AI visibility, brand mentions, and citations: what each number tells you

AI visibility, brand mentions, and citations are three different observations of the same AI answer, and each one has its own denominator. Brand visibility asks how often your brand is named in the answers that have been analysed. Citation rate asks how often your domain is linked as a source in the answers that were collected. A recommendation asks what role the answer gave your brand once it was named. Put them on one screen and they will disagree, and they can all be correct at the same time.

In short

• A mention is your brand named in an answer. A citation is your domain linked as a source. A retrieval is a page the engine pulled into its context, whether or not the answer then linked it. You can have any one without the others.
• Brand visibility divides by analysed answers. Citation rate and domain coverage divide by collected answers. That is why 40% and 15% can sit side by side.
• Share of voice is relative to the competitors you chose to track, so adding or pausing a competitor moves it while the answers stay the same.
• A missing value is "no data", not 0%, and a count under 30 observations is provisional.
• Put the denominator next to every number you report.

Start with one small dataset

The fastest way to see the differences is to hold the answers still and read every metric off the same data.

Quillstone sells document-review software to mid-sized legal and compliance teams. It tracks three competitors: Brieflane, Clausewise and Docket North. For one 28-day window, on one engine that reports the pages it retrieved, the raw counts look like this.

Two of these need explaining. A collected answer is a completed prompt execution (a Google run that showed no AI answer is not one). An analysed answer is a collected answer whose mentions have also been extracted, so the platform has read the text and recorded who it names (Metrics defined). Extraction runs after collection, so the analysed group is always a subset of the collected group: here, 10 of the 60 answers are still waiting on extraction.

Citations and retrieved pages are recorded during collection itself, which is why they divide by the collected group instead.

What does brand visibility tell you?

Brand visibility is analysed answers naming your brand, divided by analysed answers. It answers one question: when the platform has read an answer to one of your prompts, how often does your name appear?

For Quillstone: 20 of 50 analysed answers, so 40%.

The 10 unanalysed answers are in neither the numerator nor the denominator, so the metric describes the analysed population, not everything collected.

It also does not say how good the mention was. An answer that names Quillstone as a warning counts the same as an answer that names it first. For that you need position and role, below.

What does brand position tell you?

Brand position is the mean of each answer's earliest self-mention order, computed only over the answers that named you. Lower is better: position 1 beats position 9 (Brand position).

Suppose the 20 answers naming Quillstone put it at these earliest positions:

Mean position is 42 divided by 20, so 2.1. The 30 analysed answers that did not name Quillstone contribute nothing. That is what makes position a different question from visibility: it describes how prominent you are when you appear, not how often you appear.

Look at the observation count, too. Position rests on 20 observations. The platform-wide floor is 30, and a metric built from fewer is shown but is provisional (How to read any number). So Quillstone's 2.1 is a reading to watch, not a settled figure.

What does share of voice tell you?

Share of voice is distinct answers naming your brand family, divided by the sum, over every active tracked brand family, of that family's own distinct-answer count. A brand family is a top-level brand plus its sub-brands, and an answer that names a brand and its own sub-brand counts once for that family.

The denominator is a sum of family counts, not a count of distinct answers. An answer that names both Quillstone and Brieflane adds one to each family's count, so the sum can exceed the number of answers.

Quillstone's example:

Share of voice is 20 divided by 75, which is 26.7%. Only 50 answers were analysed, but the denominator is 75, because answers naming several families are counted once per family.

Now the part that catches teams out. Suppose you pause Docket North because it is no longer a real competitor. The sum drops to 65, and Quillstone's share becomes 20 divided by 65, which is 30.8%. No answer changed. The roster you build sets the denominator, and untracked brands the platform notices but you have not added are excluded entirely (Share of voice). The metrics page calls this the most misreadable number in the product, and it is worth taking seriously.

The Competitors screen also shows a second figure with the same name, built from domain citation counts instead of mentions. Report which one you mean; see Domain-citation share of voice.

What does citation rate tell you, and how is it different from domain coverage?

Citation rate is collected answers that cited your domain, divided by collected answers. Domain coverage is collected answers that retrieved your domain, divided by collected answers. Both read the collected population.

For Quillstone: citation rate is 9 of 60, so 15%. Domain coverage is 14 of 60, so 23.3%.

The gap between them is the useful part. Retrieval is not citation: an engine can pull one of your pages into its context while composing an answer and never link to it. Domain coverage counts that regardless of whether a citation followed. In this example, Quillstone's pages were retrieved in 14 answers and cited in 9. If those 9 are among the 14, that leaves 5 answers where a page was retrieved but not cited, which is worth asking about, and Your page was retrieved but not cited walks through it.

Two limits are easy to miss. Domain coverage has no value for Google AI Overviews, Google AI Mode, ChatGPT (app) and Gemini (app), because those engines report only the sources the answer cites, never a list of retrieved pages, so counting them would read as a measured zero. That is why this example uses an engine that reports both. And whether a citation is "yours" is a single check: the cited URL's registrable domain against your project's domain, so a syndicated copy of your article on another domain is attributed to that other domain (How a citation is attributed to you).

What is "answers citing sources"?

Answers citing sources is collected answers that cited at least one source of any domain, divided by collected answers: 45 of 60, or 75% in the example. It describes the engine, not you, so it has no "higher is better" direction. Use it as context for your citation rate. Some engines have no value for it, because their saved citations include every search result (Answers citing sources).

Why is the Citations screen number different again?

The Citations screen opens on one hero number: the share of the citations currently in view that point at your domain. It divides by citations, not by collected answers, so it answers a different question from citation rate and the two percentages are not meant to match (Citations).

Continuing the example, suppose 120 citations are in view and 12 point at Quillstone's domain. The Citations screen would read 10%. The 12 citations sit inside those 9 answers because one answer can link several pages, or the same page more than once.

The hero number and both Citations tabs are built from the newest 500 citations in the window that match the filters, so a busy window is a capped view, not the full count.

What does a recommendation tell you?

A recommendation is the role the answer gave your brand after naming it, such as Top recommendation, Alternative option, Warning or caveat, or Not recommended. It is a property of a stored mention record, and the sentiment screens count roles by brand (Sentiment trends). A role is independent of sentiment, and each mention record contributes once regardless of repeated name occurrences.

Quillstone's 20 mention records split like this:

This is where a mention stops being a good outcome by default. Quillstone is named in 40% of analysed answers, but 4 of those 20 mentions are a warning or a "not recommended", and only 6 are a top recommendation.

The role table gives counts, and any rate is one you calculate yourself, so state the denominator. Six top recommendations is 30% of the 20 answers that named Quillstone, and 12% of the 50 analysed answers. Both are true, they answer different questions, and neither is a built-in metric. If you put either in a report, say which one and label it as a hand calculation.

For the reasons an answer gives behind a role, see Brand reasons.

Mention, citation, recommendation and retrieval: four separate questions

The four ideas answer different questions, so an answer can be strong on one and empty on another.

So a mention without a citation is real. A citation without a mention is real too. Neither proves why an engine produced the answer. These are observations of what the answer contained, not evidence of what caused it, so avoid writing "AI ranks us lower because" in a report.

The reporting cheat sheet

Copy this table as the header of any report. Every row except the last is taken from the metrics definitions; the last is a screen figure described on the Citations page.

And a per-report checklist you can paste into your template:

REPORT CHECKLIST (paste above every AI visibility report)

Population
- [ ] Window (dates) and filters stated: engine, country, persona, language
- [ ] Collected answers: ___    Analysed answers: ___
- [ ] Collection channel(s) named; channels not pooled without saying so

For each metric
- [ ] Metric name exactly as defined, with numerator and denominator
- [ ] Count behind the percentage (e.g. 20 of 50), not the percentage alone
- [ ] Under 30 observations? Label it provisional
- [ ] "No data" reported as no data, never as 0%

Share of voice
- [ ] Tracked competitor list stated; note any added or paused in the period
- [ ] Which share of voice: mention-based or domain-citation

Mentions vs outcomes
- [ ] Mentions reported next to recommendation roles, not instead of them
- [ ] Any hand-calculated rate labelled as hand-calculated, with its denominator

Wording
- [ ] Observations described as observations, not as causes

Common mistakes and what these numbers cannot tell you

• Comparing two "observations" counts. On brand visibility it counts every analysed answer, on brand position only the answers naming you, and on share of voice the sum of family counts. Never compare across metrics.
• Reading a share of voice change as a change in the answers. Pausing one competitor moved Quillstone from 26.7% to 30.8% with no new answers.
• Treating a mention as a win. Read the role next to it.
• Reading no data as zero. A rate with no denominator reads as no data, not 0%.

What none of these numbers can tell you:

• Why the engine wrote what it did. A mention, a citation and a stated role are observations of an answer, not proof of a model's internal reasons.
• What real buyers see. The numbers describe the prompts you chose to monitor, collected in the way each channel is collected, not every question every buyer asks.
• Whether a change came from your content. Defining a metric is not the same as attributing a movement to an edit; that is a separate exercise.

Frequently asked questions

Can my brand be mentioned in an AI answer without being cited?

Yes. A mention is your brand named in the answer, and a citation is your domain linked as a source. They are counted separately, from different populations, and an answer can have one, both or neither (Key terms).

Why is my brand visibility so much higher than my citation rate?

They divide by different populations and count different events. Brand visibility divides by analysed answers and counts being named; citation rate divides by collected answers and counts being linked. A 40% and a 15% on the same screen can both be right.

Is a low citation rate always a problem?

Not by itself. Read it beside answers citing sources, which says how often the engine cites anything, and beside domain coverage. An answer with no sources is one where the engine showed none, which on ChatGPT (app) and Gemini (app) usually means the app did not search.

Why did my share of voice change when nothing else did?

Its denominator is built from the competitors you track and have not paused. Adding a competitor lowers your share, and pausing one raises it, with no change in the answers. Check whether you added or paused a competitor in the period before you explain the movement.

What counts as enough data to trust a number?

For a metric that reports an observation count, fewer than 30 observations is provisional: still shown, but too small a sample to treat as settled. Above that, keep reporting the count next to the percentage so a reader can judge it.

Which number should I lead with in a client or leadership report?

Lead with the question, not the number. Say what you are asking ("how often are we named?"), give the metric with its counts, then add the role and citation evidence that qualifies it.

Next step

Open the metrics in your own project and check each one against the cheat sheet: for every percentage, find the numerator and denominator it uses, and note whether it reads the collected or the analysed population. The definitions live in Metrics defined, and the same formulas hold in the Explorer, where you can break them down by engine, country, persona and other dimensions. To see them on your own prompts, sign in to DiscoveredBy.

If you are investigating a specific answer instead, start with Why AI recommends your competitor, and to decide where a citation gap deserves work, see Find the citation gaps that deserve your next content update.

Enlarge image

Enlarge image

Enlarge image