What is AI search visibility? How to measure it

What does AI search visibility look like for a real brand?

Here is ours. Attensira reads its own answers the same way it reads a customer's: 61 questions a buyer of an AI SEO product might ask, read on ChatGPT and Perplexity on a daily schedule. These are the 30 days from 4 September to 3 October 2026, read over the Attensira MCP on 4 October. Answers were stored on 24 of the 30 days; 4 September and 9 to 13 September have none, and ChatGPT was first read on 14 September. Those days are not measured, not zero. The table at the top of this page splits the same answers by whether the question contains our name; this one splits them by engine.

• ChatGPT

  Answers read

  2,274

  Named Attensira

  4.0%

  Cited attensira.com

  6.2%

• Perplexity

  Answers read

  2,100

  Named Attensira

  4.3%

  Cited attensira.com

  3.7%

• Both

  Answers read

  4,374

  Named Attensira

  4.1%

  Cited attensira.com

  5.0%

The band below counts the same 4,374 answers for the brands we track as competitors.

That is AI search visibility, measured. Three things in it are worth reading closely, because they are the whole subject in miniature.

The engines disagree. SE Ranking is named about three times as often on Perplexity as on ChatGPT, while Peec AI and Otterly AI are named more than twice as often on ChatGPT as on Perplexity. Our own citation rate is higher on ChatGPT, while our name rate is slightly higher on Perplexity. An average across the two engines would hide both facts.

The rates do not add up to a whole. Semrush, Profound, Ahrefs, Peec AI and Otterly AI are named in more than a fifth of the answers each. Several brands are named in one answer, and some answers name none. Each rate answers its own question: in what share of answers was this brand named?

Our number is small and honest. Every one of our mentions comes from three questions that contain our name. On every unbranded question, the answers named us zero times. We publish it because a page about measurement that hid its own baseline would not be worth reading.

Which number does "AI search visibility" mean on this page?

The definition at the top counts two events in an answer: your brand named, and your site linked. The assistants are the answer engines a person types a question into: ChatGPT, Perplexity, Claude, Gemini and Google's AI Overviews and AI Mode. The questions are the ones a buyer asks before choosing, written the way they would type them.

The phrase is vendor vocabulary. We could not trace it to a primary source, and different tools use it for different numbers. Some mean a single blended score. Some mean the share of all brand mentions that were yours. Some mean the rate at which your name appears in answers. This page uses it for the plain version, the share of answers that named or cited you, because that is the version that can be checked.

It is the measurement side of AI SEO. The work of changing it is covered there and in generative engine optimization and answer engine optimization.

Share of voice is the rate at which answers name you. An AI visibility score is the single blended number this page argues against relying on. AI brand mentions are the events share of voice counts. Citation position is where a link sits among an answer's sources.

Which AI visibility numbers can be measured honestly?

Four, each with its sample size.

Share of voice: how often you are named. Successful answers that named your brand, divided by all successful answers in the window. One answer that names you nine times counts once. In Attensira's definition it is a rate, not a slice of a pie, so the brands' rates do not add up to a whole, and adding a competitor to your list does not change your number.

Citation rate: how often your site is linked. Successful answers that attached at least one link to your own domain, divided by the same denominator. It is a different event from being named. An engine can name a brand it never links, and it can link a page on your domain in an answer that never names you.

Position: where you are named among the brands you track. Where your brand first appears in an answer, counting only you and your tracked competitors, always shown with its base, for example second of five tracked names. An answer that does not name you has no position. It is never position zero.

Sentiment: what answers say about you, when they say anything. Each answer is read for its stance on your brand. Only answers that take a position, praise, criticism, a recommendation, a reservation, are scored. An answer that names you without judging you, such as a price quote or a feature list, is left out rather than counted as neutral, so the score moves when opinions move and not when the mix of questions does.

Beside those four sits source share: which domains the answers cite, as a share of all citations. Unlike the brand rates, source shares do describe parts of a whole, because each link belongs to exactly one domain.

Which AI visibility numbers are constructed rather than measured?

A single visibility score. Blending engines, questions, mentions, citations and position into one number produces something that moves without a reason anyone can trace. Two engines can move in opposite directions and the score can stay still. If a tool shows one number, ask what it is made of and whether its parts come with their sample sizes.

A pie-chart share of voice. Your mentions divided by all tracked brands' mentions. Its denominator is the competitor list you typed, so adding a competitor lowers your share and removing one raises it. Nothing about your visibility changed. That is why Attensira computes share of voice against answers, not against brands.

Prompt volume. How many people ask a given question in an assistant. The assistants do not publish it. Attensira's own methodology says it asks prompts on a schedule and does not observe the questions real people asked, or how many asked them. A tool that shows a volume next to a prompt is estimating it from something else, and should say what.

Traffic and revenue from answers. Nothing joins an answer a model gave someone to a person who later visits your site. Referral visits from assistants are visible in analytics, but they undercount the answers that shaped a decision without a click. Pew Research Center found that Google users clicked a link inside an AI summary in 1% of visits to pages showing one. The rest of the influence leaves no trail.

Search engine reports, read as rankings. Microsoft's AI Performance report in Bing Webmaster Tools shows how often your pages are cited in Copilot and Bing AI answers. Microsoft says it does not indicate ranking, authority or a page's role within an individual answer. It is a useful citation count, not a position.

Why does AI search visibility move when nothing changed?

Three reasons, and our own data shows the biggest one.

The question set changed. Look at our ChatGPT share of voice day by day. From 14 to 23 September it read zero on every sampled day. From 24 to 29 September it read 6.8%, and 5.2% from 30 September, when more questions arrived and the same nine mentions were spread over more answers. Nothing about the answers to our existing questions changed on either day. We added three questions that contain our name, and they started producing answers. A workspace-level number moved because the sample changed, not because the world did. Any report that does not say when its question set changed cannot be trusted across that date.

Answers are not deterministic. The same question asked twice, minutes apart, can return different brands. A rate built from a handful of answers moves in large steps on its own. That is why every rate needs its sample size beside it, and why a change should only be reported when it clears a significance test. Attensira tests every change with a two-proportion z-test at 95% confidence and reports anything that does not clear it as "nothing we can prove yet".

Coverage changed. A day with no reading is not a day with zero visibility. If an engine was not read, or a reading failed, the honest value is "not measured", and Attensira reports it as a dash rather than a zero. A chart that draws missing days as zeros invents a drop that never happened.

How does mention detection affect the numbers?

Every share of voice depends on how a mention is detected, and the method has edges worth knowing. Attensira counts a mention as a case-insensitive whole-word match of your brand name, or a host-aware match of your domain, in the answer text. There is no entity resolution. A brand whose name is also an ordinary word, a brand called Linear, say, is counted when the word is used in its ordinary sense, so such a brand reads high, never low.

Two practical consequences follow. If your name is a common word, read a sample of stored answers before trusting the rate. And if a competitor's name is a common word and yours is not, the comparison between your rates is tilted in their favour. Attensira stores the verbatim text of every answer for exactly this reason: the stored answer is the evidence, and anyone can check what the model actually said.

How do you measure AI search visibility yourself?

You can run a defensible version in an afternoon and keep it up in an hour a week.

1. Write the questions. Twenty to fifty questions a buyer asks before they know your name, in their words. Add three or four that contain your brand, and keep them separate, because they measure something else.
2. Choose the engines. The ones your buyers use. Read each one separately; never average them.
3. Ask each question several times on each engine, and store the full answer text, not a screenshot.
4. Count four things per answer: were you named, was your domain linked, which tracked competitors were named, and which domains were linked.
5. Report rates with their sample sizes, per engine and per question group, for example "named in 3 of 40 ChatGPT answers".
6. Keep a log of changes to the question set, and never compare numbers across a change without saying so.
7. Repeat on a schedule, and call a difference real only when it is larger than the noise.

The GEO measurement guide is the long version, with chapters on building a prompt set and a defensible report.

How is AI search visibility different from a search ranking?

A ranking is a position in a list that stays put long enough to report. Rank three on Monday is usually rank three on Tuesday, and a rank tracker can check it once a day and be roughly right. An AI answer has no list to hold a position in. It is a paragraph, written fresh each time, naming a few brands and attaching a few links, or none.

That changes four things about measurement.

The unit is the answer, not the keyword. Visibility is counted per answer: named or not, linked or not. A question asked thirty times produces thirty observations, and the rate is the share of them that named you.

One reading is an anecdote. A rank checked once is a fact for the day. An answer read once is one draw from a distribution. You need many draws per question per engine before a rate means much.

The deciding pages are visible. A search ranking does not tell you which pages beat you. An AI answer often does: it lists its sources. That makes AI search visibility unusually actionable, because the sources behind a lost answer are a list of places to work on.

Branded and unbranded questions diverge sharply. In classic search, a brand usually ranks for its own name and for some of its category terms. In AI answers, our own data shows a brand can be named in every answer that contains its name and in none of the answers that do not. Report the two groups separately or the branded group will flatter the total.

What is a good AI search visibility number?

There is no universal benchmark, and anyone quoting one is comparing numbers built from different question sets.

A share of voice depends entirely on which questions you track. A set of easy questions, close to your brand, produces a high number. A set of hard, unbranded category questions produces a low one. Two companies with identical presence in the engines can report very different numbers because they asked different questions. That is why our own figures on this page come with the question set described and the branded questions called out.

The useful comparisons are narrower:

• You against your competitors, on the same questions and engines. The competitor table above is that comparison for us. It is meaningful because every brand in it was counted over the same 4,374 answers.
• You against yourself, on the same questions over time, with the question set unchanged and the difference tested for significance.
• One engine against another, for the same questions. A gap between engines is often the most actionable finding, because it points at the sources one engine trusts and the other does not.

A "good" number is one that is rising on a fixed set of questions that matter to your buyers, and that a significance test agrees is rising.

How many answers do you need before the number means something?

More than most dashboards show. A rate built from a handful of answers moves in large steps: with four answers, one answer changing its mind moves the rate by a quarter of the scale. Attensira's dashboard switches from a percentage to a count, such as "3 of 12", below 20 answers, and marks the count as thin, because a percentage implies a precision the sample does not have.

Sampling depth sets how fast evidence accumulates. Attensira reads each question once per engine per day. Starter samples each reading once and the higher plans three times, with the three runs issued at the same moment so the variation is the model's own. A seven-day window on one run a day gives seven observations per question per engine, which only detects large swings. Widening the window to thirty days, or pooling related questions into a topic, multiplies the answers behind each rate without changing the setup.

The rule of thumb is simple: read the sample size before the percentage, every time, and treat a movement as real only when a significance test says so.

What should an AI search visibility report contain?

At minimum, every number in it should arrive with the facts needed to check it. A report worth reading states:

• the question set, with how many questions, how they were chosen, and which ones contain the brand name;
• the engines and the surface, with each engine reported separately and whether the consumer app or the API was read;
• the window, with its dates and how many of its days were actually read;
• each rate with its sample size, for every engine and every group of questions;
• what was not measured, shown as not measured rather than as zero;
• every change to the question set during the window, with its date;
• each claimed change with its test, and "nothing we can prove yet" where the test was not cleared;
• what was done, the pages, passages and placements shipped, so a movement can be read against the work rather than in a vacuum.

The tables at the top of this page follow most of that list. The chapter on a defensible report builds a full report from our own data, including the parts that make us look bad.

Who should own AI search visibility inside a company?

Whoever owns search today, with one addition. The measurement fits naturally beside SEO reporting, because the inputs overlap: the same pages, the same crawlers, the same content calendar. The addition is the off-site half. A large share of what decides an answer is other people's pages, so the owner needs a way to work on reviews, comparisons, directories and communities, not only on the site.

In small companies that is usually the founder or the one marketer, which is exactly where the work stalls: the reading is easy, the doing takes hours nobody has. That gap is what an agency fills through people and what Attensira fills with a crew of agents that a person approves.

How should you read a vendor's AI visibility claims?

Five questions sort most claims.

1. What is the denominator? Answers, or brand mentions? A share of voice computed against brand mentions moves when the competitor list changes.
2. What is the sample size? A percentage without its n is not a measurement you can act on.
3. Which engines, and which surface? A rate read through a model's API can differ from what a person sees in the consumer app. Ask which one was read.
4. Is a missing reading shown as zero? If yes, every outage looks like a loss.
5. Was the change tested? "Up 12 points" from two small samples is one answer changing its mind.

Apply the same five questions to our numbers on this page. They should survive them.

What does Attensira do with AI search visibility?

Measuring is the first job, not the product. Attensira is an AI-native SEO agency run by software: the service of an SEO agency at the price of a tool. For this page's job, the number itself, the workspace reads your buyers' questions on each engine on a daily schedule and reports share of voice, citation rate, position and sentiment per engine, each with the number of answers behind it, and a day it could not read as not measured. The Analyst wakes every week to say what moved and whether the move is larger than the noise, including when it is not. When a rate falls, the rest of the crew takes the work that changes it, and every piece waits in your inbox until you approve it. The prompt tracking and citation tracking pages show the measurement side in the product.