How to Measure AI Visibility

To measure AI visibility, ask the AI assistants your customers use the exact questions they'd ask to find a business like yours, on a fixed schedule, and record three things every time: whether you're named, which competitors are named instead, and which sources the answer was built from. AI visibility is your presence inside those generated answers, turned from a hunch into a number you can watch move.

Most businesses have never done this once, let alone repeatably. They can see their Google ranking in a dashboard, but they have no idea what an assistant told a customer yesterday, in a private chat that never touched their website. This guide is the method for closing that gap: what to count, how to collect it without a tool, and how to trend it over time so the reading actually means something.

The five metrics that actually tell you something

A single question asked once is an anecdote. AI visibility is a pattern across many questions, tracked over time, measured on five things.

• ✓Mention rate. Across a fixed set of buying questions, on each assistant, how often are you named at all? Count it as a fraction: named in 4 of 20 prompts on one assistant is a mention rate of 20%. This is your headline number and the one to watch first.
• ✓Share of answer. When the assistant names businesses, how many of the slots are yours versus competitors'? An answer that recommends three firms and you're one of them is a very different reading from one that names three and you're absent. This is where you see who's getting cited instead of you, by name, in your market.
• ✓Citation sources. Which pages the answer was built from, your site, a competitor's, a directory, a review site, a forum thread, an industry roundup. This is the most useful metric of the five, because it tells you what to go and influence next. A mention count says where you stand; the source list says what to do about it.
• ✓Sentiment and accuracy. Being named isn't enough if what's said is wrong. Record whether the description is positive, neutral or negative, and, separately, whether it's correct. An assistant confidently recommending a service you stopped offering is its own problem, and you only catch it if you're reading the whole answer, not just scanning for your name.
• ✓Prompt coverage. How much of your buyers' real question space you're actually testing. Ten prompts covering one product line tell you little about the rest of the business. Coverage is the metric that keeps the other four honest, a high mention rate on three cherry-picked prompts is not visibility, it's a screenshot.

The method: a repeatable process you can run yourself

You can start today, for free, with nothing installed. The point of a method rather than a one-off check is that AI answers vary run to run and drift over weeks, so the value is entirely in doing it the same way, on a schedule, and comparing like with like.

1. Build your prompt set

Write down the questions a real customer would type to find a business like yours, phrased the way people actually ask an assistant, not as keyword fragments. "What's the best commercial solar installer in [your city] for a warehouse roof?" beats "commercial solar." Aim for 15 to 30 prompts spanning your main products or services, a few locations if you're regional, and a mix of stages: broad ("who are the top providers of X"), specific ("best X for Y situation"), and comparative ("X company vs Y company").

Lock this list. It's your measuring stick, the moment you change the prompts, you lose the ability to compare this month to last.

2. Run each prompt across the assistants your buyers use

ChatGPT is not the whole market. Run the same prompts through the assistants your customers actually reach for, in most markets that means ChatGPT, Gemini, Perplexity and Google's AI Overviews. Presence in one does not predict presence in another; treating a single assistant as "AI visibility" is how businesses convince themselves they're fine while three competitors own the answer everywhere else.

Because answers vary, run each prompt more than once, two or three times per assistant, so you're recording a pattern, not a fluke.

3. Log the answer, not just the mention

For every prompt, on every assistant, record the same fields in a spreadsheet: the date, the assistant, the prompt, whether you were named (yes/no), every competitor name returned, the sources cited (ask the assistant "which sources did you use?" if it doesn't show them), and a quick note on sentiment and accuracy. One row per run.

This is the step almost everyone skips, and it's the one that turns anecdote into measurement. Logging the whole answer, competitors and sources included, is what lets you calculate share of answer and see the citation supply chain, not just a bare yes/no on your own name.

4. Repeat on a schedule

Measure monthly for the trend. These systems move slowly, pages have to be re-crawled, third-party mentions have to accumulate, and the model's picture of your category shifts over weeks, not hours, so anything more frequent is mostly noise. The exception: after you publish or change something, run the affected prompts again within a week or two to see whether it moved the answer. Same prompts, same assistants, same fields, every time.

5. Trend it over time

A month of data is a baseline. Three months is a story. Chart mention rate and share of answer per assistant across each run, and the numbers start telling you what's working: this page earned a citation and your mention rate ticked up, do more of it; this one earned nothing in two months, cut it or rework it. Watch the citation sources too. When the same competitor roundup keeps feeding the answers and you're not in it, that's not a content problem on your own site, it's a page you need to get included in.

That loop, produce, measure, decide, adjust, is the entire reason to measure at all. Tracking your AI visibility and doing nothing with it is weighing yourself daily and never changing what you eat.

What the numbers actually mean

Read the reading against a competitor, not against zero. If you sell commercial solar and the answer to "best commercial solar installer near me" names three firms, none of them you, your share of answer for that prompt is 0 of 3, and you now have three named companies to study: what do their cited pages do that yours doesn't? A mention rate that climbs from 10% to 25% over a quarter while a competitor's holds flat is real progress, even if you're still behind. A single strong answer proves nothing; the trend, across a locked prompt set, is the signal.

The uncomfortable finding for most businesses running this for the first time is not that their score is low, it's that the answers are built almost entirely from sources they don't own. That's not a reason to stop measuring. It's the reason measuring matters: you can't influence a citation supply chain you've never looked at.

The honest cost of doing this yourself

None of the above is hard. It's just relentless. Fifteen to thirty prompts, four assistants, two or three runs each, logged field by field, every month, then charted and actually acted on, that's several hours of careful, repetitive work a month that only pays off if you never skip a cycle and never quietly change the prompts. The measuring is easy. Owning the loop, doing it the same way for long enough that the trend becomes decision-grade, and then making the decisions, is the part that falls off the to-do list by week three.

That's the real fork. If someone on your team has the hours and will run it rigorously, the method above is all you need, and if you'd rather it were automated, the companion guide on choosing an AI visibility tool covers what to look for in one. But if the honest situation is the one most businesses are in, producing content is easy now, knowing whether it worked is the hard part, and nobody owns the measurement, then the shortcut isn't another dashboard. It's having the whole loop run for you: the prompts tracked, the competitors and sources logged, the trend read, and the next move decided and made, so you get the outcome instead of a spreadsheet you'll stop updating.

For the background on why generated answers changed search in the first place, the explainer on generative engine optimisation sets the scene; for the specific work of becoming the source an assistant quotes, the guide on how to rank in ChatGPT covers the fixes.