How to run an AI visibility audit in an afternoon, with no paid tool

What does an AI visibility audit cover?

An AI visibility audit answers one practical question: when a buyer asks an AI assistant about your kind of business, what comes back, and why? You may also see it called an AI search visibility audit, an LLM visibility audit or an AEO audit. Those names describe much the same job.

It is wider than a check of your website. A site check looks at what you control. An audit also looks at the answers themselves and at the pages other people run, because the assistants link to those too.

You can run a free AI visibility audit in one afternoon with accounts on ChatGPT, Gemini, Claude and Perplexity, a spreadsheet and a document to paste answers into. There are five steps:

• Pick the questions: 15 to 20, mostly in the buyer’s words.
• Ask all four assistants, and log every answer on one sheet.
• Check the site half: can they read it, do they know who you are, is there an answer to quote.
• Check the outside half: which third-party pages the answers lean on, and whether you are on them.
• Decide the order of work, then repeat the same questions later and compare.

That is 60 to 80 answers. Allow an hour or two for asking and logging (a planning estimate, not a measurement). You finish with one filled-in page: the sheet, a few lines under it, and three jobs with a name next to each.

Which 15 to 20 questions should you ask?

A buyer who has not heard of you cannot type your name, so most of your questions must leave it out. Those show whether you are found. A few must name you. Those show whether what the assistants say about you is right.

“Which vet in Spokane is good with nervous cats?” is a buyer question. “Is Juniper Street Vet open on Saturdays?” (an invented name) is a brand question.

Take the wording from where buyers have already written it down: inquiry emails, sales call notes, reviews, the search terms in Google Search Console, and forum threads in your trade. Then build the set:

• 12 to 16 buyer questions, mixing choosing (“who is good for...”), cost, problems and comparisons.
• 3 or 4 brand questions, on the facts that cost you most when wrong.
• The town or region, written into every question where place matters.

Keep the questions you expect to lose: a set built to flatter gives you a number that means nothing. And freeze the wording, because a later comparison only works if the words stay the same.

How do you ask four assistants without your own history in the way?

Your accounts know you. If you have spent months asking an assistant about your business, that history can reach the answers you log, so use the cleanest session each product offers. This table is what each company’s help pages said on November 4, 2026. Menus move, so check yours when you start.

The pages: OpenAI’s “Searching the web with ChatGPT” and “Temporary chat in ChatGPT”; Google’s “Use Gemini Apps”, “View related sources” and Gemini Apps Privacy Hub; Anthropic’s “Enable and use web search”, “Use incognito chats” and “Does Claude use my location?”; Perplexity’s “Memory”, “Account & Settings” and “How does Perplexity work?”. OpenAI’s release notes of August 27, 2026 added a Personalized option to temporary chats, so check that setting first.

Three habits keep the run fair:

• Open a new chat for each question, so one answer cannot shape the next.
• Write the place into the question. OpenAI’s page gives this advice for wrong local results: include your city, neighborhood or postal code.
• Paste each raw answer into your document before you log it, and note the date, the city and the plan at the top of the sheet. Free plans have limits: Anthropic’s page says Claude’s reset every five hours and that searches count toward them.

What do you write down for each answer?

One row per question, one mark per assistant. Copy this table into a spreadsheet and add a row for each question.

• The marks: L means named, with a link to your own site. N means named, with no link to your site. 0 means not named.
• Pages linked: each site the answer links to, whoever runs it. Write “none” if it shows no links.
• Wrong fact: what the answer said, and which assistant said it. Look for hours, price, location and what you sell.

Count the two kinds of question apart. A brand question names you because you typed your name, so adding it to the total flatters you. Count buyer questions for presence: named in how many answers, linked in how many. Read brand questions for accuracy: which wrong facts.

Under the table, keep six lines: counts, others named most, pages linked most, site half, outside half, three jobs. They turn the sheet into your one page, and the worked example shows them filled in.

Can they read your site, know who you are and quote it?

The sheet shows what happens, not why. The site half is the first place to look, because it is the part you can change yourself.

• Can they read it? If your pages are closed to the bots, nothing you write reaches an answer built from a search.
• Do they know who you are? Your site, profiles and listings should read as one business.
• Is there an answer to quote? With no direct answer on your site, an assistant has nothing of yours to use.

Our GEO audit checklist lists every item behind these three, if you want to work by hand. The free AI visibility checker reads the same items in a few seconds with no sign-up. Write one line on your page for each question: pass, partial or fail, and the biggest thing wrong. A score summarizes that work: it says how ready the site is, not how often you are named, and our guide to AI visibility scores explains how to read one.

Then go back to the sheet. For each buyer question where you were not named, look for a page on your site that answers it directly. Write down the numbers of the questions with none.

Which outside pages do the answers lean on?

The “Pages linked” column shows what else the answers use. Count how many answers link to each site, sort the list, open the top five to ten and ask three things:

• Are you on it? If competitors are listed and you are not, that goes on your page.
• Is what it says about you right? Check hours, price, location and what you sell.
• Does it say what the answer claims? OpenAI’s help warns that search results and citations can be incomplete, outdated or incorrect. Google’s “View related sources” page says the links “may include content that is related to parts” of a response. A link is not proof that the page shaped the sentence next to it.

What you can do depends on who runs the page:

• Directories and review sites: claim your profile, correct it, and use the same name as on your site.
• “Best of” lists and comparison articles: write to the editor with your facts and ask what it takes to be considered.
• Forums and Reddit threads: you cannot edit them. Where the rules allow, answer as yourself and say who you are.

Some wrong facts have no link behind them. Search the web for the wrong fact itself: the old phone number, the former address, last year’s price. A page that still carries it is a likely source, though you cannot be sure the assistant used it.

None of the help pages named here says how an assistant picks its pages, or how much one page counts. A site that keeps appearing is evidence that it is used, not proof that it decides who gets named, and getting onto it does not guarantee a mention.

How much does one run really tell you?

Enough to show where to look. Answers vary from run to run and from account to account, so a single run is a snapshot.

From run to run. OpenAI’s help article on how its models are developed, read the same day, says there are multiple plausible ways to complete a sentence, which gives the replies an “inherent element of randomness”. SparkToro, with a researcher from Gumshoe.ai, a startup in AI tracking, tested this in a study published January 28, 2026: 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI Overviews or AI Mode, 2,961 runs in November and December 2025. The authors put the chance that ChatGPT or Google’s AI, asked 100 times, gives the same list of brands twice at under 1 in 100, with Claude only slightly more likely. The same list in the same order was closer to 1 in 1,000.

The authors say they are not professional researchers and did not control country, device or history. Perplexity and the Gemini app were not tested. They advise reading a visibility percentage across many prompts, and asking one prompt at least 60 to 100 times to learn its real set of recommendations. An afternoon does not get there, and that is the honest limit of this audit.

From account to account. All four companies’ help pages, read the same day, describe personal features that can shape what you see. OpenAI’s search page says that with memory on, ChatGPT may use relevant details to formulate a search, and that saved location information may influence results. Google’s past-chats page says Gemini “can learn from your chats”. Anthropic’s memory page says memory is on by default on the Free, Pro, Max and Ultra plans. Perplexity’s says it calls on memory and previous searches “depending on the question”. None of these pages says how much this changes an answer. That figure is not published.

A clean session narrows the gap, but buyers ask from their own accounts, with their own history, in their own town. Clean sessions make your runs comparable with each other. To feel the spread, ask two or three people in your market to put three of your questions to an assistant in their usual account.

Three rules for reading the result:

• Read counts, not cells. One “0” means little. A column of them means more.
• Treat a wrong fact as real the first time. Ask again, and trace it. A buyer who reads the wrong hours can be lost on that one answer.
• Run the set twice before you call it a baseline, on different days. The gap between the runs shows how big a change must be before it means anything.

What should you do about what you found?

The middle column says “may” on purpose: the help pages do not explain how a given answer is built, so these are leads to check.

Then choose three jobs, not ten, working down this order: anything that stops the bots reading your site; wrong facts a buyer would act on; the outside pages that appear most, where you are missing or wrong; missing answers on your site, starting with the questions closest to a sale. Write each job at the foot of your page, with the name of whoever does it.

When do you ask the same questions again?

Once the jobs are done and some weeks have passed: none of the help pages named here says how soon a corrected page reaches the answers. Use the same questions, setup and place, and run the set twice. Keep the first page next to the new one, with the date of each change, so a rise or fall can be read against it. To keep watching after that, read our guide to AI brand monitoring.

When does an AI visibility audit tool make sense?

The afternoon method is enough for a first look. Repeating it by hand every week is what costs the time, and tools automate that part. Of any AI visibility audit tool, ask: which assistants does it ask, and through the app or the API? How many times does it run each question? Can you read each answer word for word, with its links? Can you edit the questions? Which location does it ask from?

What does one filled-in page look like? A worked example

Here is a first run for an invented clinic, Juniper Street Vet in Spokane, with 15 buyer questions and 3 brand questions. The clinic, its rivals and every figure are illustrative. Two of the 18 rows are shown.

• Counts: 15 buyer questions, 60 answers: named in 6, linked in 2. 3 brand questions, 12 answers: 4 carry a wrong fact, all “closed Saturdays”.
• Others named most: Clinic A in 31 answers, Clinic B in 14.
• Pages linked most: a city “best vets” list in 19 answers, a directory in 12.
• Site half: can they read it: pass. Do they know who you are: partial, two spellings of the name. Is there an answer to quote: fail, no prices. Buyer questions with no answer on the site: 2, 3, 5, 6, 9, 11, 14.
• Outside half: city list: not on it. Directory: listed, with old Saturday hours.
• Three jobs: the owner corrects the Saturday hours on the directory and the Contact page. The owner and the web developer settle on one spelling of the name. The owner writes prices on the service pages.

The wrong hours line up with one directory listing, so start there, though the clinic cannot prove the assistants took them from it. And 6 of 60 is a starting point, not a grade: a second run could come back a little different with nothing changed. Clinic A’s 31 does not show that Clinic A is better. It shows which pages to open next.

What the MWY checker does and does not do in this audit

The AI visibility checker covers the site half of this audit. It reads your home page and up to 7 inner pages, plus robots.txt, sitemap.xml and llms.txt, as raw HTML without running JavaScript, and scores 24 items in three groups. The free check gives your score, the score of each group and the 5 items that matter most, in a few seconds and with no sign-up. The rules are fixed, so the same site gets the same result.

The free check does not ask any assistant a question, so the sheet stays yours to fill in. It does not read your listings, your reviews or the outside pages your answers lean on.

Pro, which opens soon, is built for the part that repeats. MWY drafts 20 buyer questions from the pages it read, you can edit them, and each week they go to ChatGPT, Gemini, Claude and Perplexity through their APIs with web search on. The report shows which answers name you, who else they name, which pages they cite and what changed since the week before. Answers through an API can differ from what one person sees in the app.

MWY does not promise that an assistant will name your business. Nobody outside the AI companies controls that. What the checker gives you is a list of what stands between your site and the assistants.