Mentions, citations and share of voice

Three questions, three metrics

Ask three people in this category what "AI visibility" means and you will get three metrics, used interchangeably in conversation and not interchangeable at all. Getting them apart is the difference between a report that supports a decision and one that supports whichever decision you already wanted.

The last column is the whole chapter. Two of these are rates measured against a fixed, external denominator. The third is a ratio measured against a denominator you selected — and that difference determines which of them can be gamed by choosing your comparison set.

Mention rate

Mention rate is the share of successful runs in which a model named your brand in its answer text: the count of successful runs where the brand was named, divided by every successful run in the window.[^attensira-mention-rate]

Three design decisions live inside that sentence, and each one is a place where two vendors' "mention rate" stop meaning the same thing.

The unit is the run, not the mention. A run either named you or it did not. An answer that names you nine times is one run in the numerator, exactly like an answer that names you once.[^attensira-run-unit] The alternative — counting mentions per answer — sounds more informative and is worse, because it makes the metric a function of how talkative the model was. A model that writes long answers would then appear to like you more than one that writes short ones.

Failed runs are outside the denominator. They are stored for debugging and excluded, so a model outage lowers your sample size rather than your rate.[^attensira-failed-runs] Include them and a provider's incident becomes your visibility crisis.

The matching rule is a definition, not an implementation detail. Whether a possessive, a plural, a misspelling or a partial match counts as a naming is a choice that materially changes the rate, especially when the brand name is an ordinary word. Any vendor should be able to state the rule in a sentence.

Mention rate is the right headline number for one reason: it answers the question the business actually asks, which is whether the assistant reaches for you when somebody asks the question you care about.

Citation rate

Citation rate is how often a model attaches a link to your own domain, measured as successful runs citing you divided by all successful runs in the window.[^attensira-citation-rate] Same denominator, different numerator. Answers that contain no links at all stay in the denominator, because a linkless answer is a real answer and a real absence of a citation.[^attensira-linkless]

It comes out lower than mention rate, and the reason is structural rather than incidental: two independent conditions must both hold. The model has to name you, and it has to decide to link your pages rather than a third party's — or link nothing.[^attensira-citation-lower]

The gap between the two rates is where the diagnosis is. Consider what each combination implies.

High mention, high citation. The model knows you and treats your pages as the source. This is the target state, and there is not much to do beyond keeping the pages current.

High mention, low citation. The model knows you and is sourcing somebody else when it needs a link — typically a review platform, a marketplace listing or a comparison article. Your brand is established; your pages are not the artefact being retrieved. This points at the content, structure and retrievability of your own site, and often at whether the crawler that feeds answers has been to the relevant pages at all.

Low mention, some citation. Unusual, and normally means you are being linked inside answers about something adjacent rather than being recommended. Worth reading the actual answers rather than the number.

Low mention, low citation. The assistant does not have you. Before treating this as a content problem, check the denominators and check the crawler logs — an absence caused by never being fetched needs a different fix from an absence caused by never being chosen.

None of those four readings is available from a single blended "visibility" figure. That is the argument for keeping the two rates separate even though one number is easier to put on a slide.

Share of voice, and why it is a different kind of object

Share of voice is your mentions divided by all brand mentions across a defined set of brands. It looks like the other two. It is not.

Mention rate and citation rate have a denominator determined by the measurement: all successful runs in the window. Competitor mention rates use exactly the same denominator,[^attensira-same-denominator] which is what makes them directly comparable — each brand's rate is an independent estimate of an independent quantity.

Share of voice has a denominator determined by you: the set of brands you decided to compare against. That makes it a framing device rather than a measurement, and it moves for reasons that have nothing to do with the world.

Work the labelled hypothetical. Suppose across a window, three brands are named at rates of 0.30, 0.20 and 0.10. Share of voice for the first is 0.30 / 0.60, or one half. Now add two more tracked competitors whose rates are 0.15 and 0.05. The first brand's mention rate has not moved at all — nothing about the world changed — but its share of voice is now 0.30 / 0.80, or 0.375. A quarter of the metric evaporated because somebody edited a list.

The reverse works too, and it is the version that appears in vendor marketing: choose a narrow set of brands you outperform, and share of voice looks commanding. It is not a lie, exactly. It is a number whose denominator is an argument.

There is a second, subtler problem. Mention rates do not sum to one across brands, because a single answer can name five brands or none, and treating a mention rate as a slice of a pie will lead to wrong conclusions.[^attensira-mention-not-pie] Share of voice manufactures the pie anyway, by dividing through by the total. The resulting figure implies a zero-sum contest — your gain is a competitor's loss — that the underlying data does not contain. In reality every brand in a category can become more visible at once, because the answers can get longer and list more names.

Use share of voice to frame a market for an audience that thinks in market share. Publish the brand set every time. Never make it the KPI a team is measured on, because the cheapest way to move it does not involve the product.

Which metric belongs in which conversation

The three numbers are not competing for the same slot. They belong to different decisions, and matching them to the decision is most of the work.

Board and executive updates want mention rate on a wide prompt set over a long window, because that is the number with the largest denominator and therefore the only one whose movement is interpretable at a quarterly cadence. Present it with its sample size and its prompt-set definition, and expect most quarters to show no statistically real change.

Content and site work wants the gap between mention and citation, sliced by topic. That gap is the closest thing this category has to a work queue: it names the questions where the assistant already trusts your category standing but sources somebody else's page, which is a solvable problem with a known owner.

Competitive and positioning conversations are where share of voice earns its place, provided the brand set is on the slide. It answers "who does the assistant reach for in this market", which is a genuine strategic question, as long as nobody treats the resulting figure as a target.

Diagnosis of a specific absence wants none of the three. It wants the answer text itself and the crawler log for the pages that should have been cited. Rates tell you that something is wrong; only the artefacts tell you what.

Ordering and position

A fourth number circulates: where your citation appears in an answer's list of sources. It is a real observation, and it is not a ranking. Average position is not a leaderboard position and not a ranking of your brand against competitors[^attensira-avgposition] — it is a property of the ordering inside individual answers, aggregated.

That does not make it useless. Movement in it can be a leading indicator, and it is worth watching alongside the rates. It makes it dangerous to present as a league table, because a table of brands sorted by average position looks exactly like a rank and will be read as one by everybody who sees it.

What not to build: the composite score

The pressure to blend these into one number is constant, and it should be resisted. We publish no visibility score and no sentiment metric, on purpose.[^attensira-no-sentiment]

A composite score fails on three counts. It hides the denominators, so a cell with four observations and a cell with four hundred contribute alike. It hides the weights, which are arbitrary and unfalsifiable — nobody can say a score is wrong, because nobody can say what it would mean for it to be right. And it cannot be reproduced by the customer from published inputs, which means it cannot be audited, which means the vendor can adjust it and call the movement a finding.

The honest alternative is not more sophistication. It is fewer numbers, each with a stated definition and a visible denominator. Two rates and a sample size will support more decisions than any index, and they will survive a sceptical executive asking where the number came from — which is the only test that ultimately matters.

The next chapter, building a prompt set that means something, covers the input that determines all three of these numbers before any of them is computed.