The 30-Minute AI Visibility Audit

The fastest way I know to end a long-running argument inside a company is to put a screenshot on the screen. Not a deck, not a benchmark. A screenshot of what Claude says when you ask it about your own organization. I have watched a room stop talking mid-sentence over that, and the reaction is visceral in a way no opinion ever manages, because everybody has been debating what the market thinks of them for six months and here is a machine that read the whole internet giving its answer in four sentences.

Sometimes the answer is fine. Often there is, as I put it to a room this summer, some cleanup to do.

Here is the audit, in one paragraph. Ask the five big assistants (ChatGPT, Google AI, Perplexity, Gemini, and Claude) two kinds of questions about yourself, in an incognito window, run each one about five times, and log three columns: mentioned, recommended, and sources cited. It is free. It is a spreadsheet and an afternoon. This page is the worksheet.

Scoping it honestly: the six checks below are the thirty minutes. The deeper five-model spreadsheet after them is an afternoon the first time and much faster every month after that. This is the take-home from two talks I gave this summer, the thing the QR code promised, kept current here instead of frozen in a PDF.

I run content and authority strategy at Product.ai (formerly Demand.io), and this is the method my team uses on our own brands before it goes to anyone else. One caveat before anything else, and it is the most useful sentence on the page: run this free method before you pay anyone. Not because paid tools are bad, but because none of them have published validation data for their accuracy claims, and you cannot evaluate a vendor’s number until you have your own.

Why run the audit before you fix or buy anything?

Because you cannot fix what you have not seen, and almost every expensive mistake in this field starts with someone guessing which problem they have.

Two things make the audit worth more than its thirty minutes. The first is the mirror effect above. Screenshots settle arguments that opinions cannot, and the hard conversations about your product, your support, and your reviews all get easier when they start from evidence somebody else generated.

The second is that it tells you where to spend. Every move you make after the audit stops being a guess, because you will know whether the machine cannot find you, cannot quote you, or simply will not recommend you. Those are three different problems with three different fixes, and people routinely buy the fix for a problem they do not have.

Now, the machines will get things wrong about you. That is not a reason to dismiss the result. There is a reason it says what it says, and that reason is usually a real page somewhere with your name on it. Use it as a compass, not a verdict. Somebody at Golden Hour pushed back on exactly this, pointing out that the AI might not be correct, and the honest answer is: no, it is not, but that is a signal too. It means you have some homework.

What are the six five-minute checks?

Print this part. Each check runs about five minutes, and together they are the thirty.

Check two is the one I would not skip. If you are blocking search bots, you are blocking your AEO, and a surprising number of sites are still blocking them because of a setting somebody flipped in 2023 and forgot. It costs you nothing to look. Which bots do what, and which ones are a genuine judgment call versus a never-block, is covered in the starter’s guide.

How do you run the deeper audit?

Open a spreadsheet with six columns: prompt, which AI, run number, mentioned, recommended, sources cited.

Open the five assistants in an incognito window or a temporary chat, so your own history does not flatter the results. Add Copilot if your buyers live inside Microsoft products.

Then write ten to twenty prompts across two buckets.

Bucket one, who-am-I questions. This is your entity check. “Who is [you].” “Is [you] legit.” “What are [you]‘s strengths and weaknesses.” And the one that stings and teaches the most: “why would someone pick [competitor] over [you].” You are looking for whether the machine knows who you are and whether it is blending you with somebody else.

Bucket two, buyer-intent questions. This is your visibility check, and the trick is to write what a customer types before they have ever heard of you. “Best [category] for [use case].” “[Competitor] alternatives.” “What to look for when buying [product type].” If you sell interior design in Los Angeles, that is “top interior designers in LA,” not your own name.

Then run each prompt about five times per assistant.

Five runs is not superstition. SparkToro ran a study in January with 600 volunteers across 2,961 runs and found that fewer than 1 in 100 repeat runs returned the same list of brands, and only about 1 in 1,000 returned the same list in the same order. One run tells you nothing. Five runs tell you a probability.

The first full pass takes an afternoon. Monthly reruns are quick, because the sheet already exists. What you are building over those months is a trend line, which is the actual thing the expensive tools sell you.

What do you log?

Three verdicts per run, and the distinctions between them carry the whole method.

Mentioned. Did your name appear at all?

Recommended. Did the machine actually tell the person to go with you? It is one thing to be mentioned and another thing entirely to be recommended, and conflating them is the most common way people misread their own results. An assistant can quote your page and still send the customer somewhere else. If a hundred reviews on Trustpilot say your brand stinks, you can be cited and not recommended all day long.

Sources cited. This column is the gold, and if you only fill in one, fill in this one. It is the exact list of third-party pages the machine trusts for your question. Which means it is your to-do list, handed to you for free: your authority files, the review sites that matter in your category, and the places your competitors show up that you do not.

Log what it says, too, not just whether you appear. Wrong category, wrong audience, stale facts, a merged identity with your name-twin. Every wrong fact has an address, meaning some record somewhere that the machine lifted it from, and that address is a row on your fix list. That is exactly how we found out the models had decided SimplyCodes, our coupon platform, was a cashback company.

How do you read the results?

Two reads, in this order.

First the binary one: is the footprint clean, or is there cleanup to do? Digital footprint is closer to binary than people expect. You usually know inside ten prompts.

Then the trend read. Count mentions and recommendations across your five runs per prompt and treat the result as a rate, not a fact. One great answer that felt good on a Tuesday is noise. What you are watching over months is three things moving: mentioned climbing means your entity work is landing, recommended climbing means your content and your actual product are landing, and showing up across more of the five models means the corroboration is widening.

Cadence is the whole edge here. What I told the room at the KINN was that “if you conduct this audit even once a month, once a quarter, you’ll be ahead of 90% of people out there,” and I stand by the spirit of it even though I made that number up on stage. Put thirty minutes on the calendar. The sheet compounds.

Two things our own recurring audit at Product.ai turned up, as an illustration of what finds look like. The assistants kept citing Product Hunt when people asked them to compare apps, and we did not have a page there, so we shipped one. And they kept pulling from a review profile that was thin, old, and skewed negative, which we had been ignoring for exactly as long as it had existed. So we stopped ignoring it and built an honest review ask into the product itself, every customer, success or failure, no cherry-picking.

Neither of those was a clever insight. Both were just the machine telling us where it was already looking, which is the part of this that keeps surprising people. Don’t be afraid to audit your own brand.

What do you do with what you find?

Match your dominant failure mode to the fix.

What you’re seeing	What it means	Where to start
Blank or confused “who is” answers; you’re mixed up with a name-twin	You’re failing the entity gate	The two gates, then claim your records
Identity is right, but your site never appears in the sources column	Your pages aren’t liftable	Why AI quotes one paragraph
You get mentioned but never recommended	Your corroboration is thin or sour	Who the machine trusts
It states something factually wrong about you	A bad record is outweighing the true one	Fix the source record, then get third parties saying the right thing

That last row deserves a sentence, because it is the one that comes with real feeling attached. You cannot flip a switch and correct what an AI believes about you. There is no edit button on a model. What you can do is crowd out the wrong version with a consistent true one: fix the facts in the canonical databases, state them plainly on your own site where they are easy to lift, and get a couple of trusted third parties saying the right thing. You are not editing the model. You are outweighing a bad signal, and that takes longer than anyone wants.

When you eventually outgrow the spreadsheet, tools exist that will track your citation share against competitors. They reprice and change constantly, which is why this page is about the method rather than the market. Ask any of them for their validation data before you pay.

Run it once this week. Honestly, the first pass will probably be unpleasant, and the sheet will tell you about work you were hoping not to have. Thirty minutes now buys you a year of not spending money on the wrong problem.

The companion piece to this one is How AI Decides What To Recommend, which covers why the machine behaves this way and what to do about each failure mode above.