Google backs a shared AI training signal, with support planned for 2027

One preference, more than one provider

Website owners should not have to mistake a growing list of crawler names for a clear policy decision. That is the practical problem behind Google’s October 30 announcement that it plans to support a shared AI training control in robots.txt. The proposed mechanism would let a site express a broad training preference across participating providers while keeping company-specific controls available.

Google’s announcement gives the idea a concrete interpretation for its own systems. If a website uses the shared policy token to opt out of training and has not explicitly specified Google-Extended, Google plans to interpret the signal as though Google-Extended had been disallowed. The qualification about an explicit provider-specific setting matters. The shared choice is not described as erasing every more specific instruction a publisher has already made.

The company says it aims to support the new control in 2027. That is a future implementation plan, not confirmation that a publisher can rely on a completed rollout today. The fresh news is Google’s stated support and intended behaviour. Foundra has not changed this site’s robots.txt, tested a live Google implementation or inferred compliance by a crawler from the announcement.

A draft is a place to develop the rule

Separately, the IETF draft record identifies the proposed machine-readable training-preference signal as an Internet-Draft. That status is important. A draft can organise technical discussion and give companies something specific to support without being a final standard or a published RFC. Describing industry interest as a completed standards process would go further than the evidence.

Google places its plan in work from the IETF’s AIPREF group and says discussions continue on the broader vocabulary. It also says existing company-specific controls will keep functioning alongside the shared signal. These points describe a transition with overlapping mechanisms, not a sudden replacement of every publisher setting.

The Verge reports support from several major AI companies. This article uses that account as context for the cross-provider ambition, while Google’s own statement establishes the narrower claims about its planned implementation. A common expression of preference becomes useful only to the extent that each participating provider implements and honours it. An announcement of support does not independently verify that behaviour.

Training is not every possible use

The scope of a preference must be understood before a site owner adopts it. Google’s post discusses AI training and also describes the relationship between Google-Extended and grounding in Gemini Apps. Those references are a reason to read the provider’s controls carefully, not to assume that a single shared training choice automatically settles every form of retrieval, citation or generated answer.

Search visibility, model training and grounding are different questions in a publisher’s decision. A control intended for one use should not be promoted as a universal switch for all of them. The reviewed evidence does not support a claim that this proposed signal guarantees inclusion in, exclusion from, or referral traffic through every AI service.

That leaves implementation details commercially significant. A publisher may want to express one broad preference and then make a deliberate exception for a particular provider. The proposed coexistence with provider-specific controls could allow that distinction, but the exact live behaviour still needs documentation and testing when the mechanism is implemented. This is not a recommendation to remove existing controls in anticipation of that future state.

Robots.txt expresses a choice; it does not sell rights

A machine-readable preference can make communication clearer without becoming a contract for the underlying material. It does not establish that the site owns every right in every page, and it does not price a licence. A publisher making a training choice still has to understand its own rights and agreements. A technical signal cannot resolve a copyright question simply by being easy to parse.

The reverse distinction matters too: blocking a use through a preference mechanism is not the same thing as proof that a particular provider has never accessed the material. Verification would need evidence about actual behaviour. The sources reviewed for this article support a proposed control and announced intentions, not a retrospective audit of crawler activity.

The useful milestone is observable support

The proposal’s attraction is practical: fewer repetitive settings could make a publisher’s policy easier to maintain and communicate. Its value should eventually be judged by clear documentation, compatible interpretation and observable implementation, not only by the number of companies named in a support statement.

For now, the distinction is manageable. There is an official October 30 Google announcement, an IETF draft and a stated 2027 aim. There is not a verified present-day rollout or a final standards outcome established by this evidence. The shared signal could simplify a choice. It has not yet made the choice, or the responsibility for understanding it, disappear.