AI visibility tracking: how to measure whether assistants name you

AI visibility tracking is measuring how often ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews name your business when buyers ask them a question, and which sources they cite when they do. It is done with a fixed list of buyer questions, run several times in each engine on a schedule, with the raw answers kept. The same question asked twice can return different companies, so a single reading is a snapshot and a repeated run is a measurement. This page is the design of a tracking programme you can defend in a board meeting.

Why two runs of the same question differ

Knowing the causes is what lets you design around them. There are six, and none of them are mysterious.

• Generation is probabilistic. These models sample the next word rather than picking one deterministically, so wording and the order of named companies move between runs.
• The decision to search is made per question. One run may fetch live pages and the next may answer from the model alone, which changes the whole source list.
• Personalisation. Memory, custom instructions and your own chat history shape the reply. Your logged in account is the least representative place to test.
• Location. The IP address and locale change which local businesses appear. Testing from your office in one city tells you about that city.
• Freshness. A searched answer reflects whatever was indexed that week, including a competitor's new page.
• Model and interface version. The app, the API and a browser based surface can route to different models and different tool settings.

Design for all six by fixing what you can and repeating what you cannot: same prompts, same locale, same interface, logged out or in a temporary chat, several runs per prompt, same window each month.

Building the prompt set

The prompt set is the whole measurement. Thirty to fifty questions is a working size, written in your buyers' words, with your company name in none of them.

Prompt type	Example shape	What it tells you
Category discovery	Who are the leading suppliers of X in Y	Whether you exist in the category at all
Comparison	Compare the main options for X for a mid sized company	Whether you make the shortlist against named rivals
Local	Best X near Y, or X in Y with Z	Local presence and directory strength
Problem led	How do I solve X, and who can help	Whether your content earns the recommendation
Price led	How much does X cost, and who publishes pricing	Whether your pricing page is usable by a model
Fact check	What does Company A do, how big is it	Accuracy of what is said about you

Only the last type names a company, and it is there to test accuracy rather than to count mentions. Keep it in a separate column of the report so it never inflates your mention rate. The reasoning behind that rule is in should you ask ChatGPT about your own company.

Run design

• Engines: at least three of the five your buyers use. Answers differ enough between them that one engine is one engine's opinion. See why one AI app is not enough to measure with.
• Runs: three per prompt per engine as a floor, in the same session window. Report the range and the count, not an average on its own.
• Conditions: logged out or temporary chat, memory off, a fixed locale, and a note of the date and the model version shown in the interface.
• Capture: the full answer text, every cited URL, and the position where your name appears in the list.
• Cadence: monthly for most businesses, with an extra run after you ship a batch of pages. The cadence argument is in how often you should measure AI visibility.

The metrics that mean something

• Mention rate. The share of runs where you are named, per engine. Report it as named runs over total runs, so the denominator is visible.
• Share against named competitors. Of all companies named across the set, how often each one appears. This is the number that moves a sales conversation, because it is relative.
• Citation share. How often your domain appears in the source list, which can move before your mention rate does and is the earliest sign that new pages are landing.
• Source inventory. Every domain cited across the set, ranked by frequency. This is your third party target list, handed to you.
• Position in the answer. First named and last named are different outcomes when a buyer reads three names and stops.
• Accuracy. On the fact check prompts, whether what is said about you is right. A wrong fact repeated across engines is a source problem you can go and fix.
• Assistant referrals. Sessions arriving from assistant domains in your analytics. Small, usually better qualified, and almost never on the dashboard.

Definitions for each of these, with worked examples, are in AI visibility reports explained. For how to read a number once you have it, see what is a good AI mention rate.

What a defensible report contains

• The full prompt list, printed, not summarised.
• The engines, interfaces and model versions used.
• The number of runs per prompt and the date window.
• Mention rate per engine as a fraction, with the range across runs.
• The named competitor set and their mention counts on the same prompts.
• Every cited domain, with counts.
• The raw answers, kept and linkable, so any figure can be traced to the text it came from.
• What changed since the last run, and what shipped in between.

If a report has a score and no prompt list, ask for the prompt list. Our own reporting format is published in full in how to report AI search visibility.

Spreadsheet or platform

A spreadsheet and a disciplined hour a month is a real measurement for a small prompt set in two engines. The reason to move to a platform is arithmetic: thirty prompts, five engines and three runs is 450 answers a month, and collecting those by hand stops happening by the third month. Tool categories are compared in AI visibility tools compared.

Tracking normally means watching a number move. Here there is no number to watch, only a reply that either names you or does not, and that differs between assistants on the same question. So the thing to track is a count across a fixed question set, taken the same way each time. AI Knows Us holds the question set steady and records each answer with its sources, which is what makes two readings comparable.

This page is the measurement design. For the work that moves the numbers, see how to write a page an assistant will quote. For a before and after protocol around a change you make, see how to measure GEO without overclaiming.

Where people get stuck on AI visibility tracking

How many prompts do I need?

Thirty is enough to see a pattern and small enough to keep running. Fewer than ten and one lucky answer moves your whole percentage.

Should I track branded prompts?

Track them separately and label them as accuracy checks. A branded prompt tells you what an engine says about you when asked directly. It tells you nothing about whether you get recommended, because you supplied the name.

Can I use Search Console for this?

For Google's AI surfaces, partly. Google states that sites appearing in AI Overviews and AI Mode are included in overall Search traffic in the Performance report under the Web search type, so clicks and impressions are there, without a separate AI breakdown. For ChatGPT, Claude and Perplexity you need the prompt test and your referral data.

What counts as a mention?

Write the rule down before you count. A workable one: your company named in the answer text counts, your domain in the citation list counts separately, and a passing reference inside a quoted competitor comparison counts as a mention with the position recorded. Consistency matters more than which rule you pick.

How soon should the numbers move?

Access fixes can show up in searched answers within days of the next crawl. New pages typically register over a few weeks, with citation share moving before mention rate. Judge a programme on a quarter.

What to do first

• Write thirty buyer questions in your buyers' words, with your company name in none of them.
• Pick three engines and run every prompt three times, logged out, in one sitting. Save the full text and every URL.
• Count named runs over total runs per engine, and do the same for your three closest competitors.
• Rank every cited domain by frequency. The top five are your next quarter of third party work.
• Put the same run in the diary for the same week next month, so you are always comparing like with like.