You can audit any local business's AI visibility by hand in about fifteen minutes, which is our own estimate for the four steps below. You need a browser, a spreadsheet, and the discipline to write down what you actually saw rather than what you expected.
This is the method. It is worth doing by hand at least once even if you plan to automate it later, because doing it manually is what teaches you which parts of the result are signal and which are noise.
What you are measuring
One number: out of N questions a real customer might ask, how many answers named this business.
That is it. Not sentiment, not position, not "brand presence". Named or not named, counted across enough phrasings that one lucky answer cannot flatter the result.
The reason to count phrasings rather than ask once is that these models are not deterministic. Ask the same question twice and the names can change. A single answer tells you almost nothing. Eight answers tell you whether there is a pattern.
Before you start: the two rules
Rule one: fresh conversation every time. Assistants use the earlier turns of a conversation as context. If you ask about Business X and then ask "who is the best plumber in this city", you have contaminated the answer. New chat, every question.
Rule two: no personalisation. Log out, use a private window, or turn off memory and chat history. Otherwise you are measuring your own account's history, not what a stranger sees.
These two rules are the difference between an audit and a vibe.
Step 1: write the question set (3 minutes)
Eight questions. Not one, not twenty. Eight is enough to see a pattern and few enough to finish.
Build them from the way people actually type. For a plumber in a city, that looks like:
- Who is the best plumber in {city}?
- I need a plumber in {city}. Who should I call first?
- Recommend a reliable plumber in {city}.
- Best emergency plumber in {city}?
- Who does water heater replacement in {city}?
- Top rated plumbers in {city} with good reviews?
- Affordable plumber in {city}?
- Which plumbing company in {city} would you use yourself?
Notice the spread. Two broad, two urgency-flavoured, one service-specific, one review-flavoured, one price-flavoured, one personal. Real demand is not one phrase, and a business can be invisible on the broad question while owning the service-specific one. The 8 prompts customers actually ask explains why each shape is on the list and what it tends to reveal.
Swap the niche and city and this template works for any local service. Keep the same eight shapes so results are comparable between businesses.
Step 2: ask, and record every name (8 minutes)
Open a spreadsheet with one row per question and one column per assistant. Ask each question in a fresh conversation.
For each answer, write down every business name mentioned, in order. Not just whether your target appeared. The competitor list is the more valuable half of this exercise, and you only get it if you record it while you are there.
A few practical notes:
- Do it across at least two assistants. They read different sources and disagree more than people expect.
- Turn web search on where the assistant lets you choose. You are measuring what it can find today, not what it memorised.
- If an answer refuses to name anyone and says "check Google", record that as zero names. It is a real outcome and it happens most often in categories with regulated advice.
- Copy the sources if the assistant offers them. That list is your roadmap later.
Step 3: score it (2 minutes)
Two numbers per business.
Mention count: how many of the eight answers named it.
Visibility score: mention count divided by answers that completed, as a percentage. Named in three of eight is 38. Named in none is 0.
Do this for your target business and for the three or four names that came up most often. Now you have a table that says, plainly: you are at 0, and these three are at 75, 63 and 50.
That table is the entire pitch, if you are doing this for a client. It needs no interpretation.
Step 4: read the sources (2 minutes)
Go back through the answers and collect the URLs the assistants cited. Count the distinct pages and note which ones come back under more than one question. The repeats are the pages doing the most work, and the count itself is worth writing down, because it tells you how concentrated the answers are.
Sort them into three buckets:
- Listings you can join today. Directories, chamber listings, trade bodies. Usually free, usually a ten minute form.
- Editorial you can pitch. Local "best of" posts, publication roundups. These take a real email but they compound.
- Aggregators you can only influence indirectly. Review platforms, Google's own data. You improve these by improving the underlying profile.
When a business is at 0 and a competitor is at 75, check the first bucket and the state of the Google Business Profile before looking for anything exotic.
If the business's own website never shows up among ChatGPT's sources, check whether the site blocks OpenAI's search crawler. How to show up in ChatGPT recommendations shows where to look.
What the result actually means
Be careful with the claim you make from this.
You measured what a specific assistant said, on a specific day, to eight specific phrasings, from a logged-out session. That is a strong signal and it is not a law of nature. Run the same audit next week and a business at 38 might read 25 or 50. A business at 0 will almost always still be at 0, which is why the zero is the finding worth acting on.
Two more honest limits.
The consumer apps are not the APIs. If you automate this later using provider APIs, you are querying the same models with search enabled, but the ChatGPT, Gemini and Perplexity apps wrap their own system prompts and product logic around those models. Treat an API-based audit as a close approximation, not a transcript. Our methodology page says the same thing in more detail, because it is the most common way this kind of measurement gets oversold.
Name matching is fuzzy. "Smith & Sons Plumbing", "Smith and Sons", and "Smith Plumbing" may or may not be the same company. When you audit by hand you can just look. When you automate it, you need a matching rule and a threshold, and you need to show the borderline cases rather than guess at them.
Turning the audit into a conversation
If you are doing this for prospecting rather than for your own business, the audit is the email.
You do not need a pitch. You need four facts: the questions you asked, the answer that did not include them, the names that were there instead, and the score gap. A business owner who reads "we asked ChatGPT who the best plumber in your city is, eight different ways, and your name did not come up in any of them. Here are the three that did" does not need to be persuaded that there is a problem.
Send the raw answers once they reply and ask for them. The rawness is the credibility. The cold email templates for AEO outreach turn those four facts into five emails.
When to stop doing this by hand
Fifteen minutes per business is fine for one. It is four hours for sixteen, and prospecting at any real volume means hundreds.
The manual method also drifts. Different people record names differently, someone forgets the fresh-conversation rule, and last month's spreadsheet has a slightly different question set than this month's, so the numbers stop being comparable.
That is where automating helps, and it is what AIgnored does, with one difference from step 1: a scan asks its own fixed set of 8 broad recommendation questions per model, listed in every scan's why drawer, so the service-specific, price and personal questions stay a manual check. It applies one matching rule with the borderline cases surfaced rather than hidden, and scores the whole competitor list in one pass. It runs on your own provider keys, so a scan costs you what the provider charges and nothing on top. The pricing page has a free plan if you want to compare its output against an audit you did by hand, which is the right way to trust any tool of this kind.
The one-page version
Eight questions. Fresh conversation each time. Logged out. Record every name, not just the one you care about. Divide mentions by answers for a score. Collect the cited URLs and sort them into what you can join, pitch, or only influence.
Fifteen minutes, and you know more about a business's position in AI answers than most of its competitors know about their own.