ChatGPT-User

It is the visit behind a live answer. OpenAI says that when users ask ChatGPT or a custom GPT a question, it may visit a web page with the ChatGPT-User agent, that users may also reach outside applications through GPT Actions, and that ChatGPT-User is not used to crawl the web automatically^[1]. A request under this name usually means that someone asked about the page.

It sits outside the search opt-out. OpenAI states that ChatGPT-User is not used to decide whether content may appear in search, and points site owners to OAI-SearchBot in robots.txt for search opt-outs and automatic crawling^[1]. How OpenAI’s three agents divide the work, training, search and a user’s request, is set out under AI crawler.

Robots.txt is a request it may not follow. Because these actions are started by a user, OpenAI says robots.txt rules may not apply^[1], and the protocol itself says its rules are not a form of access authorization^[2]. A site that must refuse these fetches does it at the server, and gives up the answer the reader was waiting for.

The name is published, and so are the addresses. OpenAI gives the full user agent string, with the token ChatGPT-User/1.0, and a JSON file of the IP addresses it uses^[1]. HTTP treats the User-Agent field as the client’s own description of itself and notes that one user agent can masquerade as another^[3], so the address is the check.

Fewer sites write a rule for it than for OpenAI’s training crawler. In QueryRev’s study, of the 1,339 sites whose robots.txt names at least one AI crawler, 43.3% close the whole site to GPTBot by name, while ChatGPT-User is blocked by name on 262 of all 6,294 sites read, 4.2%^[4].

In the server logs only. ChatGPT-User runs no tracking script, so analytics never sees it. Count its requests by URL, keep those from addresses in OpenAI’s published list, and read the status each one got.

Read the pattern as demand. The pages ChatGPT-User fetches most are pages people ask ChatGPT about; set them beside the prompts you track and the AI referral traffic those pages receive, because a fetch with no visit after it is a reader who got the answer without the click.

Check what it was served. A 403 from a firewall, a cookie wall or a page that builds its text with scripts after load can leave the answer with less to work from than the page your readers see.

Sample data: an invented site, to show the method.

An online paint shop reads a month of logs for ChatGPT-User, keeping only requests from OpenAI’s published addresses. Its colour-matching guide was fetched 1,180 times and its returns policy 410 times, all served normally. Its product pages were fetched 95 times, and 61 of those got a 403 from a bot rule on /shop/.

The returns policy is a page people ask about before they buy, so it is checked for accuracy first. The rule on /shop/ is the second fix: a reader asking about a product gets an answer drawn from other sites when the shop’s own page refuses the fetch.