Google, OpenAI and Microsoft back a shared AI-training signal

Google, OpenAI and Microsoft announced on 30 October 2026 that they will support a shared way for website owners to express AI-training preferences in robots.txt. The proposal addresses a practical problem: publishers currently have to name different crawlers or controls for different providers. A shared signal could let a site state a broad preference once while keeping company-specific rules available.

Google Search Central says the signal will use a policy token and that Google aims to support it with other providers in 2027. If a site uses the shared signal to disallow training and does not separately allow Google-Extended, Google plans to treat that as a Google-Extended disallow. Its existing controls remain in place. OpenAI says it will honor the common signal and aims to implement it in the coming months; its current crawler controls will continue to apply.

Microsoft’s Bing announcement adds separate preferences for AI training and AI grounding. The company says the new settings will take effect as it crawls pages from 1 February 2027, while older NoArchive and NoCache settings need a transition before 1 March 2027. Microsoft also says these choices do not remove eligibility for traditional Bing search results, but they can reduce visibility in AI-generated experiences. These are published plans, not evidence that every provider already processes the signal.

The technical basis is an individual Internet-Draft published at the IETF on 30 October. Kevin Kelley, whose author information lists Anthropic, proposes the policy token __usage_ai-training in an ordinary robots.txt group. A Disallow rule can express a preference against training, while the draft says the token is not itself a crawler-access directive. The IETF explicitly describes the document as a work in progress without formal standing as a standard.

Independent coverage from Search Engine Land and SiliconANGLE confirms that the announcements were made across the industry. Exact public view or reaction figures were not available, so that coverage is an interest signal rather than a popularity measurement.

The practical importance is that publishers, creators and businesses may eventually manage AI-training choices through a common policy surface instead of maintaining provider-by-provider lists. AI builders will need to distinguish technical access, model training and grounding. The proposal does not establish licensing, payment, enforcement or how previously collected material is treated. It is therefore a significant governance update, but not yet a universal opt-out or a finished standard.