What this check tests
This is one of three scored reach checks that share the same pass rule. The other two are Perplexity-User and Google-Agent. What differs is the string sent, the vendor behind it and the place in the queue.
While gathering site data, the audit requests the page you submitted (normally your homepage) once, with this header:
User-Agent: Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; ChatGPT-User/1.0; +https://openai.com/bot
That is the string OpenAI publishes for ChatGPT-User. The response decides the result:
| Outcome | Result |
|---|---|
| 2xx status and a body with at least one byte | Pass, 100 |
| Any other status, such as 403, 429 or 5xx | Fail, with the status |
| 2xx but an empty body | Fail |
| Timeout or network error | Fail, naming the outcome |
| No fetch recorded, or the time budget ran out first | Unmeasured, shown as an error and not scored |
Redirects are followed, up to a limit, so the status is from the final response. The request waits up to the audit's timeout and reads at most the first 64 KB. ChatGPT-User is fetched first, so it is the last of the three to be skipped when a slow site uses up the shared budget (the smaller of three timeouts and 15 seconds).
How this differs from the other agent checks: AI agent access reads robots.txt and never makes this request. Edge treatment of AI agent user agents compares each agent against the audit's own crawler, recognises challenge pages, and is tracked, not scored. This check is scored, has no control request, and does not inspect the body, so a 200 challenge page or a redirect to a login page passes.
Why it matters to agents
OpenAI documents ChatGPT-User as used for certain user actions in ChatGPT and Custom GPTs: when a person asks a question, ChatGPT may visit a page with this agent. OpenAI also says it is not used to crawl the web automatically, that robots.txt rules may not apply because a user initiated the action, and that it is not used to decide whether content appears in Search, which is the job of OAI-SearchBot.
So robots.txt is the wrong tool for this fetch, and a network-layer rule is the likely cause of a fail. The audit sends the string from its own address, not from OpenAI's. OpenAI publishes the address list for ChatGPT-User, so a rule that checks both can treat the audit differently from the real agent. A pass shows that your server does not reject the string; it does not prove ChatGPT reaches you.
How to fix it
- Read the status in the message. A 403 or 429 for this string usually comes from a bot-management or rate-limit rule, not from your application.
- In your CDN or WAF, find rules for AI crawlers or "AI bots" and decide how to treat ChatGPT-User. If you allow it, allow by user-agent plus the address list OpenAI publishes for it, so the rule cannot be copied by anyone who sets the header.
- Keep training crawlers separate. GPTBot is a different token, and blocking it does not affect this check.
- Re-run the audit, or request your homepage with that header from another network and read the status.
- If the result is unmeasured, the audit recorded no fetch for this agent. ChatGPT-User is first in the queue, so the time budget cannot be the cause; re-run the audit. A slow page that fails by timeout is a response time problem.
Background: AI crawler blocking mistakes and robots.txt for AI bots.