How to Measure AI Visibility in 5 Steps

Here's how to measure AI visibility: count how many AI answers name your brand out of all the answers you collected, using the same questions every time. The method is short. The discipline is the hard part.

One answer proves little. Asking ChatGPT once is like tasting one spoonful of soup and rating the restaurant. Nobody sends you a report of what the engines said, so you build the report yourself.

Five steps get you a number you can explain, and a clear view of where it stops being reliable.

Key takeaways

• Measure AI visibility by counting answers that name your brand over all the answers collected, not by reading one answer.
• Fix the question list and the conditions first, because changing either changes the reading.
• Report a count with its base ("named in 56 of 160 answers"), and call a change only after repeat runs.
• AI visibility data shows what answers say. Clicks, sales and accuracy need other tools.

How to measure AI visibility step by step

Measuring AI visibility means counting how many AI answers name your brand, and how many link your site, out of all the answers collected for a fixed list of buyer questions. You repeat the count on a schedule and read the trend, not any single run.

The method has five steps:

1. Fix the questions. Build one list of buyer questions, tag each by funnel stage and keep the wording.
2. Fix the conditions. Ask the same engines, from the same place, logged out, every time.
3. Give every answer a verdict. Named and linked, named, linked but not named, or absent.
4. Count over a base. Divide the answers that name you by the answers collected, then do the same for each rival.
5. Call a change late. Wait for repeat runs and a gap larger than the sampling range.

Each step has a reason. Questions come first because the rest is arithmetic on their answers, and conditions come second because a signed-in answer and a logged-out answer can be two different measurements. Verdicts and a base turn answers into counts, and repeat runs show whether a count moved or just wobbled.

There's no blended score in that list, and that's deliberate. Firemark reports counts instead of one blended score, because a count carries its base on its face. A score hides it.

How do you fix the questions?

Fix the questions by writing one list of buyer questions, tagging each by funnel stage, and asking the same wording on every run. The list is your measuring instrument, so changing it changes the reading.

Tag each question top, middle or bottom of the funnel. A brand can be named when buyers compare options and missing when they first research, and one blended share hides that. Firemark tags every prompt this way for the same reason.

Say you sell project-management software. "How do agencies keep track of client work?" sits at the top of the funnel, and "best project management tool for a 20-person agency" sits at the bottom. Being named for one says little about the other.

Size matters too. Firemark shows a percentage only when the list holds at least 20 prompts, so below that, report counts. Changing more than a quarter of your prompts starts a new comparison baseline, which is why you edit the list slowly.

The guide to building a prompt set covers where the questions come from and includes a template.

Then fix the conditions. Firemark asks each prompt on each engine from a US location, in English, from a logged-out session. Depending on plan that's up to five engines: ChatGPT (the web app, not the API), Google AI Overviews and Google AI Mode on every plan, Gemini from Starter and Perplexity from Growth.

A signed-in person with history can get a different answer, so a logged-out reading is a baseline, not a mirror of any one buyer.

That baseline is a real way people use the product. OpenAI's help page says people who aren't signed in can also use ChatGPT's web search, and that a Sources view, when available, lists the cited sources and other relevant links.

How do you count one run?

Count one run by giving every answer a verdict, then dividing the answers that name your brand by the answers collected. Do the same count for each rival.

Named means your brand's name is in the text of the answer. Linked means the answer's sources include your site. The two are separate, which gives four verdicts: named and linked, named, linked but not named, and absent.

A name that appears only inside a list of sources doesn't count. A failed answer is never counted as absent either: if an engine doesn't answer, the gap stays a gap.

Answers per run are prompts times engines. Take 40 prompts on 4 engines: 160 answers. The numbers below are invented sample data, not a real brand.

Named is 30 plus 26, so the brand is named in 56 of 160 answers, a 35% share. It's linked in 38 of 160, or 24%. Those are two different findings, and a single "mentions" number would have blurred them into one.

The 96 absent answers are 96 chances where the buyer was never told your name.

Now the same run by engine, each on its own base of 40 answers.

The blended 35% hides a 20-point gap between the strongest engine and the weakest.

Rivals get the same count. If a rival is named in 72 of the 160 answers, it holds 45%. Firemark shows a rival named in fewer than 10 answers as "fewer than 10" and gives it no rank.

An engine that answers under 95% of your prompts is left out of that run's totals. A broken engine then can't masquerade as a brand that's absent.

How a share is divided matters more than it looks. The guide to AI share of voice works one dataset through two denominators and shows why the totals disagree.

When is a change real?

A change in AI visibility is worth acting on only after repeat runs and a gap clearly larger than the 95% range. Firemark doesn't run a significance test. It reports a movement after three comparable runs and a change of 5 points or more, which is a rule of thumb, and on a small list 5 points can still sit inside the range.

Engines rarely give the same list twice. SparkToro's study, run with Gumshoe, an AI tracking vendor the post discloses, had 600 volunteers put 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. It found less than a 1 in 100 chance that ChatGPT or Google's AI gives the same list of brands twice. Identical lists are rare, but a brand's appearance rate was often steady: one brand appeared in 85 of 95 Google AI responses to a single prompt.

Its author still concluded that visibility percent across dozens to hundreds of prompts, run multiple times, is a reasonable metric, and that a "ranking position in AI" is not.

The study was small, left open how many prompts and runs are enough, and used volunteers' own default accounts. Its author suggests asking each prompt at least 60 to 100 times to see a full set of brands. Firemark's 20 prompts and three runs are practical floors, not statistically derived ones.

List size sets the noise. On one engine, a 40% share on 25 prompts (10 of 25 answers) has a 95% range of 23% to 59%. The same share on 500 prompts (200 of 500) narrows to 36% to 44%, figures that match the Wilson method for a proportion, which NIST's handbook describes. Firemark draws that range per engine on its own base, and it covers sampling only, not an engine changing how it answers or which prompts you chose.

Keep the conditions steady between runs. Changing more than a quarter of your prompts starts a new baseline, and a plan change that adds an engine restarts the overall trend from that run.

On a weekly schedule, three runs means two more weeks of patience after the first. That's slow, and it's cheaper than explaining to leadership why last Tuesday's three-point wobble was never a trend. Every one of these rules is written out on Firemark's method page.

By hand or with a tool?

Measure by hand for a first look, in a sheet for a list you'll keep up, and with a tool when you add engines, prompts or a history. The test is how many answers you can read and log on schedule before the logging stops.

A hand check suits ten buyer questions on one engine, reported as counts rather than a percentage. It can't reach 20 prompts on 5 engines, which is 100 answers per run, repeated three times, without taking much of your week. A sheet holds up until keeping the log becomes the job instead of reading it, which tends to happen around the second tab named "final_v2". A tool adds the schedule, and Firemark keeps full answers and sources for 13 months. The guide to tracking brand mentions in ChatGPT walks through the hand and sheet routes.

Another tool fits better than Firemark in three cases. If you need one weighted score for a board, Firemark records which third of an answer holds your first mention, and its sentiment, but ranks brands by how many answers name them and reports the share as a plain count. If you sell outside the United States, Firemark asks only from a US location in English. If you need Claude tracked today, it isn't yet.

What can't this method show?

AI visibility measurement shows what answers say, not what buyers do next or whether an answer is right. Firemark doesn't track clicks, visits or sales that follow an answer, or whether what an answer says about you is accurate.

Google covers part of its own side. Search Console has a dedicated Generative AI performance report, available to all sites worldwide as of August 31, 2026, after Google's June announcement. It shows impressions of links to your site in AI Overviews and AI Mode by page, country, device and date, and its data comes from the Web search type in the main Performance report. It counts impressions of your links, which is a different question from whether an answer said your name.

Google counts a click on a link to an external page in an AI Overview or AI Mode as a click, according to Search Console Help, but the AI report itself shows impressions only. Your analytics holds what visitors did next.

Three gaps remain. Personalized or logged-in answers aren't collected. Countries and languages other than the United States and English aren't covered. Claude is added to Pro when it ships.

Every run is a sample, too. Another phrasing can get another answer, which is why the method fixes the wording and reads the trend.

Where to start

Start with the list. Write 20 buyer questions, fix the conditions and count one run before you read anything into it. One spoonful proves little, and a full list of questions on a schedule proves a lot more.

Firemark runs this method on up to five engines, depending on plan, and its MCP server lets Claude or ChatGPT read the results. The demo shows a full report on sample data.