, and each one was checked when it arrived — against the provider's published IP ranges where they exist, by forward-confirmed reverse DNS where that is what the provider documents, and recorded as an unverified claim where our check had nothing published by the provider to compare with.
Results
| Provider | Requests | Passed the check | Failed it | Only claimed | Share verified | | --- | --- | --- | --- | --- | --- | | OpenAI | 3724 | 1406 | 2315 | 3 | 37% | | Amazon | 3052 | 1760 | 1292 | — | 57% | | Google | 2742 | 2665 | 77 | — | 97% | | Anthropic | 2522 | — | — | 2522 | — | | Meta | 1644 | — | — | 1644 | — | | Perplexity | 1345 | 611 | 728 | 6 | 45% | | Apple | 1081 | 911 | 164 | 6 | 84% | | xAI | 270 | — | — | 270 | — | | Yandex | 181 | 148 | 33 | — | 81% | | Microsoft | 166 | 108 | 58 | — | 65% | | ByteDance | 150 | — | — | 150 | — | | Common Crawl | 53 | — | — | 53 | — | | Cohere | 14 | — | — | 14 | — | | DuckDuckGo | 12 | — | 12 | — | 0% |
Across all providers that publish something to check against: 12 288 requests, 61% verified. Requests from providers that publish nothing: 4 668, or 27% of everything recorded.
What this does not say
A failed check is not proof of impersonation. Our check can be wrong: a provider may use an address before publishing it, and a reverse lookup can fail for reasons unrelated to the requester. "Only claimed" is not an accusation either — it is the accurate record for a provider that gives no way to confirm.
What the numbers do support is narrower. If you count AI crawler traffic by user agent alone — which is what analytics dashboards do — a large share of what you are counting has never been checked by anyone. The figure to carry away is not a verdict about any provider; it is that the question "was this really that crawler" has an answer, and it is worth recording it at the moment of the request rather than inferring it later.
Correction, 23.09.2026
The figures on this page were restated after every recorded visit was checked again against what the providers publish. The edition of 21.09.2026 understated verified visits; no figure moved the other way.
How to repeat this
The check runs on our own server at request time; anyone with server logs can do the same. Take the user agent, resolve the requesting address, confirm the hostname resolves back to that address, and compare it with what the provider publishes. Store the result next to the request. The series only becomes useful once it is long enough to show change, which is why this page is dated and its previous editions are kept rather than overwritten.