Bing splits AI training from grounding New preferences in robots.txt, meta tags and HTTP headers

From February 1, 2027, Bing will let you decide separately whether your content can train AI models and whether it can be used to build answers in Copilot and other Microsoft products. Today both are handled by the same NoArchive and NoCache directives, which Microsoft introduced in 2023. The new settings have no effect on classic search results. If you use the old directives and want to keep your restrictions, you have until March 1, 2027.

The change was announced on October 30, 2026 by Krishna Madhavan, Principal Product Manager at Microsoft Bing, and Barry Schwartz covered it the same day on Search Engine Roundtable.

How training differs from grounding

People mix these two up all the time when they talk about “blocking AI”, so I’ll start with the definitions Microsoft uses:

• AI Training means using content to change the learned parameters of a generative model. Microsoft follows the IETF working definition, under which training a model that is used only for non-generative tasks falls outside this category, even if that model could generate.
• AI Grounding means using content as input that a model bases a specific answer on. When Copilot answers a question and footnotes the pages it used, that is grounding.
• Classic indexing and ranking are not grounding. Microsoft states plainly that the new settings don’t change whether a page can appear in traditional Bing results.

In practice, you may not want a model to learn from your articles while still wanting Copilot to cite them. The old directives couldn’t express that clearly, while the new ones give you two independent switches.

What exactly Microsoft is introducing

The training preference in robots.txt and on the page

You express consent to training in a shared syntax that grew out of the IETF AI Preferences working group. Microsoft says it is introducing it together with other industry players but doesn’t name them. According to Search Engine Roundtable, Google, OpenAI and Anthropic are on board. The IETF standard itself is still an Internet-Draft, which Microsoft openly acknowledges.

In robots.txt it looks like this:

# Block AI training for the whole site
User-Agent: __usage_ai-training
Disallow: /

# Or only for a specific path
User-Agent: __usage_ai-training
Disallow: /example/

With no rule at all, your content may be used for training, because the default is to allow it. If you want to state that explicitly, you can use Allow: / under the same User-Agent. Microsoft also extends this preference to individual pages, and the page setting overrides robots.txt:

<meta name="Bingbot" content="ai-training:disallow">

X-Robots-Tag: Bingbot: ai-training:disallow

The grounding preference, page level only

Grounding is a Microsoft-only setting and has no robots.txt rule. You set it per page, with a meta tag or an HTTP header, and pick one of three options:

Setting	What it means
Allow (default)	Content may help build answers.
DisallowExtended	Content is used only in Microsoft-operated products (e.g. AI Mode, AI Overview, Copilot, Microsoft 365), not in the grounding services Microsoft provides to third parties (Web IQ, grounding through the API in Foundry).
Disallow	Content is not used to build answers.

The syntax mirrors the training one:

<meta name="Bingbot" content="ai-grounding:disallow">
<meta name="Bingbot" content="ai-grounding:disallowextended">

X-Robots-Tag: Bingbot: ai-grounding:disallow

The grounding preference doesn’t cover the Bing search syndication API. According to Microsoft, that API returns links for traditional search and isn’t meant for AI-generated summaries.

Timeline

The move away from the old directives has several dates, and some of them only apply to sites using NoArchive or NoCache:

Date	What happens
December 15, 2026	Temporary grounding preferences open in Bing Webmaster Tools, only for sites that use NoArchive or NoCache on that date.
February 1, 2027	Microsoft starts honoring the new preferences as it crawls.
March 1, 2027	Deadline for sites with NoArchive/NoCache that want to keep their restrictions. After that, the old directives are no longer read as a training or grounding preference.
September 30, 2027	Temporary grounding preferences expire. For a given page, the temporary setting ends earlier if Bing recrawls it and finds a new preference.

Once Bing detects the new preferences on a site, it ignores any NoArchive and NoCache left there.

What this means in practice

If you publish nothing, you allow everything. No rule means consent to both training and grounding. A site that has never used NoArchive or NoCache doesn’t need to change anything. It’s still better to make this choice on purpose, since robots.txt tends to be written once and then left alone for years.

Sites using NoArchive or NoCache need to act. To keep blocking grounding, you have to do two things: publish the preference on your pages and record a temporary preference in Bing Webmaster Tools. Microsoft asks for the second step because it may not recrawl your pages between February and March 1. If you skip it, pages with the old directives are treated as “Allow” until Bing recrawls them. Training has no such temporary setting, so you need to publish the preference in robots.txt or on your pages by March 1, 2027. Settings published after that date still work, but only once the page is recrawled, and there’s no telling when that will be.

Blocking grounding costs you AI visibility. Microsoft warns that DisallowExtended and Disallow will significantly reduce your content’s visibility in AI answers, without changing your position in classic results. Allowing grounding gives you a chance at links and citations, though Microsoft is clear that no option guarantees inclusion, traffic or payment.

An example setup for publishers. Microsoft’s own example is training blocked and grounding allowed: your content doesn’t train models but can still be cited. Whether that suits your site depends on how much traffic from AI answers is worth to you. The AI Performance report in Bing Webmaster Tools helps here: it shows the number of citations, the cited pages and the queries your content was used to ground (Grounding Queries).

The AI Performance report for one of my sites, catalog.polo.blue: 7.7K citations across Microsoft Copilots and partners over the last three months.

On uper.pl the same report shows 2.5K citations over that period, and the queries include two Polish searches asking how a training bot differs from an AI search bot (22 and 14 citations). This is exactly the kind of use that ai-grounding:disallow would cut off. Google is testing a similar report in Search Console, which I covered in the article on AI Search reports and the opt-out.

Limitations and pitfalls

• Preferences govern how content is used; robots.txt still governs access. “Allow” doesn’t lift a robots.txt block on Bingbot. If you block the crawler, Microsoft sees neither the page nor its preferences.
• A setting takes effect once Bingbot visits. Microsoft has to crawl the page to read it. IndexNow raises the chance of a faster recrawl but doesn’t guarantee a date.
• Preferences aren’t a license. They create no agreement with the publisher and no right to payment.
• Grounding is Microsoft’s own mechanism. The shared __usage_ai-training rule is meant to work more widely, but grounding, DisallowExtended and the Bingbot meta tags are Bing-only. I wrote about the difference between AI bots and search crawlers in AI crawlers vs search crawlers, and about the file itself in the robots.txt guide.
• The IETF standard is still a draft. Details of the syntax may still change in the final version.

Summary

Bing separates consent to model training from consent to being cited in AI answers, and leaves classic ranking alone. Most sites don’t need to do anything, though it’s a good moment to decide which of the two you’re willing to allow. Sites using NoArchive or NoCache have less room: by March 1, 2027 they must publish the new settings and record a temporary preference in Bing Webmaster Tools, or their restrictions will stop working.

Frequently Asked Questions

Will blocking AI training in Bing affect my rankings?

No. Microsoft says the new preferences do not change whether a page can appear in traditional Bing results, as long as it meets the usual crawling and indexing requirements.

What is the difference between training and grounding?

Training changes the parameters of a generative model. Grounding uses a page as input that a model bases a specific answer on, usually with a link to the source.

What do I put in robots.txt to block AI training in Bing?

Use the shared syntax: User-Agent: __usage_ai-training followed by Disallow: / on the next line. For a single page you can instead use the ai-training:disallow meta tag for Bingbot or the X-Robots-Tag header.

Will NoArchive and NoCache stop working?

As training and grounding preferences, yes: after March 1, 2027 Bing will stop reading them that way. To keep your restrictions, publish the new preferences by that date, and for grounding also record the temporary setting in Bing Webmaster Tools.

Will blocking grounding reduce my citations in Copilot?

According to Microsoft, yes: Disallow and DisallowExtended reduce your content's visibility in AI answers. Allowing grounding does not guarantee anything either.

Sources

1. Microsoft Bing Blog: Your content, your choice: clearer preferences for training and grounding https://blogs.bing.com/search/2026/10/your-content-your-choice-clearer-preferences-for-training-and-grounding/
2. Search Engine Roundtable: Microsoft Bing Separates AI Training From AI Search Grounding With New Controls https://www.seroundtable.com/microsoft-bing-ai-controls-42429.html
3. Bing Webmaster Tools: New AI preferences: help documentation https://www.bing.com/webmasters/help/new-ai-preferences-2t1cc16d