Microsoft Copilot SEO: the citation is decided in Bing's index

Three places Copilot answers, one index behind them

Microsoft Copilot reaches people in three places a site owner should care about. Each shows sources a little differently and sends clicks under a different name, but all three take their web answers from the same place: Bing.

The consumer app now takes personal and work accounts in one place (Microsoft Support, August 2026), and since the end of August Microsoft requires everyone using it to sign in, while Copilot Search on Bing still doesn’t, per Microsoft staff quoted by Search Engine Roundtable. Copilot Search itself launched on April 4, 2025, linking “the entire sentence or passage” to its source (Bing). Since late September 2026 Bing’s search tab and the page’s own title read AI Mode instead, a change first spotted as a test and since seen everywhere. Microsoft hasn’t announced the rename, and the page’s description still calls it Copilot Search, so expect both names in reports and docs.

Underneath, Microsoft’s documents agree. Copilot’s transparency note, updated August 18, 2026: “For certain conversations where users seek information, Copilot is grounded in web search results,” and it “centers its response on high-ranking content from the web.” The work-account docs name the service: the query goes to “the Bing search service.” And Bing’s Webmaster Guidelines: “Bing and Copilot search experiences rely on the same core crawling, indexing, and ranking foundation as traditional search.”

From prompt to grounding query: what Copilot actually searches

Copilot doesn’t paste the user’s question into Bing. Microsoft’s web search documentation says the query it sends “consists of a few words informed by the user’s prompt,” and the full prompt goes only “if the prompt is very short (for example, ‘local weather’).” Those few words are what Bing calls a grounding query, and they decide which pages Copilot can cite.

1. 1

   Copilot decides the answer needs the web

   Information-seeking conversations are grounded in web search. Copilot also has a Search mode, and on work accounts IT can switch web search off for a whole tenant, while each user has a Web content toggle.

2. 2

   It writes a few-word Bing query

   Drawn from the prompt, shortened, sometimes several per answer. On work accounts, Copilot Chat’s citation section shows “the exact web search queries” it sent to Bing.

   Your move: Use the plain words such a query would contain in your titles and opening lines: the category, the task, the team size, the comparison.

3. 3

   Bing retrieves content it can ground on

   Grounding uses “the same crawlers, the same quality signals” as search, Bing wrote in May 2026, but asks a different question: “what information can an AI system responsibly use to construct a response?”

   Your move: Be indexed, canonical and fresh in Bing: a sitemap with honest lastmod, and IndexNow on every change.

4. 4

   Your robots meta tags set the depth

   noarchive keeps the page out of Copilot responses and grounding results. nocache limits Copilot to your URL, title and snippet.

   Your move: Keep both off every page you want quoted.

5. 5

   The answer links its sources

   The transparency note says Copilot “provides hyperlinked citations following generated text responses.” Since November 2025, Show all opens the full list of references.

   Your move: Write titles that still say who you are and what the page answers when they appear alone in a source list.

Why ranking for the prompt isn't the goal

In Ahrefs’ study of 15,000 long-tail prompts (July 2025), 16.6% of the URLs Copilot cited ranked in Bing’s top 10 for the prompt, 6.3% ranked lower, and 77.1% didn’t rank for it at all. That’s still the closest match to Bing of any assistant tested; Perplexity’s unranked share was 80.5% and Gemini’s 92.9%. Copilot cites pages that rank for its own short queries, not for the user’s sentence.

Here’s the shape of it, with an invented CRM brand. The queries are illustrative; the format of the report row is Bing’s.

PROMPT (what the user typed)  Which CRM is easiest to set up for a five-person sales team? GROUNDING QUERIES (a few words each, written by Copilot)  easy crm setup small team  crm pricing 5 users  pipefolk vs quotaloft AI PERFORMANCE, LATER (Bing Webmaster Tools)  Grounding query   crm pricing 5 users  Cited page        https://pipefolk.io/pricing  Citations         +1

You can read real ones in two places. AI Performance’s Grounding queries tab lists them as “grouped representations” that “are not full user questions or prompts.” And if you have a work account, Copilot Chat prints the exact queries under its answers. One agency that published three months of its own AI Performance data (19,717 citations) found the queries “read like machine-generated retrieval queries” (Search Influence, February 2026).

Bingbot is the only door in

Bing’s crawler list names five bots: Bingbot, AdIdxBot, BingPreview, MicrosoftPreview and BingVideoPreview. None of them is a Copilot fetcher. Vercel had to leave Copilot out of its December 2024 crawler study because it “lacks a unique user agent for tracking,” and Bing’s removal help describes Copilot as experiences “that rely on Bing’s index.” Block Bingbot and you leave Bing and Copilot together.

• BingbotSearch indexHonors robots.txtAllow

  Bing’s standard crawler. It feeds the index behind Bing search, Copilot, Bing’s AI summaries and its grounding partners, and renders pages with the latest stable Microsoft Edge.

• BingPreviewPreviews and adsHonors robots.txtAllow

  Generates page snapshots for Bing. Its documented user-agent strings are Bingbot’s own, so the two can’t be told apart in logs.

• MicrosoftPreviewPreviews and adsHonors robots.txtAllow

  Generates page snapshots for other Microsoft products, under its own MicrosoftPreview token.

• AdIdxBotPreviews and adsHonors robots.txtYour call

  Microsoft Advertising’s crawler, which checks ad landing pages. It matters only if you advertise, and Microsoft’s ads can appear inside Copilot answers.

One Bing quirk in robots.txt catches people: “On finding a specific set of instructions for itself, Bingbot will ignore the directives listed in the generic section” (Bing). A bingbot group replaces User-agent: * for Bing, so repeat every generic rule in it. Changes “can take a few hours” to apply because Bing caches the file.

# Bing's crawler: the only route into CopilotUser-agent: bingbotAllow: /Disallow: /cart/Disallow: /account/ # A bingbot group replaces the * group for Bing,# so every generic rule is repeated above.User-agent: *Disallow: /cart/Disallow: /account/ Sitemap: https://www.example.com/sitemap.xml

Rendering is less of a worry than with most AI engines: Bing says it updates its rendering engine “to the most recent stable version of Microsoft Edge.” Its guidelines still tell you to avoid “hiding critical content behind client-side rendering,” though, and warn that “content that cannot be reliably rendered may not be indexed or selected for grounding results.” Treat JavaScript-only text as a risk, not a blocker.

Firewalls are the more common failure. Since September 15, 2026, Cloudflare’s Block setting for AI crawlers also stops Bingbot, “search included.” Before you allowlist Bingbot anywhere, verify it: reverse-resolve the IP to a host ending in search.msn.com, then resolve that host back to the same IP, or match against bingbot.json, which Bing says to refresh daily.

# 1. Is the key sentence in the HTML Bingbot downloads?curl -s -A "Mozilla/5.0 (compatible; bingbot/2.0; +http://www.bing.com/bingbot.htm)" \  https://yoursite.com/pricing | grep -c "Starter plan" # 2. Is anything telling Copilot to stay away?curl -sI https://yoursite.com/pricing | grep -i "x-robots-tag"curl -s https://yoursite.com/pricing | grep -ioE '<meta[^>]+(robots|bingbot)[^>]*>' # 3. Is that "Bingbot" in your logs real? Expect *.search.msn.com,#    then check the host resolves back to the same IP.host 157.55.39.1 # 0 in step 1 = client-rendered text or a WAF challenge. Fix it.

noarchive, nocache and the other dials on Copilot

Robots.txt decides whether Bingbot fetches a page. What Copilot may do with a page Bing has indexed is set by Bing’s robots meta tags. Each works as a meta tag or an X-Robots-Tag header, and writing name="bingbot" instead of name="robots" scopes it to Bing alone. Bing documents a Copilot effect for most of them:

<!-- Out of Copilot answers and Microsoft AI training; stays in Bing search --><meta name="bingbot" content="noarchive"> <!-- Copilot may show only your URL, title and snippet --><meta name="bingbot" content="nocache"> <!-- Keep one passage out of snippets and AI summaries --><div data-nosnippet>Internal pricing notes</div> <!-- The same rule as a response header, for PDFs and other files -->X-Robots-Tag: bingbot: noarchive

Three details trip people up. A page with both nocache and noarchive “will be treated as NOCACHE,” so it stays in Copilot at snippet depth. noindex only works if Bingbot can read it, so “do not block access to the page” in robots.txt. And the data-nosnippet rows come from Bing’s October 2025 announcement: marked text “will be excluded from snippets and AI summaries” but is “still indexed normally.” See snippet controls for how these compare with Google’s.

Audit templates, not just pages. Paywalls, member areas and some CMS themes add noarchive site-wide, and a header set at the CDN won’t show up in your page source. Check both before assuming Copilot has simply passed you over.

IndexNow: the fast lane between your CMS and Copilot

Freshness matters more to Copilot than to a results page. “In grounding, a stale fact produces a misleading response,” Bing wrote in May 2026, and its guidelines say timely IndexNow submissions “reduce outdated or incorrect URL references in Copilot responses and grounding results.” They add a detail most setups miss: “Avoid batch submissions when possible. Streaming submissions provide faster updates.” Ping as each change happens, not in a nightly job.

1. 1Verify your site in Bing Webmaster Tools. You can import sites already verified in Google Search Console. AI Performance, URL Inspection and the IndexNow report all live there.
2. 2Submit a sitemap with honest lastmod. Bing wants ISO 8601 date and time reflecting “the true last modification time of the page content,” revisits sitemaps “at least once per day,” and takes up to 50,000 URLs per file (Bing, July 2025). List canonical URLs only.
3. 3Host an IndexNow key and ping on every add, update and delete. Many CMSs and CDNs have it built in; otherwise it’s one POST per change, shown below.
4. 4Read the IndexNow report. It shows URLs submitted, crawled and indexed, and lists Important URLs Missed: pages Bing indexed that never came through IndexNow (Bing).
5. 5Check key pages in URL Inspection. Confirm the answer is in what Bingbot fetches, not behind a tab, a click or a script.
6. 6Send freshness signals too. Bing says accurate lastmod values and HTTP validators such as ETags “help Bing detect content changes more reliably.”

IndexNow is an open protocol, not a Bing API. A submission is “automatically shared with all other participating search engines” (IndexNow), currently Bing, Yandex, Seznam, Naver, Yep, the Internet Archive and Amazonbot. Host the key as a text file at your root, then POST changes as they happen:

POST https://www.bing.com/indexnowContent-Type: application/json; charset=utf-8 {  "host": "www.example.com",  "key": "3f6c9a1e2b7d4c58",  "keyLocation": "https://www.example.com/3f6c9a1e2b7d4c58.txt",  "urlList": [    "https://www.example.com/pricing",    "https://www.example.com/blog/new-post"  ]} # Up to 10,000 URLs per POST. A single URL can go as a GET:# https://www.bing.com/indexnow?url=<url>&key=<key># 200 = OK, 202 = key validation pending, 403 = key not valid,# 422 = URL not on this host, 429 = too many requests.

Correcting or removing a page Copilot cites

Wrong facts in a Copilot answer usually trace back to a stale page. Fix the page, update lastmod, and send it through IndexNow. To take a page out entirely, return a 404 or 410 or add noindex, then ping it. Bing’s Block URLs tool hides a URL “for 90 days” while that takes effect, and its removal help covers Copilot experiences “that rely on Bing’s index.”

Writing pages Bing will ground on

Microsoft is unusually specific about what grounding prefers. Its guidelines devote a run of sections to content that is “easy to understand without external context,” and its advertising team published a companion guide in October 2025. The independent data adds a twist: Copilot’s citations lean hard toward retailers and reference sites.

What gets a site demoted

Bing’s guidelines, rewritten in February 2026, now carry a “Prompt Injection and AI Manipulation” section: “Attempts to add content designed to manipulate or interfere with language models used by Bing or Copilot may result in reduced visibility or removal from search experiences.” Content made “to manipulate ranking systems or trigger citations or AI responses,” large-scale content without oversight, and cloaking are on the same list. Bing also says plainly that “GEO does not guarantee grounding or citations in AI experiences.”

Bing's AI Performance report: Copilot's citation receipts

Most AI engines never tell you when they cite you. Microsoft does. Bing Webmaster Tools’ AI Performance report launched in public preview on February 10, 2026, gained Intents, Topics, Citation Share and Compare on June 16, and added a Topics Portfolio beta on October 13. It covers “Microsoft Copilot; AI-generated summaries in Bing; Select partner AI integrations,” pooled together.

What it never shows is traffic. A citation “does not represent traffic, clicks, or user engagement,” says the help page, and you can’t tell whether Copilot, a Bing AI summary or a partner cited you.

A monthly routine with the report

1. 1Export grounding queries and pages for the last 30 days, and keep each month’s file. Sampling makes month-to-month comparisons more useful than any single number.
2. 2Map each query to the page you want cited. Where Copilot cites a weaker page for a query, strengthen the right one and link it from the weaker one.
3. 3Work the Opportunity Queries. For each, check you have one page whose opening lines answer it in the query’s own words. If not, that’s your next brief.
4. 4Republish and ping IndexNow for every page you change, then watch whether it appears under the target query over the next few weeks.
5. 5Read Citation Share on the queries that matter commercially, not on everything. Intent labels help you find them.

Expect swings. In the Search Influence data, one page earned 69% of all citations, and daily citations fell from about 1,520 in early December to 34 by February. That’s one site, but it’s a fair warning against reading a single week as a trend.

Copilot clicks in GA4, Clarity and your logs

Copilot has passed a copilot.microsoft.com referrer since at least March 2024. As of October 2026 the web app also runs at copilot.com, which copilot.microsoft.com names as its canonical host, and both send a host-only referrer (origin-when-cross-origin and strict-origin-when-cross-origin policies). Expect either hostname as the source, never the conversation URL. Clicks from Bing’s AI Mode, Copilot Search, carry bing.com and land with your organic Bing traffic.

Check where copilot sessions land in GA4. If any sit in Referral, add a custom channel above it that matches both hosts. Clicks from the Windows and mobile apps aren’t documented, so expect some in Direct. More on reading AI referral traffic.

# Session source: Copilot, either host^(www\.)?copilot\.(microsoft\.)?com$ # Session source: the major AI assistants togethercopilot\.microsoft\.com|copilot\.com|chatgpt\.com|perplexity\.ai|claude\.ai|gemini\.google\.com

Clicks will understate Copilot. Bing’s guidelines say content “may appear as impressions, citations, or grounding references in Copilot responses,” and that “a decline in clicks does not always indicate a loss of visibility.” Read GA4 next to AI Performance, and use your logs to confirm the plumbing: a fresh Bingbot hit after an IndexNow ping means Copilot’s source of truth has your update.

# Bingbot hits per page, most-crawled firstgrep -i "bingbot" access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20 # When did Bingbot last fetch the page you just pinged?grep -i "bingbot" access.log | grep " /pricing " | tail -1

Six Copilot assumptions that cost people citations

• The assumption

  Copilot has its own crawler you need to allow.

  What's actually true

  Microsoft documents none. Copilot grounds in Bing’s index, so Bingbot is the whole story, and blocking it removes you from Bing and Copilot at once.

• The assumption

  Ranking in Bing for the question is what gets you cited.

  What's actually true

  Only 16.6% of Copilot’s cited URLs ranked in Bing’s top 10 for the original prompt in Ahrefs’ data. Copilot searches its own short queries, and AI Performance shows you which.

• The assumption

  noarchive is a harmless AI-training opt-out.

  What's actually true

  It also keeps the page out of Copilot answers and grounding results. Until Microsoft’s robots.txt no-training preference arrives, reportedly in early 2027, you can’t opt out of training and stay in Copilot.

• The assumption

  Copilot Search is gone now that Bing's tab says AI Mode.

  What's actually true

  The page still runs at bing.com/copilotsearch, and its own description still calls it Copilot Search. The label changed; the Bing grounding behind it didn’t, and its clicks still arrive as bing.com.

• The assumption

  AI Performance tells you how much traffic Copilot sends.

  What's actually true

  It counts citations, sampled and pooled across Copilot, Bing’s AI summaries and partners. Microsoft says a citation “does not represent traffic, clicks, or user engagement.” Traffic lives in GA4.

• The assumption

  Buying Microsoft ads gets you cited.

  What's actually true

  Ads in Copilot are separate placements “automatically created using your campaign’s existing assets,” and advertisers can’t opt out of them (Microsoft Learn). Nothing Microsoft publishes describes a paid route into organic citations.

The Copilot setup, in order

Everything above as one run, highest leverage first. Tick steps off as you go; your progress stays in this browser.

Sources

Every claim above links to one of these. Microsoft’s own pages come first; studies and reporting are marked as observed in the text.