You can check your AI visibility with a fixed set of about 20 prompts, three assistants and a spreadsheet, in roughly 45 minutes a month. The goal is not to find a ranking, because AI answers do not have one. The goal is to see, month over month, whether assistants mention you, cite you, ignore you or say something wrong.
Paid trackers exist and some are useful at scale. But for a small business or a single brand, a manual routine gives you something a dashboard rarely does: you read the actual answers, notice how you are described and see which competitors keep appearing next to you. Below is the exact process we use, adapted so one person can run it.
Why can’t you just check once and be done?
AI assistants do not return identical answers every time. The same question can produce a different list of businesses on Tuesday and Friday, and answers differ by assistant, by location, by whether the assistant is allowed to search the web and sometimes by account history. A single screenshot proves very little.
What does carry information is a repeated, consistent test. If your business appears in 6 of 20 prompts this month and 9 of 20 next month, with the same prompts and the same method, that is a signal, though still an imperfect one. Think of it as a thermometer, not a ruler.
How do you build the prompt set?
Write 20 prompts in three groups. Use the words your customers use, not your internal jargon.
Group 1: brand prompts (5)
These test whether assistants know you exist and describe you correctly.
- “What is [your business name] and what do they do?”
- “Is [your business name] a good choice for [your main service]?”
- “What do people say about [your business name]?”
- “How much does [your business name] cost?”
- “[Your business name] vs [closest competitor]”
Group 2: category prompts (8)
These test whether strangers can discover you.
- “Best [service] for small businesses in [country or city]”
- “Who should I hire for [service] if my budget is [amount]?”
- “Recommend [type of provider] that works with [your niche]”
Vary the budget, location and customer type across the eight so you cover your real market.
Group 3: problem prompts (7)
These test whether you appear when someone describes a problem instead of naming a category. They are the most valuable and the hardest to win.
- “Why are my Google Ads getting clicks but no sales?”
- “How do I get my blog to show up in ChatGPT answers?”
Write yours around the questions your customers genuinely ask on sales calls. If you do not know them, your last ten enquiry emails are a good source.
How do you run the check?
- Use three assistants. ChatGPT, Gemini and Perplexity cover the main behaviours: one with a large user base, one tied to Google, one that shows sources prominently. Add Copilot if your audience is corporate.
- Use a clean session. Log out or use a private window where possible, and avoid memory or personalisation features that bend the results toward you.
- Paste the prompt exactly. Do not rephrase between months.
- Copy the answer’s key lines into your sheet: whether you appear, how you are described, which sources are cited and which competitors appear.
- Run each prompt once, or twice for category prompts, since variation is highest there.
At about 20 prompts and three assistants, expect 60 answers. You do not need to read each one in depth; you scan for your name, the competitors and the cited links.
What should the spreadsheet look like?
| Column | What to record |
|---|---|
| Date and assistant | When you ran it and where |
| Prompt group and text | Brand, category or problem; exact wording |
| Result | Cited, Mentioned, Absent or Wrong |
| How described | One line quoting the assistant’s wording |
| Sources cited | Domains, especially if yours appears |
| Competitors named | Who appears alongside or instead of you |
Definitions matter. Cited means your own site is linked as a source. Mentioned means the assistant names you without linking you. Absent means you do not appear. Wrong means it says something inaccurate about you, such as an outdated price, a wrong location or a service you do not offer.
How do you read the results?
Resist the urge to total everything into one score on day one. Look at it in layers:
- Wrong first. Inaccurate statements are the most damaging and the most fixable. If any appear, work out where the assistant learned them. Our correction playbook for when ChatGPT gets your business wrong walks through that process.
- Brand prompts next. If assistants do not know you or describe you vaguely, the issue is basic entity clarity: a clear About page, consistent business details and a few independent mentions.
- Category prompts. Here you compete with everyone. Note which domains the assistants cite repeatedly; those are the sources shaping the answers in your market.
- Problem prompts. If you are never cited, look at whether you have any page that answers that exact question directly in the first two paragraphs.
What are the common traps?
- Treating one lucky answer as success. Look at trends across months.
- Changing prompts to get a better result. That destroys the comparison. Add new prompts, but keep the old ones.
- Ignoring the sources. The cited domains tell you what to do next, for example a directory or review site you are missing from.
- Measuring only mentions. Visibility that does not lead to visits or enquiries is not the goal. We cover that in how to tell if your AI search visibility is making you money.
What does a realistic result look like?
Here is an illustrative picture, not a benchmark. A small service business with a thin web presence runs the 20 prompts and finds: brand prompts mostly work, with two outdated details; category prompts show it in perhaps one in five answers, usually mentioned without a link; problem prompts show it almost never. That is a normal starting point. It tells you the business exists in the assistants’ knowledge but is not yet the answer to a stranger’s question.
The fix is rarely mysterious. It tends to be a mix of clearer pages that answer the question outright, consistent business details across the web, and a few independent sources that describe you. Progress usually shows first in the brand and category groups and last in the problem group, which is why it pays to keep the whole set.
How do you keep it to 45 minutes?
Batch the work. Run all prompts for one assistant in a single sitting, paste the answers into a document, then fill the sheet in one pass. Do not read every sentence; scan for names, links and anything that sounds wrong. Put the check on the first Monday of the month and treat it like bookkeeping: small, boring and regular.
What do you do with the findings?
Turn each month’s check into three actions at most. For example: fix one wrong statement, publish one page that answers a problem prompt where you were absent, and request or update one listing on a source the assistants keep citing. Smaller, repeated actions beat a big annual push. If you want a structured way to prepare your site for these actions, our GEO audit checklist lists the eight technical and content checks we start with.
When is a paid tool worth it?
Consider a tracker when you need more than about 50 prompts, when you cover several countries or languages, or when you report to clients or a board and need history that is not stored in a personal spreadsheet. Until then, the manual routine costs nothing but 45 minutes, and the reading itself teaches you how your market is described.
The bottom line
Pick your 20 prompts today, run them once as a baseline, and put a repeat date in your calendar. In three months you will know more about your visibility in AI answers than most competitors, and you will have a clear short list of what to fix.