OpenAI announced on October 30, 2026, that it plans to honor a shared AI-training preference signal for websites. For publishers, independent creators and small-business owners, the aim is to communicate the same choice to several participating AI providers without maintaining a separate declaration for each one.
This is a commitment, not a universal opt-out available today. OpenAI targets implementation in the coming months, has not given an exact activation date and says its existing controls will remain in place.
Training and search are already separate choices at OpenAI
OpenAI already lets website owners decline training while allowing its search crawler. Its official crawler documentation treats GPTBot, which collects potential training material, and OAI-SearchBot, which serves ChatGPT search, as independent settings.
For an owner who wants search visibility without training use, that distinction matters now. Blocking OAI-SearchBot instead would exclude the site from ChatGPT search answers, although navigational links could remain. Allowing search crawling is not a promise of inclusion or traffic. OpenAI documents those search controls separately.
Nor does a GPTBot restriction block every ChatGPT visit. OpenAI says ChatGPT-User handles certain user-initiated requests, for which robots.txt rules may not apply.
There is a third question: may a provider use the page as a source when generating an answer? That is called grounding. Microsoft’s planned controls distinguish grounding from training and from ordinary search eligibility. For website owners, these are separate decisions even when a provider bundles some of the settings.
What the shared signal would change
The proposal uses robots.txt, the file websites publish to communicate rules to automated visitors. A special label, __usage_ai-training, would express a training preference for an entire site or selected paths. It is a usage-policy label, not the name of a new crawler. The proposal remains an individual Internet-Draft, not an adopted IETF standard.
Under the draft, the training preference does not itself prohibit fetching a page; crawler rules remain separate. Existing catch-all rules do not automatically become this new declaration: the special label must be present. The draft also permits provider-specific preferences and private agreements to take precedence. These details are set out in its design sections.
For website owners dealing with OpenAI, the prospective gain is simpler cross-provider administration, not the first opportunity to separate training from search. How much work it saves will depend on provider support and the exceptions an owner still needs.
The providers are on different timelines
Provider | Announced timing |
|---|---|
Targets implementation in the coming months; exact date unannounced. | |
Aims to support the shared control in 2027. | |
Plans to honor new training and grounding preferences from February 1, 2027, as it crawls. | |
Kevin Kelley says implementation has begun, but support across services will take time; no completion date stated. |
The same declaration may also have different product effects. Google says that, when Google-Extended is not explicitly specified, it plans to interpret a shared training opt-out as a Google-Extended disallow.
Google’s current documentation says Google-Extended covers both Gemini training and certain grounding uses in Gemini Apps and Vertex AI. It does not affect inclusion or ranking in Google Search.
Read together, those policies suggest that a shared training opt-out could also affect eligibility to inform some Gemini answers through Google’s planned fallback. Preserving ordinary Google Search eligibility is not the same as preserving every AI-answer appearance.
What website owners can do now
Keep a record of current provider-specific rules and retain working controls during rollout. OpenAI says the new signal will complement its existing controls.
Write down the desired outcome for each use: page retrieval, model training and AI-answer grounding. Before changing settings, check which uses a provider combines; Google-Extended is a concrete example.
If your site uses Microsoft’s NoArchive or NoCache preferences, review its transition instructions before March 1, 2027. Microsoft says those tags will no longer express training or grounding preferences after that date. Its guide covers replacement settings and the additional Bing Webmaster Tools step for eligible sites restricting grounding.
Finally, a preference declaration is not a security barrier: the robots.txt standard says these rules are not access authorization. And the training proposal associates preferences with retrieval time; it does not provide a mechanism to erase material from past training.