A dentist in Bellingham had spent eleven years building the kind of practice that fills its own schedule by word of mouth. Then a new patient mentioned, almost in passing, that she'd asked ChatGPT for "a good dentist near downtown Bellingham" before booking — and it had named the practice two blocks away, not his. He hadn't lost his reviews. He hadn't lost his ranking on the map. He'd lost the recommendation, in a channel he didn't know existed, to a competitor he'd never worried about.
That conversation is happening in every vertical right now, and most of your clients haven't had it yet. Which is exactly why the audit you run in the first week of onboarding is the most persuasive thing an agency can put in front of them. Not a pitch. A mirror.
Why the first-week audit carries so much weight
When a prospect signs on for AI visibility work, they're skeptical in a specific way. They believe in Google. They can see their map pack. What they can't see is the conversation an assistant has with a customer who never lands on their website at all — the customer who asks Perplexity to compare three orthodontists and books the one it described most confidently. Your onboarding audit exists to make that invisible conversation visible, in their own market, with their own name in it or absent from it.
So the audit is diagnostic, not decorative. It's the difference between telling a real estate broker in Tempe that "AI matters now" and showing her that when you ask Gemini to recommend a listing agent for a first-time buyer in her ZIP code, it names two agents from the brokerage down the road and never says hers. The first is a slide. The second is a client for three years.
Step one: the baseline scan, run like a customer
Everything downstream depends on getting the starting picture right, and the only honest way to get it is to stop thinking like an SEO and start thinking like the person asking the question. Real people don't type domain names into assistants. They type situations. "Emergency dentist open Saturday." "Personal injury attorney who handles motorcycle accidents." "Realtor who knows the school districts in Round Rock."
Run those situational queries across all seven surfaces — ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Grok, and Copilot — because they don't agree with each other. One assistant might name your client warmly while another has never heard of them, and that spread is itself a finding. You're not looking for a single score here. You're building a map of where the client appears, where they're skipped, and, critically, who gets named in the gaps. (Our own AI Visibility Index evaluates tens of thousands of these answers across verticals, and still counting, and the pattern that holds is that appearing on one platform tells you almost nothing about the other six.)
If you want the shared vocabulary for what that baseline is measuring, what an AI visibility score actually is is worth reading before your kickoff call, because clients will ask.
Step two and three: the competitor and the ground truth
The competitor step is where the audit gets teeth. Once you know which situational queries skip your client, you look at who fills the vacuum. It's rarely random. The same two or three names tend to recur across a client's gaps, and those names are your urgency argument made concrete. "This isn't hypothetical. When AI gets asked, it's already saying Dr. Ruiz. Here's the screenshot."
Then you go to the ground truth — the actual information about the business scattered across the web that models read to form their answers. And here's the mechanism worth understanding deeply, because it's the whole game.
AI models don't take a business's word for anything. They can't. A model has no way to verify that a firm is "the leading estate planning practice in the county" just because the firm's homepage says so. What it can do is check whether the same facts show up, described the same way, in more than one independent place. Agreement across sources is what converts a claim from an assertion into something a model treats as true. When a model finds that agreement — the address, the practice areas, the hours, the specialties all lining up across sources it already trusts — its confidence climbs, and confident models recommend. When it finds contradictions, it hedges. And a hedging model, faced with a customer who wants a clear answer, tends to route that customer to the business it can describe without a caveat.
That's why a directory audit and a description-consistency audit aren't busywork. They're an audit of the raw material a model's confidence is built from. A NAP discrepancy isn't a typo problem. It's a trust problem, and why consistency matters this much to AI search goes deeper than the map-pack version of the same idea most agencies still carry around in their heads.
Take Martindale-Hubbell as the illustration
For a firm audit, it's worth understanding one source in real depth, because it makes the whole trust mechanism click. Martindale-Hubbell is one of the oldest names in legal directories, and its Peer Review Ratings have a particular quality that AI models find useful: they're not self-reported. The rating comes from other attorneys and judges evaluating a lawyer's ethical standards and legal ability. That third-party origin is exactly what makes the signal credible to a model — it's the opposite of a business describing itself. A firm can't simply assert an AV rating into existence. Someone else conferred it.
So when an assistant is assembling an answer about an attorney and encounters a Martindale profile that's been claimed, kept current, and corroborates what the firm's own site says — the same practice areas, the same bar admissions, the same office — the model isn't just adding one more citation. It's finding an independent, hard-to-fake source agreeing with the firm's own story. That agreement is what a model reaches for when it decides whom to name with confidence. A firm that ignored the profile years ago, let it drift out of date, or never claimed it at all has left that corroboration on the table, and the model quietly notices the absence.
That's the concept, worked through one example. Not a checklist of fields to fill. The point isn't Martindale specifically; it's understanding what kind of information earns a model's trust and why, so you can recognize it wherever it shows up in a given vertical — Healthgrades and Doximity for physicians, ZocDoc for a booking-heavy practice, Zillow and Realtor.com for agents, each carrying its own version of hard-to-fake corroboration.
The rest of the workflow, and why the roadmap is the deliverable
From there the audit moves through structured data — whether a machine can cleanly parse the business's own pages — and a content gap review, comparing what the site says against the situational questions the baseline scan showed people actually asking. If customers are asking assistants about sedation dentistry and the practice's site never addresses it in language a model can lift, that's a gap with a name on it.
But the deliverable that closes the client isn't the findings. It's the 90-day roadmap that turns findings into a sequence with owners and dates. Clients don't buy problems. They buy a way out, staged so they can watch it work. The early moves are chosen for visible momentum — the corrections most likely to shift how an assistant answers within the first month, so the client sees the mirror change before the retainer renews. If you're walking a skeptical client through why this deserves budget, how to frame that first conversation pairs well with the audit itself.
What actually separates the agencies winning this work from the ones losing it isn't which directories they know about. It's that they treat the audit as a living baseline and keep re-running it, because a model that re-reads the web next month can change its answer without warning, and a snapshot from onboarding is worth less every week it sits still.
The window is open now, and it's narrowing
Your client spent years earning their reputation the slow way. The assistant answering their prospect's question doesn't know any of that history — it knows what it can read and corroborate today, and if a competitor's information reads as trustworthy while your client's reads as ambiguous, the recommendation goes to the competitor, this afternoon, in a conversation neither of you will ever see. That's not a slow decline you can address next quarter. That's a specific patient, a specific client, a specific buyer, choosing someone else right now because an AI couldn't confidently name your client instead.
RankCommander gives your agency the cross-platform scan, the competitor map, and the continuously-tracked score that turns that fear into a diagnosis your client can act on before the gap gets wider. See how agencies are running this onboarding at /agencies — and be the one who shows the client the mirror before their competitor's name is the only one the assistant knows how to say.