Mention rate vs citation rate: two different questions

The short answer

A mention is your brand name appearing in the words of an answer. A citation is one of your URLs appearing in the list of sources attached to that answer. They are produced by different mechanisms, recovered from different places, and they fail in different ways. Reporting one and calling it the other is the most common measurement error in this category.

Four outcomes, not two

Every sampled answer lands in one of four states, and each one means something different.

Named and cited. The model both states your brand as a fact and attributes a page of yours. This is the strongest single signal available, because it says the model had your material in front of it and was willing to say your name out loud.

Named, not cited. The model knows about you without reaching for your site. That knowledge came from somewhere — training data, or a third-party page that was cited instead of yours. It is a good outcome for reputation and a bad one for attribution, and it is the state most brands with any history are actually in.

Cited, not named. Your page was useful enough to attribute and your brand was not the subject of the sentence. A documentation page answering a technical question does this constantly. It builds domain authority with the retrieval system without building brand recall with the reader.

Neither. The uninteresting majority, and the only one of the four that is genuinely unambiguous.

A single number cannot express four states. That is the whole argument for tracking both.

The measurement asymmetry

The two metrics do not have equal evidential standing, and pretending they do is how dashboards mislead people.

Citation data is handed over. OpenAI's web search tool returns a url_citation annotation object carrying the URL, title and location of each cited source. Perplexity's chat completions response returns a search_results array whose objects carry a title, a url, a publication date, a last-updated date and a snippet. Gemini's Google Search grounding returns annotations whose url_citation entries carry the URL, the title and the text indices of the span they support. Three vendors, three shapes, one common property: the platform is telling you what it used.

Mention data is not handed over. No assistant returns a field listing the brands it named, because from the model's point of view no such object exists — there is only text. So a mention has to be recovered, and recovery means matching a string. Attensira's own published limitations are blunt about the method: a mention is a case-insensitive substring match of the brand name or domain against the answer text, with no entity resolution and no word-boundary check.

The consequence is not subtle. A brand named after a common word inherits noise that no amount of averaging removes, because the error is systematic rather than random. A brand with an invented name inherits almost none. Two competitors can therefore have mention rates that are not fairly comparable, while their citation rates are.

Why the fix is storage, not cleverness

The honest response to a noisy detector is not a better detector. It is keeping the evidence.

Attensira stores the verbatim answer text for every successful run, which means the rate is a derived number over a corpus that still exists. When detection improves, it can be re-run across the same history without losing the past, and in the meantime a human can read ten answers and find out in two minutes whether a rate is measuring their brand or the dictionary.

That is a modest claim and it is the right one. A metric you can audit by reading is worth more than a metric with a confident label.

Reading each number without over-reading it

A rate arrives as a value between 0 and 1 alongside n, the number of successful runs it was built from. The n is not decoration. At a small n, a single different draw swings the percentage a long way, and the percentage on its own invites a conclusion the sample cannot carry.

Competitor rates deserve one specific warning. They are computed on the same denominator as your own and are independent rates rather than slices of a pie, so one answer naming you and three rivals raises four rates from the same run. Adding them produces a number that means nothing. There is no metric here that divides a fixed total between brands, and the field named ShareOfVoice in the API is the brand's own mention rate under an older name[^mention-share-of-voice] — not a share.

The citation side has its own trap. Position within a citation list is averaged per source domain: it answers "when this domain is cited, how far down does it sit". It is not a leaderboard, not a rank of your brand against a competitor, and not comparable to a search result position. A domain that appears third in every citation list it appears in is not losing to two other companies. It is appearing third.

And neither metric carries tone. There is no visibility score and no sentiment metric here, and nothing scores whether an answer spoke about you kindly. If you want to know how you were described, the answer text is stored; read it.

What to do when they diverge

Divergence is the reason to run both, so it is worth having a reading for each direction.

Mentions up, citations flat. Something off your own domain is doing the work — a roundup, a forum thread, a competitor's comparison page. You are winning recall and losing attribution. The fix is off-site: be in the sources that get cited, or publish the material those sources are paraphrasing.

Citations up, mentions flat. Your pages are useful and your brand is not the subject. Common for documentation and reference content. The fix is on-page: make the sentence a model would quote say who you are, not just what is true.

Both flat while traffic moves. Neither metric measures traffic, and neither should be asked to. There is no identity join between an answer a model gave someone and a person who arrived at your site afterwards. Treat these as measures of presence in answers, connect them to commercial outcomes with your own analytics, and say out loud that you are doing so.

Both numbers are estimates, and only one of them admits it

Neither metric is a reading off an instrument. Ask a model the same question twice and you can get two answers, so any figure derived from sampled answers is an estimate of a distribution, not a measurement of a state. That is a property of the models rather than a fault in collection, and it applies identically to both sides of this comparison.

Where they differ is in how visible the uncertainty is. A citation is a discrete, platform-supplied fact about one answer: the URL was in the list or it was not. Aggregate enough of those and the noise is ordinary sampling noise, the kind a confidence interval handles. A mention carries that same sampling noise plus a second, nastier layer — detector error. If your brand name is a substring of an ordinary English word, some share of your mentions were never mentions, and that error does not shrink as the sample grows. It is a bias, not a variance.

This is why the two metrics deserve different amounts of trust at the same sample size, and why a chart that plots them on one axis with one confidence band is quietly lying about one of them.

The cost of collecting each

Cost matters, because it decides how often you can afford to look.

Citation rate is cheap once you are already making the call. The annotation object arrives in the same response as the answer; parsing it is a field access. The expensive part is coverage — you can only collect citations from surfaces that return them, and the surfaces a buyer actually uses are not the same set as the surfaces with a documented API. A consumer chat product with no API returns nothing at all, however visible it is to your market.

Mention rate is cheap in a different way and expensive in one specific place. The matching itself is trivial. The expense is the audit: someone has to read stored answers periodically and confirm the detector is still counting the right thing, especially after a rebrand, a product rename, or the arrival of a competitor with a confusable name. That work is small, unglamorous, and the only thing standing between a mention rate and a number nobody should quote.

Budget for it explicitly. A team that collects both metrics and audits neither has two numbers and no evidence.

Do not average them into one score

The temptation, once you have two numbers, is to blend them into a single index and put it on a slide. Resist it, for a mechanical reason rather than a philosophical one: the two inputs have different denominators, different error models and different collection coverage, so any weighting you choose is an opinion dressed as arithmetic. Worse, a blended score hides exactly the divergences that made tracking both worthwhile — the week your citations climbed while your mentions did not is the week the composite stayed flat and told you nothing.

Report them side by side, each with its sample size, and let the gap between them be the finding.

The one-line version

Citation rate tells you which page was used. Mention rate tells you whether you were part of the answer at all. The first is better evidence; the second is closer to what a reader actually sees. Track both, report the sample size beside each, and never quote either as a score.

[^mention-share-of-voice]: Attensira, "Limitations", retrieved 4 September 2026.