The AI visibility audit checklist

You don't need us to know whether AI names your firm. You need a coffee, an afternoon, and this checklist. Run it once by hand — you'll understand exactly what the diagnostic does at scale, and you'll have a real number instead of a guess. Here's the whole procedure.

Before you start: what you'll produce

By the end you'll have a filled grid of yes/no marks and a single score out of 100, scored the same way the Citewise engine scores it. Budget about an hour for a 12-answer version (3 questions × 4 engines), longer if you go deeper.

Step 1 · Write category-level questions, not brand questions

This is the step people get wrong. Do not ask "Is [my firm] a good advisor?" That tests whether the model remembers your name — useless. Ask the question your prospect actually types, which never contains your name:

Good (category-level):

"Best fee-only financial advisor for physicians in Austin"
"Who should a startup founder use for cross-border tax?"
"Top estate-planning attorneys for blended families in Denver"

Bad (brand-recall):

"Is Bluff Creek Advisors any good?" ✗

Write 3 for a quick pass or 12 for a real one. Vary the axes a buyer varies: specialty, fee model, client type, and locality. If you serve physicians and dentists, write one of each. These are the exact axes the diagnostic uses to generate its 12 questions.

Step 2 · Ask each question in all four engines

Open ChatGPT, Claude, Gemini, and Perplexity. Paste the same question into each, unedited. Four engines matter because they retrieve differently — a firm can be an A in Gemini and an F in Perplexity for the same question. A one-engine check is an anecdote.

Two practical notes so your audit is honest:

Turn off "memory" / personalization where you can, and don't sign in as yourself. You want the answer a stranger gets, not one shaped by your own history.
Ask each question fresh in a new chat. Prior turns bias later answers.

Step 3 · Mark two things per answer

For each answer, record two yes/no marks — the same two the engine scores:

Mark	Rule
Mentioned	Does your firm's name appear anywhere in the answer? (Ignore spacing and punctuation — "Bluff Creek" = "BluffCreek.")
Cited	Does your website/domain appear as a linked or named source?

While you're there, jot the names it gave instead — the two-to-four firms that won the answer you lost. That "who wins instead" column is the most actionable thing in the whole audit; it tells you precisely which competitors are on the surfaces you're missing from.

Step 4 · Score it

Total your marks and apply the formula:

score = round( (mentions + citations) ÷ (answers × 2) × 100 )

Worked: if across 12 answers you counted 5 mentions and 2 citations, that's (5 + 2) ÷ (12 × 2) × 100 = 7 ÷ 24 × 100 ≈ 29. Drop any answer where the engine errored or refused from both counts. Map to a grade: 65+ is a B or better (strong), 50–64 a C, 35–49 a D, under 35 an F. The scoring guide has the full band table and why one mark can swing a grade on a small sample.

Step 5 · Read the pattern, not just the number

The score tells you whether there's a problem. The grid tells you where. Look for:

A whole question you lose in every engine. That's a category you're simply not associated with — usually a missing answer-shaped page plus missing third-party presence.
One engine that's much worse than the others. Points at that engine's specific source diet — often review platforms (Perplexity) or the open-web directories.
Mentions without citations. The model recalls you but won't link you — a corroboration gap. Being on the surfaces engines cite is what converts a mention into a citation.
The same one or two competitors winning repeatedly. Go look at where they appear that you don't. That's your directory and listing to-do list, handed to you.

The honest limits of a DIY audit

Running this once is genuinely worth it — it turns "I think we're fine" into a number. But know what the manual version can't do:

Small samples are noisy. 12 answers means each mark is worth ~4 points; one lucky run can move a grade. The paid diagnostic measures 48 for stability.
It's a snapshot. Answers drift with model updates; a one-time audit can't tell you the direction of travel. Re-running monthly is where the signal is.
It's slow to repeat. Doing this by hand every month across a dozen questions and four engines is a chore most firms abandon by month two — which is the whole reason the Monitor exists.

If the DIY number comes back strong (70+), good — you probably don't need us right now, and we'd tell you the same. If it comes back a C, D, or F, the grid you just built is the exact input the fix pack works from.

Skip the afternoon — get it free.

Send us your site and we run steps 1–4 for you across the major engines, then email you the answers with the "who wins instead" column filled in — usually within one business day. No payment, no call.

Get my free scan Full diagnostic — $490

Next: Where the citations come from → · Ship an llms.txt →