Measurement

Microsoft will tell you your share of AI citations. It won't tell you who has the rest.

By Arnav Mukherjee, founder of TofuBofu · September 7, 2026

I've spent a lot of this year telling founders that no AI engine will show them their own citation data, so they have to go and measure it from the outside. That was true when I started saying it. Microsoft has now made it partly false, and I'd rather say so than keep repeating a line that's aged.

On 16 June 2026, four Microsoft product managers published an expansion to the AI Performance Report in Bing Webmaster Tools: Intents, Topics, Citation Share and Compare. The one that matters is Citation Share, and it's the first time an engine has handed a site owner its own slice of the citation space for free. I went looking for the catch. There is one, and the reason I trust the feature more than I expected is that Microsoft states the catch itself, on the same page, in a sentence most vendors would have buried.

So this is what the report gives you, what it can't give you, and where the two gaps sit. If you own a site and you haven't opened Bing Webmaster Tools in three years, this is the reason to.

What Citation Share actually measures

Microsoft's definition is precise, so here it is whole: "While total citation counts show how often your content appears in AI-generated answers, Citation Share shows how much of the citation space your site receives for a specific grounding query. It is calculated as the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query."

That's a share-of-voice metric with a real denominator, which is rarer than it sounds. Most AI visibility numbers, ours included, are built from the outside: we ask an engine a question and record what came back. Microsoft is reading its own logs. When it says you hold 8 percent of the citation space for a grounding query, that 8 percent is counted against every citation the system actually showed, not against a sample somebody assembled by asking.

To see how unusual that is, put it beside Google's equivalent. Google Search Console has its own generative AI performance report, and its documented dimensions are Pages, Countries, Dates and Devices. There's no query dimension in it at all, and no share of anything. Google tells you that links to your site were shown; Microsoft tells you which grounding query it happened on and what fraction of that query's citations were yours. Of the two free first-party AI reports on offer right now, Bing's is the more useful one, and I don't think many people have noticed.

The other three features are supporting cast and worth ten minutes each. Intents classifies grounding queries into categories including Informational, Commercial, Navigational, Learn and Solve, Research, Creation and Local. For a B2B firm the Commercial bucket is the one to read first, because that's the closest thing in the report to a buying question. Topics clusters related grounding queries into themes. Compare overlays a previous period on the current view. Microsoft calls all four "early preview innovations" and, writing about Topics, warns that some labels "may still be broad, especially for highly specialized or niche domains", which is a caveat worth taking at face value rather than discovering three months in.

The limitation Microsoft states itself

Here's the sentence: "Importantly, Citation Share is designed as an observational metric, not a ranking system or a competitive scoreboard. It does not expose competitor domains, represent traffic share, or assign quality scores to content."

Read what that does to the number. You learn you hold 8 percent of the citation space for a grounding query. You don't learn whether the remaining 92 percent belongs to one competitor who owns the topic, or to forty sites that each got a sliver. Those two situations demand opposite responses, and Citation Share reports them identically. The metric tells you your size and withholds the shape of the field.

I don't think that's a flaw Microsoft is hiding, and it would be cheap of me to frame it as one. It's a deliberate product decision, stated up front, and it's the honest position for a search engine that doesn't want to hand every publisher a rival-tracking console. But the consequence for you is real and it doesn't go away because the disclosure was fair: the competitive half of the question stays unanswered, and it's the half that decides what you write next.

A grounding query is not your buyer's question

This is the part I expect most people to get wrong, because the report looks like Search Console and Search Console rows are things humans typed. These rows aren't. Microsoft defines the term on the same page: "In AI-generated answers, grounding refers to the source material and web evidence the system uses to support and cite its response."

A grounding query is the search the AI system composed for itself while assembling an answer. Your buyer asked something in their own words, Copilot decided what to go and look up, and the row you're reading is that second thing. The buyer's actual question sits upstream and you never see it. So you can't take a grounding query, put it in a content brief, and call it a keyword your market uses, because there's no guarantee any human ever phrased it that way.

That doesn't make the data weak. It makes it a different kind of evidence, and arguably a more useful one: it's a window into how the machine goes looking for sources, which is exactly the thing you're optimising for. Just don't confuse it with demand. We hit the same distinction from the other side when we pulled AI Mode questions out of Google Search Console, where the AI report has no query dimension at all and the real question data lives somewhere else entirely.

Copilot doesn't speak for the other five engines

The scope line is Microsoft's own: the report covers "Microsoft Copilot, Bing, and select partner AI experiences". One engine family. If you read a healthy Citation Share and conclude you're in good shape with AI, you've generalised from one surface to six.

We can put a number on how bad that generalisation is, from our own corpus. Across 34 questions that both Google AI Mode and Bing Copilot answered, drawn from 4 reports and 3 brands, the two engines shared no cited domain at all on 12 of the 34, and mean domain overlap over the whole set was 0.065. Those are counts from a small sample and we publish them as counts, not as a rate you should extrapolate. The direction is still unambiguous: on more than a third of the questions where both engines answered, the webs they drew from didn't intersect once.

The volume gap points the same way. In that corpus AI Mode carried 683 cited sources against Copilot's 120, averaging 20 citations per answer against Copilot's 2.5, and a median of 3 among the Copilot answers that cited anything at all. Copilot cites few sources and cites different ones. A metric built on it is a real measurement of a narrow surface, which is worth having and is not a proxy for the rest. We wrote up the full paired comparison in two engines, the same questions, and almost no shared sources.

What to do with it this week

Verify your domain in Bing Webmaster Tools if you haven't, open the AI Performance Report, and filter Intents to Commercial. Those rows are the closest thing Microsoft gives you to a buying question, and your Citation Share on them is the number worth writing down. Then use Compare against the prior 30 days, because a share that's falling while your citation count holds steady means the field around you grew, and that's a signal you can't get from a raw count.

Then treat the result as one engine's answer and go get the other five, because the questions that decide a B2B deal are being asked somewhere Microsoft can't see. That's the whole argument for measuring from the outside as well as the inside: an engine's own logs are the only place your true share exists, and no engine will ever show you a competitor's. Bing just proved both halves of that sentence on the same page.

Frequently asked questions

Sources and further reading

• Microsoft: New AI Visibility Insights in Bing Webmaster Tools, by Krishna Madhavan, Meenaz Merchant, Saral Nigam and Trishna Shah, Product Managers, Microsoft AI, published 16 June 2026, read at source 7 September 2026: the announcement carrying the Citation Share definition, the scope line, and the competitor-domains limitation quoted above
• Microsoft: Introducing AI Performance in Bing Webmaster Tools, Public Preview, February 2026: the original report these four features expand
• Google Search Console Help: Generative AI performance report, read at source 29 August 2026: the counterpart report on Google's side, whose dimensions are Pages, Countries, Dates and Devices, with no query dimension and no share metric
• Our own paired citation corpus: 34 questions answered by both Google AI Mode and Bing Copilot, across 4 completed reports and 3 brands. Every one of those brands ran a scan because somebody already suspected they were missing from AI answers, so the sample is selected on the problem being measured