A CMO at a B2B software company asks a fair question in a Monday meeting: we spent a quarter restructuring our content for AI search, so is any of it working? The SEO lead opens Search Console and shows impressions and clicks, which say nothing about whether an answer engine ever quoted the company. The paid tools that sample prompts give a visibility score nobody can audit. For most of 2025 there was no first-party number to point at. That changed when Microsoft shipped an AI Performance report inside Bing Webmaster Tools, a search-engine dashboard that reports first-party data on how often a site's pages are cited as sources in AI-generated answers. It is free, it comes straight from the engine doing the citing, and it has sharp limits that most write-ups gloss over. This is how to read it, how to build a repeatable GEO loop around it, and which questions it structurally cannot answer.
What is the Bing Webmaster Tools AI Performance report?
It is a dashboard that shows how often your pages are displayed as sources in AI-generated answers across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations. Microsoft introduced it as a public preview in February 2026, and per the launch post it reports five things:
- **Total citations:** how many times your content was displayed as a source in AI answers during the selected period.
- **Average cited pages:** the average number of unique pages from your site shown as sources per day, aggregated across AI surfaces.
- **Grounding queries:** the key phrases the AI used when retrieving content that ended up referenced in answers. Microsoft states these are a sample of overall activity, not a full log.
- **Page-level citation activity:** citation counts for individual URLs.
- **Visibility trends over time:** a timeline of citation activity across supported AI surfaces.
What is a grounding query, and why does it matter more than a keyword?
A grounding query is the retrieval query an AI system generates internally to fetch documents it can cite, and it is usually not what the user typed. When someone asks Copilot a long, conversational question, the system decomposes it into shorter retrieval phrases, runs them against a search index, and writes the answer from the passages that come back. It is the same mechanism we unpacked in our piece on Google AI Mode's query fan-out, and it is why a page can be cited for phrases that appear nowhere in its title tag.
That makes grounding queries the closest thing to a keyword report for answer engines, with one important difference: they show you the retrieval vocabulary that already leads to you, which is a better editorial brief than a keyword tool's guess about what people might search. If a grounding query is a specific, commercial question your sales team hears daily and your page is cited for it, you have found a page worth protecting and extending. If your best commercial page has no grounding queries at all, the retrieval layer isn't matching it to the questions that matter.
What did the June 2026 update add: intents, topics, citation share and compare?
In June 2026 Microsoft extended the report with four features, rolling out in preview globally. Each one converts raw counts into something you can act on.
| Feature | What it does (per Microsoft) | How to use it |
|---|---|---|
| Intents | Classifies grounding queries into categories such as Informational, Commercial, Navigational, Learn and Solve, Research, Creation and Local | Check whether your citations skew informational while the revenue-bearing queries are commercial |
| Topics | Groups related grounding queries into thematic clusters, for example several solar queries into one Solar Energy topic | Plan content by topic cluster, and spot subjects where you are cited thinly. Microsoft notes labels may be broad for niche domains in preview |
| Citation Share | Percentage of citations attributed to your site out of all citations shown across all sites for the same grounding query | Find queries where you are cited but sharing the answer with several others. It does not expose competitor domains |
| Compare | Overlays a prior period, for example the last 30 days against the 30 before, or custom ranges | Tie a citation change to a specific content update or release date |
The four features added to Bing Webmaster Tools' AI Performance report in June 2026, summarised from Microsoft's announcement.
How do you turn the report into a repeatable GEO workflow?
Treat it as a weekly loop rather than a dashboard you glance at. This is our working process, not a Microsoft-prescribed method:
- **1. Verify the site and let data accrue.** Add the domain to Bing Webmaster Tools, confirm ownership, and note the start date. You need a baseline of at least one full 30-day window before Compare means anything.
- **2. Export the top grounding queries and map each to a URL.** Build a sheet with query, cited page, intent label and topic. Pages cited for queries they were never written to answer are your accidental winners.
- **3. Find the zero-citation pages.** List every commercially important page with no citations. Before rewriting anything, confirm Bing has indexed it (our Bing indexing and IndexNow guide has the checks), then look at structure: an answer-first layout with clear headings and tables is what Microsoft recommends.
- **4. Ship one change at a time.** Update a page, note the date, and use Compare 30 days later. If you change five things at once you learn nothing.
- **5. Watch Citation Share on your money queries.** A share that falls while your absolute citations rise means answers are citing more sources per query, so your slice of each answer is shrinking even as volume grows.
- **6. Reconcile against revenue.** Log Copilot or Bing referrals in your own analytics and compare against citation counts. The report tells you that you were cited, not that anyone clicked.
What does the AI Performance report not tell you?
Three limits matter, and Microsoft states them itself. First, the data does not indicate ranking, authority, or the role a page played in an individual answer: a citation could be the primary source for the answer or a footnote, and the report doesn't distinguish. Second, grounding queries are a sample, so a query missing from the list is not proof it never triggered a citation. Third, Citation Share is explicitly an observational metric, not a competitive scoreboard.
There is a fourth limit that comes from scope rather than method: the report covers Microsoft's surfaces only. It says nothing about ChatGPT, Perplexity, Gemini or Google's AI Overviews, whose retrieval indexes and citation behaviour differ, as our comparison of how each assistant cites sources explains. Use it as one instrument in a panel, alongside a fixed set of buyer prompts run across assistants, as described in our guide to tracking AI citations. And because personalised AI answers make classic rank tracking unreliable, first-party citation data like this is worth more than a third-party visibility score you can't audit.
What does Microsoft say actually earns citations?
The launch post gives publishers five directions: strengthen depth of subject expertise, improve structure with clear headings and tables, support claims with evidence and data, keep content current, and align information across formats. None of that is exotic, and all of it maps to work you can measure with the report: restructure a page, watch its grounding-query coverage widen. It also confirms that Bing respects content owners' robots.txt preferences, so crawler access decisions you make in robots.txt for AI bots apply here too.
Read the list as a diagnostic, not a checklist. A page that already has the right structure but no citations usually has an indexing or entity problem, while a page cited only for one narrow query usually lacks depth on the surrounding subtopics.
How AIBOOTSTRAPPER helps
GEO results come from building discoverability in at launch and measuring it afterwards, in that order. When we built PropLock, a blockchain-backed real estate platform for a UK property firm, the challenge was slow, manual listing creation and leads leaking to faster competitors. We shipped an AI engine that writes SEO/GEO-optimised listings and a GEO website that ranks for local intent, and the published results were 12k monthly organic visitors within 90 days, 47% more qualified viewings and 5x faster listing creation.
The Bing report is the measurement half of that discipline: it shows which of those pages answer engines actually cite and for which retrieval queries, so the next content decision is driven by data. AIBOOTSTRAPPER's process also pings IndexNow on every publish so new pages reach Bing quickly. If you want a GEO programme with a measurement loop attached, see our AI marketing services or talk to us.
Want this done for you?
Book a free strategy call and we'll show you how to build and market your business with AI.