Direct answer
On October 30, 2026, Microsoft Bing announced two independent content-use preferences. AI Training controls whether content may be used to modify the learned parameters of generative AI models, while AI Grounding controls whether content may be used as input when generated answers are created. A publisher can disallow training while allowing grounding to preserve eligibility for citations, or use DisallowExtended to permit grounding in Microsoft-operated experiences while excluding Microsoft grounding services provided to third parties. Both preferences default to Allow, do not change eligibility for traditional Bing search results, and do not control crawl access. Microsoft will begin honoring them as pages are recrawled from February 1, 2027, with a specific transition for sites that use NoArchive or NoCache.
Facts and background
Microsoft Bing announced on October 30, 2026 that web publishers can set independent preferences for generative AI training and for grounding generated answers. The announcement covers web content collected by Microsoft crawlers and used for future model training or web-grounded answers in Bing, Copilot and Microsoft 365. The preferences do not remove eligibility for traditional Bing search results, although normal crawling, indexing, quality and ranking requirements still apply.
The practical change is a separation of four controls: crawl access, AI training use, AI grounding use and traditional search eligibility. A publisher can now choose “do not train, but allow grounding” instead of treating AI model use and answer visibility as one decision.
Keep four control planes separate
| Control plane | Purpose | Covered by the new preferences? | What it does not prove |
|---|---|---|---|
| Crawl access | Whether Bingbot can request a URL and read page signals | No; use Bingbot robots rules, access control and HTTP responses | Crawl access is not training or grounding permission |
| AI Training | Whether an asset may modify learned parameters of a generative AI model | Yes, at site/path or page level | It does not control traditional search or automatically change grounding |
| AI Grounding | Whether content may inform a generated answer | Yes, at page level with Allow, DisallowExtended or Disallow | Allow does not guarantee a citation, traffic or payment |
| Traditional Bing Search | Crawling, indexing and ranking eligibility in ordinary results | Not removed by the new content-use preferences | Eligibility is not a guarantee of indexing or ranking |
Both content-use preferences default to Allow when no applicable setting is published. Microsoft cautions that DisallowExtended or Disallow for grounding will significantly reduce visibility in AI-generated experiences without affecting Bing search eligibility. No choice guarantees inclusion, traffic or payment, and the preferences do not create a licensing or payment arrangement.
Configure AI Training and AI Grounding independently
AI Training uses the shared robots.txt purpose token __usage_ai-training. It reuses ordinary Allow and Disallow path semantics rather than creating a new directive. Publishers can express a site-wide or path-level preference and may also use a Bingbot meta tag or X-Robots-Tag for a page. Microsoft says a page-level training preference overrides the robots.txt training preference.
# Disallow generative AI training site-wide
User-Agent: __usage_ai-training
Disallow: /
# Or scope the preference to a path
User-Agent: __usage_ai-training
Disallow: /licensed-research/
Allow: /press-kit/
A page-level training restriction can use <meta name="Bingbot" content="ai-training:disallow">. Non-HTML assets and centralized edge policies can use X-Robots-Tag: Bingbot: ai-training:disallow. If no setting applies, training defaults to Allow.
AI Grounding has no site-wide robots.txt preference. It is page-level through a meta tag or X-Robots-Tag. Allow permits content to inform generated answers. DisallowExtended limits grounding to Microsoft-operated experiences and excludes Microsoft services provided to third parties, including Web IQ and Foundry grounding API access. Disallow prevents grounding use.
<!-- Prevent this page from grounding generated answers -->
<meta name="Bingbot" content="ai-grounding:disallow">
<!-- Allow only Microsoft-operated experiences -->
<meta name="Bingbot" content="ai-grounding:disallowextended">
A page can combine the two preferences, for example disallowing training while allowing grounding. Teams should use Microsoft's current syntax and must not substitute crawler names, Content Signals or directives from other providers.
Choose by content value instead of applying one site-wide rule
| Content tier | Typical goal | Possible combination | Main tradeoff |
|---|---|---|---|
| Public product and service facts | Search discovery, AI citations and accurate source links | Set training by policy; usually keep grounding Allow | Grounding eligibility still does not guarantee a citation |
| Public brand analysis | Remain citable without allowing training | AI Training Disallow + AI Grounding Allow | Content can still inform generated answers |
| Licensed or paid summaries | Keep Microsoft-operated visibility while limiting third-party services | AI Training Disallow + AI Grounding DisallowExtended | Visibility in third-party AI experiences falls |
| Public pages excluded from generated answers | Retain traditional search eligibility without grounding | AI Grounding Disallow; decide training separately | AI-answer visibility and citations decline significantly |
| Private or unpublished content | Prevent unauthorized access | Authentication, network control and correct status codes | Usage preferences are not security controls |
Global brands should not collapse legal rights, distribution policy, search discovery and AI visibility into one site-wide choice. Classify content by rights and business purpose, then define the training and grounding defaults for each tier.
NoArchive and NoCache have a dated transition
| Date | Microsoft plan | Publisher action |
|---|---|---|
| December 15, 2026 | A temporary grounding transition preference opens in Bing Webmaster Tools only for sites using NoArchive or NoCache on that date | Inventory affected sites and pages and prepare page-level settings |
| February 1, 2027 | Microsoft begins honoring the new preferences as content is recrawled | Verify live responses, logs and page source |
| March 1, 2027 | Critical deadline for maintaining old grounding restrictions; NoArchive and NoCache stop serving as training or grounding preferences | Publish page-level grounding settings and, if eligible, record the temporary BWT preference |
| September 30, 2027 | The temporary grounding preference expires, or is replaced earlier when Microsoft recrawls a page and discovers its new setting | Do not rely on the temporary preference as a permanent substitute for recrawl |
When Microsoft detects a new AI Training or AI Grounding preference on a site, it will disregard remaining NoArchive and NoCache preferences on that site. There is no temporary Bing Webmaster Tools training setting. The temporary grounding path applies only to sites using the old tags as of December 15, 2026.
Publishing a preference is not the same as activation
Microsoft must crawl a page to discover and honor its new preference. A template change, edge-header update or robots.txt deployment proves only that the setting is live. IndexNow can notify participating search engines that a URL changed and improve the chance of a recrawl, but a successful response means the URLs were received; it does not guarantee a crawl date or indexing.
Keep evidence from the live HTML or HTTP header, the robots.txt response, Bingbot logs or URL inspection, and Bing Webmaster Tools AI reporting. For CDN or market-specific rules, test multiple regions so one canonical URL does not return conflicting preferences by geography, device or cache layer.
Measure the effect in layers
Bing Webmaster Tools AI Performance reports visible citations, cited pages and grounding queries at an aggregate level. Microsoft says these metrics are not ranking, authority, clicks, traffic or the page's role in an individual response. Demand, content changes, system updates and partner refresh cycles can all move the trend.
| Evidence layer | Record | It can establish | It cannot establish |
|---|---|---|---|
| Configuration | Page, path, value, deployment version and header | The preference is published correctly | Microsoft has crawled it |
| Discovery | Bingbot logs, URL inspection, IndexNow submission and timestamps | Crawl or update signals occurred | Every Microsoft system has activated the setting |
| AI visibility | AI Performance citations, pages, topics and grounding queries | Aggregate citation movement in supported experiences | Ranking, clicks, revenue or causality |
| Owned-site outcome | Identifiable referrals, landing behavior, leads and conversions | The post-click business path | Unclicked AI exposure or third-party grounding use |
Build a pre-change baseline, record the first confirmed Bing recrawl, and compare like-for-like page tiers after deployment. A published tag without recrawl evidence is not a valid experiment start, and a citation trend without a control or time boundary is not proof of causal impact.
Business impact
The update gives global brands a Microsoft-specific way to separate “do not train” from “remain eligible for AI-answer citations.” Sites that rely on NoArchive or NoCache must migrate because those signals stop serving as training and grounding preferences after March 1, 2027. For publishers with product documentation, research, knowledge bases and licensed content, this becomes joint governance across content rights, distribution and brand visibility. Page-level grounding also creates an operational scaling challenge. Language, market, content type and rights state must map consistently to page markup or response headers. Conflicting CMS, CDN and template rules can either broaden use unintentionally or suppress AI visibility.
Zhihe Growth assessment
Zhihe Growth sees the separation of training, grounding, crawl access and traditional search as the most important part of the announcement. Treat these signals as content-use governance, not a ranking tactic. Public facts, FAQs and evidence pages generally benefit from citation eligibility; licensed research may need tighter distribution; private content still requires real access control. Use a five-step model: content classification, policy mapping, machine configuration, recrawl evidence and layered measurement. Build a URL-level ledger, generate robots, meta and X-Robots-Tag settings, verify discovery through Bingbot evidence, and then measure AI citations separately from owned-site outcomes. Default Allow is not a citation guarantee, and DisallowExtended is not a license or payment mechanism.
Recommended actions
- Inventory current NoArchive, NoCache, noindex, nosnippet, data-nosnippet and Bingbot robots rules by URL and template.
- Classify public product facts, brand knowledge, licensed content, paid content and private content, then assign separate training and grounding goals.
- Configure User-Agent: __usage_ai-training with Allow and Disallow for sites or paths that need a training preference.
- Use page-level Bingbot ai-training settings for exceptions and verify their precedence over the robots training policy.
- Select DisallowExtended or Disallow for each page that needs a grounding restriction; do not invent a site-wide grounding robots rule.
- Keep Bingbot able to read pages that carry meta or X-Robots-Tag preferences; protect private content with authentication and access control.
- If a site uses NoArchive or NoCache on December 15, 2026, prepare the temporary Bing Webmaster Tools grounding preference and complete page migration by March 1, 2027.
- Submit changed URLs through IndexNow and record submission, crawl and response evidence without treating HTTP 200 as a crawl or index result.
- Test HTML, PDFs, images and API output separately, including CDN cache and regional response consistency.
- Measure Bing AI Performance, server logs, identifiable referrals and owned-site outcomes in layers after a verified recrawl and baseline.
Limitations
This article uses Microsoft Bing, Bing Webmaster Tools, IndexNow and IETF material accessible on October 31, 2026. Microsoft plans to begin honoring the preferences as pages are crawled from February 1, 2027, and part of the Bing Webmaster Tools transition does not open until December 15, 2026. This is therefore a preparation and migration framework, not a claim that every page is already processed under the new rules. The __usage_ai-training token is documented in an Internet-Draft dated October 30, 2026. The IETF page states that it is an active individual draft with no formal standing or IETF endorsement, not an RFC. It may change, be replaced or become obsolete. Microsoft has announced adoption of the current syntax; support by other providers must be verified separately. Content-use preferences are not access control, copyright licenses, contracts, payment mechanisms or legal advice. Allow does not guarantee crawling, indexing, ranking, citation, traffic or payment. Disallow and DisallowExtended do not replace authentication, network isolation, removal or legal enforcement. Global brands should review policy jointly across content, legal, SEO, data and infrastructure teams.