Write 15 prompts a customer would actually type, run each one five times across ChatGPT, Perplexity, Google AI Mode and Copilot, logged out, from a fixed location, on the same day each month. Record whether you were named and which sources were cited. Five runs because one answer is noise.
Every tool on page one for this question will sell you a dashboard, and some of them are good. But you do not need one to start, and if you buy one before you understand what it is counting, you will not know whether the number it gives you means anything.
Below is the protocol we run for clients every month, written out in full so you can run it yourself for free.
Why this is harder than it looks
Search had fixed rankings, a shared results page, and free analytics. AI answers have none of those. The same prompt produces different answers to different people at different times, there is no position one, and no assistant gives you a dashboard of when it named you.
That means any honest measurement has to control for variance rather than pretend it does not exist. Which is where most casual checking falls down: somebody asks ChatGPT once, sees a competitor, and concludes something. One answer is an anecdote.
The protocol, step by step
1. Write 15 prompts, in customer language
Not keywords. The sentences a person actually types when they are trying to choose someone. A useful spread:
- Five shortlist prompts. “Best commercial cleaning companies in Manchester, give me a shortlist with reasons.”
- Five problem prompts. The situation, not the service. “Our office cleaner keeps missing days, what are my options in Manchester.”
- Three comparison prompts. “How do I choose between two commercial cleaning contractors.”
- Two brand prompts. Your own name. “What is [your business] like to work with.” This catches what the model already believes about you, which is sometimes wrong and worth knowing.
2. Fix the conditions
| Condition | Setting | Why |
|---|---|---|
| Account state | Logged out | A logged in session is personalised by your own history. You are measuring yourself, not the market. |
| Location | Fixed, and recorded | Local answers change by city. Pick the one your customers are in and never move it. |
| Date | Same day each month | Models update. A monthly cadence separates real movement from release noise. |
| Runs per prompt | Five | The same prompt gives different answers. One run tells you nothing. |
| Engines | Four | ChatGPT, Perplexity, Google AI Mode, Copilot. They cite very different sources. |
3. Record three things, not one
Most people only count mentions. Mentions alone will not tell you what to do next.
- Named or not. Binary, per run. This is your visibility rate.
- Who else was named. The same three competitors appearing everywhere is a different problem from a different name each time.
- What was cited. The URLs the answer leaned on. This is the actionable column, and it is the one people skip.
4. Do the arithmetic
15 prompts, five runs, four engines is 300 data points a month. Your visibility rate is simply the share of those 300 in which you were named.
| Metric | How it is counted | What it tells you |
|---|---|---|
| Visibility rate | Named runs / 300 | The headline. Track the direction, not the absolute. |
| Engine spread | Rate per engine | Strong in Perplexity and absent in ChatGPT is a source problem, not a site problem. |
| Competitor set | Names, by frequency | Who actually owns the category answer today. |
| Cited domains | URLs, by frequency | The single most useful output. These are the pages worth being on. |
The cited domains column is the one I would keep if I could only keep one. Mentions tell you where you stand. Citations tell you what to do on Monday.
Run it for a month and you will notice the same handful of sources coming up over and over. Directories, a couple of local publications, a trade body, sometimes a single well written comparison page. That list is your actual to do list. Being on those is the work. Everything else is decoration.
The sheet you actually keep
One row per run. Three hundred rows a month sounds like a lot until you realise it is a two hour job with a timer and a browser window.
| Column | Values | Why it earns its place |
|---|---|---|
| Date | Fixed day each month | Makes months comparable. |
| Engine | ChatGPT, Perplexity, AI Mode, Copilot | Lets you break the rate down per engine. |
| Prompt ID | 1 to 15 | Shows which prompts you never win, which is the useful signal. |
| Run | 1 to 5 | Separates variance from movement. |
| Named | Yes or no | The headline metric. |
| Position | 1, 2, 3 or unranked | Being named third in a shortlist is not the same as being the recommendation. |
| Competitors named | Names, comma separated | Tells you who owns the category answer. |
| Sources cited | Domains | The action list. The most valuable column on the sheet. |
The engines do not behave the same, and it matters
Splitting the rate by engine is the difference between a number and a diagnosis. The four behave differently enough that a single blended figure hides what is going on.
| Engine | Tends to lean on | What a low score there suggests |
|---|---|---|
| ChatGPT | Its own index plus web search, with fewer visible citations | Weak general presence. Slowest to move and the hardest to influence directly. |
| Perplexity | Live search with heavy, visible citation | A source problem. If you are absent here you are absent from the pages it reads. |
| Google AI Mode | Google’s own index and local data | Classic SEO and Google Business Profile issues. The most fixable of the four. |
| Copilot | Bing’s index | A Bing indexing problem. Check Bing Webmaster Tools before anything else. |
Decent in Perplexity and Copilot, absent in ChatGPT. That combination almost always means the business exists on the open web but nothing about it is quotable enough to have been absorbed. The fix is content and citations, not technical work.
A worked month
To make the arithmetic concrete, here is what a single month looks like for a fictional but realistic case: a commercial cleaning contractor in one city, 15 prompts, five runs, four engines.
| Engine | Runs | Named | Rate | Reading |
|---|---|---|---|---|
| ChatGPT | 75 | 6 | 8% | Barely present. |
| Perplexity | 75 | 21 | 28% | Cited sources mention them. |
| Google AI Mode | 75 | 18 | 24% | Local signals working. |
| Copilot | 75 | 14 | 19% | Indexed in Bing. |
| Total | 300 | 59 | 20% | The headline figure. |
Illustrative figures showing the shape of a typical first month, not measured results from a specific client.
A blended 20 percent tells you almost nothing. The split tells you the business is visible where sources are cited and invisible where they are not, which points straight at the citation work rather than at the website.
What to do with the citation column
After one month, count the domains in that column by frequency. You will typically get a list like this, and it is the most actionable thing the whole exercise produces.
- Directories and listings. Usually the top of the list. Check your entry exists, is accurate, and is not a duplicate.
- Review platforms. Frequently cited when the prompt involves choosing between options.
- Local or trade publications. Slower to get onto and worth more when you do.
- Comparison and roundup pages. Someone else’s article listing suppliers in your category. Being added to one of these is often a single email.
- Your own site. Usually further down than owners expect.
The first month is always uncomfortable because the number is lower than people expect and the competitor column has names they have never heard of. Both of those are normal and neither means anything on its own.
What I would actually watch is the spread between engines. A business that scores zero everywhere has a content problem. A business that scores well in Perplexity and nothing in ChatGPT is doing better than it looks and just has not accumulated enough yet. Those two situations need completely different work, and a single blended percentage would have told you neither.
What the tools do and do not add
Paid trackers do three things the manual version cannot: they run at a scale you would not sustain by hand, they keep history without you maintaining a spreadsheet, and they do it on a schedule whether or not you remember.
What they cannot do is decide which 15 prompts matter for your business. That judgement is the part that determines whether the number means anything, and no dashboard will make it for you.
Run it by hand for two or three months first. You will learn which prompts actually matter and what your citation sources look like. Then buy a tool to take the repetitive part off your hands, if you still want to. Buying first usually produces a number nobody trusts and nobody acts on.
Two mistakes worth avoiding
- Measuring while logged in. The most common one. Your own chat history skews the answer toward things you have already talked about, including your own business. It makes you look more visible than you are.
- Treating a bad month as a failure. These systems update in steps, not curves. A flat or slightly negative month inside a rising six month trend is normal. Judge on the trend line.
Common questions
How often should I measure AI visibility?
Monthly, on the same day. More often produces noise you will misread as movement. Less often and you cannot tell what caused a change.
Why five runs of the same prompt?
Because the same prompt does not give the same answer twice. One run is an anecdote. Five gives you a rate you can compare against next month’s rate.
Can I see AI visibility in Google Analytics?
Only partly. GA4 will show referral traffic from assistants when someone clicks through to you. It cannot show the far larger number of times you were named and nobody clicked, which is most of them.
Do I need a paid tool to do this?
No. The protocol above is free and takes a couple of hours a month by hand. Tools save time and keep history. They do not decide which prompts matter, which is the part that determines whether the measurement is worth anything.
What is a good visibility rate?
There is no benchmark worth quoting, because it depends entirely on how contested your category is. Measure your own direction over six months. Anyone quoting you an industry average for this is guessing.
Want the first pass done for you?
Before any of this matters, an assistant has to be able to read your site at all. Our free checker scores the nine things that decide whether a machine can understand and quote your homepage. Ten seconds, no email.