The User-Agents of AI apps fetching a pasted URL: ChatGPT-User, Claude-User, the single word "Google", Firefox on Linux (1 to 17 October 2026)
The answer first: we pasted pages of our lab site into ChatGPT, Claude, Gemini, DeepSeek and Perplexity and read the User-Agent (what the visitor calls itself) in the server log. There were four patterns. ChatGPT sent ChatGPT-User/1.0, Claude Claude-User/1.0, Gemini the single word "Google", and DeepSeek Firefox/149.0 on Linux. Perplexity sent no request at all. Only two of the four apps can be found by name.
Measured in the request log of 1 to 17 October 2026; the check against Google's published lists was run on 11 November 2026. This is the case where a user pastes a URL and the app fetches it, not a crawl. It is one observation on one site of ours.
How we measured
Each pasted link carried a token unique to the run, so a request in our server log could be matched to the app that opened it. We counted only requests carrying such a token. The log was read on 17 October 2026 at 02:40 UTC, and the experiment record ties each run to its app.
The table: User-Agent and requests (1 to 17 October 2026)
| App | User-Agent (the key part) | Requests | Runs |
|---|---|---|---|
| ChatGPT | contains ChatGPT-User/1.0 | 173 | 143 |
| Claude | contains Claude-User/1.0 | 150 | 122 |
| Gemini | the whole string is Google | 155 | 153 |
| DeepSeek | an ordinary browser string, Firefox/149.0 on Linux | 445 | 149 |
| Perplexity | no request | 0 | 0 |
- Gemini's requests came from six addresses (
192.178.21.x: 192.178.21.1, 192.178.21.5, 192.178.21.6, 192.178.21.7, 192.178.21.9, 192.178.21.10), plus one request from another address (108.177.69.102, on 1 October). - Perplexity: in the 7 runs of 2026-10-16 where our page was pasted, Perplexity said in 6 that it could not fetch the page.
- Requests with the User-Agent "Google" and no run token: 0 since 1 October 2026. The last Gemini request with it was on 2026-10-14.
"Google" is on none of Google's lists (checked 11 November 2026)
On 11 November 2026 we checked Gemini's six addresses against Google's published IP lists. All six sit inside Google's general address range (192.178.0.0/15; that list is dated 10 November 2026). None sits in Google's lists of common crawlers, special crawlers, user-triggered fetchers (the fetchers that run when a person asks), the Google-specific user-triggered fetcher list, or user-triggered agents. The common crawlers were checked in our registry snapshot of 9 November 2026 (Google's own update of 6 November).
Google's page on user-triggered fetchers (updated 19 August 2026, read 17 October 2026) names Google-Read-Aloud, Google-Agent, Google-GeminiNotebook and others; none is a User-Agent made of just "Google". So Gemini's fetch is Google infrastructure without a name Google lists for it. By the official lists we cannot call it a Google bot; we only know the requests carried our Gemini test tokens.
Basis and limits
- One site, one account per app, our lab's machine, our test pages, 1 to 17 October 2026. The User-Agent and the addresses can change at any time; we have not looked at them since 17 October, apart from the address check on 11 November.
- We do not know whether Gemini uses this User-Agent for every user, or only for the way our lab asks. On our site, no request with the User-Agent "Google" came in this period without one of our test tokens.
- We cannot prove the requests are Google's. We know the addresses are in Google's general range and that the requests carried our Gemini tokens.
- Not a count of how often AI visits real sites. Every request here was caused by our own test.
- The zero for Perplexity is what we saw on one host of ours.
- Sources:
knowledge/claims/ai-apps-fetch-pasted-urls-with-distinct-uas.mdandknowledge/claims/perplexity-user-never-reaches-pw.md(level: observed in our lab). The Google lists are the ones Google publishes; we did not make up any address range of our own. - What would show more: the same check on a site with real visitors, and on a later date.