One line to refuse AI training: at Google, that same line also costs you Gemini

On October 30, 2026, Google and Microsoft announced on the same day that they would honour the same robots.txt line for refusing to let your site train AI models, and at Google that line also removes you from Gemini's answers. The syntax is identical at both companies. The consequence is not: Microsoft separates training a model from using your content in an answer, while Google treats the two as a single decision.

What did Google and Microsoft announce on October 30, 2026?

Google and Microsoft announced support for a shared token written in robots.txt, named __usage_ai-training, which expresses one refusal of AI training that applies across several AI providers at once. The token goes where a crawler name normally goes, on the User-Agent line, and reuses the existing robots.txt grammar without adding a single new directive. Until now, refusing AI training meant naming every provider one by one: Google-Extended for Google, then the equivalent at each of the others.

The reference text is an Internet-Draft filed on October 30, 2026 with the IETF (Internet Engineering Task Force), the body that standardises web protocols, by Kevin Kelley, an engineer at Anthropic. Google published its post the same day, signed by John Mueller, and says it is aiming for 2027 to support the control. Microsoft, through Krishna Madhavan, will begin honouring the new preferences on February 1, 2027, as it recrawls.

Why does the same line behave differently at Google and at Microsoft?

Because Google collapses that training refusal onto a setting that also governs citations. Google writes that when a site uses the shared token to opt out of training and Google-Extended is not explicitly specified, Google will read that line as a Google-Extended disallow. In the same post, Google states that websites which specifically allow Google-Extended are eligible to appear in responses and receive clicks via grounding in Gemini Apps. Refusing AI training in one line, with nothing else specified, therefore means leaving Gemini's answers.

Microsoft went the other way and split the decision in two. At Microsoft, a training refusal can be written in robots.txt or page by page, while the use of your content inside a generated answer, which Microsoft calls AI Grounding, can only be set page by page, through a meta tag or an HTTP header. A site can therefore refuse AI training at Microsoft and remain citable in Copilot. Microsoft even adds a middle setting, DisallowExtended, which permits grounding inside Microsoft's own experiences and withholds it from the grounding services Microsoft sells to third parties.

What is the difference between training a model and feeding an answer?

Training a model permanently changes its parameters, while feeding an answer merely hands the model a source at the moment it replies. Microsoft states plainly that indexing and ranking pages in traditional search results is not grounding, and that grounding is what brings links and citations into a generated answer. That is where the commercial difference sits: training returns nothing you can trace, whereas use inside an answer is what can produce a citation, and then a click.

That slot inside the answer has a measurable value. Ahrefs compared 300,000 keywords between December 2023 and December 2025 and found the click-through rate of the top organic position is 58% lower when an AI overview appears above it. Once the answer absorbs the click, appearing inside that answer stops being a bonus. Microsoft puts it without promising anything: allowing grounding makes your content eligible to inform those responses, with no guarantee of inclusion, traffic or payment.

What does this change for a small business?

Nothing before February 2027 if you have never touched these settings, because allow is the default on both sides. A site that has written nothing stays exactly where it is: training allowed, use in answers allowed. The October 30, 2026 announcement takes nothing away from anyone automatically, and requires no immediate action.

Two checks are worth doing now, however. The first is what your robots.txt already says about Google-Extended, because a great many sites closed it in 2023 or 2024 as a precaution, often through a checkbox in a plugin, and that line is already costing them eligibility for citations in Gemini Apps. We described that trap in "Blocking AI bots with one click can cost you Google: what changes on September 15": the decision takes one click, and the bill arrives months later, in lost visibility, with no alert.

The second check is for the NoArchive and NoCache tags, which old SEO plugins still set without anyone remembering. Microsoft will stop reading them as training and grounding preferences after March 1, 2027, and the affected pages will revert to allow by default until they are recrawled. In other words, a site that believed it was restricting something is about to stop restricting it, without anyone changing a thing. Microsoft opens a temporary setting in Bing Webmaster Tools on December 15, 2026, available only to sites using those tags on that date, and it expires on September 30, 2027.

For anyone trading in Europe, the regulatory backdrop makes this less abstract than it looks. Google writes that its proposal is consistent with the EU Code of Practice for General Purpose AI model providers under the EU AI Act, which recognises robots.txt as a primary method for publishers to express machine-readable preferences. A file of a few lines at the root of your domain is now a document that regulators read.

Who will not be able to apply these settings, even if they want to?

Sites on platforms that give access to neither HTTP headers nor the meta tags of individual pages. Microsoft's decision to put grounding exclusively at page level has a direct consequence: the setting that matters commercially lives in the one place many site owners do not control.

• A training refusal is written in robots.txt, which already assumes you can edit a file at the root of the domain.
• Microsoft's grounding setting runs through a meta tag or an HTTP header, page by page, with no site-wide equivalent.
• On locked platforms, the meta tag is sometimes still reachable when the HTTP header is not, and the root file not at all.

This is the same fault line described in "Shopify, Wix, Webflow, Framer: what your platform actually lets you do for AI". Per-platform feasibility is one of the things we check, and the full list of signals is published on /en/methode. A setting you cannot write is not a setting you declined. It is a setting your platform made on your behalf.

Is this actually a standard?

Not yet, and that deserves saying plainly. The document filed with the IETF on October 30, 2026 carries an explicit notice that it is not endorsed by the IETF and has no formal standing in the IETF standards process. It is an individual submission rather than a working group document, and it expires on May 3, 2027 unless something replaces it. The Google and Microsoft posts are dated intentions, not shipped features.

That does not make the text toothless. Robots.txt itself was nothing more than a convention for twenty-eight years: Martijn Koster defined it in 1994, and it was only published as an internet standard in 2022, as RFC 9309. On the web, practice arrives before the standard, and two companies the size of Google and Microsoft announcing the same syntax on the same day carry more weight than an RFC number. What matters for you is not the document's status. It is that two engines will soon read this line, and that they will not read it the same way.