A patient in Sacramento opens ChatGPT and types "my crown fell off and it's the weekend, what do I do." No location in the query. No practice name. Just a person in a small panic. The assistant walks through the immediate steps, then names two local dentists who handle emergencies. One of them has run a practice in that neighborhood for eleven years. The other opened two years ago. Guess which one got named.
The eleven-year practice didn't get skipped because it's worse at dentistry. It got skipped because the AI never built a clear picture of what it knows. That gap — between being good at something and being understood to be good at something — is the whole game now, and most local professionals don't know it's being played.
Keywords Were a String-Matching Problem
For twenty years, search rewarded a specific kind of thinking. You found the phrase people typed, you put that phrase on the page, you repeated it a sensible number of times, and you earned position for it. "Emergency dentist Sacramento" was a string. Your job was to own the string.
Concept understanding breaks that model completely. When someone types "my crown fell off and it's the weekend," a modern AI assistant doesn't hunt for pages containing those exact words. It recognizes the situation — an urgent dental repair, outside normal hours — and understands that "crown fell off," "emergency dental repair," and "tooth broke in half what do I do" are all expressions of the same underlying need. The literal words dissolve. What's left is meaning.
This is why keyword density stopped being the thing to optimize. An AI isn't counting how many times "emergency dentist" appears on your homepage. It's asking a harder question: does this practice actually understand emergency dental care? And it answers that question by reading everything it can find about you and forming an impression.
What the Model Is Actually Building
Every AI assistant that names a local business is working from an internal representation of that business — a compressed sense of what you are and what you're about. Think of it as the model's mental model of your practice. It's assembled from your website, your directory profiles, your reviews, and any place your name shows up in text the model was trained on or can retrieve.
That representation has a shape. For some practices it's sharp and specific: this is the implant practice in this part of town. For others it's a smear: this is a dentist, probably, who does dentist things. The sharp representation gets named when a matching question comes up. The smear gets passed over, because the model has nothing precise to hold onto.
Here's the mechanism that governs which one you get. AI systems don't trust a single assertion. When your website says you specialize in something, that's one voice making a claim. The model treats it as exactly that — a claim, not a fact. What moves the model from "this is asserted" to "this is true" is agreement: the same specialization showing up, described consistently, across sources that don't answer to each other. Your site says implants. Your Healthgrades profile says implants. Your reviews mention implants by name. Now the model isn't taking your word for it — it's seeing a claim corroborated from several angles, and corroboration is what confidence is made of. A model that finds that agreement recommends you readily. A model that finds your site saying one thing and your profiles saying another hedges, and a hedging model reaches for a name it's more sure about.
Why Depth Beats Breadth, Concretely
Say you're a family dental practice that wants to be known for everything — cleanings, whitening, implants, aligners, root canals, pediatric care, emergencies, the full menu. So you build a thin page for each. A couple of paragraphs, the service name a few times, a call-to-action.
To a concept-understanding model, that practice reads as shallow on all of it. Not because any single page is bad, but because the model averages. It sees a business that names a lot of things and explores none of them, and the representation it forms is "generalist" — which is almost never the answer anyone gets recommended for, because the questions people ask AI assistants are specific.
Now take the same effort and point it at three things. A periodontist in Portland who wants to own gum grafting doesn't publish one page that says "we offer gum grafting." They publish pieces that explore why grafts fail and how to avoid it, what recovery genuinely feels like week by week, how connective tissue grafts differ from the alternatives, what happens if you leave recession untreated. Each piece approaches the subject from a different angle. To a model reading all of it, this isn't a service listing. It's evidence of understanding — the practice keeps returning to this subject and keeps saying something real about it. That's what topical authority looks like from the inside of an AI. Not keyword coverage. Demonstrated depth on a bounded subject.
The model can tell the difference between mentioning a topic and knowing it. Coverage across thirty subjects makes you legible on none. Coverage in depth across three makes you the obvious answer for three specific kinds of question.
One Directory, Read Closely
Directories carry weight in this because they're independent voices the model can cross-check against your own claims. Take Healthgrades, since it's the one most medical and dental searches eventually touch.
What makes a Healthgrades profile useful to an AI isn't that it exists. It's what it contributes to that agreement-across-sources picture. Healthgrades is structured — it separates a provider's specialties, procedures, and conditions treated into their own fields, and it carries patient reviews written in natural language by people describing what actually happened to them. When someone's review says "Dr. Okafor rebuilt my front tooth after a bike accident and I was seen the same afternoon," that's a completely independent voice confirming, in a stranger's words, something your own site might only assert about itself. The model reads that alongside your homepage and your other profiles, and if they all point the same direction, its confidence in your specialization climbs. If the Healthgrades profile is half-empty, or describes a different focus than your website does, the model gets a mixed signal and trusts the whole picture less. The value isn't the listing. It's whether the listing corroborates or contradicts everything else the model already thinks it knows about you.
That's one illustration of a general truth: consistency across independent sources is what turns a claim into something an AI will stake a recommendation on.
The Part That Should Worry You
None of this is theoretical. AI assistants are naming specific local businesses right now, today, in response to exactly the questions your future patients and clients are typing. When RankCommander evaluated visibility across verticals for its AI Visibility Index — thousands of real answers assessed across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Grok, and Copilot — the pattern held across all seven: the practices with sharp, consistent, deeply-covered topical representations got named, and the ones with vague or fragmented ones got skipped, regardless of how long they'd been in business or how good they actually were.
Which means the eleven-year practice losing the emergency crown patient in Sacramento isn't losing a "visitor." It's losing a specific person who was ready to walk in the door that weekend, who instead walked into the practice down the street because an assistant knew that one clearly and yours vaguely. That's not a slow erosion you'll notice in a quarterly report. It's happening one named recommendation at a time, and every one of them is a real appointment that went somewhere else.
The competitor two years into practice, the one the assistant named instead of you? They didn't out-work you at dentistry. They just got understood, and you didn't. Yet.
Where This Leaves You
The uncomfortable part is that the businesses winning here aren't necessarily the best ones. They're the legible ones — the practices whose knowledge is visible to a machine that reads everything about you before it decides whether to say your name. What separates them isn't a trick. It's that a model can actually tell what they're for.
The catch is you can't see your own representation from the inside. You don't know what picture ChatGPT has of your practice, or where Gemini gets your focus wrong, or which platform names your competitor when someone asks the question you should own. That blind spot is exactly what costs you the patient you never knew was looking.
RankCommander closes it. Run the scan and see what all seven assistants actually think you do — where you're named, where you're skipped, and where the business you've spent years building is quietly losing ground to a name an AI finds clearer than yours. You can't fix a gap you can't see. Start by seeing it.