Bing AI Performance: the first first-party report on AI citations

For two years, “AI visibility” meant estimating. Tools ran prompt panels, counted how often a brand appeared in the answers they happened to see, and called the result share of voice. Useful — but a sample of a black box. In February 2026 Microsoft opened the box a crack: AI Performance, a report inside Bing Webmaster Tools that tells a verified site owner how often Copilot actually cited their pages. It is the first first-party AI-citation data any engine has shipped, and it changes what a GEO report can honestly claim.

What the report measures

Three things, in ascending order of usefulness.

What you see	What it means	Why it matters
Total citations	Times pages from your site were used as a source in AI answers over the period	The volume signal — is the site being drawn on at all?
Average cited pages	Daily average of unique URLs from your site that got cited	The breadth signal — one hero page or a working corpus?
Grounding queries	The internal queries Copilot generated to go find sources	The intent signal — what the model was actually trying to answer

The date picker offers 7 days, 30 days, 3 months and a custom range, with history reaching back to November 2025. The June 2026 update added intent labels, topic groupings and citation share, so the report now groups those grounding queries into themes instead of dumping a flat list.

Grounding queries are the part to read twice

Total citations is the number people screenshot. Grounding queries are the number that changes what you write.

When someone asks Copilot a messy human question, the model does not search with that question. It reformulates — often into several tighter queries — and those are what retrieval runs on. AI Performance shows you the reformulations that pulled in your pages. That is a keyword list nobody could have guessed from a keyword tool, because it is machine-authored: the phrasing a model believes will surface an answer.

Read it as an editorial brief. Recurring grounding queries with few citations are gaps — the model is asking, your corpus is not answering well enough. Recurring queries with many citations are the topics to defend and deepen.

What it can’t tell you

Four honest limits, worth stating in any client report that quotes these numbers:

• Citations are not clicks. The report counts references, not visits, so there is no AI click-through rate to compute. Referral traffic from assistants lives in your analytics, not here.
• Microsoft surfaces only. Copilot, Bing’s AI answers and selected partner experiences. ChatGPT, Perplexity, Gemini and Google’s AI Overviews are all outside it.
• No API — yet. Microsoft has said API access is on the backlog with no timeline. Until then, the export button is the integration.
• Preview-grade history. The lookback starts in late 2025, so year-over-year comparisons do not exist and won’t for a while.

A weekly rhythm that actually uses it

The report rewards routine more than analysis. What a working cadence looks like:

1. Check total citations against the previous period — trend, not absolute value. The absolute number depends on how much Copilot traffic your topic gets, which you don’t control.
2. Check average cited pages. Rising citations on a flat page count means one asset is carrying the site; that is fragility, not a win.
3. Skim the grounding queries for new phrasings. Anything recurring that you answer badly goes into the content plan this week, while the phrasing is current.
4. Export the CSV and append it to your own history. The dashboard’s window will roll; your record shouldn’t.

That fourth step is the one teams skip, and it is the one that makes the other three worth anything six months from now.

Getting it out of the dashboard

Because there is no API, every honest integration today starts with the Download button. In seo·matrix we wired that path directly: the Rankings tab takes the Bing AI Performance CSV, parses it into your project history, and from then on the citation trend sits next to the position data, feeds the weekly digest, and can be alerted on like any other signal. Same data, kept in one place, on a timeline longer than the dashboard’s own.

When Microsoft ships the API, that import becomes a connector and nothing else changes — which is exactly why it is worth starting the history now rather than waiting for it.

In short

• AI Performance is the first first-party AI-citation report from a search engine — provenance, not estimation.
• It gives total citations, average cited pages, and the grounding queries Copilot used to find you.
• Grounding queries are the editorial gold: machine-authored phrasings that reveal what the model was trying to answer.
• It cannot tell you clicks, competitors, or anything outside Microsoft’s surfaces.
• There is no API yet; export the CSV into your own history so the record outlives the dashboard’s window.

Sources

• Search Engine Land, “Bing Webmaster Tools officially adds AI Performance report” — searchengineland.com/bing-webmaster-tools-ai-performance-report-468751: the February 2026 rollout, metric definitions and Microsoft’s stated use cases.
• CXL, “Tracking AI visibility with Bing Webmaster AI Performance reporting” — cxl.com/blog/bing-ai-visibility-tracking-webmaster-tools: date-range behaviour, lookback limits and the citations-are-not-clicks caveat.
• Otterly.ai, “Bing Webmaster Tools AI Performance report” — otterly.ai/blog/bing-webmaster-tools-ai-performance-report: grounding-query analysis and the June 2026 expansion (intent labels, topic groups, citation share).
• Related: AI visibility metrics: how to measure GEO in 2026 and AI share of voice.