AI visibility metrics: which numbers actually matter (and which are decoration)
· by boneLau
Short answer: seven things get sold under the name "AI visibility". Four of them survive scrutiny: mention rate with a stated denominator, citation ownership, per-answer accuracy, and an interval on every percentage. The other three — a blended score, share of voice without a denominator, sentiment without a sample — cannot tell you what to do next, which is the only test that matters for a metric.
I build one of these dashboards, so assume the usual bias. Every number below is one we actually compute, and where our own product falls short I say so.
The test: can this metric change a decision?
A metric earns its place only if a different value would lead you to do something different. Apply that to each one:
| Metric | If it moves, do you act differently? | Verdict |
|---|---|---|
| Unbranded mention rate | Yes — it decides whether you work on content or on distribution | Keep |
| Citation ownership | Yes — own domain vs third party implies opposite fixes | Keep |
| Per-answer accuracy | Yes — a wrong category needs a definition page, silence needs a source | Keep |
| Interval / sample size | Yes — it tells you whether to believe the other three | Keep |
| Blended "AI visibility score" | Not really — you cannot tell which component moved | Decoration |
| Share of voice without a denominator | No — you cannot reproduce it | Decoration |
| Sentiment without a sample | Rarely — tone rarely changes what you publish | Weak |
The four that matter
1. Unbranded mention rate — with its denominator
This is the only metric that predicts whether new buyers can find you. It asks: when someone who has never heard of you types a category question, does the answer include your brand?
The denominator is the whole metric. Three vendors can report the same week as 6%, 25% or 60% by choosing differently:
- All answers (6%) — including the many answers that recommend no product at all
- Answers naming any product (25%) — excluding the "it depends, here are general factors" responses
- Answers whose prompt already contains your brand (60%) — which measures recall, not discovery
On our own brand the same run produced 0/27 unbranded and 17/18 branded. Blended, that is a healthy-looking 38%. Either half alone tells you the truth; the average tells you nothing.
What to ask a vendor: "What is your denominator, and do you report branded and unbranded separately?" If they cannot answer in one sentence, the headline number is not measuring what you think.
2. Citation ownership
A mention is not a citation, and a citation is not necessarily yours. Three cases produce identical-looking URL lists:
| Cited source | What it means | What to do |
|---|---|---|
| Your own domain | You are the only thing feeding this answer | Get third-party corroboration — a single source is fragile |
| A third party writing about you | Someone else is doing your explaining | Keep that page current; it is your real front door |
| A different company with the same name | The "mention" is not yours | Fix entity disambiguation, not content |
We hit the third case on our own brand: an answer that appeared to mention BrandScope was citing a same-name competitor's pricing page. Any dashboard that counts that as a mention is inflating your visibility.
3. Per-answer accuracy
Being named and being described correctly are separate failures. A model that calls a monitoring platform an "SEO tool" has mentioned you and damaged you in the same sentence — and that wrong category tends to persist across runs, because it came from the same stale source.
Accuracy has to be judged against your own positioning, per answer, with the model's stated reason attached. It cannot be inferred from sentiment: positive tone about the wrong category is still wrong.
On our own brand, the first properly-measured run came back 38% accurate — meaning most mentions described us partially or wrongly. That is a more useful number than any visibility score, because it points at one specific fix.
4. The interval (and the sample size behind it)
Every AI answer is a sample from a non-deterministic process. The same prompt asked twice routinely names different products.
So a percentage without an interval is decoration. Consider 42% from 7 answers: the Wilson interval is roughly ±30 percentage points. The true rate could be 12% or 72%. A dashboard showing "42%" and nothing else is selling precision it does not have.
Practical rule: with n under 5, report "too few to bound" instead of a number. With n between 5 and 30, print the interval. Above 30, the interval narrows enough that week-over-week moves become readable — which is also why small question sets cannot detect real change in a week.
The three that are decoration
A blended score. Weighted averages are useful for a single glance and useless for diagnosis. If the score moves, you cannot tell whether discovery improved, accuracy dropped, or the weighting changed between versions. Any vendor that shows a headline score should also show its components and the version of the formula — otherwise a scoring change looks like a visibility change. (We have shipped a scoring version bump ourselves; the number moved, the world did not.)
Share of voice without a denominator. "You have 12% share of voice" is unfalsifiable unless you know which questions count as contests and which engines are in the pool. Include questions that never name a product and everyone's share inflates together. Ask for the numerator, the denominator and the engine set, in that order.
Sentiment as a headline. Tone matters for reputation, but it rarely changes what you publish next. It becomes actionable only when attached to a specific quote from a specific answer — "Perplexity says we are 'expensive for what it does', on the objection question about pricing" is actionable; "sentiment: 62% positive" is not.
What a good dashboard actually looks like
Four numbers, each with its denominator and its sample:
``` Discovery mention rate 0/27 unbranded answers (0%, interval too wide at n=27 to call a trend) Branded mention rate 17/18 branded answers (94%) Citation ownership 4 own / 2 third-party / 1 same-name Accuracy 8 of 21 judged accurate (38%, model reason attached) ```
Note what is missing: no total, no grade, no letter. Those would require collapsing four independent findings into one number that cannot be acted on.
Three questions to ask any vendor
- "Do you report branded and unbranded separately, and what is the denominator?" — separates a measurement from a marketing number.
- "Do you classify the cited domain as mine, third-party, or same-name?" — separates your visibility from a competitor's.
- "What is the interval on that percentage, and how many runs is it based on?" — separates a signal from a single sample.
If a vendor answers all three in specifics, the product is probably real. If any answer is "it depends" or "roughly", the headline number is the product.
Last updated 2026-11-11. Our figures come from weekly runs on brandscope.dev (run 2026-11-02, five core surfaces, 55 answers). Method and known limits: [how it works](/how-it-works). Definitions behind these metrics: [what is AI visibility monitoring](/blog/what-is-ai-visibility-monitoring).