How do we audit our website for what AI assistants say about us?

You audit your website for what AI assistants say about you by asking the assistants first and reading your site second. Open ChatGPT, Claude, Perplexity and Google's AI Overview, type the questions your buyers actually type, and record the answers word for word: how you get described, which competitors get named beside you, and which pages get cited. Then go back to your own site and find the sentence that produced each answer. The gaps show up where no such sentence exists.

That's the whole method, and it takes an afternoon. What makes it hard isn't the procedure. It's that most audit advice points you at the wrong layer. Search "AI visibility audit" and you'll get a crawl checklist: schema markup, robots.txt, an llms.txt file, render-blocking JavaScript, whether your content sits behind a client-side render. Those checks tell you whether a machine can read your site. They tell you nothing about whether it found anything worth repeating.

Do we even need to run this audit?

Answer one question before you spend the afternoon. Take the sentence your best customer used when they described their problem to you, before they knew your company existed. Type that into ChatGPT as a question. If your name doesn't come back, run the audit. If it does come back, run the audit anyway, because being named and being recommended are different outcomes and the difference is the whole game.

Skip it if you sell to a market of forty named accounts who all know you personally, or if you're pre-revenue with no public footprint for an engine to read. Everyone else in the $5M to $75M range is being summarized by these systems right now, whether or not anybody on the team has looked.

What are we actually auditing, the site or the answer?

Both, in that order, and the sequence matters more than any single check. An engine doesn't rank your pages the way search did. It compresses everything public about you into a few sentences, keeps whatever is distinctive, and drops whatever is vague. We've written before about why AI describes B2B companies inaccurately, and the mechanism is the same one at work here: generic material gives the model nothing to hold onto, so it fills the space with category boilerplate or a competitor's language.

That's why this is a comprehension audit. You're measuring what survived the compression. A crawl audit asks whether the machine could reach your words. This one asks what it did with them.

“An AI audit measures one thing: whether your website gave the machine anything worth repeating.”

... Greg Rosner, PitchKitchen

How do we run the audit in an afternoon?

Six steps. Use a spreadsheet, one row per engine per question. Don't delegate the reading, because the value is in seeing your own company described by a stranger who owes you nothing.

1. 1Write down five questions in your buyer's words, not your keywords. Category-level, no brand names. "Who helps a healthtech company fix messaging that isn't landing with hospital buyers" beats "B2B messaging consultant." If you can't write five, that's a finding on its own.
2. 2Run each question in a logged-out or temporary session across ChatGPT, Claude, Perplexity and Google's AI Overview. A logged-in session carries your history and will flatter you.
3. 3For each answer, log three things verbatim: how you were described or that you were absent, which companies were named beside you, and which URLs the engine cited inline. Inline citations matter more than what merely got crawled.
4. 4Now ask the branded version at each engine: "What does [your company] do, who is it for, and when should someone choose it over alternatives?" This separates a recall gap from a description gap. Most companies pass this one and fail step 2, and the reason that happens is covered in detail here.
5. 5Trace every claim in every answer back to a specific sentence on your site. Open the page, highlight the line. Where you can't find a source sentence, mark the answer as invented or borrowed.
6. 6Score each page the engines cited against the pages you expected them to cite. The gap between those two lists is usually the most useful thing the audit produces.

What do the results actually mean?

Four patterns cover nearly everything you'll see. Find yours in the left column.

Notice that only the third row is anything close to a technical problem. The other three are messaging findings wearing a technical costume, which is why teams that run a schema audit and then wait for the numbers to move usually wait a long time. If you're also being handed dashboard percentages by a vendor, read them carefully first, because most AI visibility scores are inflated by design.

What did this audit find when we ran it on ourselves?

We track 50 buyer prompts daily across five engines. Over a five-week window we pulled the ten pages the engines retrieved most often from pitchkitchen.com, out of more than 250 published posts.

Nine of the ten were listicles or comparison pages. The tenth was the homepage. Not a single long-form essay made the list. We'd spent months writing the essays.

That's what a real receipt from this audit looks like. It contradicted something we believed. If your audit confirms everything you already thought, you probably asked the engines your own marketing questions instead of your buyer's questions.

What do we fix first?

Fix the sentence before you fix the schema. If the audit showed a distinctiveness or recall gap, the work is on the words that state who you serve, what changes for them, and when someone should pick you over the alternative. That's the raw material every engine compresses, and it's also what a list-maker, a new sales rep, and a buyer's internal champion all need. One fix, four audiences.

Verbal identity is the half of your brand that machines can read at all, and most companies have never written it down. Once it exists, the technical layer starts earning its keep, because there's finally something specific for the crawler to carry.

If you'd rather not run the six steps yourself, we do this as a free AEO audit: we run your category questions across the engines and send back how you're described today, where you're invisible, and how your closest competitor is positioned in the same answer. If you want the messaging-side version, the Brand Signal Score scores your homepage on narrative clarity, trust, AI-readiness and conversion. Either one gets you a real finding this week.

And if the audit turns up the pattern we see most often, where every engine describes you competently and none of them recommend you, the problem isn't discoverability. Your positioning hasn't given anyone, human or machine, a reason to name you. That's fixable, and it's the work we do.