An AI visibility score is only as useful as the decisions behind it. Before any number appears, someone has to decide what counts as being visible in an AI answer, which answers matter more than others, and how much uncertainty to admit. This article explains the signals we observe in each answer, how we weigh them into a single score, what we deliberately leave out of that score, and what our team is still working to improve.

We will describe the reasoning, not every constant. The meaning of the score, the conditions it is measured under, and its limits are public on our Method page. The exact weights and statistical settings stay internal, partly because they change as we learn and partly because they are the product of a great deal of measurement work. What follows should be enough to read the score correctly and to judge whether it measures the right thing.

Update, 26 October 2026: the AI score is now a site’s place against the site the sampled answers mention most, read from the same ranking as its rank. It is no longer the top of an estimated range. We revised the sections on the weighing, the range, and what stays out of the score, and the FAQ. The Method page always describes the current score.

What does an AI visibility score measure?

It measures how often a site appears when an AI assistant answers questions that a potential customer might ask, and how prominently it appears, read against the other sites in the same answers. That sounds simple. In practice it raises three questions that every generative engine optimization tool has to answer, whether or not it says so.

Which questions? We separate questions that do not name the brand from questions that do. A question such as “what is the best invoicing tool for freelancers” is a discovery question: the asker has a need and no answer yet. A question such as “is this invoicing tool reliable” already names the company. The first kind is where new customers are won, so it is the population behind the AI score. The second kind feeds a separate Brand score, because a site that is always mentioned when its own name is in the question has not shown that AI systems recommend it to strangers.

Which answers? We measure answers the engine produces when it searches the web. That is the setting in which a site can be cited with a link and send a visitor. The score does not claim to describe every AI conversation, including answers produced from the model’s memory alone.

Appearing how? This is the part that most deserves explanation, because “the site was in the answer” covers very different outcomes.

What do we observe in each answer?

For every answer, we record how each site appears in it. Four kinds of observation matter.

  • Linked. The answer cites a page on the site, with a link a reader can follow. This is the strongest form of visibility, because it can send traffic and it shows the engine treated the site as a source.
  • Named. The answer mentions the site or its brand in the text, without a link. The reader learns the name, but the answer does not hand them a way in.
  • Read. The engine retrieved a page from the site while building the answer but did not cite it. This is invisible to the reader, but it tells the site owner that the engine found the page and chose something else.
  • Placement and treatment. Where the site first appears, whether it is in the opening of the answer, its place in any list, which third-party page is cited next to it, and whether the answer recommends it, simply lists it, compares it with alternatives, or speaks against it.

Recognizing a name sounds trivial and is not. Sites share words with ordinary language, brands collide, and an answer full of links contains domain names that are not mentions at all. Our detector is built to count only names that a reader would take as a mention, and each of its rules came from finding a real answer where a simpler rule was wrong.

How are the signals weighed into one score?

We combine these observations across many answers to decide how much each answer gives each site. Three weighting decisions shape it.

A link counts more than a name. A linked citation counts in full. A name without a link counts for less. A mention has value, since the reader now knows the name, but it does not generate a visit and it shows weaker endorsement from the engine. Pages that are read but not cited do not count toward the score at all, because the reader never sees them. They are reported separately as a diagnostic.

Popular questions count more than rare ones. Not every discovery question is asked equally often. Questions that people type more often carry more weight than questions few people ask, so that a site cited for the questions its market actually asks scores higher than a site cited only for obscure phrasings. We estimate popularity from public search behavior rather than from any private user data, and we cap how far one question can outweigh another, so a single popular query cannot dominate the result.

Engines are weighed by their role. Each AI engine carries its own weight, and its answers count in proportion to it. Today the measurement uses OpenAI’s API with web search. Our scope follows the assistants people use: every engine with 5% or more of the AI user share, plus Google’s AI features, which run inside Google Search and have no share of their own. Engines in that scope are added one at a time.

With those weights, every site that appears in the answers has a total. The AI score compares that total with the leader’s: a number from 0 to 100 that says how close the site comes to the site the answers give the most. The leader scores 100, and a site the answers never mention scores 0.

Why do we show a range instead of a single number?

Because AI answers vary. Ask the same question twice and you can get a different set of sites, in a different order. Any score built from a sample of answers is an estimate, and pretending otherwise would mislead.

So every AI score comes with an estimated range. The displayed number is our best single reading of the site’s place, and the range shows how far sampling alone could move it. A result near either end of the range is just as consistent with the measurement.

Until 26 October 2026 the displayed number was the top of the range. We chose that because site owners found that a point estimate, which usually sits below the top, did not match what they saw when they asked the assistant themselves. We replaced it because the number should agree with the rank. The site ranked first now reads 100, and every other site reads by its distance to that one, which also answers the original complaint: a site no longer reads low while it leads.

When a site has few answers, its estimate is less certain, and the range widens to say so. Sites measured in a shared market draw on the answers of every site in that market, which gives each of them a tighter estimate than they could get alone.

What is deliberately left out of the score?

Several things we measure carefully do not enter the AI score, and the reasons matter as much as the formula.

Rank and share of voice, until 26 October 2026. For every site, we show its rank among the competing sites that appear in the same answers and its share of voice, its slice of all mentions in that set. Your site and every competitor are counted by the same rule. Until that date we kept these figures out of the score, so that a competitor’s change could not move it. Since then the AI score is read from this same ranking: the site ranked first scores 100, and every other site scores by its distance to it. We changed this because a score that sat apart from the rank left sites that rank first with a low number, and sites nobody mentioned with no rank at all. The cost is that the score now depends on the field. If the leader is mentioned more often, or a stronger competitor appears, your score can fall without your site changing. If every competitor disappears from AI answers, your score can rise without your visibility changing. That is why we show the rank, and how many sites it is out of, next to the score.

Comparisons follow the method used for your latest measurement. When that method changes, your existing comparison remains until your site is measured again; results from the previous method are not mixed into the new comparison.

How you are presented. Whether an answer recommends a site or merely lists it is valuable context, and we record it for every site in every answer. Treatment is judged by a language model, one reading per answer, so it has its own error rate. Until we can show how it relates to real outcomes, it stays in the analysis rather than in the score.

Page readiness. We also assess how well a site’s pages match what cited pages tend to have. That assessment is about likely causes, and the score is about observed outcomes. Mixing the two would make it impossible to tell whether a change helped.

The effect benchmark. Our public benchmark is a separate measurement. It asks which changes actually raised citations across many sites, comparing sites that made a change with comparable sites that did not. It uses its own published method and version, and it does not alter anyone’s score.

How do we keep the score honest?

A score earns trust by holding up when someone checks it. A few commitments follow from that.

  • Same rules for everyone. You and your competitors are counted by identical rules over the same answers, so a comparison is a comparison.
  • Fixed language and region. A site is measured on questions in its own language, and sites are never ranked across languages, because answers in different languages come from different conversations. Searches run from the United States.
  • A deadline, not a delay. A measurement started from the browser returns within a minute. If the deadline cuts some answers off, the score is computed from those that arrived and marked as partial, rather than making the visitor wait.
  • Versioned detection. When we change how a mention is recognized, the detector gets a new version, and the public benchmark gets a new version with it, so numbers from different methods are never silently mixed.

What is our team working to improve?

The measurement is not finished, and we would rather say so than imply otherwise. Most of what we changed this autumn came from finding cases where the score was technically correct and practically misleading.

Counting competitors the way we count you. Until recently, share of voice counted a competitor only when it was linked, while your own visibility also credited unlinked mentions. That made the comparison uneven. Both sides now use the same rule.

Recognizing names more strictly. Our third version of the mention detector removed several classes of false mentions we had found in real answers.

Reading the whole answer. We now record every site in every answer, how it appears, where, through which page, and how the answer treats it. None of it enters the score yet. The next step is to compare these readings with real traffic from sites that connect their analytics, and to give a signal weight only once it predicts something.

Next. Measuring more of the AI surfaces our customers care about, and calibrating our weights against the referral traffic that participating sites share with us. Each site that takes part sharpens the estimate for everyone in its market. That is the main reason we keep measurement free.

FAQ

A linked citation can send a visitor and shows the engine treated the page as a source. A name alone tells the reader the brand exists but offers no path to the site. Counting a name for less than a link credits awareness without treating it as equal to a citation.

Why is my score different from what I see in ChatGPT?

The score summarizes many sampled answers to questions that do not name your brand, measured through the engine’s API with web search from the United States. What you see in a consumer app is one answer to one question, possibly shaped by your history, settings, or location. A single answer is one draw. The score describes the distribution.

How does my rank relate to my AI score?

They are read from the same ranking of sites in the same answers. The site ranked first scores 100, sites tied for first all score 100, and every other site scores by how close it comes to the first. A higher rank therefore means a higher score. Because the score is a place against the other sites, it can also change when the field changes, even if your site does not.

Why don’t you publish the exact weights?

We publish what the score means, what it covers, the conditions it is measured under, and its limits, so that anyone can read a result correctly. The exact weights and statistical settings change as the method improves, and they are the result of our own measurement work. Our effect benchmark, which makes public claims about what works, is the exception: its method is published in full so the claims can be checked.