GPTBot

OpenAI’s crawler for gathering web content used to train its models.

17.0% of the 2,228 domains Domduck read on disallow it, against 16.7% on .

Operated by
OpenAI
Blocked by
17.0%
378 domains
Allowed by name
5.4%
the file names it
Not named
80.4%
the file is silent
Sample
2,228
domains read
Token
GPTBot
in robots.txt

What blocking GPTBot costs

Your pages stop being added to OpenAI’s training data from the moment you block it. Content already collected is not withdrawn, so this is a decision about the future and not about the past.

What it does not cost

It does not remove you from ChatGPT’s live browsing. That is OAI-SearchBot, a separate agent with a separate name, and blocking one does nothing to the other. Sites that block GPTBot expecting to disappear from ChatGPT answers are usually surprised.

OpenAI publishes its crawler IP ranges, so a request claiming to be GPTBot can be verified rather than trusted.

OpenAI’s own documentation

How to block it

or allow it explicitly
User-agent: GPTBot
Disallow: /

To allow it and say so, use a bare Disallow: with nothing after it, which RFC 9309 defines as “nothing is disallowed”. That matters more than it looks: a file that never names an agent leaves it in a third state, neither allowed nor blocked, and a catch-all rule added later will block it without anyone deciding to.

Block rate over time

2 days
2026-08-022026-08-03

The percentage of domains in the corpus whose robots.txt disallows GPTBot, measured once per domain per pass. The denominator is how many domains were read that day and is stored with each point, so the line is not restated when the corpus grows.

The other seven

Anthropic
ClaudeBot
Perplexity
PerplexityBot
Common Crawl
CCBot
ByteDance
Bytespider

How this page is made

The explanation above is written by a person. The numbers are computed from Domduck’s own daily reads of each domain’s robots.txt, parsed to RFC 9309 rather than searched for a string: a rule binds to the nearest preceding group of user-agent lines, the longest matching pattern wins, and Allow breaks a tie. No prefix matching is done, so a group naming Google is not treated as a policy about Google-Extended. What the collectors have read.

GPTBot: what it is, and who blocks it | Domduck