What is ChatGPT-User and what does it read?
ChatGPT-User is the user agent OpenAI uses when ChatGPT fetches a web page live to answer a person's question. It is not a crawler. It appears on your site at the moment someone asks ChatGPT something your page answers, and it reads the page in that moment. On tollbit.com it made about 10.5K successful reads between September 15 and October 14 2026, the second largest AI reader of our site after Meta's indexer, and the single largest source of human referrals back from an AI answer.
How it declares itself
OpenAI documents three agents. ChatGPT-User is the retrieval one. Its user agent string contains the token ChatGPT-User/1.0 and a link to https://openai.com/bot, inside a Mozilla-compatible prefix:
Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; ChatGPT-User/1.0; +https://openai.com/bot
OpenAI publishes the IP ranges it fetches from at https://openai.com/chatgpt-user.json, so a request can be verified against the list rather than trusted on the string alone. The robots.txt token is ChatGPT-User. OpenAI's documentation describes it as fetching on behalf of a user and not for crawling the web in an automatic fashion.
How it differs from GPTBot and OAI-SearchBot
| ChatGPT-User | OAI-SearchBot | GPTBot | |
|---|---|---|---|
| Purpose | Live fetch for one user's question | Build the index behind ChatGPT search | Collect pages for model training |
| When it arrives | When a person asks | On a schedule, sitemap-driven | On a schedule, bulk |
| Reads on tollbit.com, 30 days | 10.5K | 1.4K | 0.7K |
| Top URL on tollbit.com | Homepage and product pages | robots.txt, about 1,000 fetches | Spread across the site |
The three are governed separately. Disallowing GPTBot in robots.txt stops the training crawler and does nothing to live retrieval. Disallowing OAI-SearchBot removes you from ChatGPT search results. ChatGPT-User's behaviour toward robots.txt has been uneven: our Q3 and Q4 2025 State of the Bots report measured a 42% rate of fetches against disallowed paths, the highest of the agents we tracked at the time.
What it did on tollbit.com
The 10.5K reads were concentrated on pages a buyer would ask about. The homepage led, followed by the publisher solutions page, the licensed RAG page and the bot paywall page, and then our one explainer post about the OpenAI Apps SDK. That pattern is retrieval, not crawling: a person asks ChatGPT what TollBit does or what a bot paywall is, and the agent fetches the page that answers.
Two more things from the same logs:
- Reads to referrals. ChatGPT sent 37 human visitors to tollbit.com in the period, the most of any AI assistant (Claude 7, Perplexity 3). Against 10.5K reads that is roughly 280 reads per human click, and across all agents on our site the ratio is about 1,000 to 1.
- Noise. A share of requests arriving under AI-looking user agents were vulnerability probes for
/.git/config,/wp-adminand similar, not content gaps. Verifying against OpenAI's published ranges separates the real agent from imitators.
How to treat it
ChatGPT-User is the agent that puts your facts into answers your customers read today, so for most commercial pages the right response is to serve it well rather than block it.
- Serve it readable text. It executes no JavaScript. A page that is an app shell gives it a few dozen words. Server-rendered HTML with the same facts as the human page gives it the answer.
- Keep facts current. It fetches live, so a price or stock change on the page reaches the next answer within whatever the platform's cache allows, from hours to weeks.
- Measure it in server logs. Google Analytics cannot see it. Count reads by page, then segment
chatgpt.comas a referrer to see the clicks that come back. - Route it. Send identified ChatGPT-User traffic to an agent site on your own domain. Same facts, cleaner format, bots only, and humans never notice.
Publishers can apply terms instead: an HTTP 402 response with a per-fetch rate charges the retrieval that powers answers, which is the recurring use of content, rather than blocking it.
Frequently asked questions
Does blocking ChatGPT-User remove me from ChatGPT? It removes your pages from live answers where the agent honours the rule, so ChatGPT falls back to training data and third-party pages about you. You stop being the source of your own facts.
Is ChatGPT-User used for training? OpenAI's documentation assigns training collection to GPTBot and live retrieval to ChatGPT-User. The two have separate user agents, IP lists and robots.txt tokens.
Why does ChatGPT-User read my site far more than GPTBot? Because people ask questions far more often than models are trained. On tollbit.com the ratio was 10.5K to 0.7K in 30 days. Retrieval is the recurring visitor.