Microsoft announced new Bing publisher preferences on October 30, 2026 that separate two uses of web content: training generative AI models and grounding AI-generated answers. The changes are announced, not yet in effect. Microsoft plans to begin honouring the new preferences on February 1, 2027 as it recrawls pages, with a March 1 transition deadline for sites relying on the older NoArchive or NoCache controls.
The distinction gives publishers a choice they could not make as clearly under those older Bing controls. A website may disallow model training while remaining eligible to inform AI answers, or limit the kinds of AI services allowed to use its content for grounding. None of these settings is a way to block crawling, guarantee a citation, or change ordinary Bing Search eligibility. Microsoft set out the new rules and examples in its October 30 announcement and implementation guide.
Two independent decisions, with different places to set them
Training means using content to alter the learned parameters of a generative AI model. Grounding means using content as input when producing a generated response. Microsoft says each preference defaults to Allow if no applicable setting is published. A decision about one use does not automatically decide the other.
| Preference | Where Microsoft will read it | Available choices |
|---|---|---|
| AI Training | A shared robots.txt signal for a site or path, or a Bing-specific page-level meta tag or HTTP header |
Allow or Disallow |
| AI Grounding | A Bing-specific page-level meta tag or HTTP header, not a site-wide robots.txt rule |
Allow, DisallowExtended or Disallow |
DisallowExtended is the important middle option. Microsoft says it permits grounding in Microsoft-operated experiences, including AI Mode, AI Overview, Copilot and Microsoft 365, while excluding grounding services that Microsoft supplies to third parties, such as Web IQ and grounding API access through Microsoft Foundry. Disallow is broader: it expresses a preference not to have the content used to inform generated responses.
This is not a blanket permission system for every AI provider. The training signal draws on work in the IETF AI Preferences working group, which has not finalised its standard. The page-level extensions and grounding controls described here are Microsoft’s own implementation. Other providers’ support for the shared training signal should not be mistaken for support for Microsoft’s grounding settings.
The configuration difference that can change a publisher’s outcome
Microsoft documents a shared training-preference group for robots.txt. Its whole-site opt-out example is:
User-agent: __usage_ai-training
Disallow: /
For individual pages, Microsoft provides Bingbot-specific meta tags and equivalent X-Robots-Tag response headers. These are two independent examples of different choices, not a required pair of tags:
<meta name="Bingbot" content="ai-training:disallow">
<meta name="Bingbot" content="ai-grounding:disallowextended">
The first example declines training for a page. The second limits that page’s grounding to Microsoft-operated experiences. Publishers can instead choose ai-grounding:disallow to express a broader grounding opt-out, or leave a preference unset for the documented default of Allow. Microsoft’s official guide also specifies the corresponding HTTP header form, including X-Robots-Tag: Bingbot: ai-grounding:disallow.
There are two implementation traps. First, a page-level training preference overrides the shared training setting in robots.txt for that page. Second, these controls govern how collected content may be used, not whether Bingbot can crawl it. An Allow preference cannot make an inaccessible or crawler-blocked page available. The examples are for planning against Microsoft’s announced 2027 implementation; publishing markup now is not evidence that Bing already honours it.
The 2027 transition is mainly for sites with legacy controls
The dates matter most to publishers that previously relied on NoArchive or NoCache to restrict Microsoft’s AI use of content. Microsoft introduced those earlier controls in 2023. The October 2026 announcement does not mean every site must open Bing Webmaster Tools before March 2027.
- December 15, 2026: Temporary grounding preferences are scheduled to become available in Bing Webmaster Tools, but only for sites using
NoArchiveorNoCacheas of that date. Microsoft says other sites do not need this transition action. - February 1, 2027: Microsoft is scheduled to start applying the new training and grounding preferences when it crawls or recrawls the relevant content. A published rule is not necessarily active before discovery.
- Before March 1, 2027: Legacy-control users wishing to preserve their restrictions must publish the relevant new training preference. Those wishing to withhold content from grounding must publish the page-level grounding preference and record a temporary grounding preference in Bing Webmaster Tools.
- September 30, 2027: Any temporary Bing Webmaster Tools grounding preference expires, unless a recrawl discovers the page-level preference first. The temporary setting ends at the earlier of those events.
After March 1, Microsoft will no longer interpret NoArchive and NoCache as preferences for AI training or grounding. That statement should not be expanded into a claim that every historic use of those directives disappears. Microsoft also says it will disregard remaining legacy preferences once it detects a new training or grounding preference on the site.
The temporary grounding control is a safeguard against recrawl delays. If a qualifying publisher skips it, Microsoft says pages with old directives will default to Allow for grounding until the new page-level preference is discovered. There is no comparable temporary training setting in Bing Webmaster Tools. Publishing a new rule after March 1 may still affect a page on recrawl, but it does not provide the same uninterrupted transition protection.
Search eligibility remains separate from AI-answer visibility
Microsoft says these training and grounding preferences do not change a page’s eligibility for traditional Bing search results, assuming it still meets ordinary crawling and indexing requirements. That is narrower than claiming any publisher’s rankings or traffic will remain unchanged for other reasons. Microsoft’s guidance also states that choosing DisallowExtended or Disallow can significantly reduce visibility in AI-generated experiences.
For a publisher that wants citations in Copilot but does not want its content used for generative-model training, the documented combination is training Disallow and grounding Allow. A publisher comfortable with Microsoft-operated experiences but not its third-party grounding services could choose training Disallow and grounding DisallowExtended. These describe policy choices, not guarantees that a service will cite the publisher or send visitors.
Measurement needs the same separation. Bing Webmaster Tools provides the verified-site workflow, while its AI Performance report shows citation counts, cited pages and sampled grounding queries across supported experiences. MOCOBIN’s earlier report on Bing’s AI visibility features covers those metrics in more detail. Citation data can help describe observed visibility; it does not by itself prove when a newly published directive was discovered, whether every downstream service obeyed it, or how many users clicked through. Bing search impressions and clicks remain a different dataset.
Microsoft also notes that IndexNow can notify Bing that a URL changed, but it cannot promise a recrawl date. The preferences do not create a licensing or payment agreement.
What to decide before the rollout
For sites with no legacy NoArchive or NoCache configuration, the immediate task is a content-use decision, not a mandatory December registration. Decide separately whether the organisation permits model training, Microsoft-operated grounding and third-party grounding services. Document that choice for the relevant site sections and individual pages.
For sites using the older directives, the operational priority is higher: identify affected pages and hosts, verify the Bing Webmaster Tools property, plan both the page-level setting and temporary grounding transition, and leave time to check which URLs Bing has actually recrawled. Review the official Bing Webmaster Tools FAQ before changing live configurations, because Microsoft’s implementation details and the IETF draft may evolve ahead of February 2027.
The real decision is not simply whether to opt out of AI. It is which uses to permit, which Bing controls express that choice, and whether older restrictions can be carried through the transition without a gap.








