AI visibility has an awkward measurement problem: there is no results page to inspect. Nobody can open a URL and see where you placed. Every number in this discipline is produced by asking an engine a question and reading what came back, which means the metric is only as good as the question set behind it and the honesty of the thing recording the answers.
The unit of measurement is the prompt, not the keyword
Classic SEO measures keywords because a keyword maps to a results page. AI visibility measures prompts, because a prompt maps to an answer. The difference is not cosmetic. Keywords are short and typed; prompts are long, conversational, and often carry the asker’s role and constraints (“I run a two-person marketing team, which tool should we use”).
This is why prompt-set design decides the quality of every metric below. Twenty questions your buyers genuinely ask will produce numbers you can act on. Twenty questions written to include your brand name will produce excellent numbers that mean nothing.
The six metrics
| Metric | What it answers | How fast it moves |
|---|---|---|
| Citation rate | How often does an answer LINK one of my pages as a source? | Weeks. The most actionable metric. |
| Mention rate | How often does an answer NAME my brand at all? | Months. Tracks reputation and training data. |
| Share of voice | Of all brands named across these answers, what share is me? | Weeks to months. Zero-sum and competitive. |
| Answer position | When I appear, am I named first or fifth? | Weeks. Often overlooked; first mention carries most of the value. |
| Engine coverage | How many of the engines cite me at all? | Weeks. Reveals single-engine dependence. |
| Citation sources | Which third-party sites do engines lean on for my category? | Slow to change, fastest to act on. |
Six numbers, not one. A blended 'visibility score' hides which of these moved and why.
The last one is the least discussed and often the most useful. If the engines answering your category keep citing three specific roundup sites, a forum and a video channel, then your citation rate is downstream of whether you appear on those sources. That is a concrete, finite piece of work, and it is invisible if you only track your own numbers.
AI citation rate: definition and how to measure it
AI citation rate is the share of AI answers to your tracked questions that cite your domain, worked out for each engine separately and counted only over the runs where that engine actually answered. A cited link is the only part of an answer that can send a visit, so citation rate is the closest this field gets to a click-through number, and it is the metric that most directly shows whether your pages are the ones engines choose to read.
The formula
For one engine over one period:
citation rate = answers that cite your domain ÷ answered runs
Answered runs are the runs you attempted minus the ones that produced no usable answer (an error, a refusal, a rate limit) and the ones still waiting for a result. Two choices in that formula decide whether the number means anything.
- The denominator is answered runs, not attempted runs. A run that failed measured nothing, so it leaves the denominator. Dividing by every attempt turns a provider outage into an apparent drop in visibility, which is the problem described in measured zero vs not measured below, in arithmetic form. An answer that came back and cited nobody is different: it is a real miss and stays in. On Google AI Overviews, a search that showed no Overview at all counts this way, as an answer that cited nobody; only a results page that could not be read is left out.
- It is worked out per engine, then compared, never pooled first. ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode each choose sources differently and cite very different numbers of them per answer. If you do report one total, add up the cited and answered counts and divide once; averaging per-engine percentages weights a thin engine the same as a busy one.
Citation rate vs mention rate
The two are read from the same answers but measure different things, and the difference between a mention and a citation is worth keeping in view whenever either number moves.
| Citation rate | Mention rate | |
|---|---|---|
| Counts an answer when | It links one of your pages as a source | It names your brand anywhere in the text |
| What drives it | Whether a page of yours is the best source the engine found for that question | What the model and its sources already believe about your category |
| What moves it | New or rewritten pages, crawlability, being quotable in one passage | Reputation, third-party coverage and reviews, over months |
| Can it send traffic | Yes, through the linked source | Only if the reader goes looking for you |
Some tools also report a wider third measure: the share of answered runs whose answer names you or links your domain. Cituna’s per-engine count works this way, which is why the three states further down this page describe “cited” as linked or named. It uses the same answered-runs denominator, but it is a different number from link-only citation rate and the two should never share a column or a chart without a label saying which is which. An answer can mention you without citing you (the engine knows you but used a roundup as its source), or cite you without recommending you (it read your comparison page and then named a rival). That second case is why citation rate is read alongside mention rate and share of voice, never instead of them.
An illustrative example
The numbers below are invented and rounded to show the arithmetic. They are not readings from any engine or any brand. Suppose one month of daily runs on a fixed prompt set gives these counts on four engines:
| Engine | Runs attempted | No usable answer | Answered runs | Answers citing you | Citation rate |
|---|---|---|---|---|---|
| Engine A | 200 | 0 | 200 | 60 | 60 / 200 = 30% |
| Engine B | 200 | 0 | 200 | 10 | 10 / 200 = 5% |
| Engine C | 200 | 0 | 200 | 4 | 4 / 200 = 2% |
| Engine D | 200 | 40 | 160 | 8 | 8 / 160 = 5% |
Illustrative only. Real rates depend on the category, the prompt set and the engine.
Three things follow from the same counts.
- The per-engine split is the finding. One engine cites the site on almost a third of answers and the other three on 5% or less, so the work for the coming weeks is on those three.
- Pooling hides it. Added together, the counts give 82 cited answers out of 760 answered runs, about 10.8%. Averaging the four percentages gives 10.5%. Neither figure describes any engine, and both look healthy while three engines barely cite the site at all.
- The wrong denominator invents a drop. Divide Engine D by its 200 attempts instead of its 160 answers and its rate reads 4% instead of 5%. Nothing about its answers changed; 40 runs simply failed.
None of this says what a “good” rate is for anyone else. The useful question is whether the next window, on the same prompts and the same engines, is higher.
How often to measure
Engines resample the web and change models often enough that one run per question is a coin toss, not a reading. Run the same fixed prompt set on a schedule, daily where you can, and read the rate over a rolling window, a week or a month of runs, rather than day to day. Compare windows only when the prompt set and the list of engines were the same in both; if either changed, follow the prompt-set rule in the next section and start a new baseline. For when a change is big enough to act on, see alert thresholds for AI visibility.
Which metrics deserve a target
Set targets on citation rate and share of voice. Both are responsive to work you can plan, both are comparable over time against a fixed prompt set, and both fail loudly when you game them, which is a property worth having in a metric.
Track, do not target, mention rate and answer position.Mention rate is largely a function of how well known you are, so a target on it is a target on brand awareness wearing a technical costume. Answer position is genuinely useful but too noisy at low volumes to hold anyone to.
Treat citation sources as a work queue, not a metric. It is a ranked list of places to go, and the right response to it is outreach and content, not a number on a dashboard.
The prompt-set rule:
Measured zero and not measured are different numbers
The most consequential detail in AI visibility measurement is also the least visible. When an engine errors, refuses, rate-limits or simply returns nothing, that is not a measurement of absence. It is the absence of a measurement.
Rendered on a dashboard, the two look identical: a zero, or an empty cell. Treated identically in a score, they are corrosive, because a run where two engines failed will read as a visibility drop, and you will spend a week looking for a cause that does not exist.
Any measurement worth building on carries three states per question and engine, not two:
- Cited, the answer linked or named you.
- Not cited, the engine answered and did not include you. A real, informative measurement.
- Not run, the engine produced no usable answer. Not a miss, and it must be excluded from the denominator rather than counted as a failure.
When you evaluate any tool in this category, ask which of those three it records and how the third is treated in the score. It is the fastest way to find out whether the numbers are measurements or decoration.
For the individual metrics in depth, see AI share of voice and mentions vs citations. For what to do once the numbers exist, why your brand is not showing up in AI search covers the usual causes in order.
Drafted with AI assistance from our own research and Search Console data, and reviewed by Rahul A before publishing. Rules and prices change; check the linked official source before you act.