The Short Version
On October 30, 2026, OpenAI, Google, Microsoft and Anthropic said they will honour one shared robots.txt token, __usage_ai-training, as a preference against model training. They agree on the token. They differ on when they start, how it interacts with their existing bot rules, and whether anything beyond training is affected. Here is what each has said.
Side by Side
| Microsoft (Bing) | OpenAI | Anthropic | ||
|---|---|---|---|---|
| Starts | "In 2027" | February 1, 2027 | "In the coming months" | Reported as live |
| Existing controls | Keep working alongside it. The shared token with no Google-Extended line counts as disallowing Google-Extended | NoArchive and NoCache are replaced as AI preferences during a transition | GPTBot and OAI-SearchBot stay. The signal "will not replace" them | Not detailed in what we read |
| Effect on search | Not addressed in Google's post | Bing search eligibility unchanged | Not addressed | Not addressed |
| Grounding (AI answers) | Not addressed. Claims that it blocks Gemini grounding are unconfirmed | Separate page-level control, ai-grounding | Not addressed | Not addressed |
| Page-level option | Not stated | Yes: ai-training:disallow meta tag or header, which overrides robots.txt | Not stated | Not stated |
| Primary source read | Yes | Yes | No, via coverage | No, via coverage |
Google's Search Central post calls it a shared "policy token" and links the IETF draft it is based on, draft-kelley-aipref-training-preference-signal. It says Google aims to support it in 2027 and that company-specific controls keep working. It does not mention Search, AI Overviews or rankings. See Google-Extended for the current control.
Microsoft
Microsoft published the most detail: the exact syntax, a February 1, 2027 start, a page-level override, and a separate grounding control with its own deadlines. It is also the only one to say clearly that search eligibility is unchanged. See setting Bing's AI grounding controls.
OpenAI
OpenAI says it aims to implement the signal in the coming months and that its existing robots.txt controls remain. It hasn't said how the shared token ranks against a GPTBot rule if the two conflict. Until it does, keep them consistent. See GPTBot and OAI-SearchBot.
Anthropic
One report says Anthropic already honours the token, and coverage credits an Anthropic engineer with the IETF draft. We didn't find Anthropic documentation of the details, so check its crawler page before relying on it. See ClaudeBot.
What to Do
If you want to opt out of training, add the shared group and keep your per-company rules matching it. If you want to stay in AI answers, leave search and user-agent bots allowed and, for Bing, leave grounding at its default. The steps are in how to set the shared AI training signal.