AI visibility tracking: how to measure your business in AI answers

AI visibility tracking means keeping a written record of whether AI tools such as ChatGPT name your business when someone asks for what you do. Without that record you can’t tell whether things are getting better or worse. You can keep it in a spreadsheet. This guide shows you how.

Here’s an example. Say you’re an electrician in Marrickville and you ask ChatGPT who can upgrade a switchboard near you. It names a few of your competitors but not you. Ask again tomorrow and the list can change. One look tells you very little while a few months of saved answers start to tell you something.

This guide gives you a method you can run in a spreadsheet, the formulas for the numbers and a monthly routine. It follows the method behind our own Sydney trades study. We apply it as an AI SEO agency.

Improving how your business appears in generated answers is called generative engine optimisation (GEO). Our guide on what is GEO explains the work. Tracking gives you a record of the answers you saw. It is a starting point for deciding which page or profile to look at.

Definition: AI visibility is your business or website showing up in an AI answer that you saw. A mention names your business. A citation links to one of your pages. A visit is a person arriving on your site. An enquiry is a person asking for your service. Each is a separate event.

In this guide an engine means any AI product you ask, such as ChatGPT, Perplexity or Google’s AI features.

What do you need to do?

Here is the whole routine in five steps.

1. Write 10 to 20 questions a customer would ask, with the job and your suburb in each one.
2. Ask each question in ChatGPT, Perplexity and Google’s AI features. Start a fresh conversation every time and save the full answer, the date and every link.
3. Mark two things for each answer: is your business named and is your website linked?
4. Work out two rates: how many completed answers name you (your mention rate) and how many link to your website (your citation rate).
5. Repeat every month with the same questions and compare.

What should AI visibility tracking measure?

It should measure five things separately: mentions, citations, accuracy, visits and enquiries. Each answers a different question. One number that blends them hides which part is working.

The first three come from the answers themselves. The last two come from your own records. GA4 is Google’s website analytics tool. Our guide to tracking AI traffic in GA4 covers the visit side.

Which customer questions should you track?

Track the questions your customers ask when they have a job and no business in mind. Write each one with the job and the place in it, then lock the list before you collect a single answer.

Our Sydney trades study used a locked list of 50 questions across 10 trades, with five questions per trade. We wrote them from household jobs and recommendation requests in 27 public Reddit and Whirlpool discussions, so they read like real requests. A switchboard upgrade is a different question from finding someone to replace a porch light. The engines answered them differently.

For your own business, start smaller:

1. Choose one service and one area you want enquiries from.
2. Write 10 to 20 questions a customer would ask, in plain words.
3. Keep most questions unbranded, meaning they leave out your name. People rarely type your name before they know you.
4. Add three or four questions that include your name, to check accuracy.
5. Save the list with a version number and the date.

Changing the questions later starts a new series. Keep the original list so you can compare month to month.

How do you run a repeatable check?

Ask each question once in each engine you care about, start a fresh conversation every time and save the complete answer. The aim is a record that another person could read and get the same count from.

Make a sheet with one row per question and engine. Give it these columns:

Write down the conditions each time. In our study Google showed the location as Inner West Council, Sydney NSW, which it worked out from our connection’s network address. ChatGPT and Perplexity showed no location label like that. A fresh conversation didn’t reset the account’s state, meaning whatever the account already carried into the chat, so a second person running the same questions can still see different answers. Write down what you can see and don’t claim more.

Keep the outcome column honest. Some results are not cases where your business was left out. A blocked check is a login request or a usage limit. A failed capture is a recording problem. A Google result may show no AI answer at all. All three are missing data, so they stay out of the totals.

How do you work out mentions, citations and share of voice?

Do the sums for each engine on its own. Divide by the number of completed answers for that engine, including completed answers that named no business. The number you divide by is called the denominator.

• Mention rate = completed answers that name your business ÷ all completed answers for that engine
• Citation rate = completed answers that link to your own website ÷ all completed answers for that engine
• Share of voice = your once-per-answer appearances ÷ all once-per-brand-per-answer appearances for the brands on your declared list

Share of voice means your slice of the names that appeared, among the brands you chose to count. It is not market share. State the number you divided by beside every figure.

Here’s an example of the sums. Say you saved 20 completed answers. Your business was named in 5 and your website was linked in 3. Your mention rate is 5 ÷ 20, which is 25%. Your citation rate is 3 ÷ 20, which is 15%. If those 20 answers named 80 brand appearances in total and 5 were yours, your share of voice is 5 ÷ 80, about 6%.

Keep blocked checks, failed captures and missing AI answers out of the number you divide by. Flag wrong-market answers, which are answers about the wrong place, separately. If you take those out of a local-only calculation, show the new denominator beside the result.

Three rules keep the numbers honest. Count a business once per answer, however many times the text repeats it. Use the same names for the same business, because “Smith Electrical Pty Ltd” and “Smith Electrical” are one operator. Never compare a month with 20 completed answers to a month with 12 without showing both denominators.

How do you check that an answer describes you correctly?

Read the answer against the facts on your own site and listings. An accurate description gets the services, suburbs, hours and contact details right. A wrong detail can send a customer to the wrong place.

Add three checks to the sheet. Does the answer describe the right service? Does it place you in the right area? Is any detail out of date? Then open the source link beside the claim. Where the wrong detail comes from a directory or an old page, fix it there. Changing your own website won’t correct a listing you don’t control.

Kavera note: Location errors happen. In our study Perplexity read a question about changing locks in Canterbury as Canterbury in England, because our locked question didn’t name Sydney. We kept that answer as a wrong-market response and did not count it as a Sydney locksmith being absent. Write the place into every question.

What did our 50-question study show?

It showed that the engines often disagree and that the number you divide by matters. Of the 200 question and engine checks we planned, 197 completed. ChatGPT and Perplexity each completed 50 answers, Google AI Mode completed 49 and Google AI Overviews completed 48. We kept the one failed capture and the two results with no AI Overview apart from those answers. The picture below shows where the 200 checks ended up.

Count completed answers

Sydney trades · 16–17 October 2026

Planned checks

200

↓

Completed answers

197

Completed answers by engine

ChatGPT

50

Perplexity

50

Google AI Mode

49

Google AI Overviews

48

Keep outside the denominator

Failed capture

1

No AI Overview

2

Missing checks do not mean business absence.

One convenience sample of Sydney trades.

Among the 47 questions that all four engines answered, only 6 had a business name in common across every answer. ChatGPT returned 122 distinct business names across its 50 answers. The same question could therefore send a customer to different shortlists depending on the engine they used.

The full method, the question set and the downloadable records sit in our Sydney trades AI search index, which uses the same 50-question AI visibility method this guide describes.

Kavera note: In two checks about two months apart, our client SHORE went from absent to first in one Sydney builder shortlist. That is one question and two saved answers, so we record it as an observation and not a guarantee. Our ChatGPT SEO guide shows the screenshots.

What do Google and Bing report for free?

Both publish AI reports inside their webmaster tools. They measure something different from your hand-built sheet, so read them beside it and don’t merge the figures.

Google’s Search Console, the tool Google gives website owners to check how their site shows up in Google, has a Generative AI performance report. It counts impressions, which means the number of times links to your site were shown in AI Overviews and AI Mode, split by page, country, date and device.² It doesn’t report clicks. Google adds that clicks from its AI features are also counted inside your Web search totals.¹ The report may not appear until your site has enough impressions.²

Bing Webmaster Tools has an AI Performance report. It covers Microsoft Copilot, AI summaries in Bing and select partner integrations.³ Its Total Citations figure counts how many times your pages were shown as sources. Microsoft says it doesn’t show where in an answer the link sat.³ The report also samples the phrases the AI used to find the pages it referenced and labels them Grounding queries.³

Bing has two more figures, called Citation Share and Topics Portfolio. The fine print near the end explains them.

How often should you repeat the check?

Once a month is a sensible starting point for a small business. It is our suggestion and not a proven best schedule. Answers move, so a single check proves little and a trend across several months tells you more.

Keep the same questions, engines and competitor list. If you change one, label the new series. When a rate changes, read the saved answers before you choose a fix. Check whether the question, the location, the sources or the businesses named have changed. Write down what you can show and leave the cause as unknown when the evidence doesn’t point to one.

How do you connect visibility to enquiries?

Put the monthly answer sheet next to your analytics and your enquiry log. Visibility is only the first step because a visit and an enquiry are what the business needs.

Check three things each month. How many visits came from AI assistants? How many of those sessions ended in a completed enquiry? Did any new customer mention an AI assistant? The first needs the setup in our guide to tracking AI traffic in GA4. The second needs a form-submission event, which is a way for your analytics to count each completed form. The third needs you to ask.

Your saved answers only cover the questions you tested. Your visit report only covers the activity it recorded. If the two don’t line up, that doesn’t tell you which question a customer asked or prove that nobody clicked. Keep the two records apart and go back to the original answers and visit records to find out why.

When is a paid tool worth it?

A tool is worth it when repeating the questions by hand costs you more than the subscription. It is also worth it when you need many questions across several engines. A sheet is enough for one business with a short question list.

If you decide to buy, compare engines, locations, question limits and exports first. Our guide to AI visibility tools does that, so we won’t repeat it here. Whatever you choose, keep your own saved answers for the questions that matter most. They are the evidence you can check.

The fine print

• Our study is a convenience sample. That means we didn’t pick its questions or trades at random. It is a sample of Sydney trades. It doesn’t rank workmanship or measure how many people search.
• Citation Share. Microsoft measures your site’s portion of all citations for one retrieval phrase, called Citation Share.⁴ It divides by citations across sites for that phrase. That differs from the named-brand appearances in your own question set, so the two numbers aren’t the same measure.
• Topics Portfolio. This beta groups those phrases by subject and reports approximate figures and opportunity ranges. Microsoft advises using them for planning and may leave out sites with too little data.⁵

What should you do next?

Write your first 10 questions today and run each one once. Save the answers in your sheet and work out your mention rate and citation rate. Write down the date. That is your baseline.

Our SEO audit is free. Implementation is separate unless retained. See the SEO packages for terms. Monthly plans start at Suburb SEO, $890/month + GST. If you want the tracking and the website work done together, talk to our AI SEO agency. No SEO package guarantees rankings, leads or revenue.