Back to Blog
Measurement 7 min read

Microsoft Clarity AI Citations: What Grounding Queries Do and Do Not Tell You

Clarity has started splitting AI citation data into branded and non-branded queries. It is a genuinely useful addition, and it sits on top of a distinction that is easy to misread in a way that will flatter your reporting.

Paul Byrne August 2026 7 min read
The short version: A grounding query is a search the AI system ran to retrieve your content, and Microsoft says these may differ from the exact phrases users type. So a grounding query marked "branded" tells you your name appeared in the retrieval step. On its own it does not establish that a person asked for you. Of the four headline numbers on the Clarity citation dashboard, three describe machine behaviour and one describes human behaviour. Knowing which is which changes what you can honestly report.

What Microsoft actually shipped

On 3 August 2026, Microsoft Clarity added branded and non-branded segmentation to its AI Citations dashboard. Individual queries are now marked as branded in the queries view, the Share of Authority card breaks results out by branded and non-branded, and the whole dashboard can be filtered either way.

Microsoft frames the purpose as separating "direct brand visibility" from "category discovery", which is a fair description of the problem it solves. Before this, a rising citation count told you that your pages were being referenced more often without indicating whether that was happening on brand-specific retrieval or on generic category retrieval. Those are different commercial situations and they deserve different responses.

The dashboard needs domain verification through the Clarity tracking code, Google Search Console or Bing Webmaster Tools, which is a one-time setup by a project administrator.

What a grounding query is

Everything else on the dashboard depends on getting this right. Microsoft's documentation defines it plainly:

"The Grounding queries table displays the queries that AI systems used to retrieve your content before generating an answer. These queries may differ from the exact phrases users type, but they provide insight into how AI systems discover and map your content."

The second sentence is the one that matters. The table records the retrieval step, and Microsoft is explicit that it may not match what the user typed. Assistants commonly compose their own searches rather than passing a question straight through, so the two often will not match.

That makes grounding queries genuinely valuable. They show you how the retrieval layer thinks about your subject, which is information you cannot get from your own analytics or from a rank tracker. We have written before about how ChatGPT rewrites your question into its own searches, and this is the same mechanism surfaced by a different vendor.

It also makes them an unreliable proxy for demand. If a grounding query reads "SearchIntel AI visibility pricing", that string is what the retrieval system searched for. Whether a person used those words, or anything like them, is not something the dashboard reports.

Branded does not mean somebody asked for you

Microsoft does not publish how a query gets classified as branded, so what follows is reasoning rather than documented mechanics. Two quite different situations would each produce a branded grounding query.

In the first, a person typed your brand name and it carried through into retrieval. The branded grounding query would then reflect real brand-level intent, which is demand and worth counting.

In the second, a person asked a category question containing no brand at all, and your name entered at the retrieval stage because the system already associates you with that category. The grounding query would be branded. The question would not have been. That is not brand demand, though it is a useful signal that you are already in the consideration set.

The dashboard reports the retrieval query rather than the prompt behind it, so it does not separate these two cases, and Microsoft does not claim it does. The risk sits in the reporting slide that turns "branded grounding queries up 40 per cent" into "brand demand up 40 per cent".

The four numbers, and what each one rests on

The citation dashboard gives you four headline metrics. They are not the same kind of evidence and they should not carry the same weight.

Three of those describe what machines did. One records a human arriving. All four are worth watching, as long as the first three are not presented as though they were the fourth.

Read Share of Authority carefully

Microsoft documents a methodological choice in Share of Authority that is going largely unmentioned elsewhere:

"The calculation is applied at a daily level. If your domain is cited for a query on a given day, all citation instances for that query on that day are included; if not, that day is excluded. As a result, this approach focuses on active participation and may lead to higher Share of authority values compared to broader calculations."

Days when you were not cited at all for a query are dropped from the denominator rather than counted as zero. So the metric measures your share on the days you showed up, and Microsoft says it may run higher as a result.

Measuring competitiveness on the days you actively participate is a legitimate thing to want, so this is not a flaw. It does mean the figure is not comparable to a share-of-voice number calculated the ordinary way. Quote it with the caveat attached.

What to do with it

Four things worth doing in the first week you have this data.

  1. Read the non-branded grounding queries first. They tell you which category questions retrieval associates you with. Nothing else on the dashboard gets that close to mapping where you are already a candidate.
  2. Watch the branded and non-branded lines separately over time. Non-branded citations growing while branded stays flat suggests you are being retrieved for the category rather than for your name. Track the two lines apart and you can see which of the two is moving.
  3. Pair it with the traffic number. Citations without referral traffic means you are being read into answers but not clicked through, which is a different problem from not being cited.
  4. Fix the denominator before you report. If Share of Authority goes into a slide, put Microsoft's own caveat next to it.

We keep coming back to the same question. Before you present any number as demand, ask what it rests on: a machine query string, a citation count, a human session, or money. Clarity now puts three of those four on one screen. That is useful, and it also makes the weakest of them easier than ever to promote into a claim it cannot support.

If you want to see which specific pages the engines actually pull from when they mention you, we cover that in which pages AI pulls from when it cites you, and the broader measurement picture in what is AI search visibility.

Sources

Find Out Whether the Machines Actually Recommend You

A citation dashboard tells you that you were referenced. It does not tell you who gets named when somebody asks the buying question. Get a free AI visibility score across ChatGPT, Claude, Gemini, Perplexity and Google AI Mode, and see who is being recommended instead.

Related reading