Microsoft Splits AI Training From Grounding — Copilot Controls Every Brand Team Should Map Before March 2027

Microsoft’s October 31, 2026 announcement is easy to bury under the headline trio — Google, Microsoft, and OpenAI all pledged support for a shared robots.txt AI training line. Microsoft’s post goes further. It is the only vendor update that replaces controls publishers already use, splits training from grounding, and names a third-party API grounding tier (Web IQ) that Copilot users never see but answer pipelines do.

Search Engine Journal’s coverage of the shared opt-out walks the full cross-vendor picture. This article focuses on Bing, Copilot, and Bing AI Mode — where Microsoft’s rules diverge from Google’s Google-Extended bundle and from OpenAI’s GPTBot/OAI-SearchBot split.

If you manage AI visibility for B2B brands that still block NoArchive on press releases, help centers, or comparison pages, March 1, 2027 is a hard deadline.

The two preferences Microsoft wants you to use

Microsoft (Krishna Madhavan, principal product manager, Bing) defines:

Training can be overridden at the page level with a meta tag or HTTP header. Grounding is always page-level.

Grounding values:

• Allow — content may inform AI answers broadly
• Disallow — content should not inform AI-generated answers
• DisallowExtended — allow grounding in Microsoft-created experiences (Copilot, Bing AI Mode) but block grounding through Microsoft-provided grounding services used by other businesses — referred to as Web IQ in Microsoft’s materials

Microsoft states that either grounding restriction limits visibility of your brand in AI answers using artificial intelligence. Neither restriction affects classic Bing search results listings.

That last sentence is the same structural promise Google makes for Google-Extended vs Googlebot — but Microsoft gives you two levers where Google bundles training and Gemini Apps grounding under Google-Extended unless you opt out via separate Search Console controls for AI Overviews.

NoArchive and NoCache are going away for AI

Microsoft will:

• Begin honoring AI Training and AI Grounding preferences on February 1, 2027, as pages are recrawled
• Stop recognizing NoArchive and NoCache as training or grounding controls after March 1, 2027

Sites that relied on those tags to keep content out of Copilot need an explicit migration — not an assumption that “we already block archiving.”

Bing Webmaster Tools temporary preference (available from December 15, 2026 only to sites still using NoArchive/NoCache on that date) bridges the gap until recrawl applies new rules.

Microsoft’s FAQ adds two easy-to-miss footguns:

1. Pages without a new preference by March 1 default to allowing grounding until Microsoft recrawls them.
2. Once Microsoft detects either new preference anywhere on a site, it disregards all remaining NoArchive/NoCache tags on that site.

Legal teams often add NoArchive to licensing or syndication pages. SEO and AEO owners need a joint inventory before December 2026.

How this connects to the shared __usage_ai-training line

Kelley’s IETF draft uses:

User-Agent: __usage_ai-training
Disallow: /

Microsoft adopts that token for training only. Grounding is not covered by the shared line — you must set page-level grounding preferences explicitly.

Google, by contrast, plans to treat a shared training opt-out (without a separate Google-Extended directive) as a Google-Extended block — which affects both Gemini training and grounding in Gemini Apps. See Shared AI Training Opt-Out in robots.txt for the full comparison.

For multinational brands, one robots.txt file cannot express one global AI visibility policy anymore — at least not on Microsoft surfaces.

DisallowExtended: when “Copilot yes, partners no” matters

DisallowExtended is the option most GEO playbooks have not caught up to. It encodes a product split:

• In — Microsoft-owned answer surfaces (Copilot chat, Bing AI Mode)
• Out — third-party products built on Microsoft grounding infrastructure (Web IQ)

Use cases that might choose DisallowExtended:

• You want consumer Copilot to cite your product docs but do not want embedded partner assistants reusing the same passages in their own UI
• Competitive or channel conflict — your content grounds your Copilot experience through official partnership, not every reseller’s white-label bot
• Incremental rollout — allow Microsoft surfaces while you evaluate API grounding partners

Use cases that need full Disallow:

• Licensed content that must not appear in any synthesized answer
• Medical, financial, or regulated copy where any generative paraphrase is unacceptable

Microsoft is explicit: grounding restrictions reduce AI answer visibility. That is the point — but it is also why blocking grounding on /pricing while leaving /blog open is a page-type strategy, similar to question intent determines which page types AI cites.

Copilot vs ChatGPT vs Google: three different control planes

Microsoft’s split aligns conceptually with OpenAI’s GPTBot (training) vs OAI-SearchBot (search for answers) — documented in Training Crawlers vs Search Crawlers.

Copilot citation behavior still differs from ChatGPT Search in which index and fan-out queries get used — see Bing AI performance reporting and Why each AI engine reads a different search index.

Google bundles more under Google-Extended and uses Search Console for Search generative AI opt-out on AI Overviews/AI Mode — see Google AI Mode self-citation shift.

Platform-specific strategy is not optional. Only 2.7% of domains are cited by all five major engines. Control-plane changes on Microsoft alone will not synchronize mention rates on ChatGPT or Perplexity.

Measurement: grounding policy ≠ mention rate

Blocking grounding should reduce citations to your URLs in Copilot answers over time. It does not automatically remove parametric mentions of your brand from model weights — Graphite’s October 2026 work on ChatGPT brand mentions with and without search shows training-time entity strength still shapes who gets named when search is off.

After you change Microsoft preferences:

1. Track buyer prompts where Copilot is in your stack (B2B SaaS, developer tools, Microsoft-centric IT categories).
2. Separate brand mention rate from owned-site citation rate — Itera’s 1.2M citation analysis found a sevenfold gap in naming frequency when own-site share of sources crossed thresholds; see source-to-mention gap on Google AI Overviews for the same structural issue on Google surfaces.
3. Watch Bing Webmaster Tools generative reporting as it evolves — classic rank tracking will not show Copilot paraphrases.

Obsurfable records prompts, answers, brands, and citations in a public corpus so teams can compare policy changes to answer-level outcomes without guessing from crawler logs alone.

Migration playbook (Q4 2026 → Q1 2027)

By December 15, 2026

• List every URL with NoArchive or NoCache
• Classify business intent: block training only, block all grounding, block third-party grounding only (DisallowExtended), or allow both
• If still on legacy tags, set temporary WMT preferences on December 15

By February 1, 2027

• Deploy robots.txt __usage_ai-training group (when Microsoft begins honoring it) plus page-level grounding tags
• Align with legal on press releases, investor relations, and partner PDFs

By March 1, 2027

• Verify no page relies on NoArchive/NoCache for AI — Microsoft will ignore them after detection of new prefs
• Confirm defaults: undisclosed pages may allow grounding until recrawl

After migration

• Re-run 30–50 fixed Copilot-relevant prompts weekly for four weeks — grounding recrawl is not instantaneous
• Compare to ChatGPT and Google panels so you do not over-index Copilot fixes while losing share elsewhere

Bottom line

Microsoft’s announcement is the clearest vendor statement yet that AI training and AI answers are separable policy choices — with a third nuance for partner grounding APIs. The shared robots.txt training line is only one column in that matrix.

Treat March 2027 like a certificate rotation: invisible until it breaks citations. Map training vs grounding vs search bots before you simplify robots.txt for privacy — and measure answers, not just crawler allowlists.