A dentist in Tucson has a 2,400-word page about sedation dentistry. It's genuinely good. A hygienist wrote most of it, a copywriter polished it, and it ranks fine in traditional search. Then someone asks ChatGPT, "who does sedation dentistry near me in Tucson," and the model names a practice two miles away with a thinner, plainer page. Not because that practice is better. Because a machine could pull a clean answer out of their page and couldn't out of his.
That gap — between content that reads well and content that gets quoted — is the whole game now. And most local professionals are on the wrong side of it without knowing it.
The thing AI does that search never did
Traditional search sent people to your page. The click was the point. An AI assistant does something different: it reads your page, decides whether it can safely repeat something from it, and then answers the person itself. You may never get the click. What you get instead is the sentence — your name, spoken by the model, presented as the recommendation.
So the question stops being "does my page rank" and becomes "can a machine lift a usable claim out of my page and stand behind it." Those are not the same skill. A page can be thorough, well-written, accurate, and still be nearly impossible to quote — because everything on it only makes sense in relation to everything else around it.
That's the trap long-form content falls into. Word count feels like authority. It reads like effort. But an assistant answering "what's the recovery time for a dental implant" doesn't want your 2,000-word journey through the topic. It wants one clean, standalone sentence it can trust and hand over. If that sentence doesn't exist on your page as a discrete, liftable unit, the model goes looking for a page where it does.
Structure is what makes a sentence liftable
Here's the mechanism underneath all of it. When an AI model builds an answer, it isn't reading your page the way you wrote it. It's scanning for passages that survive being removed from context — a sentence that still means the same thing standing alone as it did in the paragraph. Extractability, in a word.
A well-formed question-and-answer pair is almost engineered for this. The question defines the scope. The answer resolves it. Lift the pair out and it still reads as complete, because it was complete to begin with. Now take the same facts buried in the fourth paragraph of an essay, where the meaning depends on a "this" pointing back two sentences and a "which is why" pointing forward one. Pull that out of context and it's gibberish. The model knows it's gibberish. So it doesn't use it.
This is why a plain 700-word page organized around real questions people ask will often get cited more than a polished 2,500-word feature on the same subject. The short page hands the machine finished thoughts. The long one makes the machine do reconstruction work — and models, given a choice, route around reconstruction work every time. They quote the source that requires the least of them.
Depth still matters. It just isn't the same thing as length. A focused page dense with specific, concrete claims outperforms a sprawling one that circles a topic without ever committing to a quotable fact. The test that actually predicts citation isn't "how much did I write." It's "does any single sentence here stand on its own." Read your own page and find the sentence a model would steal. If you can't find one, neither can the model.
Why "who did you hear that from" decides everything
Extractability gets you into the running. It doesn't win. Because a model that can lift a claim from your page still has to decide whether to believe it — and that decision doesn't happen on your page at all.
AI systems don't take a business's word for it. When your page asserts something about who you are or what you do well, a model treats that as a claim to verify, not a fact to repeat. It looks for whether the same thing shows up, independently, somewhere it doesn't control. Agreement across sources is what turns an assertion into something a model will say out loud. When the model finds your claim echoed in places you don't own, its confidence climbs and it starts recommending you by name. When it finds silence — or worse, contradiction, your name spelled one way here and another way there, one specialty listed on your site and a different one elsewhere — it hedges. And a hedging model is a model that recommends someone else, because naming the wrong business is a mistake it's built to avoid.
Take Healthgrades, as one worked case of how this plays out for a medical practice. On its own, a Healthgrades profile is not magic. What makes it matter to an AI model is that it's an independent surface — the practice doesn't write it, doesn't control it, can't quietly edit it to say whatever's convenient. So when a model sees a physician's name, specialty, and location stated the same way on their own site and on Healthgrades and in a couple of other places the practice doesn't own, that consistency reads as corroboration. The claim isn't just asserted anymore. It's confirmed by sources that had no reason to collude. That's the quality a model is hunting for when it decides whose name is safe to put in an answer. A profile that's stale, half-filled, or subtly at odds with the practice's own site does the opposite — it introduces the exact contradiction that makes a model back away and hedge toward a competitor whose story lines up cleanly everywhere it looks.
Notice what that means. Your best-structured page can be perfectly extractable and still lose, because the corroboration a model needs to trust it lives out in the world, not on your site. Format gets your sentence readable. Consistency across independent surfaces gets it believed. You need both, and they're built in different places.
This is already happening, across every assistant
None of this is a forecast. Ask ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Grok, and Copilot the same local question — "best estate planning attorney in Naples," "pediatric dentist that takes my insurance in Boise" — and you'll watch them name specific practices, skip others, and disagree with each other. Each is running its own version of the same read-verify-decide loop, which is why the same firm can be the confident answer on Perplexity and invisible on Gemini.
RankCommander's AI Visibility Index tracks exactly this, across verticals, evaluating thousands of real AI answers to see who gets named and who gets skipped — the full methodology and disclosed sample sizes are at /ai-visibility-index. The pattern that shows up again and again isn't subtle. In any given market, a small number of practices get recommended repeatedly while everyone else is functionally invisible to the assistant, regardless of how good their actual work is. The model isn't judging the quality of the dentistry. It's judging the quality of the evidence.
If you want the deeper mechanics of how machines read structured pages, our guide to schema markup for AI search goes further, and the generative engine optimization guide and our piece on how to rank in ChatGPT cover the wider picture.
What separates the named from the skipped
The practices winning here aren't the ones who wrote the most. Some of them wrote surprisingly little. What they have in common is quieter than word count: they made themselves easy to quote and easy to confirm, so that when a model reaches for a name, theirs is the one it can say without flinching. Where you should start depends entirely on which half of that you're currently failing — and you can't fix what you can't see.
That's the real problem. You can't watch this happen. There's no dashboard in your practice showing that ChatGPT recommended the group across town to a patient who would have been yours, or that Gemini has been quietly answering "who should I call" with a competitor's name for months. The client just never arrives, and you never learn why. Years of building a reputation people trust, and an assistant hands the introduction to someone else in a sentence you'll never hear.
RankCommander shows you that sentence. Our scan checks how you actually appear across all seven assistants right now, whether they can quote you, whether they trust you enough to name you, and who's getting named in your place while you wait. The competitor who's already the answer isn't gaining ground slowly. In AI recommendations there's often room for one name, and right now it may not be yours. Run your scan and find out what the machines are saying about you — before the next patient asks and hears somebody else.