Bing Webmaster Tools now includes an AI Performance report showing which of your pages are cited in Microsoft Copilot answers: and, more valuably, the "grounding queries" Copilot generated internally to find you. No other AI platform gives you first-party access to its internal fan-out. The data is sampled rather than complete and covers Copilot only, but it is free, it is real, and almost nobody in the industry is using it.
Key Takeaways
- Copilot grounds on the Bing index, which is why Bing Webmaster Tools can report on it at all.
- The AI Performance report gives first-party citation data: which of your pages Copilot cited in its answers.
- Grounding queries are the internal search strings Copilot generated to find your content: the model's own fan-out, exposed.
- This is the only free, first-party window into an LLM's internal query expansion that exists anywhere. ChatGPT, Claude and Perplexity offer no equivalent.
- The data is sampled, not complete, and Copilot-only. Treat it as a directional sample, not a census.
- Pair it with IndexNow for fast Bing indexation, so new and updated pages become groundable quickly.
- Nearly all published Copilot content is generic filler. The grounding-query layer is genuinely under-exploited.
Why This Report Matters More Than Its Traffic Share
The reflexive objection is fair: Copilot's usage share is smaller than ChatGPT's, so why invest attention here?
Because the value is not primarily the Copilot traffic. It is the diagnostic.
Every modern AI search system uses query fan-out: expanding one user question into multiple internal sub-queries, retrieving for each, and synthesising. Google does it in AI Mode and AI Overviews. ChatGPT does it. Perplexity does it. Every one of them treats the internal query set as invisible plumbing.
Bing shows you theirs.
That means you can see how an LLM reformulates questions in your category: what vocabulary it reaches for, how specific it gets, what angles it splits a broad question into. Those patterns are not identical across platforms, but they are far from random. The fan-out behaviour Copilot exposes is the closest thing to a readable proxy for how the others are likely thinking about your topics.
You are getting a free sample of a signal every other platform charges nothing for and shows nobody.
What the Report Shows and What It Does Not
Being clear about the boundaries is what makes this data usable rather than misleading.
| What you get | What you do not get |
|---|---|
| Which of your pages Copilot cited | Citation data for ChatGPT, Claude, Gemini or Perplexity |
| The grounding queries Copilot generated to find you | The original user question that triggered them |
| First-party data, direct from Microsoft | A complete census, the data is sampled |
| Free access with a verified Bing Webmaster Tools account | Real-time reporting; there is processing lag |
| Page-level citation visibility | Whether the citation drove a click or a conversion |
| Evidence of how one LLM expands queries | Guaranteed transferability of those patterns to other engines |
| Signal on which topics you are groundable for | Ranking-position data inside the AI answer |
The two limits that matter most: sampled, so absolute counts are not reliable and proportions are more trustworthy than totals; and Copilot-only, so cross-platform inferences are inferences and should be labelled as such when you present them.
Setting It Up
Twenty minutes, most of which is waiting for verification.
Step one: verify your site in Bing Webmaster Tools
Go to Bing Webmaster Tools and add your site. The fastest route for most people is importing from Google Search Console, which carries verification across in a couple of clicks. Otherwise use DNS or file verification as normal.
Step two: confirm Bingbot can actually reach you
Same edge-layer problem that affects every other crawler. Check robots.txt, then check your CDN and WAF rules. Bing's own crawl diagnostics will tell you if requests are failing.
Step three: submit a sitemap
Standard, but skipped surprisingly often by teams who did it for Google years ago and never repeated it here.
Step four: turn on IndexNow
IndexNow lets you ping search engines the moment a URL is published or updated, rather than waiting for a scheduled recrawl. Bing supports it directly, and most major CMS platforms and CDNs have a plugin or native integration.
For AI grounding this matters more than it does for classic SEO. If Copilot grounds on the Bing index, then the delay between publishing and being indexed is the delay before you are groundable at all. IndexNow compresses that window.
Step five: wait, then check AI Performance
Data accumulates over time and lags. Do not draw conclusions from the first week.
How to Read Grounding Queries
Once you have a few weeks of data, this is where the actual work is.
Sort by frequency first
The queries appearing most often are the reformulations Copilot reaches for repeatedly in your space. These are your highest-leverage phrasings, the language the model itself uses when thinking about your topic.
Compare against your own keyword list
This comparison is where the surprises are. In every account I have looked at, a meaningful share of grounding queries use phrasing nobody on the marketing team had in their keyword research. The model's vocabulary and the marketer's vocabulary are not the same vocabulary.
Those unmatched queries are content gaps handed to you by the system itself.
Look at specificity
Note how narrow the grounding queries are relative to what a user would plausibly have typed. Fan-out generally decomposes broad questions into much more specific sub-questions. Seeing the actual granularity in your category tells you how deep your content needs to go to be retrieved.
Map queries to cited pages
Which pages are being found by which queries? Two useful patterns emerge. A page cited across many distinct grounding queries is doing heavy lifting and deserves investment. A grounding query that finds a weak or tangential page is a signal to build the page that should have been found instead.
Watch for questions you do not answer
Grounding queries touching your category where your cited page is a poor fit are the clearest content briefs you will ever get. The demand is demonstrated, the phrasing is given, and you know the model is already looking.
Using Copilot Data as a Proxy for Other Engines
The honest framing, since this is where overclaiming starts.
What transfers reasonably well
The shape of fan-out: that broad questions decompose into specific sub-questions, roughly how many, roughly how specific. That is an architectural property of the approach, not a Bing quirk, and it holds across systems.
The vocabulary around your topic: the terms a language model reaches for when reformulating questions in your category are likely to be similar across models trained on broadly similar corpora.
What does not transfer
The specific query strings. Different systems, different fan-out implementations, different retrieval. Do not build a keyword strategy that assumes ChatGPT generates the same internal queries Copilot does.
Ranking outcomes. Copilot grounds on Bing; the others do not. Being groundable in Bing tells you nothing about being retrievable in Perplexity's curated index or Claude's Brave-backed search.
How I would actually use it
As a hypothesis generator, not a measurement system. Grounding queries tell you what to test and what to write. They do not tell you your cross-platform performance. Anyone presenting Copilot grounding data as an AI visibility score is stretching it well past what the data supports.
Why Almost Nobody Is Doing This
Three reasons, and none of them are good ones.
Copilot has low mindshare. The AEO conversation is dominated by ChatGPT, so a Microsoft report gets skipped by default.
The report is new and lightly covered. Most published content on Copilot optimisation is recycled general AEO advice with the platform name swapped in. Actual walkthroughs of the grounding-query data are rare.
It requires setup and patience. You have to verify a property you may not otherwise use and then wait weeks for meaningful data. That loses to anything that produces a chart today.
The result is a free first-party data source with almost no competitive attention on it. That combination does not last long once people notice, which is a reason to start now rather than later.
A Four-Week Plan
Week one: setup
Verify in Bing Webmaster Tools, confirm Bingbot access at every layer, submit sitemaps, enable IndexNow. Then leave it alone.
Week two: baseline
Check AI Performance. Export what exists. Do not act yet. You are establishing a baseline, and early data is noisy.
Week three: analysis
Pull grounding queries, sort by frequency, map to cited pages, and compare against your existing keyword research. Flag every query using phrasing you had not considered. That list is your output.
Week four: act and instrument
Build or improve pages against the highest-frequency unmatched queries. Ping them through IndexNow on publish. Set a monthly recurring check so this becomes a habit rather than a one-off project.
Frequently Asked Questions
What are grounding queries?
The internal search strings Copilot generates to find sources for an answer, its own query fan-out. Bing Webmaster Tools exposes a sample of the ones that surfaced your pages.
Is this data available for ChatGPT or Claude?
No. Neither OpenAI nor Anthropic publishes an equivalent first-party report. Bing's is currently unique.
Is the AI Performance report free?
Yes. It requires only a verified Bing Webmaster Tools account.
Is the data complete?
No, it is sampled. Treat proportions and patterns as informative and absolute counts as approximate.
Does the report show the original user question?
No. You see the grounding queries Copilot generated, not the prompt that triggered them.
Does Copilot use the Bing index?
Yes, that grounding relationship is what makes this reporting possible at all.
What is IndexNow and why does it matter here?
IndexNow pushes a notification to search engines when a URL is published or updated, instead of waiting for a scheduled recrawl. Because Copilot grounds on the Bing index, faster indexation means faster groundability.
Can I use Copilot grounding queries to optimise for ChatGPT?
Partially and carefully. The general shape and vocabulary of fan-out are informative across systems; the specific query strings are not transferable. Use it to generate hypotheses, not to report cross-platform performance.
How long before I see useful data?
Expect a few weeks. Data accumulates and lags, and a single week is too noisy to interpret.
What should I do first with the data?
Sort grounding queries by frequency, then find the ones whose phrasing does not appear in your existing keyword research. Those are content gaps the system has identified for you.
If you want to go through your own AI Performance data together: or figure out whether the Copilot signal is worth building into your reporting at all, I would be glad to help. I have spent 4+ years in marketing helping edtech and startup brands grow organically, including the work behind Masai School's Instagram going from 26K to 117K and LinkedIn from 50K to 160K. There is more of that work, along with a contact form, at younusfardeen.com. Reach out whenever it suits.