AI visibility is how often, and how well, AI assistants name your business when someone asks them a question you could answer. If a buyer asks ChatGPT for "the best payroll software for a ten person agency" and the answer lists four products, each of those four has AI visibility for that prompt. Everyone else has none, however well they rank on Google for the same words.
Measuring it is less obvious than it sounds, because an assistant does not return a ranked list of ten links the way a search engine does. It writes an answer, names a few brands, and sometimes links to a few sources. This guide covers what to count, why a single score hides more than it shows, and how to measure it in a way you can defend.
We make Seen42, an AI visibility tool, so read our product notes with that in mind. Everything else in this guide works without it.
Mentioned and cited are two different things
An answer can name you without linking to you, and link to you without naming you. Keep the two apart.
A mention means the answer names your brand. It is what a buyer notices first, so most people care about it most. A citation means the answer lists one of your pages among its sources. Citations matter because they show the assistant actually read your page, and because some assistants put the links where people click.
The two move independently. A review site can be cited while your brand is only mentioned. Your pricing page can be cited while a competitor is mentioned first. Tools that fold both into one number make it hard to tell which problem you have.
What a visibility score usually means
Most tools report a percentage: the share of answers, across your tracked prompts and the assistants you check, that mentioned you. Some add position (were you the first brand named or the fourth), share of voice against named competitors, and a separate citation rate for your own pages.
A score of 40 percent on ten prompts across five assistants means you were named in about 20 of 50 answers. That is useful. It is also fragile, for three reasons.
First, assistants answer differently from one run to the next. Ask the same question twice and the list of brands can change without anything on the web changing. Second, assistants fail: a request can time out, return an empty answer or get cut off. Counting a failed answer as "not mentioned" drags the score down for no reason. Third, your score depends entirely on which prompts you chose. Ten branded prompts will give you a flattering number that says nothing about new buyers.
Which prompts to measure
Measure the questions a buyer asks before they know you exist. For most small companies that is three groups.
- Category prompts: "best [category] for [audience]" and "which [category] should a [type of company] use".
- Comparison prompts: "[competitor] vs [competitor]" and "alternatives to [competitor]".
- Problem prompts: the job the buyer wants done, with no product named at all.
Add two or three branded prompts ("what are the pros and cons of [you]") so you can see how answers describe you once they name you, but keep them out of the headline number. On a check we ran on 2026-10-20, asking five assistants "What are the best AI visibility tools for small businesses?" produced lists in which three of the same tools were named by four of the five assistants. Category prompts like that are where a small brand is usually missing, and where measurement earns its keep.
Which assistants to measure
ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews behave differently. In our checks Perplexity cites ten sources for most answers, ChatGPT often cites two or three, and Claude sometimes answers from its own knowledge with no sources at all. Google AI Overviews only appears for some questions, and in our checks two of ten requests to it failed outright.
Measure the assistants your buyers use, and look at them separately. A brand can be named in every Perplexity answer and in no ChatGPT answer, and the fix is different for each.
How to measure it without fooling yourself
- Fix the wording. Write each prompt once and never edit it in place. If you must change the wording, retire the old prompt and start a new history for the new one.
- Repeat on a schedule. Weekly is enough for most small companies. Compare trends over four to six checks, not two single results.
- Count mentions and citations separately, and show the sample size next to every number.
- Leave out failed answers. A timeout, an empty answer or a cut-off answer is missing data, not evidence of absence.
- Keep control prompts. Pick a few prompts you will not build pages for. If your targeted prompts improve and the controls do not, you have a reason to believe your work did something. If both move together, the assistants changed, not you.
- Record the sources. For each answer, store which pages were cited. The list of sites assistants trust for your prompts is your publishing plan.
This is observational measurement. It tells you what the assistants said on the days you asked, under the conditions you asked. It does not prove that a page caused a change. It is still far more than most teams know today.
Doing this by hand
You can do all of the above in a spreadsheet. Write 20 prompts, ask each assistant in a fresh session every Monday, and record for each answer whether you were named, in which position, which competitors were named, and which URLs were cited. For five assistants that is 100 answers a week, which takes an afternoon and is worth doing at least once to see the shape of the problem.
Where Seen42 fits
Seen42 checks the prompts your buyers ask on ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews, shows who is named and cited instead of you, and builds and hosts comparison, facts and guide pages from your own facts. Paid plans start at $39 a month for 15 prompts; the Free plan checks 10 prompts once on ChatGPT and Gemini. It keeps mentions and citations separate, excludes failed answers from the denominator, keeps control prompts out of the headline number, and marks on the trend when each page went live, so you can read a change in context. The methodology page explains each of those choices.
