How Microsoft Copilot chooses sources: what Microsoft documents, and what it does not

The short answer. Copilot answers that cite the web are grounded in Bing’s index, so to be eligible as a source a page must first be crawlable by Bingbot and indexed by Bing. Microsoft says Bing and Copilot “rely on the same core crawling, indexing, and ranking foundation as traditional search”, and documents three page-level controls that change what Copilot may use: noindex, noarchive and nocache. It recommends IndexNow to keep the index current. It does not document how Copilot chooses or ranks the sources it cites.

Everything below comes from Microsoft’s Bing Webmaster Guidelines, its pages on robots meta tags and Bing’s crawlers, and three posts on the Bing blogs, as read on 15 October 2026.

Where does Copilot get its sources?

From Bing. The guidelines open by saying they describe how Bing “discovers, crawls, indexes, evaluates, and surfaces content across Bing search experiences, Copilot, and grounding API results”, and that following them helps URLs stay eligible for “grounding results and citations”. Microsoft calls the step that connects a model to the web “grounding”: in its own words, “the system that connects AI to current, authoritative information”, and it says Microsoft grounding “powers nearly every major AI assistant in the market”. It does not name those assistants.

The practical consequence is simple: there is no separate Copilot index to submit to. If a page is not in Bing’s index, Microsoft’s documentation gives no other route into a Copilot answer.

Which Microsoft crawler matters?

Bingbot. Microsoft describes it as “our standard crawler” that “handles most of our crawling needs each day”. The crawler page also lists AdIdxBot (Bing Ads quality control), BingPreview, MicrosoftPreview (page snapshots for Microsoft products) and BingVideoPreview. None of them is described as a Copilot-specific crawler, and Microsoft lists no separate robots.txt token for Copilot or for model training.

Two points from that page are worth acting on:

1. Verify, do not trust the user agent. Microsoft says “user agent strings are easy to spoof” and points to its Verify Bingbot tool for checking the IP.
2. Do not block it. The guidelines list “Blocking Bingbot in your robots.txt file” among things to avoid, and warn that “content that cannot be reliably rendered may not be indexed or selected for grounding results”. Our Bingbot reference page has the documented user-agent strings.

How do noindex, noarchive and nocache affect Copilot?

This is the one area where Microsoft is specific. The guidelines say “robots.txt controls crawl access, not indexing”, and then set out what each meta directive does.

Directive	What Microsoft says it does for Copilot	Bing search results
noindex	Use it “when a URL should NOT appear in Bing search, Copilot experiences, or grounding API results”	Removed
noarchive	“Prevents content from being used in Copilot responses and grounding results”; the meta tag page adds “Do not link in Chat and Copilot”	Still shown, per the 2023 announcement
nocache	“Limits Copilot to using only the URL, title, and snippet, reducing citation depth and answer quality”	Still shown, per the 2023 announcement
Neither	The default: “No action is needed to remain in Bing Chat”	Shown

Three details matter:

• Both tags means nocache. The meta tag page says: “If content has both NOCACHE and NOARCHIVE tags, we will treat it as NOCACHE.” So adding nocache to a page that already has noarchive loosens the restriction rather than tightening it. The 2023 announcement recommends exactly that for paywalled articles, “since many paywall sites use only the NOARCHIVE tag”.
• Training is tied to the same tags. Per the meta tag page, noarchive content is not used “for training Microsoft’s generative AI foundation models”, and for nocache content only “URLs, Titles and Snippets” are. There is no separate training opt-out token in these pages.
• Snippet controls cost citation quality. The guidelines say nosnippet and data-nosnippet “may limit Copilot citation quality”, and that to control the text Bing may display or cite you can use the data-snippet attribute.

Each directive can also be sent as an X-Robots-Tag header, and name="bingbot" in place of name="robots" limits a meta tag to Bing. Remember that Bing has to crawl a page to see any of these tags: a robots.txt Disallow hides them.

1. 1

   Discovered

   Found through IndexNow, XML sitemaps and crawlable links.

2. 2

   Crawled and rendered

   Bingbot fetches and renders the page.

   Stopped by: a robots.txt block, or content that cannot be reliably rendered

3. 3

   Indexed

   In the Bing index that Copilot is grounded in.

   Stopped by: noindex

4. 4

   Grounded and cited

   Used as a source in a Copilot answer.

   Not documented: how sources are chosen. noarchive excludes; nocache limits to URL, title and snippet

Measured separately

Bing Webmaster Tools’ AI Performance report counts citations after the fact.

What does Microsoft ask site owners to do?

The guidelines are long, but the parts that mention Copilot or grounding fall into four groups.

1. Keep the index current with IndexNow. Microsoft asks site owners to notify Bing when URLs are added, updated or removed, saying timely notifications “reduce outdated or incorrect URL references in Copilot responses and grounding results”, and that streaming submissions are better than batches. The AI Performance announcement repeats the advice.
2. Keep sitemaps and URLs clean. Canonical URLs only, accurate lastmod, 301 redirects for permanent moves, and one URL per piece of content. Microsoft says duplicate URLs “reduce Bing’s confidence in selecting a URL for grounding results or citations”.
3. Write pages that stand on their own. Microsoft says URLs “are more likely to be selected for grounding queries and citations when content stands on its own”, and asks for explicit facts and definitions, one primary topic per URL, key information near the top, and consistent names for people, organisations and products.
4. Use structured data honestly. It “may support clearer grounding” but “does not guarantee visibility or grounding traffic”, and must match visible content.

The guidelines also name a Copilot-specific abuse: content “designed to manipulate or interfere with language models used by Bing or Copilot” may lead to “reduced visibility or removal”.

What does Microsoft not document?

• Selection and ranking. The guidelines describe what makes a URL “more likely” to be selected, but not how Copilot chooses between eligible sources, how many it cites or in what order. They are explicit that “GEO does not guarantee grounding or citations in AI experiences”.
• Which partners use Bing grounding. Microsoft says its grounding powers “nearly every major AI assistant” and the AI Performance report covers “select partner integrations”, but neither page names them.
• Special files. Nothing in these pages asks for llms.txt or a Copilot-specific sitemap or token.

What Microsoft does offer is a first-party report. The AI Performance report counts how often your pages are “displayed as sources in AI-generated answers” across “Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations”, with a sample of grounding queries. It is worth reading with that sample caveat in mind. In our own case studies it can move the other way from Google: on Bagasse Packaging UK, Bing AI citations rose “from 40 in May to 188 in September” while Google AI-feature impressions fell.

For the answers themselves, measure from the outside: the sampling method works for Copilot the same way as for the engines in our Lab.

A checklist

1. robots.txt does not block Bingbot (check the live file), and your firewall lets verified Bingbot requests through.
2. Important content is in the server HTML, not only rendered client-side.
3. The site is verified in Bing Webmaster Tools, with a sitemap of canonical URLs and accurate lastmod values.
4. IndexNow pings go out when pages are added, updated or removed.
5. You have audited noindex, noarchive, nocache and nosnippet on pages you want cited, including X-Robots-Tag headers and bingbot meta tags, and you know that noarchive plus nocache behaves as nocache.
6. Key pages state their facts explicitly, near the top, with one topic per URL and consistent entity names.
7. The AI Performance report is checked monthly, alongside a sampled prompt set that includes Copilot.
8. Calendar a quarterly re-read of the guidelines and the meta tag page.

Our AI crawler access tool fetches a URL as Bingbot and reports what it receives.

What we don’t claim

• We do not claim to know how Copilot selects or ranks sources. Microsoft does not document it in the pages linked here.
• We do not claim that following the guidelines leads to citations; Microsoft itself says it does not.
• We do not claim the details are permanent. The AI Performance report is a public preview and vendor documentation changes without notice; re-read the sources before relying on them.