Definition
The AI training preference signal is a single robots.txt group that tells every participating AI company not to use your pages for model training. Instead of writing separate rules for GPTBot, Google-Extended, ClaudeBot and others, you address a shared user-agent token. OpenAI, Google, Microsoft and Anthropic announced support on October 30, 2026. It is based on an IETF Internet-Draft, draft-kelley-aipref-training-preference-signal, which is a draft and not an adopted standard.
Syntax
Microsoft's Bing post gives the form:
User-Agent: __usage_ai-training
Disallow: /You can disallow a path instead of the whole site, and add Allow: lines the usual way. Google's post calls it a shared "policy token" without printing the token, so confirm against your own crawler logs and each company's docs once support goes live.
When Each Company Honours It
| Company | Stated timing | Notes |
|---|---|---|
| Anthropic | Reported as live now | Single secondary report. Check Anthropic's crawler docs |
| OpenAI | "In the coming months" | Existing GPTBot and OAI-SearchBot controls stay in place |
| Microsoft (Bing) | February 1, 2027 | Bing search eligibility is unchanged. Grounding has its own page-level control |
| "In 2027" | Using the token with no Google-Extended line counts as disallowing Google-Extended |
What It Does Not Do
It is a preference, not a block. Nothing technically stops a crawler that ignores it. It also covers training only. Microsoft says it does not change eligibility for Bing results, and Google's announcement says nothing about Search, AI Overviews or rankings. Some coverage claims Google will also treat the token as a block on Gemini grounding. Google's post does not say that, so treat it as unconfirmed.
Retrieval for AI answers is a separate decision. Microsoft handles it with a page-level ai-grounding directive, covered in how to set Bing's AI grounding controls. For the other bots, see robots.txt for AI.
Should a Brand Use It?
Usually not across a whole brand site. A model that never trained on your pages knows less about you when it answers without searching, which is how most local and offline models work. The signal makes most sense for publishers and for paid or licensed content. The step-by-step decision is in how to set the shared AI training signal.