Ask ChatGPT which orthodontist in Scottsdale handles adult Invisalign cases. Then ask Gemini. Then Perplexity. You'll notice something that should worry any practice owner who spent the last fifteen years building a website and chasing backlinks: the assistant names two or three offices with confidence, skips right past the rest, and never once explains why. The ones it named didn't necessarily have the most links. Some of them barely rank on page one of a traditional search. What they had was something the old playbook doesn't measure.
They got talked about, in the right places, in plain text.
The assumption that just got inverted
For most of the search era, the hierarchy was settled. Backlinks from high-authority domains were the currency. A link from a respected publication passed authority to your site, moved your rankings, and a whole industry grew up around acquiring them. Mentions without a link were the consolation prize — nice, but you'd email the writer and politely ask them to "add the link" because a mention alone didn't do much.
That instinct is now working against local professionals, and it's worth being precise about why.
A language model was not trained on a link graph. It was trained on text — an enormous corpus of sentences about the world, including sentences that name businesses in context. When you ask ChatGPT or Claude or Copilot to recommend a family law attorney in Tampa, the model isn't traversing hyperlinks the way a crawler does. It's drawing on patterns in the text it absorbed and, for the live-retrieval assistants like Perplexity and Google AI Overviews, on the passages it pulls in the moment. In both cases the unit that matters is the sentence that names you. Whether that sentence contains a clickable link is close to irrelevant to what the model reads.
So the unlinked mention — the thing the old playbook treated as a near-miss — is often doing more work for AI visibility than the backlink you fought to acquire.
Why a model trusts a mention it can't click
Models don't take your website's word for it. Your homepage can claim you're the leading pediatric dentist in Boise, and that claim is worth roughly nothing on its own, because every practice site says a version of the same thing. What a model is actually looking for is corroboration — information that shows up the same way across more than one independent source. Agreement across sources is what turns a claim from an assertion into something the model treats as probably true. When it finds that agreement, its confidence rises and it's willing to name you. When it finds contradictions, or silence, it hedges. And a hedging model reaches for a safer answer — usually someone else's name.
This is where the link-versus-mention distinction stops being academic. A backlink is a machine-readable pointer; it tells a crawler where to go. A mention is a piece of testimony; it tells a reader — human or model — that some independent party, in the course of writing about a topic, named your business as part of the real world. The second one is what corroboration is made of. Ten independent sources mentioning that a Portland estate-planning firm handles complex trusts, none of them linking anywhere, will shape a model's picture of that firm more than a single high-authority link ever could, because the model is reading ten separate confirmations of the same fact.
It's not that links do nothing. It's that mentions punch far above the weight the old scoreboard assigned them, and most local professionals have never once optimized for the thing that's now carrying the load.
Take Super Lawyers, for a moment
To see how a source earns the trust a model extends to it, it helps to sit with one example rather than a list of them.
Super Lawyers is a rating service that has covered attorneys for decades. What makes it useful as an illustration isn't the logo on a lawyer's website. It's the structure of how a listing there comes to exist. The selection runs through peer nomination and independent evaluation before an attorney appears, which means the resulting listing reads, to a model, as a third party vouching for a fact — this person practices this kind of law, in this place, at a level their peers recognized. That's categorically different from the attorney's own site saying so.
Now consider how that surfaces when someone asks Grok or Gemini for a personal-injury attorney in Denver. The model has, somewhere in what it absorbed and what it can retrieve, the attorney's name sitting inside a source built around independent verification, associated with the practice area and the metro. It has, quite possibly, the same attorney's name in a bar association note and a local news quote about a case. The listing isn't valuable because it's a directory with a high domain rating in the SEO sense. It's valuable because the way it's assembled makes each name in it read as corroborated rather than self-declared — the kind of independent testimony that raises a model's confidence enough to actually say the name out loud.
That's the concept worth internalizing. Not "get listed on Super Lawyers." The point is what kind of source produces the sentences a model believes: places where your name arrives through someone else's judgment, in context, consistently. Understanding why that quality of source moves an AI matters more than any single place to be listed, because the reasoning generalizes and the checklist doesn't. (For more on how models assemble these associations, our piece on the entity graph behind AI business relationships goes deeper, and how LLMs decide who to recommend traces the decision itself.)
The mentions you already have — and the ones your competitor has
Here's the part that turns theoretical into personal.
Right now, without you doing anything, models are forming a picture of your practice from whatever text about you already exists. If you've been quoted in the local paper, thanked in a community event roundup, written up in a professional newsletter, that's already shaping how an assistant answers. And if you haven't — if your entire online footprint is your own website plus a handful of review-site listings — then the model has very little independent testimony to work with, and it defaults to the businesses that gave it more.
The RankCommander AI Visibility Index makes this concrete. Across tens of thousands of real assistant answers evaluated and still counting in local professional verticals, the gap between practices that appear consistently and those that never surface isn't primarily a website-quality gap. Plenty of invisible practices have excellent sites. The pattern that separates them tracks how legibly they exist in the text a model reads — the coverage, the context, the corroboration — more than how many links point at their domain. You can see the vertical benchmarks at the AI Visibility Index.
Which raises the question you should sit with. When someone in your city asks an assistant for a recommendation in your category, and the model names two practices and moves on, is one of those names your competitor's? Not because they're better at what they do. Because they got written about, in credible places, in plain text — and the model read it.
What years of good work has to lose
This is the loss-aversion part, and it's real. You built a practice. You earned the reputation the honest way, patient by patient, client by client, over years. That reputation lives in the heads of people who know you and in the reviews of people you served. What it does not automatically live in is the text a language model absorbed about your category — and increasingly, that text is the room where the recommendation happens.
The client who would have found you is now asking Copilot instead of scrolling a search page. If the model doesn't have enough corroborated, in-context material to say your name with confidence, it says someone else's. That's not a slow erosion you'll notice over years. That's a specific person, this week, who needed exactly what you do, handed to a competitor by an assistant that simply had more to read about them. You never see that person. You never know they existed. They just don't call.
The good news is that this is diagnosable and fixable, and it's early. The businesses winning right now aren't the ones with secret knowledge. They're the ones who noticed the shift before it became obvious to everyone. What actually separates them isn't a tactic you can copy in an afternoon — it's that they started treating the text a model reads as an asset worth building, while their competitors were still emailing writers to ask for the link.
See what the assistants already say about you
You can't fix what you can't see, and right now you're flying blind across seven different assistants — ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Grok, and Copilot — each with its own picture of your category and its own list of who gets named. A rank-tracking tool won't show you any of it. It's measuring the old scoreboard.
RankCommander's scan looks at how you actually surface when real people ask real questions in your city and your specialty, across all seven platforms, and it reads the same signal the models read — where and how your name appears in the text they trust, not just where your links point. Run it, and you'll see the answer to the question that should be keeping you up: when the assistant names someone, is it you, or is it the practice down the street that got there first? Find out before the next client asks. Run your visibility scan.