AI Visibility Tools: What They Measure and What They Miss

A tool emails you a number. Your AI Visibility Score is 34. Or 61. Or a letter grade, or a dial sitting in the amber band. There is no prompt list, no engines named, and the only obvious next step is a demo call. The reasonable question is whether any of it means anything.

Mostly it is real, in the sense that something was genuinely counted. The problem is that what was counted is often not what you assumed, and the gap between those two things is where the money gets wasted. So rather than ranking tools, this is a guide to the three ways these numbers are produced, which of them constitutes measurement, and the questions that separate a credible figure from a decorated guess.

The short answer

That single test kills most of the market. It is not a question of sophistication or price. A simple tool that shows you forty prompts and the answers they produced is doing real work. An expensive one that returns a composite figure with no visible prompt set is not, however good the dashboard looks.

Three ways tools produce an AI visibility number

1. Querying an index of AI answers

Some platforms maintain a standing index of AI-generated responses, collected continuously across engines, and then search that index for brand mentions. Ahrefs Brand Radar works this way, and it is the approach we use ourselves.

The strength is scale and consistency: because the index is built independently of you, the same question asked in March and in September is answered against comparable data, which is what makes period-over-period comparison meaningful. The limitation is equally important and rarely stated. You are measuring appearances within that index, not across every AI answer generated worldwide. It is a sample, and a sample is a legitimate basis for a trend but not a census.

2. Running prompts live

Other tools send a defined list of prompts to the engines on a schedule and record what comes back. This is the most direct method available, and when the prompt list is yours it is also the most useful, because the questions match how your buyers actually ask.

Two things complicate it. AI answers are probabilistic, so the same prompt can produce different responses on different days. Run one prompt three times in a single week and it can swing from naming you in none of the three to naming you in two, with nothing having changed at your end. A tool reporting from a single run is reporting noise.

Answers are also personalised and localised, so a result captured from one location with one account history is not universal. Credible implementations run each prompt repeatedly and report a frequency rather than a yes or no.

3. Inferring from traditional signals

The third method does not touch an AI engine at all. It takes inputs the tool already has, such as your rankings, backlink profile, schema coverage, domain authority and page speed, applies a weighting, and presents the output as an AI visibility score.

This is where most free scores come from, and it is the category to be wary of. The inputs are real and they genuinely correlate with being cited, so the number is not arbitrary. But it is a prediction dressed as an observation. A site can score well on every input and still be absent from the answers that matter, because the engine chose a competitor’s page, or the question is answered without citing anyone, or your category simply is not being asked about. None of that is visible in an inferred score, and that is exactly the information you were trying to buy.

What is real measurement

Strip away the presentation and there are only four things worth knowing, all of them countable:

• Appearance rate. Of the prompts tracked, in what share does your brand appear at all? Per engine, counted separately, because appearing in Google AI Overviews and appearing in ChatGPT are different achievements earned in different ways.
• Source breadth. How many distinct pages of yours are being drawn on? One page carrying every mention is fragile. Twelve pages is a position.
• What is being said. Presence is not endorsement. An answer that names you as the expensive option is an appearance and a problem at the same time.
• Direction of travel. The same prompts, the same engines, measured again. A score without a prior period is a reading, not a finding.

Everything else on a typical dashboard is a recombination of those four. We cover how to establish each one in what an AI Visibility Score is and how it is measured.

What the tools miss

Even a well-built tool has blind spots, and knowing them is what stops you making bad decisions from good data.

Mentions without links. An engine can describe your firm accurately and send nobody to your site. That is a real outcome with real value, and it is invisible in analytics. We looked at this specifically in the citation gap.

Attribution to the wrong source. An answer can use your research and credit a publication that quoted you. The work travelled; the name did not. See ghost citations.

Prompts nobody tracks. Buyers ask long, specific, situational questions. “Best project management tool for a 12-person design agency already using Figma and Slack” is the kind of prompt that converts, and it appears in nobody’s generic category list. A prompt set of twenty category terms will miss the enquiry that was actually going to turn into money.

Why you are absent. This is the big one, and the reason a score on its own is close to useless. Every tool can tell you that you did not appear. Almost none can tell you which of four things caused it, and the four need completely different responses:

• The page is not crawlable. A firewall rule or a robots directive is blocking the AI user agent, so nothing you publish can be cited. This is an afternoon’s fix and it blocks everything else.
• The content is not extractable. The answer is on the page but buried in narrative, with no sentence an engine can lift and attribute. This is an editing job, not a writing job.
• Your entity is ambiguous. The engine is not confident who you are, so it declines to name you rather than risk being wrong. This is a schema and consistency problem, and it shows up first in branded prompts.
• A competitor is simply a better source. Everything works and you lost on merit. This is the only one of the four that costs quarters rather than weeks.

The figure is identical in all four cases. That is the whole problem with buying a number.

Seven questions that expose a weak tool

Ask these before paying for anything. A good vendor answers all seven without hesitation.

1. Which prompts did you run? Show me the list. If there is no list, there is no measurement.
2. Which engines, reported separately? A single blended figure hides which platform you are winning and which you are absent from.
3. How many times was each prompt run, and over what window? Once is noise.
4. Is this observed or inferred? Did you query an engine or index, or model the score from my SEO metrics?
5. What is the denominator? Thirty-four out of what? A percentage with no base is decoration.
6. Can I see the actual answers? The verbatim responses are more useful than the score, because they tell you how you are being characterised.
7. What would make this number go up? If the answer is a list of their services rather than a diagnosis of your absence, the score was a sales instrument.

Question seven is the one that matters most, and it is the same instinct that catches a weak agency pitch. We set out the broader pattern in warning signs of a bad SEO company.

What a credible number looks like

A defensible AI visibility figure is unglamorous. It reads something like this:

Nobody puts that on a landing page, because it is not a single satisfying number and it admits to zeroes. It is also actionable in a way that “34/100” never is. You can see that Gemini is untouched, that the content base is dangerously narrow, and that the trend is upward.

That is the honest state of this field in 2026. The measurement is young, the samples are partial, and anyone presenting a precise single figure with total confidence is telling you more about their sales process than about your visibility. Treat the score as the opening question, and insist on the diagnosis underneath it.

FAQ