How to measure AI search visibility: metrics, KPIs and formulas

How to measure AI search visibility

Write down the questions a buyer asks before they know your name, put each one to ChatGPT, Claude, Gemini and Grok several times under the same settings, and count the answers that name you. Divide by every answer received and you have the AI Recommendation Rate. Report it per assistant with a 95% interval, track citations and accuracy beside it, and add the free reports from Google, Bing and your analytics for the traffic side.

The method in seven steps:

1. Write 10 to 40 buyer questions in your customers' words and language, such as "which accountant in Leeds is good for a small limited company". Keep any question that contains your name in a separate group, because it inflates the result.
2. Fix the settings and write them down: which assistants, app or API, web search on or off, signed in or out, language and location. Change one and this month stops being comparable with last month.
3. Ask each question to each assistant at least three times. One answer is a single draw: in the AIGROW Index, an assistant asked the same question again repeated 56% of its names.
4. Save every answer in full with the date, the assistant and the settings, so anyone can recount it later.
5. Mark each answer: named you or not, linked to your site or not, described you correctly or not, and which competitors it named. An answer that names nobody stays in the count; a call that failed is left out.
6. Work out each rate per assistant first, then overall, each with its 95% interval.
7. Repeat every month with the same questions and settings, and put the Google, Bing and analytics figures next to the rates.

The metrics and their formulas

Each metric answers a different question, so keep them on separate lines instead of blending them into one score.

Three counting rules decide whether the numbers mean anything. Count an answer once, however many times it repeats your name. Keep answers that name nobody in the denominator, because a buyer who got no name was sent to no one.

Leave failed calls out, because no answer was given. These are the rules behind the AI Recommendation Rate definition, and they are why the rate and share of voice differ: share of voice only divides among the businesses that were named.

Citations need their own set of answers. They exist only when the assistant searched the web and showed its sources; with web search off, an answer carries names and no links.

Many tools also report your position inside the answer. Treat it as secondary: an answer that lists you fourth still names you, and an average position over a handful of answers rests on a small sample.

Worked example: one month of answers

An accounting firm writes 10 buyer questions and asks them to four assistants, three times each: 120 planned answers. Two calls fail, so 118 answers come back. The numbers are invented to show the arithmetic.

From the same 118 answers:

• AI Recommendation Rate: 25 ÷ 118 = 21%.
• Empty-answer rate: 39 ÷ 118 = 33%. Assistant D named no one in 22 of its 30 answers, so its low rate says more about how that assistant answers than about the firm.
• Share of voice: the answers named businesses 260 times in all and the firm 25 times, so 25 ÷ 260 = 10%. It sits below the rate because answers that listed several rivals dilute it.
• Prompt coverage: the firm was named at least once on 6 of the 10 questions, so 60%. The four questions with no mention are where to start, and the GEO vs SEO guide for small businesses covers what to change on the pages behind them.

The next month the firm is named in 31 of 118 answers, 26%. The new interval, 19% to 35%, overlaps the old one, 15% to 29%, so the fair report is "no clear change yet", even though the headline number went up five points.

How many answers you need

A rate from a small sample swings widely. Here is the same 25% at different sample sizes, with its 95% Wilson interval, the method described on the AIGROW Index method page:

Twelve answers tell you whether you are named at all. To see a ten-point move clearly from one month to the next, plan for about 480 answers a month, for example 40 questions × 4 assistants × 3 runs. At 120 answers, a rise from 25% to 35% still leaves the two intervals overlapping.

The AIGROW Index shows why one check is a weak guide. Its October 2026 edition holds 80,059 answers on 2,580 ranking pages, with each question asked up to three times:

• an assistant asked the same question again repeated 56% of its names;
• any two assistants shared 21% of the names they gave;
• 61% of the businesses in a ranking were named by only one assistant;
• a page gathered 29 different names on average, and 45% of them were named only once.

So show each assistant on its own line before you show a total, and treat a business named once as noise until it comes back. The full figures are in the AIGROW AI Visibility Report.

What Google, Bing and your analytics already show

These reports cost nothing and come from the platforms themselves. They count appearances and visits on each platform's own surface, so they sit beside the answer counts above rather than replacing them.

Google rolled its report out to all websites on 31 August 2026. Bing added intents, topics, citation share and a compare view to its report as previews in June 2026. For what makes a page eligible to show in Google's AI features, see the guide to Google AI Overviews for business owners.

Set up an assistant channel once. In your analytics tool, make a segment or channel whose source matches the assistants' domains, such as chatgpt.com, perplexity.ai, gemini.google.com, claude.ai and copilot.microsoft.com, then check your own referral list for any others that turn up. Add a "How did you hear about us?"

question with an AI assistant option to your enquiry form and call notes, so the people who never clicked are counted too.

How to check the numbers are accurate

Visibility and accuracy are separate questions: an assistant can name you and still get your prices, area or opening hours wrong. Check both, and check the measurement itself.

1. Recount a sample yourself. Take 20 stored answers from any tool or sheet and mark them by hand. If your count and the tool's differ, find out why before you trust its totals.
2. Check the denominator. Failed calls should be left out and answers that name nobody kept in. A count that drops empty answers reports a higher rate than buyers experience.
3. Check the settings. In the Index, with web search switched on, 43% of the names were ones the model also gave without it, across 3,260 answers. A rate measured with search on and one measured with it off are different numbers.
4. Open every cited link in the sample: does it load, does the page back up the sentence, is it current? If a page you expect is never cited, confirm it is indexed and allowed a snippet, which Google lists as the requirements for its AI features.
5. Score facts against a short fact sheet of your own: services, area served, where prices are listed, opening hours and contact details. Count only facts you can check against that sheet.

Personalised answers set the limit of any measurement. A signed-in customer with chat history and a location can get a different answer from the one in your sheet. Fixed settings, with no history and no location, make one month comparable with the next; they measure a fair sample of answers, and one customer's screen can still differ.

Want to know where your business stands? The free checker shows whether ChatGPT, Claude, Gemini and Grok name it.

One number for a client report or a board

For a client report or a board slide, lead with the AI Recommendation Rate on your fixed question set, per assistant and overall, with its interval and last month's figure beside it. It reads in plain words ("named in 21 of every 100 answers"), it makes sense without a list of competitors, and it is the same definition the AIGROW Index, free scan and audit use.

Then add one line for traffic, one for enquiries and the next fixes:

If you report this for several clients, the agency plan with client seats and white-label services keeps the reports under your own brand.

Connecting AI visibility to enquiries and revenue

Answers rarely leave a clean trail from question to sale, so build the link from three pieces and say which ones are counted and which are estimated:

1. Visits: sessions that arrive from assistants, from your analytics.
2. Enquiries: form fills and calls from those sessions, plus enquiries that answer "an AI assistant" to how they heard about you.
3. Value: your own share of enquiries that become clients, and what a new client is worth.

A worked example with invented numbers: 5 enquiries a month come from assistants, one in four becomes a client, and a client is worth $1,500 in the first year. The expected value is 5 × 0.25 × $1,500 = $1,875 a month. If the work costs $600 a month in time or fees, the return is ($1,875 - $600) ÷ $600 = 2.1 times the cost.

Read that figure as an estimate. People who read an answer and then searched your name show up as Google traffic, and only the "how did you hear" question catches them. When a change goes live, note the date, keep the questions fixed and compare the months before and after, checking for seasons, campaigns and assistant updates.

A rise after a change is a correlation until a controlled test shows more.

Set up a tracking sheet in an afternoon

A spreadsheet is enough to start. Use one row per answer and these columns:

A pivot table on the "Named you" column gives the rate per assistant and per question. A count of all names gives share of voice, and the rows with no name give the empty-answer rate. Re-run the same questions on the same day each month.

When collecting the answers takes more time than reading them, a tracker can automate it. Before you pick one, check it reports the metrics above with their formulas, not a single blended score.

The buyer guide to the best AI visibility tools and the comparison of AI SEO tools and trackers list what each covers and what it costs.

Where AIGROW fits

AIGROW reports the same rate, under one definition, in five places:

• The free AI scan derives five questions from your website, puts them to the models behind four assistants without web search, and reports how many answers named you, inside a score out of 100 where the rate is worth up to 45 points. It is free, with no account and no card, and takes about three minutes after it reads your homepage. It is a first reading and cannot guarantee an AI mention.
• The AI visibility audit measures 20 to 40 of your own questions across four assistants, with every answer dated and re-runnable, a visibility score from 0 to 100 per assistant and overall, the competitors named in your place and a fix list ordered by impact.
• The free ChatGPT visibility tracker and Gemini visibility tracker check, with no sign-up, whether that assistant names your business, next to what the Index recorded about it.
• AI visibility monitoring plans ask the same questions again on a schedule, so you see when you start getting named and who is named instead.
• The AIGROW Index publishes the rate for businesses on 2,580 ranking pages, each with a 95% interval, under CC BY 4.0. It is a dated research sample.

AIGROW measures itself the same way. In the study of whether AI assistants recommend AIGROW, 30 questions in 11 languages went to ChatGPT, Claude, Gemini and Grok three times each with no web search, on 4 October 2026. Of 358 answers, 4 named AIGROW, a rate of 1.1%: ChatGPT 3 of 90, Claude 1 of 88, Gemini 0 of 90 and Grok 0 of 90.

Semrush, named in 76 answers, was the name they gave most. That is the baseline the next measurement, due in November 2026, is compared with.

Sources

Official pages read on 20 October 2026:

• Google Search Console Help: Generative AI performance report
• Google Search Central: AI features and your website
• Bing Webmaster Blog: AI Performance in Bing Webmaster Tools, February 2026
• Bing Search Blog: new AI visibility insights, June 2026

AIGROW data:

• AI Visibility Report, October 2026 edition
• AIGROW Index method
• AI Recommendation Rate definition
• Does AI recommend AIGROW?

The worked examples use invented numbers. Every interval in them is a 95% Wilson interval.