Bing Splits AI Training From AI Grounding: The 2027 Deadline

As of November 2026, Microsoft's answer to site owners is that training and grounding are two permissions, not one. Bing will read an AI Training preference and an AI Grounding preference independently, both default to Allow when you publish nothing, and the old NoArchive and NoCache tags stop counting for either purpose after March 1, 2027. Anyone working on generative engine optimization singapore for a client site therefore has a migration task with a fixed calendar, on top of the ordinary question of whether blocking training costs visibility.

Key takeaways

• Training and grounding are separate switches. Blocking training does not remove a page from AI answers, and blocking grounding does.
• Both default to Allow. Doing nothing is a decision, and after March 1, 2027 it is the only decision the old tags leave in place.
• The transition has four dates, and the December 15, 2026 window in Bing Webmaster Tools applies only to sites that still use NoArchive or NoCache.
• Nothing here changes eligibility for traditional Bing search results, and no setting guarantees inclusion, traffic or payment.
• A published setting takes effect at the next recrawl, so on a site bought by a generative engine optimization singapore retainer the measurement starts after that crawl, not on the day the robots file is edited.

Table of contents

• What changed, in one paragraph
• The four dates your calendar needs
• The three real choices, and who should pick which
• What each setting actually covers
• The update trap: two signals that need a workaround
• What the multi-provider signal means for a Singapore site
• How to verify the change took effect
• The Topkee view on training, grounding and visible citations
• Recommended next action
• FAQ
• Sources

What changed, in one paragraph

Microsoft has been offering publisher controls since 2023 with NoArchive and NoCache tags. Those tags covered training and grounding together, and they are now being replaced. The new arrangement splits the decision: AI Training is set in robots.txt for a site or path, or at page level; AI Grounding is a page-level setting only, with no site-wide robots.txt equivalent. A publisher can now prevent model training while keeping an article eligible to be cited in a Copilot answer, which was not expressible under the old tags.

The wording of the announcement is careful on one point. These preferences govern how content is used, not whether Bingbot may crawl it. An Allow setting does not override a robots.txt rule that blocks the crawler outright.

The four dates your calendar needs

The practical reading: February 1, 2027 is when the new rules begin to apply, and March 1, 2027 is when the old ones stop protecting you. A site that does nothing keeps the defaults, which are Allow for both. A site that had a restriction under NoArchive needs to republish it, or the restriction lapses.

The September 30, 2027 date matters most for grounding. A temporary preference recorded in Bing Webmaster Tools holds until that date or until Microsoft recrawls the page and finds the new on-page setting, whichever comes first. The temporary setting exists because a recrawl between February and March 1, 2027 is not guaranteed.

The three real choices, and who should pick which

In practice a publisher faces three positions, and the right one depends on what the content is worth and who is licensing it.

The second row is the option the announcement makes possible for the first time. Under the old tags a publisher had to give up AI-answer visibility to stop training. That trade is now avoidable, and for a media site with a paywall it is probably the default answer.

The bottom row is the middle setting: DisallowExtended keeps grounding inside Microsoft's own products, Copilot among them, and leaves out the retrieval services Microsoft sells on to other applications, such as Web IQ and the grounding API in Foundry. It is a middle setting for a brand that is comfortable with the major surface but not with its content being resold as retrieval fuel to other applications.

What each setting actually covers

AI Training covers the use of an asset to modify the learned parameters of a generative model, under the IETF working definition. Training of models used only for non-generative tasks sits outside that category even when the model can generate.

AI Grounding covers content used as input to help inform a generated response. Indexing and ranking pages in traditional search results is not grounding, which is why the two settings do not touch each other.

A few limits are worth stating plainly, because they decide how much the exercise is worth:

• The preferences do not change eligibility for traditional Bing search results, subject to normal crawl and indexing requirements.
• They do not establish licensing or payment arrangements. A Disallow does not create a negotiating position by itself; it only removes the default use.
• Neither choice guarantees inclusion, traffic or payment.
• Selecting DisallowExtended or Disallow for grounding significantly reduces a site's visibility in AI-generated experiences, and Microsoft says so in the guidance.
• The current Bing search syndication API returns links for traditional search scenarios and is not affected by the grounding preference.

The update trap: two signals that need a workaround

For a site that already has pages in place, the interesting part of this change is not the three options. It is that grounding has no site-wide robots.txt switch.

Training is straightforward to publish at site scale: a robots.txt block with the user-agent __usage_ai-training and a Disallow line covers the whole domain, and a path-level rule covers a section. Training also follows the shared IETF syntax, and Microsoft extends the same preference to page level so an individual page can override the robots.txt setting.

Grounding has no such lever. You set it page by page with a meta tag or an X-Robots-Tag header, and you also record a temporary transition preference in Bing Webmaster Tools if the site used NoArchive or NoCache. That is the update pattern this article puts first: a site with ten thousand pages and no template-level control will do the grounding part badly unless the CMS template is changed, and the change is recorded in the webmaster tool before the recrawl window.

There is an asymmetry worth planning around. Once Microsoft detects a new AI Training or AI Grounding preference on a site, it disregards any NoArchive and NoCache preferences that remain there. A site that ships the training rule first and the grounding tags later can briefly run with a looser grounding position than it had before.

What the multi-provider signal means for a Singapore site

The AI Training preference is not a Bing-only format. Microsoft describes it as a shared preference expressed in robots.txt for a site or path, built on the work of the IETF AI Preferences working group.

OpenAI has said it will honour a common signal in robots.txt and aims to implement it in the coming months, while keeping the crawler controls it already provides. OpenAI's own framing of the benefit is administrative: previously a site operator identified individual crawlers and maintained separate rules for each provider, and one shared signal reduces that work.

For an SME with a single marketing site, that is the argument that matters. The robots file becomes a place where a training decision is written once. It is not a blanket permission across every engine, because each provider still has to honour it, and honouring it is voluntary.

How to verify the change took effect

Measurement comes after a recrawl, and the signal to look for is a change in which pages support AI answers, not a change in rank.

1. Record the current state in Bing Webmaster Tools before you touch robots.txt, including any pages that still rely on NoArchive or NoCache.
2. Publish the training rule and the page-level grounding tags in the same release if you want a tighter grounding position, so the old tags are not dropped first.
3. For a site using the old tags, record the temporary grounding preference in Bing Webmaster Tools between December 15, 2026 and March 1, 2027.
4. Log the recrawl date per page. A published setting is not an active one until Microsoft crawls the page and discovers it.
5. Sample the AI answers your buyers actually see, before and after, on a fixed question list and the same engine and language.

The last step is where most reports go wrong. Bing's AI Performance reporting shows citations, cited URLs and grounding queries, which tells you what Microsoft saw. It does not tell you what your buyer saw in a consumer product, and it does not tell you whether a citation was visible in the answer or only used behind it.

The Topkee view on training, grounding and visible citations

Topkee's GEO service is built around a distinction the Bing change makes unavoidable: being retrieved, being cited and being shown to a reader are different states, and a report that merges them describes none of them. The service's AI browsing analysis looks at the retrieval sources and the cited pages behind an answer, with a check on whether the site itself can be read, and its citation summaries keep retrieval sources that were not inline separate from visible citations, because a summary that counts them together cannot be called a visible citation rate.

What the service does not do is cover the training layer. Nothing in Topkee's GEO work measures what entered a model's weights, and the pricing reflects a narrower scope than a training audit. Each engagement covers one brand on one primary website, in one market and one content language, starting at an entry monthly fee quoted in HKD, with tracked questions and a monthly knowledge-content quota that differ by plan. What that fee buys is capacity, not an audit of model weights.

Two limits belong in any proposal built on this news. Topkee's GEO service does not guarantee AI recommendation, fixed exposure or a set number of citations, and its monitoring samples come from contracted API environments rather than from what every user sees in a consumer AI product. That second limit is exactly why a Bing setting change is hard to read from a visibility dashboard alone: the API sample and the consumer surface can disagree, and only site log data settles whether a crawler visit happened.

For a Singapore buyer assessing a geo agency singapore, the test of the offer is whether it can tell the three layers apart in a monthly report, and whether a generative engine optimization singapore retainer starts with a baseline taken before the setting changes. A dashboard that reports one blended citation number cannot show the effect of a grounding setting, because it never separated the retrieval source from the visible citation. Topkee's technical SEO work, which covers site architecture, load speed, mobile adaptation and robots.txt availability in its audits, is the other half of the practical question here: the crawler has to reach the page before any preference on it means anything.

One more boundary, stated because it will come up in a client call. The GEO service documentation is written for the Hong Kong market, with HKD pricing and Hong Kong contact details, and the service notes hold no Singapore-specific package. A Singapore enquiry is answered from that same offer sheet.

Recommended next action

Before February 1, 2027, do one inventory pass rather than a policy debate. List every page where the site currently relies on NoArchive or NoCache, note which CMS template prints the robots or meta tags, and decide per section whether the content is being trained on or merely cited.

The geo agency singapore service page is the place to take that inventory if you want a fixed question list and a baseline before the recrawl, since the first measurement has to be recorded before the setting changes to be worth anything.

FAQ

Does blocking AI training remove my site from Copilot answers?

No. Training and grounding are independent settings. Set AI Training to Disallow and AI Grounding to Allow, and the content stays eligible to inform generated responses while being closed to model training. Blocking grounding is the setting that reduces AI-answer visibility.

What happens if I do nothing?

Both preferences default to Allow, so the site stays open to training and grounding. If you previously relied on NoArchive or NoCache, that protection ends after March 1, 2027, because Microsoft will no longer read those tags as a training or grounding preference.

Can I set grounding once for the whole site in robots.txt?

No. There is no site-wide robots.txt grounding preference. Grounding is set page by page with a meta tag or X-Robots-Tag header, plus a temporary transition preference in Bing Webmaster Tools for sites that used the old tags. Training, by contrast, can be set for the whole site with the __usage_ai-training user-agent.

Will this change my Bing search rankings?

The preferences do not change eligibility for traditional Bing search results, subject to normal crawl and indexing requirements. They regulate the use of content in training and in generated responses, not its place in the traditional results page.

When does a new setting start working?

When Microsoft crawls the page and discovers it. Publishing a setting is not the same as activating it, so the date to record for each page is its recrawl date, not the day the tag went live. Settings published after March 1, 2027 can still take effect, but a recrawl before the temporary preference expires is not guaranteed.

Sources

1. "Microsoft Bing Separates AI Training From AI Search Grounding With New Controls", Search Engine Roundtable, 2026-10-31. https://www.seroundtable.com/microsoft-bing-ai-controls-42429.html
2. Krishna Madhavan, "Your content, your choice: clearer preferences for training and grounding", Microsoft Bing blog, October 2026. https://blogs.bing.com/search/2026/10/your-content-your-choice-clearer-preferences-for-training-and-grounding/
3. "Making it easier for websites to express AI training preferences", OpenAI, undated. https://openai.com/index/helping-websites-express-ai-training-preferences/