There are four numbers on the dashboard and one question sitting in your inbox. The numbers are mention rate, citation share, share of voice, and something labelled sentiment, and nothing on the screen tells you which of them answers the question you were actually asked.
They are not four views of the same thing. Each has a different denominator, each answers a genuinely different question, and each fails in a specific, predictable way. This page gives you the formula, an illustrative worked example, and the failure mode for each, then a method for picking one primary number and defending it.
Disclosure: RankRift publishes vendor-neutral criteria for how AI-visibility trackers compute their numbers. That is a real conflict and it is stated on every page where it applies. The metric definitions below are the ones we publish and hold ourselves to, and the honest position is that many teams should compute these in a spreadsheet and buy no tool at all. Our disclosure and conflicts page sets out how we handle this.
The Question Behind Each Metric
Before any formulas, the mapping. Pick the row that matches what you were asked, not the row whose number looks best.
| The question you were asked | The metric that answers it | The unit it is computed over |
|---|---|---|
| "Do assistants know we exist for this kind of question?" | Mention rate | Runs of a prompt |
| "Are our pages among the sources being used?" | Citation share | Cited sources across runs |
| "How much of the category conversation is ours versus theirs?" | Share of voice | Brand mentions across a prompt set |
| "When we are named, does it help us?" | Sentiment and framing | Individual mentions |
Two of these are about you in isolation. Two are competitive. Reporting a competitive metric when you were asked an existence question, or the reverse, is how a report ends up technically correct and completely unhelpful.
All numbers in the worked examples below are illustrative, not measured. They are invented to demonstrate arithmetic. No study, no run, and no real brand sits behind any of them.
Mention Rate: Formula, Example, Failure Mode
What it answers. Out of the times we asked, how often did the assistant name us at all?
Formula.
Mention rate = (runs in which the brand is named) ÷ (total runs of that prompt)
The denominator, stated explicitly. Total runs of a specific prompt on a specific surface in a specific period. Not prompts. Not sessions. Runs. If you aggregate across a prompt set, the denominator becomes total runs across all prompts in the set, and you must say which you did, because the two numbers are not interchangeable.
Illustrative worked example. You run one prompt 60 times against one surface. Brand A is named in 18 of those runs. Mention rate is 18 ÷ 60 = 30%. If the same 18 mentions came from 60 runs spread across six different prompts, that is a set-level mention rate of 30% and it can conceal a prompt that never mentions you at all alongside one that always does. Illustrative figures.
Failure mode. Mention rate is the metric most often reported without its sample size, and it is unstable at small samples. A brand named in 3 of 10 runs and a brand named in 30 of 100 runs both read as 30% on a dashboard, and only one of them supports a decision. It also treats every mention as equal, including mentions that damage you. Always publish the run count beside the rate, and never report a difference between two mention rates as a change without checking whether it exceeds ordinary run-to-run variation.
Citation Share: Formula, Example, Failure Mode
What it answers. Among the sources the assistant actually surfaced, how often was one of ours among them?
Formula.
Citation share = (citations to your domain) ÷ (total citations returned across the measured runs)
The denominator, stated explicitly. Total citations returned, across only those runs that returned citations. This exclusion is the whole ballgame. If a surface returns sources on some runs and not others, including the citation-free runs in the denominator produces a different and much smaller number, and both versions can be described as "citation share" in plain English.
Illustrative worked example. Across 40 runs, 25 returned citations, totalling 90 cited URLs. Six of those 90 point to your domain. Citation share over cited runs is 6 ÷ 90 = 6.7%. If you instead divide by an all-runs denominator that a different tool might use, you get a different figure from identical data. Illustrative figures.
Failure mode. Citation share is only meaningful on surfaces where citations are near-universal. On a surface that retrieves occasionally, you are computing a share over a small and non-random subset of runs, and the number moves for reasons unrelated to your visibility. It also conflates being cited with being recommended: your page can be cited as a source for a fact in an answer that then recommends someone else. When comparing across assistants, check first whether citation behavior is even comparable, using the sourcing comparison.
Share Of Voice: Formula, Example, Failure Mode
What it answers. Of the brand-naming that happens in our category's answers, how much is ours?
Formula.
Share of voice = (your brand mentions) ÷ (total mentions of all tracked brands) across the prompt set.
The denominator, stated explicitly. Total mentions of every brand in your defined competitive set, across all runs of the prompt set, in the period. Two decisions inside that denominator determine the entire number: which brands count as in-set, and whether repeat mentions of the same brand inside one answer count once or multiple times.
Illustrative worked example. Across a prompt set, brands are named 300 times in total. Brand A accounts for 45, Brand B for 120, Brand C for 90, and everything else for 45. Brand A's share of voice is 45 ÷ 300 = 15%. Now add two more competitors to the tracked set. The denominator grows, Brand A's count does not, and its share of voice falls without a single answer having changed. Illustrative figures.
Failure mode, and why two vendors can both be right. Share of voice is the most manipulable metric on the dashboard, not through misconduct but through defensible modelling choices. Consider what a tool must decide before it can compute one number:
- Which brands are in the competitive set, and who decides.
- Whether a brand named three times in one answer counts as one mention or three.
- Whether mentions in a list of options weigh the same as a mention in a direct recommendation.
- Whether unnamed or long-tail brands go into the denominator or get dropped.
- Whether each prompt in the set is weighted equally, or weighted by assumed buyer volume.
- Whether runs are pooled across surfaces or kept separate.
Every one of those has a reasonable answer and an equally reasonable opposite. Two vendors making different choices on identical raw data will publish different share-of-voice figures, and both can be internally consistent and honestly derived. This is not a scandal. It becomes one only when a vendor publishes the number without publishing the choices.
The practical consequence: share of voice is not comparable across tools, ever. Never put a figure from one source next to a figure from another in the same chart. If you change tools, you have started a new time series, and the correct thing to do is say so on the slide.
Sentiment And Framing: Why It Isn't Optional
The three metrics above all count mentions. None of them asks whether being mentioned helped.
An answer can name your brand as the budget compromise, as the option with the difficult onboarding, as the one people leave, or as an example of what to avoid. Every one of those increments mention rate and share of voice exactly as much as an enthusiastic recommendation does. A team that reports only count-based metrics can show a rising line while the substance of what is being said about them deteriorates.
You do not need a sophisticated classifier to fix this. A coarse three-level classification applied consistently is enough to catch the pathological cases:
- Recommended — named as a suggested option for the asker's stated need.
- Neutral mention — named as existing, or as context, without endorsement.
- Negative or qualified — named with a caveat, a warning, or as a contrast case.
Record which of the three each mention falls into, along with the sentence it appeared in, and report the split alongside the headline rate. The related distinction between merely appearing in the text and being genuinely put forward as an option is worked through in when a mention isn't a recommendation.
Two cautions. Classification is a judgement call, so write down your rules and apply them consistently rather than by feel. And do not convert the three levels into a weighted composite score without publishing the weights, because that is exactly the move that makes share of voice incomparable.
Choosing One Primary Metric And Defending It
There is no universally correct primary metric, and any recommendation of one would be a recommendation to ignore your actual question. What there is instead is a procedure.
Start from the decision the number will inform. If the decision is "do we invest in this category of work at all," mention rate is usually the honest primary, because it answers an existence question and it is the least dependent on modelling choices. If the decision is competitive positioning, share of voice fits, with its denominator published. If the decision is about which of your pages or which third-party sources are doing the work, citation share fits, but only on surfaces where citations are near-universal.
Write the definition down before you compute it. One paragraph: formula, denominator, competitive set, counting rule for repeat mentions, weighting, surfaces included, sample size per cell. If you cannot write that paragraph, you do not yet have a metric, you have a number.
Pick one primary and demote the rest to context. Reporting four co-equal headline numbers means whoever reads the report will pick whichever one supports the argument they already had. One primary metric, three supporting, and a stated reason for the choice.
Freeze it. A metric definition that changes between quarters produces a time series that means nothing. If you must change it, re-compute history under the new definition or start a fresh series, and say which you did.
Do not reintroduce false precision. These are rates estimated from samples. Report them with a range, not to two decimal places, and resist any request to convert them into a single composite visibility score. The reasoning for that refusal is set out in why answer position isn't a ranking.
FAQ
Can I add these into one overall AI visibility score? You can, and we would advise against it. A composite requires weights, and no principled universal weighting exists. If you build one anyway, publish the weights alongside every use of the number.
Why do my tool's numbers not match what I see in the chat window? Different surface, different sample, different denominator, different date, and non-determinism on top. A tool sampling through official APIs and a person typing into a consumer app are not measuring the same thing.
Which metric should I put in the board deck? The one that answers the question the board asked, with its sample size beside it. If the board asked something these metrics cannot answer, such as revenue impact, say so rather than substituting a metric that looks close.
Do I need a tool to compute these? No. All four are arithmetic over a run log, and many teams should stay in a spreadsheet until the manual effort genuinely outgrows it. A tool buys you scheduling, storage, and consistency, not better definitions.
In Plain English
Mention rate asks whether you show up. Citation share asks whether your sources are being used. Share of voice asks how much of the category's brand-naming is yours. Sentiment asks whether any of it helped. They have different denominators, so they are not interchangeable, and share of voice in particular is not comparable between tools because every tool makes defensible but different modelling choices.
This is for whoever has to choose a headline number and then defend it to people who will not read the footnotes. Use it when you are designing a report or auditing one you inherited. Do not use it to build a single blended score.
Your next step is to write the one-paragraph definition of your chosen primary metric, denominator and all, before you compute anything. The measurement discipline underneath all four metrics is on how to measure AI visibility.
Grab the AI visibility scorecard and the metric-definition worksheet from resources and downloads to get that paragraph written in the right shape the first time.