What this audit tells you, and what it does not
This audit answers a narrow question: when buyers in your category ask an AI assistant for help today, are you named, who is named instead, and which sources are shaping the answer? One person can complete it in an afternoon with a spreadsheet and access to the assistants your buyers are likely to use.
It is a snapshot. Answers vary between runs and change over time, so an afternoon's work shows direction and gaps and stops short of a precise score. For ongoing tracking, use the panel method in how to measure AI search visibility. The audit is the first pass that tells you whether tracking is worth setting up and what to fix first.
Step one: choose prompts from the buying journey
Write between twelve and twenty prompts. Fewer gives too thin a picture and more will not fit in the afternoon. Draw them from how buyers speak: sales call notes, enquiry forms and the questions your team answers every week. Write them as a buyer would, with the context a real person includes, such as company size, location or a constraint.
Cover the whole journey. A list made only of prompts that ask for the name of a provider misses the earlier moments when a buyer is still working out what kind of help they need, and the later ones when they are checking a shortlist.
Resist the urge to write prompts that flatter you. A prompt containing your own tagline will name you and teach you nothing. If you are unsure whether a prompt is realistic, read it to someone in sales and ask whether a customer would say it.
- Problem prompts: the situation described, with no category named.
- Category prompts: which providers or products to consider.
- Comparison prompts: you against named alternatives.
- Evaluation prompts: cost, risks, fit and what to ask a supplier.
- Brand prompts: what the assistant says about you by name.
Step two: run each prompt several times
Pick two or three environments: the assistants your buyers plausibly use, plus the AI answers in search results for the same questions typed as searches. Run every prompt at least three times in each, starting a fresh conversation each time so earlier answers do not colour later ones.
Use a clean session where you can, signed out or with history and personalisation switched off, and note the date and your location. Do not rephrase a prompt between runs. The point of repetition is to see how much the answer moves when nothing else changes.
A single run misleads in both directions. Being named once can be luck, and so can being absent once. Three runs will not give you a reliable rate. They will separate always, sometimes and never, which is enough to act on.
Step three: record who is named and which sources are cited
Use one row per run. The columns matter more than the tool, and the five below are enough.
Copy the description of your company word for word. An assistant that names you and places you in the wrong category, or credits you with a service you dropped, has revealed a problem you can trace to a source. Record the cited sources with the same care, because they are the part of the answer you can influence.
Where an environment shows no sources, say so in the row and move on. The absence is itself information: that answer is coming from what the model already holds, and it will be slower to change.
- Prompt, environment, date and run number.
- Whether you were named, and how you were described.
- Which competitors were named, in the order given.
- Which sources were cited or linked.
- Anything stated about you that is wrong or out of date.
Step four: find the gaps
Read down the sheet and sort what you see into four kinds of gap. Each has a different cause and a different remedy, which is why sorting comes before fixing.
An absence gap is a prompt where competitors are named and you never are. A source gap is a publication, directory, review platform or comparison page that is cited repeatedly and does not mention you. An accuracy gap is a wrong or stale statement about you. A content gap is a question where the cited pages belong to others because no page on your site answers it.
Look for patterns across prompts. If the same two or three sources appear behind most answers in your category, they deserve attention before anything else. If you are named for brand prompts and never for problem prompts, the system knows who you are and has no evidence linking you to the problems you solve. The mechanisms behind these patterns are explained in how AI assistants decide which businesses to name.
Step five: decide the first fixes
Order the fixes by how far they are within your control and how close the prompt is to a purchase. Accuracy comes first: correct the wrong facts at their source, whether that is your own site, a business profile or a directory entry. These fixes are cheap, and an error repeated to buyers does damage every day it stands.
Content gaps come next, because the remedy is yours to publish. Write or rewrite the page that answers the question, following the approach in how to write a page an AI answer can quote. Source gaps and absence gaps take longest, since they depend on earning coverage and reviews from other people. Choose the one or two sources that matter most and start there.
Keep the list short. Three to five fixes, each with an owner and a date, are worth more than a long backlog nobody starts.
What to do with the result
Save the sheet and the exact prompts. After a couple of months, run the same prompts the same way and compare. Movement from never to sometimes on a commercial prompt is a real result. A small shift on a single run is noise.
If the audit shows wide gaps on prompts close to a purchase, the work is worth resourcing properly, either in house or through AI SEO services. If you would like a second opinion on what you found, bring the sheet to a growth audit conversation.
Whatever you decide, keep the first sheet. It is the baseline against which every later claim about AI visibility, from your own team or from a supplier, can be checked.
