How to Measure AI Visibility Across ChatGPT, Claude, Gemini and Perplexity

Key takeaways

The short version of this page, in the order the sections below develop it. Each point is stated here in full so it can be read without the rest of the page, and the section that argues it gives the measurement and the source behind it.

• Measure it as a rate, not as a screenshot: take a fixed list of unbranded buying questions, ask each one several times to several assistants, and…
• Two from the first run, Mention Rate and Model Agreement, with two more recorded alongside them.
• In eight steps: five that hold conditions still, then three that produce numbers you can compare.
• Because the assistants do not name the same companies, so a result from one of them does not tell you what the others say.

How should a B2B company measure its visibility across AI assistants?

A B2B company should measure its visibility across AI assistants as a rate, not as a screenshot: take a fixed list of unbranded buying questions, ask each one several times to several assistants, and count how often each assistant names your company. We use two numbers. Mention Rate is the share of your buyers' real questions in which an AI assistant names your company. Model Agreement is, of the companies named by at least one assistant, the share named by all of them. The first says how often you are on the shortlist; the second says how far a result from one assistant holds in the others.

Each Mention Rate here is computed over repeated runs: every run counts as one asking, and the number of runs is written next to the rate. To see only what one assistant says today, the ten-minute manual check is enough.

Which AI visibility metrics are worth tracking?

Two AI visibility metrics are worth tracking from the first run, Mention Rate and Model Agreement, with two more recorded alongside them. All four come from a single log of answers rather than four separate exercises, which is what makes them affordable to keep. Mention Rate is the share of your buyers' questions in which an assistant names you; Model Agreement is the share of named companies that every assistant names. The table below gives each one's formula and, as importantly, what it cannot tell you.

How do you run the measurement, step by step?

You run the measurement in eight steps: five that hold conditions still, then three that produce numbers you can compare. This is the procedure behind our S&P 100 study, scaled down to one company.

1. Write the question list. Collect what a buyer asks when choosing a vendor in your category: "which companies should we consider for X". No company names, yours included: a question that names a brand returns an answer about that brand.
2. Fix the assistants. We use ChatGPT, Claude, Gemini and Perplexity. Note which ones search the live web; a browsing answer and one from learned knowledge are different measurements.
3. Ask in a clean session. A fresh conversation with memory and custom instructions off, or a logged-out window.
4. Repeat every question. At least three times per assistant.
5. Log every answer. One row per answer: question, assistant, model name, date, full text, and whether it was usable. An error, a refusal or an empty reply is a failed measurement, not a result.
6. Count names. Decide in advance which name variants count, then mark which companies each answer names.
7. Compute the two metrics. Mention Rate per assistant, over all runs, written with its conditions, for example "named in 18 of 72 answers; 24 questions, three runs each". Then Model Agreement, by the procedure under the study table.
8. Re-run unchanged. Same questions, same assistants, same number of runs, new date.

Twenty-four questions, four assistants and three runs is 288 answers.

Why is one assistant not enough?

One assistant is not enough because the assistants do not name the same companies, so a result from one of them does not tell you what the others say. In our S&P 100 study, 32 of the 48 companies that at least one of four assistants named were named by all four. That is the study's Model Agreement. One limit belongs beside that figure: in the industries analysed, ChatGPT returned usable text in 54 of its 96 answers, against 96 of 96 for each of the other three. Fewer ChatGPT answers pushes agreement down, so 32 of 48 likely overstates the disagreement, and we cannot say by how much.

The study asked four unbranded buying questions per industry, three times each, to one model from each of ChatGPT, Claude, Gemini and Perplexity, in September 2026, and matched the answers against the S&P 100 list. Eleven industries were asked and eight could be analysed, with 342 usable answers in those eight. The assistants also differed in how many distinct S&P 100 companies they named, including the three that answered every question.

Two more limits apply. The models were called through an API, not the consumer apps, and only Perplexity's searched the live web, so this measures learned association, not what a consumer app with browsing returns today. Financials, Health Care and Information Technology are excluded because too many of their answers failed.

What does one company do with Model Agreement?

Model Agreement describes a set of companies, so one company cannot have one. Compute it for your category from your own log, then read your own row. This procedure is our method, not an industry standard.

1. List every company named in any usable answer.
2. Mark a company as named by an assistant if at least one usable answer from that assistant, across all its runs, names it.
3. Divide the companies named by every assistant by the companies named by at least one. That is your category's Model Agreement.
4. Write down which assistants name you and which do not.
5. Report your highest and lowest Mention Rate per assistant, never only the average.
6. Count usable answers per assistant before reading a gap as real. An assistant that failed often looks like one that disagrees, as ChatGPT did in our study.

Low agreement means a reading from one assistant cannot stand in for the rest. High agreement with your name missing from one assistant is a specific gap to investigate. We have no tested threshold for "low"; compare against your category's previous run.

How many questions and repeats do you need?

You need at least two dozen questions, each run at least three times per assistant. The first number is the low end of SparkToro's "dozens to hundreds of prompts"; the second is the floor we used in our study, not a tested threshold. Two runs of the same question rarely return the same list. In a SparkToro experiment from November and December 2025, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI results a combined 2,961 times.

The source is a blog post, not a peer-reviewed paper, and was run with an employee of Gumshoe, an AI tracking startup; the post raises that conflict of interest itself. Its author writes that "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric", and that a ranking position in AI is not. So we count whether you are named and ignore where.

We know of no published answer to how many runs are enough; the post lists it among its open questions. To learn one assistant's set of recommendations for one question, its author's estimate is "usually at least 60-100X". If a decision rests on one question, run it far more than three times. By the same standard our study was thin, at four questions per industry: enough to show that assistants disagree, not enough to rank one company against another.

What can analytics and Search Console show, and what can they not?

Analytics and Search Console show where links to your site were displayed and which visits arrived; they cannot show an answer that named you, or a competitor, without a link and without a click. Google's guide to generative AI features, last updated 2026-07-10, says:

The report's help page says it covers AI Overviews and AI Mode and, as of August 31, 2026, is rolled out to all websites worldwide. It counts impressions, defined as "how many times links to your site were shown to a user in a generative AI feature on Google Search", grouped by page, country, date or device. Two limits follow:

• It counts links, not names. The page describes no count of answers that name a company without a link.
• The page lists no grouping by query. Its dimensions show which of your URLs were shown, not which buyer question produced the answer.

Microsoft's AI Performance in Bing Webmaster Tools, announced as a public preview on February 10, 2026, "provides a consolidated view of when your site is cited in AI answers".

Neither page mentions ChatGPT, Claude or Perplexity, and neither says its report shows which competitors were named beside you.

Referral visits in analytics undercount too, because clicks are rare. Pew Research Center tracked 900 US adults across 68,879 Google searches in March 2025; when an AI summary appeared, users clicked a link inside it in 1% of visits. Pew measured Google's summaries, so we do not know the rate for ChatGPT or Claude.

When does this measurement mislead you?

This measurement misleads you in four situations, each one making a rate look more precise than it really is. The questions can be too broad for your category, so nobody in it is ever named. The session can be unclean, with memory and location shaping the answer. The model itself can change between runs, moving a number you then attribute to your own work. And a mention can be read as a recommendation, which it is not.

• The questions are too broad. None of our study's Industrials questions asked about construction equipment, so a maker of it could look absent without being absent.
• The buyer's session is not clean. Memory, location and earlier turns change answers; your rate estimates the default one.
• The model changes. A jump between two runs can be the model, not your work. Log the model name every time.
• A mention is read as a recommendation. Being named is the start of a shortlist, not a ranking, a click or a sale.

To have the list re-run for you, the options are a tracking product, covered in our comparison of tracking tools, or a service. UPAI is a service: we run this check by hand and hand over every question and answer.

Questions people ask next

These are the follow-ups we are asked most often once the answer above has landed. The same rule applies to each: where a figure appears, its source and year sit in the same sentence, and where we do not know something we say so rather than estimate it.

How do I track AI visibility over time?

Re-run the identical question list on the same assistants with the same number of runs, and store every answer with its model and date.

What is a good AI visibility score?

We know of no honest public benchmark. A Mention Rate depends on how broad the questions are, so compare only against competitors named in the same answers and against your own previous run.

Does Google Search Console show AI visibility?

Partly. Its Generative AI performance report counts the times links to your site were shown in AI Overviews and AI Mode. Its help page describes no count of answers that name you without a link, and the report covers Google Search only.

Related

Other questions in this library that cover the ground next to this one. Each is answered the same way — the direct answer first, then the measurement behind it — and each carries its own sources, so you can check any of them on its own page without taking this one on trust.

• How do I know if ChatGPT mentions my brand?
• What are the best tools to track brand visibility in AI search?