A dentist in Portland asked ChatGPT to recommend a good implant specialist in her own zip code. It named three practices. Hers wasn't one of them. She'd been in that neighborhood eleven years, had the reviews, had the referrals. The model had simply never learned to associate her name with the question a patient was actually asking. Two of the three it named had been open less than half that time.
That's the phenomenon, and it's already live. AI assistants answer "who's the best" questions with specific names every day, in every city, across dentistry, law, medicine, and real estate. Somebody gets recommended. The uncomfortable part is how little the recommendation tracks with who's actually best. It tracks with who the model can talk about confidently — and confidence, for a language model, is a very particular thing.
Keyword research is dead weight here
For twenty years the job was to find what people typed and rank for it. Volume tools told you exactly how many people searched a phrase each month, and you built pages against the big numbers. That entire apparatus assumes a search box and a list of blue links.
AI recommendation doesn't work that way, and there's no volume tool for it. Nobody can tell you how many people asked Perplexity "what's a good estate attorney in Austin who handles blended families" last month. You can't buy that data. You have to go find the questions by asking them yourself, across every platform, the way a real person would — messy, conversational, specific to a situation rather than a keyword.
And the questions come in shapes that don't map onto keywords at all. A prospect asks a model to recommend someone. Or to explain whether they even need the service yet. Or what it typically costs. Or whether a given practice handles their specific case. Same category, completely different question, and a model answers each one from a different corner of what it knows about you. Content that satisfies one can be invisible to the next.
Why the model names someone else
Here's the part worth slowing down on, because it's the whole game.
A language model recommending a business isn't retrieving a ranked list from a database. It's assembling an answer from everything it has absorbed, and it's making a running judgment about how much it can trust each thing it's about to say. That judgment is the mechanism. A model doesn't take your website's word for who you are. It looks for whether the same facts about you show up, stated the same way, in more than one independent place. Agreement across sources reads as true. A claim that appears only on your own site, and nowhere else, reads as merely asserted.
When a model finds that agreement, it gets confident, and a confident model volunteers a specific name. When it finds contradictions — your address listed three different ways, your specialty described one way here and another way there, a credential you claim that nothing else corroborates — it hedges. And a hedging model reaches for the safer answer. It names the practice it can speak about without qualification. That practice might be your across-town competitor. The one who's been quietly consistent everywhere a model looks.
Take Healthgrades as a worked example of how this consistency actually functions. For a physician, a Healthgrades profile isn't valuable because it's one more link. It's valuable because of what it is to a model reading the web: an independent, structured, widely-referenced record of a specific person's name, specialty, location, and credentials. When the facts on that profile line up cleanly with the facts on your own site, and with your hospital affiliation page, and with the way you're described elsewhere, a model reading all of it in sequence stops hedging. The claim "Dr. Reyes is a board-certified endocrinologist in San Diego" has been corroborated by something that isn't Dr. Reyes. That corroboration is what converts a maybe into a name spoken out loud. The failure mode is just as instructive: a profile that's half-filled, or contradicts your own site on something as small as which suburb you're in, doesn't just fail to help. It introduces the exact contradiction that makes a model back away from you and toward someone cleaner.
Notice what that isn't. It isn't a checklist of fields to fill in. It's a principle about why verifiable, agreeing information earns a model's confidence, illustrated through one place that happens to embody it well. The principle travels — it's the same reason an attorney's Super Lawyers listing or a realtor's Zillow presence can either reinforce or undermine what a model believes, depending entirely on whether the story agrees with itself.
Writing so a model can actually use it
There's a second layer, and it's about the content on your own site.
Models work in fragments. When one assembles an answer, it lifts a sentence here and a fact there — it rarely quotes a whole paragraph. So a sentence that only makes sense in the context of the page around it is hard for a model to use, even if the information is good. "We've proudly served the community for years with a patient-first philosophy" tells a model nothing it can repeat. A plainly stated, checkable fact about what you do, where, and for whom — a sentence that stands entirely on its own — is something a model can lift into an answer without guessing.
That's what "AI-native content" means in practice. Writing with the awareness that the reader might be a model that will quote you. Most professional service sites are the opposite: fluent, warm, and impossible to cite. They read beautifully and give a model nothing to hold onto. The question worth asking of any page you own is whether a single sentence, pulled out and dropped into a stranger's answer, would still be true and still make sense. If not, no model will use it, no matter how many people it charms.
The gaps are the opportunity
When you actually run the questions — thirty of them, phrased the way real prospects phrase them, across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Grok, and Copilot — a pattern surfaces fast. Some questions return confident, specific names. Others return a shrug: "I don't have enough information to recommend a specific provider." Those shrugs are open ground. Nobody owns the answer yet. The practice that publishes a clear, verifiable, quotable answer to a question the models are currently ducking has a real shot at becoming the name they start giving.
This is where our own data gets specific. Across the AI Visibility Index, RankCommander has evaluated tens of thousands of real assistant answers to real category questions, and still counting — and the through-line is how uneven visibility is even within a single city. In a given vertical, a handful of names recur across platforms while dozens of qualified practices never surface once. You can see the vertical benchmarks and the disclosed sample sizes at the AI Visibility Index. The gap between "excellent business" and "recommended business" is wide, and it's not about quality. It's about whether the model can talk about you.
If you want the mechanism in more depth, we've written about how ChatGPT, Claude, Gemini, and Perplexity actually decide who to recommend, about why content format changes citation rates, and about how to measure AI visibility with real benchmarks.
The part that should keep you up
Every one of those questions is a patient, a client, a buyer, mid-decision. Someone with a real problem, asking a machine who to trust, right now. When the model names three practices and skips yours, that person doesn't see the omission. They just pick from the three. You never find out. There's no bounce rate for a recommendation you were never in.
That's the loss to sit with. Not abstract traffic. A specific person in your city who needed exactly what you do, asked the question out loud, and was handed your competitor's name instead. It's already happening, quietly, across your whole category. The practice down the road that got consistent early is compounding that advantage while you read this — every clean, agreeing mention makes the next answer more confidently theirs. Years of building a reputation the old way can lose ground to that in months, because the model isn't grading your history. It's grading what it can verify today.
Find out what the models say about you
You can keep guessing, or you can see it. RankCommander runs the exact questions your prospects are asking across all seven assistants and shows you where you appear, where you don't, and who's getting named in your place — the specific competitor already occupying the answer that should be yours. That picture, tracked continuously rather than glimpsed once, is the difference between hoping you're visible and knowing. Run your free scan and find out what an AI assistant tells the next patient who asks — before the answer hardens around someone else's name.