Reading Bing's AI Performance Report: Citations, Grounding Queries and Citation Share

Microsoft launched the AI Performance report in Bing Webmaster Tools as a public preview on 10 February 2026 and expanded it on 16 June 2026. It counts citations of your site in Microsoft Copilot, AI-generated summaries in Bing, and what Microsoft calls select partner integrations.

That scope matters for three reasons:

• It is Microsoft's surfaces only. It is not a ChatGPT report. If you want the details, see whether the report includes ChatGPT citations.
• It counts citations, not outcomes. The report shows citation activity. It does not tell you clicks, traffic or a position inside an answer.
• It has no API. As of October 2026 the data lives in the Bing Webmaster Tools interface only. Plan to export it by hand.

If you have not verified your site yet, start with setting up Bing Webmaster Tools for AI reporting. The report only shows data for verified properties.

Why bother with a Microsoft-only view? Copilot grounds its answers in the Bing index, and Bing can behave very differently from Google. In Otterly's 2026 experiment, Google deindexed two AI-generated test sites while Bing kept them, and Microsoft Copilot produced 2,718 of the 3,934 citations Otterly tracked for one of them. One engine's health tells you nothing about another's.

Total Citations

The number of times AI answers on Microsoft's surfaces cited your site in the period. Treat it as your headline volume number. It moves with how many relevant questions people asked, not only with anything you did.

Average Cited Pages and page-level activity

These show how your citations spread across your URLs. The page view is the most useful part of the report for a marketing lead: it tells you which pages Copilot actually quotes. Check the in-product description for Microsoft's exact definition of the average before you report it upward.

Grounding queries

Microsoft describes these as "the key phrases the AI used when retrieving content that was referenced in AI-generated answers." They are not what users typed. AI systems rewrite prompts before searching: OpenAI says ChatGPT search rewrites a query into one or more targeted queries, and Google's Gemini API generates its own search queries. Bing is the only operator that shows you these retrieval queries. Microsoft scopes its sampling note to this metric: the data shown "represents a sample of overall citation activity." The other metrics carry different caveats that do not mention sampling. More on that in is Bing's AI Performance data complete or sampled.

Intents and Topics

Microsoft groups activity into intents such as Informational, Commercial, Navigational, Learn and Solve, Research, Creation and Local, plus topic groupings. Use them to see whether you are cited for buying questions or only for background reading.

Citation Share and Compare

Microsoft frames citation share as "an observational metric, not a ranking system or a competitive scoreboard," and says it does not expose competitor domains, represent traffic share, or assign quality scores to content. The denominator is dominated by whoever already wins a query. So your share can fall while your absolute citations rise. The citation share definition covers the arithmetic.

1. Set the date range to the last full month. Partial months make every comparison noisy.
2. Record Total Citations and Average Cited Pages. Write both into a running sheet with the date.
3. Open the page-level view. List your top ten cited URLs. Mark which ones you actually want quoted (pricing, product, comparison pages) and which are incidental (old blog posts, help articles).
4. Read the grounding queries. Sort them into three buckets: matches what you sell, adjacent, and irrelevant or wrong. Remember this is a sample, so read it for patterns, not exact counts.
5. Check intents. Note the split between Commercial and Informational activity. A business that sells something wants presence in the commercial bucket.
6. Look at citation share last, and alongside Total Citations. If share fell but citations rose, more sites are being cited on those topics. That is not a loss.
7. Export or screenshot everything. With no API, your sheet is the only history you will have.
8. Write one action per finding. For example: "pricing page is never cited; confirm the prices are in the HTML before JavaScript runs."

There is no target number. Microsoft publishes no benchmark, and no one can promise a citation count. Read the shape of the data instead.

A sudden drop to zero is usually an access problem, not a content problem. A cited page that states an outdated price is a correctness problem: the engine is quoting you, just the wrong version. A falling share on its own is ambiguous, which is why a falling citation share needs the absolute count beside it.

Work through the access gates before you touch content. These are the levers with documented mechanisms behind them.

• Bing index status. Use URL inspection in Bing Webmaster Tools on the pages you expect to be cited. Copilot cannot cite what Bing has not indexed.
• bingbot permission in robots.txt. Confirm no blanket rule blocks it. Training-bot rules for other vendors are separate and do not affect Bing.
• Firewall and CDN responses. A robots.txt that allows a crawler means nothing if the firewall returns 403. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
• Snippet controls. Search templates for data-nosnippet, nosnippet, max-snippet:0 and noindex. Bing says content marked data-nosnippet is still indexed but excluded from snippets and AI summaries.
• Rendering. Check that prices, product names and key claims appear in the raw HTML, not only after JavaScript runs.
• Facts on the page. Make sure the cited pages state current, specific facts consistently with the rest of your site.

What you cannot control: which queries people ask, how Copilot rewrites them, and which retrieved page it decides to cite.

Run the full read once a month, on the same day, with the same date range. Monthly is frequent enough to catch a silent access failure and slow enough to avoid reacting to noise in a sampled metric.

Also re-check within a week of any of these events:

• A site migration, redesign or framework change.
• A robots.txt edit.
• A CDN, firewall or bot-management change, including vendor-side default changes you did not make.
• A pricing or product launch, so you can see whether the new page gets cited.

Pair the Bing read with Google's equivalent. The Search Console generative AI report covers AI Overviews and AI Mode, which Bing's report does not. Keep the two in separate columns. Blending them hides exactly the engine-by-engine differences you need to see.

The report is the best first-party AI data most sites can get, and it still has limits.

• It does not cover other engines. In Otterly's test, ranking in Bing produced effectively zero ChatGPT citations even with ChatGPT's crawlers visiting.
• It is not comparable with prompt trackers. Otterly reports that Microsoft's report credited one test site with 42,817 citations, against 3,934 logged by its own prompt-based tracker. The time windows may differ, so this is not a clean ratio. It does show that a chosen prompt set is a small sample. The Otterly experiment write-up covers the method and limits.
• It does not show competitors. See whether the report shows competitor citations.
• Cited is not the same as used. Research on retrieval systems shows citations are a noisy proxy for which sources shaped an answer.

If you want a quick outside view of what AI engines say for your own buyer questions, the free checker on this site runs up to three of them as a snapshot. Treat it as a complement to Bing's data, not a substitute.

Are grounding queries the same as the prompts users type into Copilot? No. Microsoft describes grounding queries as the key phrases the AI used when retrieving content it then cited. AI systems rewrite user prompts into one or more search queries before retrieving pages, so grounding queries show the retrieval input, not the user's words. That is why they are useful: no prompt tracker can observe them. Microsoft also labels this data a sample, so read it for patterns rather than exact volumes.

Does a high citation share mean I rank first in Copilot? No. Microsoft says citation share is an observational metric, not a ranking system or a competitive scoreboard. It does not expose competitor domains, represent traffic share or score content quality. No position data exists for AI answers. Read citation share next to Total Citations, and never report it to leadership as a rank.

Why does my report show zero citations? Start with access. Confirm the site is verified and indexed in Bing, that robots.txt allows bingbot, that your firewall or CDN does not return 403 to crawlers, and that templates do not carry data-nosnippet or nosnippet by accident. If all of those pass, zero may simply reflect that people are not asking questions your pages answer. No change guarantees citations.

Can I pull this data into a dashboard automatically? Not as of October 2026. The AI Performance report has no public API, and the Bing Webmaster API reference does not cover citations or grounding queries. The practical workaround is a monthly manual export or screenshot into a sheet you control. Keep the date range consistent so month-over-month comparisons mean something.

Should I change my content to match the grounding queries I see? Use them to check topical fit, not to stuff phrases. Research such as C-SEO Bench found most citation-oriented rewriting tactics ineffective or harmful, and Google says you do not need to chase every keyword variant. The defensible move is making sure the page for each real topic states the specific facts plainly and can be fetched without JavaScript.