← All guides · Published November 14, 2026 · Reviewed November 14, 2026 · By the Ingenious AI team · 7 min read
How Do You Block GPTBot Without Losing ChatGPT Citations?
The short answer: you treat OpenAI's three agents as three separate decisions, because they are three separate programs. GPTBot is the training crawler — block it and your pages stay out of training crawls. OAI-SearchBot indexes pages for ChatGPT answers and ChatGPT-User fetches a page when a user asks — allow those two and ChatGPT can still cite you. All three are verified entries in AgentBlock's database, so a kit can block exactly one and allow exactly two.
Build my block kit — free · Test a policy in the simulator first
What is GPTBot?
OpenAI's training crawler, a verified entry in AgentBlock's database. It collects pages for model training data. It is a different agent from OAI-SearchBot (search indexing for ChatGPT answers) and ChatGPT-User (user-triggered fetching), and each of the three can be blocked or allowed independently.
Three OpenAI agents, three jobs
OpenAI publishes three user-agents in the database. GPTBot crawls to build training data — the visit happens whether or not any user ever sees your page. OAI-SearchBot crawls to build the index ChatGPT searches when composing answers. ChatGPT-User does not crawl at all in advance; it fetches a specific page at the moment a ChatGPT user asks about your content, which is why its note in the database reads: blocking removes your pages as citable live sources.
Most site owners who say 'block OpenAI' mean the first job only. They object to training use; they do not object to being cited in answers, which sends readers. A single 'block OpenAI' rule that sweeps all three tokens achieves the training goal and silently pays for it with every future citation — the most common own-goal in AI crawler policy.
What blocking GPTBot does and does not do
Blocking GPTBot stops that crawler's future fetches of your pages. It does not remove copies already collected, does not affect pages about you on other sites, and does not stop ChatGPT-User, which is a retrieval agent with its own token. It also does not stop OpenAI traffic that arrives without the GPTBot name — the kit's standing disclaimer is that sophisticated spoofing bypasses any user-agent block.
GPTBot also has limited JavaScript support, a detail the database records because it affects what the crawler sees on script-heavy sites. Server-rendered content is fully visible to it. If your objection is to training use of your full text and images, the block belongs at the edge or server, where it applies regardless of rendering.
The citation side of the trade
OAI-SearchBot and ChatGPT-User are the citation supply chain. The searchbot's index is what an answer draws on; the user fetcher is what lets ChatGPT read a page fresh when someone asks about your pricing, your guide, or your product by name. Sites that allow both regularly appear as cited sources; sites that block both disappear from that surface while competitors remain.
There is a legitimate reason to block retrieval too: some publishers do not want their content summarised at all, even with attribution. That is a posture, not a mistake — as long as it is chosen. The failure mode this guide exists to prevent is choosing training-only blocking and receiving full blocking because the policy treated a vendor as a single switch.
Building the split policy
In AgentBlock's builder the split is three flips: GPTBot to block (its training-class default), OAI-SearchBot and ChatGPT-User to allow (overriding the retrieval-class default of block). The builder shows the consequence per agent before generating anything, and the pre-deploy simulator runs the draft policy against every known agent and your key paths, so the citation paths can be checked before the kit ships.
Deploy on the artifact that matches your stack — the free Cloudflare Worker at the edge, the WordPress plugin on your server, or the JS snippet as a deterrent layer — then verify with the kit verifier, which checks the deployed artifact against the policy you built: the blocked list, the allowed list, the kit ID and the kit hash.
Watching the result in your logs
After deployment, the evidence is in two places. Your kit's block log should show GPTBot requests refused, with counts falling as the crawler learns the refusal pattern. Your server log should still show OAI-SearchBot and ChatGPT-User requests succeeding — if retrieval visits drop to zero too, something broader than your policy is at work and the enforcement-gap audit, which compares robots.txt and meta tags against the edge policy, is the tool that finds the drift.
Re-check quarterly. OpenAI's agent set has already changed once in this database's short life, and a policy built in October is a snapshot. The stale check diffs your kit against the current database so an agent added later is a decision you make, not a gap you discover.
Five actions to take today
- Separate the three OpenAI decisions. GPTBot, OAI-SearchBot and ChatGPT-User each get their own block/allow call — never one vendor switch.
- Block GPTBot. Set the training crawler to block; it is the only one of the three whose visits build training data.
- Allow both retrieval agents. Flip OAI-SearchBot and ChatGPT-User to allow if you want ChatGPT citations to keep flowing.
- Simulate before deploying. Run the draft policy in the simulator and confirm retrieval paths stay open on your key pages.
- Verify after deploying. Check the kit verifier, then watch one week of logs for blocked GPTBot and successful retrieval fetches.
Worked example: a tax guide publisher splits the OpenAI decision
Ledger Lane, a fictional publisher of small-business tax guides, wanted its 340 guides out of training sets after finding its checklists reproduced in model outputs. Its first instinct — block everything from OpenAI — would have been expensive: its analytics showed ChatGPT referrals had grown to 612 sessions a month, driven by citations in answers about quarterly filings.
The split policy blocked GPTBot and allowed OAI-SearchBot and ChatGPT-User. In the first 30 days the Worker's block log recorded 8,417 refused GPTBot requests across the guide library. Retrieval traffic held: 1,193 OAI-SearchBot fetches and 286 ChatGPT-User fetches returned 200s in the same window, and ChatGPT-referred sessions measured 598 — statistically the same as before the change.
The one surprise was a legacy robots.txt rule disallowing all OpenAI-named agents, written a year earlier, which the enforcement-gap audit flagged as drift against the new edge policy. Removing it mattered: robots rules and edge rules that contradict each other produce a policy nobody actually chose.
Frequently asked questions
Does blocking GPTBot remove my content from ChatGPT's training data?
It stops future GPTBot crawls of your site. It cannot recall copies already collected in past crawls, and it does not affect your content quoted on other sites. Think of it as closing the door from today forward, which is what crawler policy can honestly promise.
Will ChatGPT still cite me if I block GPTBot?
Yes, if you allow the retrieval agents. Citations come from OAI-SearchBot's index and ChatGPT-User's live fetches, both separate verified agents. Block GPTBot alone and the citation supply chain stays intact.
What happens if I block ChatGPT-User by mistake?
Your pages stop being fetchable at the moment a user asks about them, so live citations and fresh reads disappear even if the search index still holds older copies. In AgentBlock the fix is one flip in the builder and a regenerated kit — then verify the new kit hash.
Does GPTBot respect robots.txt?
OpenAI documents robots.txt controls for GPTBot, and many sites start there. AgentBlock still recommends edge or server enforcement for the block you care about, because robots.txt is advisory and other layers of your stack can contradict it — the enforcement-gap audit exists to catch exactly that.
Is the GPTBot / retrieval split possible on the free plan?
Yes. Building the policy and downloading the Cloudflare Worker are free, and per-agent decisions are the builder's core function. Pro adds the WordPress plugin and JS snippet downloads; Team adds the AgentGate export and classify API.
Related guides
- How to Block AI Scrapers and Bots from Your Website
- Which AI Crawlers Should You Allow?
- How Do You Block AI Bots on WordPress, Cloudflare, or Plain JS?
- Pre-deploy breakage simulator — test a policy before it ships
- Impact preview — your server log against a draft policy
- What Is an AI Crawler?
- How Do You Block ClaudeBot and Keep Claude Citations?
- How Do You Block CCBot and the Common Crawl Training Feed?
- Google-Extended vs Googlebot: How Do You Opt Out of AI Training Only?
- How Does Applebot-Extended Opt You Out of Apple Intelligence Training?
- Block evidence — parse a log, build an evidence pack
- The dated agent database — 25 entries, graded