What an AI visibility audit is
An AI visibility audit is a structured look at what AI assistants say when your buyers ask about your category, and at the sources those answers rely on. It ends with a written finding, not a score on a dashboard.
That definition is narrower than it sounds, and the narrowness is deliberate. A one-time look at whether ChatGPT knows your name is a check. If you want to know how to run one, our page on running the check yourself covers it. An audit goes further in three ways. It is wider, because it covers many questions and several assistants. It is deeper, because it traces each answer back to the pages and sites that shaped it. And it is written down, because the point is a document someone can act on and later compare with a second audit.
The word audit is used loosely in this field. Some suppliers sell a crawl of your website under that name. Some sell a screenshot of three prompts. What follows describes the version worth paying for, or doing yourself, so you can tell the difference when a proposal arrives.
It audits answers, not pages
The most common confusion is between two kinds of audit that sound alike. A site audit asks whether machines can read your pages, whether each page says what it should, and whether the markup is sound. We cover that on our AI SEO audit page, and it is useful work. But a site audit starts from your pages and asks what is wrong with them.
An answer audit starts from the other end. It asks what the assistants are saying today, and works backward to why. Your site may be in perfect technical shape and still be absent from answers, because the assistants lean on review sites, forums, local directories and other businesses' pages that mention you or fail to. Equally, a messy site can be named often because a respected third party wrote about it.
That is why the answer audit usually comes first. It tells you which parts of the site audit matter. If the answers show you are missing for questions about pricing, there is little point polishing your about page.
The practical test for any audit proposal is the evidence it cites. If every finding is a fact about your own website, it is a site audit with a new label. If findings quote what an assistant said, on which date, to which question, it is an answer audit.
What it should cover
Five areas, each producing its own part of the finding.
1. The question set
The audit is only as fair as its questions. They should come from how your real buyers talk, spread across broad category questions, specific need questions, comparisons and questions about your own name. Our page on writing the question set covers how to build it. An audit with ten questions, all written by the supplier from your homepage, is measuring the supplier's imagination.
2. Coverage across assistants
At minimum the assistants your buyers use. ChatGPT, Perplexity, Claude and Gemini behave differently, draw on different sources, and show links differently. Google's AI answers are a further case. The audit should report each separately.
3. Presence and treatment
For each answer: absent, named, recommended, or cited with a link, and the actual wording used about you. This is the core table of the audit.
4. Who is named instead
The businesses that appear when you do not. This set is your real competition in this channel, and it is often different from the competitors you track in search. Our page on the competitor check goes deeper on recording it.
5. The sources behind the answers
Where the assistants appear to get their information. Cited links are the easy evidence. For answers with no links, the audit infers sources by comparing the wording to pages that exist. This is the part that separates an audit from a check, and it is where the most useful findings come from.
How to trace the sources behind an answer
Source tracing is the step most do-it-yourself audits skip, so it is worth describing in practice.
Start with answers that show links. List every domain cited across your whole question set. Count how often each appears. Within a day you will see a pattern, and it is usually a small number of sites doing most of the work: a handful of review platforms, a local publication, a forum, a few comparison pages, and the websites of the businesses that are named often.
Then sort those domains into types. Your own site. Review and directory sites. Forums and community threads. News and trade publications. Competitors' own sites. Government or institutional pages. The split tells you where the assistants look when they are asked about your category. Our overview of what AI engines cite describes the general pattern, but your category will have its own version, and the audit is how you find it.
For answers with no links, use the wording. If an assistant says a business is known for same-day service and free estimates, search for those exact phrases. They often come from one page, usually the business's own page or a review. That is a lead on a source, not proof, so record it as inferred.
Finally, look at what the cited pages have that yours do not. Is it a clear answer in the first paragraph, a table, customer quotes, a price, a location detail? You are not copying them. You are finding the specific thing the assistants found worth citing.
What an audit should hand you
An audit that ends with a slide of charts has not finished. A useful one produces five things you can put on a desk.
The baseline table. Every question, every assistant, the outcome, and the quoted sentence. This is the permanent record. Every later check is compared with it.
The shares. The percentage of answers where you were named, cited and recommended, per assistant, with the number of answers behind each. If you want the method, the page on how to measure AI visibility sets it out.
The competitor map. Who is named, how often, and for which questions.
The source map. The domains and page types the answers rely on, ranked by how often they appear, with a note on which ones you could realistically be on.
A short list of findings, in order. Five to ten statements of the form: for this group of questions, we are absent, the answers draw on these sources, and the page type that is cited is this. Then, beside each, the action it points to. Nothing on the list should be there only because it sounds good.
The list should include at least one finding that is uncomfortable. If all ten are good news, the audit has been shaped to please.
Running one yourself in a weekend
A small business can do a credible audit without software. It takes more time than a quick check but not a great deal.
Saturday morning: the questions. Write twenty questions in your buyers' words. Mix broad, specific and comparison questions. Add two or three that name your business. Freeze the list.
Saturday afternoon: the runs. Use a logged-out browser. Run each question three times in each assistant you care about. Save the full text of every answer and every link into a spreadsheet. A plain spreadsheet with one row per answer is enough.
Sunday morning: the reading. Mark each answer absent, named, recommended or cited. Copy the sentence about you. List who else appears.
Sunday afternoon: the sources. Count cited domains. Sort them by type. Open the top ten and look at what they contain. Write the findings, five to ten lines, and the action each points to.
You will make some judgment calls along the way, and that is fine. Write down your rules for them so the next audit follows the same ones. Then read how to read AI visibility results before you decide what the findings mean.
If you are buying an audit
Four questions separate a real audit from a repackaged one.
Who writes the questions? The right answer is you and the supplier together, from your actual customer language. A supplier who generates them from your website alone is measuring your marketing, not your market.
How many answers is each figure based on? Ask for the denominator. A percentage with no count behind it can mean anything.
Will I see the raw answers? You should be able to read every answer that went into the findings. Refusal is a warning sign.
What do I get when it is done? The five items above. If the answer is a score and a deck, ask what you would do on Monday morning.
Also ask how often the audit is repeated and whether the question set stays fixed. One audit is a snapshot. The value comes from the second and third, set against the first. If the audit is a one-off with no plan for repeating it, you are buying a photograph of a moving thing. For how to judge the repeat reporting, see our page on what a good AI visibility report contains.
What an audit cannot tell you
Be clear about the limits so nobody oversells the result.
An audit is a sample taken on certain days. Assistants vary from run to run, so it describes a pattern rather than a fixed truth. It cannot tell you how many real buyers saw those answers. It cannot show revenue from them. And it cannot predict what any assistant will do next quarter.
It can tell you where you stand against a fixed set of questions, how you are described, who is named instead, and what the sources have in common. That is enough to decide what to write, fix or earn first. It is also enough to tell, six months later, whether the work changed anything, provided you ran the same set again.
What a well-written finding looks like
Findings are where audits tend to go soft, so it is worth seeing the difference between a weak one and a useful one.
A weak finding reads: your AI visibility is low and should be improved. It cannot be checked, and it points nowhere.
A useful finding reads more like this: for the six questions about choosing between two kinds of provider, you were absent from every assistant. In the answers that showed links, the cited pages were comparison guides from three independent sites and one competitor's page that sets the two options side by side in a table. You have no comparison page. Action: write one, in plain words, with the table near the top.
Notice what the second version does. It names the group of questions. It states the outcome. It names the evidence, the cited pages. It says what you lack. And it ends in an action small enough to start this week. If a finding cannot be written in that shape, the audit has not dug far enough, or the thing it found is not actionable and can be dropped.
Order the list by how many of your buyer questions each finding touches, not by how interesting it is. The finding that covers a third of the set outranks the clever one that covers a single odd question.
What to do between audits
An audit is a heavy exercise, and nobody should run one every month. Between audits, do something lighter. Pick the five questions where you were closest to winning, the ones where you were named but not recommended, and run only those, three times each. If a change you made is working, it will show up there first.
Keep a short log of what you changed and the date. Pages written, pages rewritten, reviews gathered, listings corrected. When the next audit shows a movement, the log tells you what to credit. Without it, every movement becomes a guess, and every disappointment becomes an argument.
How RunOctopus relates to an audit
RunOctopus is a service that does the work an audit points to, rather than a free audit tool. There is no free automated AI visibility checker on the site, and the free site read does not test AI answers.
What it does is this. Ollie, the company's AI, researches your business and then writes and publishes pages on your own website, with a team that plans, writes, reviews and publishes alongside it. It works on Shopify, WordPress, WooCommerce and Wix. In the first week you get eight long-form articles, six pages and a custom interactive tool built for your niche, and after that 175 or more pieces of content a month. Customers receive reports on where ChatGPT, Perplexity, Claude and Gemini name the business.
If you want to see what Ollie would plan for your site, paste your address into the free site read at app.runoctopus.com/start. It reads the site and shows the plan. No account and no call.