TL;DR: An AI visibility score is the share of a fixed question set for which an engine cites your brand. Three things get called visibility, and they are not the same: appearing in the search results the model consulted, being named in the answer text, and being cited as a source. A single check is one sample from a probabilistic system, not a measurement. And a score is only as honest as its denominator: a rate that cannot tell you what it failed to measure cannot tell you when it is wrong. This is the method, and the questions that separate a usable score from a decorative one.
Contents
- What an AI Visibility Score Actually Is
- Three Things Called Visibility
- How to Run the Measurement Yourself
- Why One Run Is Not a Measurement
- Reading a Series Instead of a Score
- Why the Denominator Decides the Score
- What Actually Moves the Number
- Key Takeaways
- FAQ
What an AI Visibility Score Actually Is
An AI visibility score is the share of a defined set of buyer-relevant questions for which an AI engine cites or mentions your brand in its answer. That is the whole definition, and everything contentious about it lives in three words: defined, cites, and share.
There is no standard. Every vendor computes this differently, which means a score quoted without its method is not a number you can act on. Before you trust one, whether it comes from a tool or from your own spreadsheet, you need to know what counted as a hit, how many questions were asked, and how many of those questions actually produced a usable answer.
That last one is the part nobody checks, and it is the one that bit us.
Three Things Called Visibility
These get used interchangeably in vendor marketing and they are three different measurements.
Cited. Your URL appears as a linked source underneath or inside the generated answer. The engine chose you as evidence for what it said. This is the strongest signal and the hardest to earn.
Mentioned. Your brand name appears in the answer text with no link. Valuable, because the person asking reads it, but it does not send traffic and it is harder to attribute.
In the consulted set. The engine ran a search, and your page was among the results it looked at before writing its answer. This is the weakest of the three and the most commonly reported as visibility, because it is the easiest to detect and it produces the biggest number.
The distinction is not academic. It is routine for a domain to appear in the results an engine consulted for a query and be absent from the sources it actually cited in the answer. On a tool that reports the consulted set, that is a hit. In reality, nobody asking the question saw the name.
If a vendor cannot tell you which of the three their score counts, it counts the easiest one.
How to Run the Measurement Yourself
You do not need a platform to start. You need discipline about four things.
Fix the question list before you start, and do not edit it. Twenty questions is enough to be meaningful and small enough to sustain. Write them as a buyer would type them, including the messy ones. A question list you revise between runs measures your revisions.
Ask every question on every engine separately. ChatGPT, Claude, Perplexity and Google’s AI surfaces behave differently enough that an aggregate across them hides the entire story. Across the properties we monitor it is common for a single engine to account for most of a brand’s citations while another produces none at all for months. An averaged score erases that, and each engine implies different work.
Record the model, not just the engine. Model versions change and are retired. A result recorded against “ChatGPT” is not comparable to a result recorded six weeks later against a different underlying model. Store the model identifier with every single result.
Keep the raw responses. Not just the yes or no. The sources an engine cited when it did not cite you are the most useful output of the entire exercise, because they tell you who owns the answer you want.
Why One Run Is Not a Measurement
Generative engines are probabilistic, and most of them run a live web search whose results move between queries. Ask the same question twice and you can get different sources, different phrasing, and a different answer to whether you were cited.
So a single check tells you almost nothing. Ask each question several times per run and record the hit rate rather than a binary. Then compare run to run, not sample to sample. A brand that goes from cited to not cited in one check has not lost visibility. It has produced one sample from a noisy distribution.
This is also why the honest form of this metric is a series, not a number. A single score with no history is a screenshot of a process that moves.
Reading a Series Instead of a Score
Once you have several runs, the shape tells you more than any single figure.
A flat line is information, not failure of measurement. If the rate does not move across runs while you are publishing, the content you are shipping is not the content the engines are reaching for. That is a content signal, and it arrives long before any ranking report would show it.
Per-engine divergence is the most actionable pattern. One engine carrying most of a brand’s citations while another produces none for months is common, and it is not noise. Engines differ in which sources they trust and how aggressively they browse, so a gap on one engine points at a different fix than a gap across all of them.
Watch the direction, not the decimal. These are small numbers on most properties, and small numbers are jumpy. Three citations against one is not a trend. Three consecutive runs moving the same way is.
Record what changed between runs. A citation rate with no change log underneath it cannot be attributed to anything. Note the content you shipped, the schema you added, and the model versions you probed, so that when the line moves you can say why.
Why the Denominator Decides the Score
The question nobody asks a visibility tool is how many of its probes actually produced an answer.
Engines reject model identifiers when a version is retired. Rate limits bite. Requests time out. Any of these produces a probe that returned nothing, and how a tool treats that probe silently rewrites the score.
There are only two defensible options and one common mistake.
Counting a failed probe as “not cited” is wrong. It deflates the rate with data that does not exist, and the more unstable the engine, the worse your number looks for reasons that have nothing to do with your content.
Dropping it silently is worse. The denominator shrinks, and a smaller denominator makes the rate look better. One out of sixty reads higher than one out of eighty. A tool that quietly excludes what it could not measure reports its best numbers exactly when it is least able to see.
The only honest treatment is a third category. A failed probe is unmeasured, counted separately and reported alongside the rate. One in sixty means something completely different depending on whether twenty probes are missing, and if your reporting has no unmeasured count it structurally cannot tell you when it is wrong.
That is the question to put to any vendor: how many probes went into this number, and how many came back empty. If they cannot answer, the score is not one you can act on.
What Actually Moves the Number
From the properties we monitor and from reading who does get cited on the queries that matter, strongest first.
Do not block the AI crawlers. This sounds too basic to mention and it is the single most common defect we find. Managed platforms inject robots.txt rules that disallow AI user agents by default, and it is invisible until someone reads the file. We audit every property we run on a schedule for exactly this reason, having previously found it live on properties whose entire purpose was AI visibility.
Write paragraphs an engine can lift whole. The unit of citation is a self-contained passage that answers a question without needing the sentences around it. Content written to flow reads better to humans and cites worse.
Publish something only you can publish. This is the one that decides whether the rest matters. When we looked at who gets cited for our own category term, the sources were technology publishers, community forums and encyclopedic references. Not one competitor’s marketing page appeared. An engine has no reason to cite a fifth summary of material it already has from a better-known source. First-hand data, measured results, and specifics from real work are the only things that change that calculus, which is why this article contains our actual numbers rather than a description of them.
Make the claim machine-readable. Structured data that states the author, the date, and the question being answered is not a ranking trick. It is how a system decides whether a passage is attributable.
Key Takeaways
- An AI visibility score is a share of a fixed question set. Without the question count and the definition of a hit, it is not a number you can act on.
- Cited, mentioned, and present-in-the-consulted-set are three different measurements. Tools that do not distinguish them report the most flattering one.
- One check is one sample from a probabilistic system. Measure hit rates across repeats, compare run to run, and keep a series rather than a score.
- Record the model identifier with every result. A model change breaks comparability as completely as changing the questions.
- A failed probe is unmeasured, not un-cited. Any rate that cannot report what it missed cannot tell you when it is wrong.
- Judge any vendor’s score by its method, not its size: the question count, the definition of a hit, and how many probes came back empty.
Need help getting cited by AI engines?
We run this measurement on our clients’ properties every week. If you want to know where you actually stand in AI answers rather than where a dashboard says you stand, that is what our answer engine optimization service is built around.
Frequently Asked Questions
What is an AI visibility score?
An AI visibility score is the share of a defined set of buyer-relevant questions for which an AI engine cites or mentions your brand in its answer. There is no industry standard, so every vendor computes it differently: some count brand mentions in the answer text, some count your URL appearing as a linked source, and some count whether you appear anywhere in the search results the model consulted. Those are three different numbers and they can differ by an order of magnitude on the same brand, so a score is meaningless without knowing which one it is and how many questions it was measured across.
How do you check if AI mentions your brand?
Define a fixed list of questions a real buyer would ask, ask each one on each engine your customers actually use, and record whether your domain appears as a cited source, whether your brand is named in the answer text, and whether you are recommended as an option. Repeat on a schedule and keep every raw response. The discipline that matters is fixing the question list in advance: a list you edit between runs measures your editing, not your visibility.
How often should you measure AI visibility?
Weekly is enough for most brands and matches how quickly the engines re-evaluate sources. More important than frequency is consistency: the same questions, the same engines, the same models, on the same cadence. A model version change between runs breaks comparability just as thoroughly as changing the questions, which is why the model identifier should be recorded with every result rather than assumed.
Is being in the search results the same as being cited?
No, and conflating them is the most common way an AI visibility number gets inflated. Modern engines run a search, consult a set of pages, and then generate an answer that cites a subset of them. Appearing in the consulted set means the engine saw you. Being cited means the engine chose you as a source for what it actually said. Only the second one puts your name in front of the person who asked, and tools that report the first as visibility are measuring reach into a process rather than presence in an answer.
Why does the same question give different AI answers?
Generative engines are probabilistic, and most of them run a live web search whose results move. The same question asked twice can produce different sources, different phrasing, and a different answer to whether you are cited. This is why a single check is not a measurement: it is one sample from a distribution. Ask each question multiple times, record the hit rate rather than a yes or no, and treat any single run as noisy.
What actually improves AI visibility?
Four things, strongest first. Being crawlable by the AI agents themselves, which means not blocking them in robots.txt, a mistake that is far more common than people expect. Writing self-contained, quotable paragraphs that answer a question completely without requiring surrounding context, because that is the unit an engine lifts. Publishing specific, verifiable, first-hand information rather than a rewrite of what already ranks, since an engine has no reason to cite a fifth summary of the same material. And structured data that makes the claim, the author, and the date machine-readable.