How we measured it
Every practice owner has run the experiment. Open ChatGPT, type “best dentist near me,” scan the answer for your own name. If it’s there, relief. If it’s not, panic. Either way, it feels like you learned where you stand.
Our data says you learned almost nothing, because the answer changes on the next ask.
Since early August, every AI visibility scan we run at Glace asks the questions patients actually ask about their healthcare needs. Things like “best dentist in Fremont,” “emergency dentist near me,” “which dentist takes PPO dental insurance.” Nothing branded; a practice’s name never appears in a question. Each question goes to both ChatGPT and Gemini with live web search on, and each is asked repeatedly at intervals, because a single ask proves nothing. Across 19 markets, that added up to 1,140 answers, each naming five to six practices on average.
Across those 1,140 answers, four patterns kept repeating. None of them look like a ranking.
Ask twice, get a different list
Take the same question, the same city, the same assistant, minutes apart. We compared every pair of repeat asks in the dataset, 1,138 pairs in all, after normalizing practice names so formatting differences don’t count as churn.
Here is what a typical pair looks like. The first answer names five or six practices. The second answer also names five or six. Only two of those names appear in both. Roughly half of the first list vanishes on the very next ask, replaced by different practices. Measured formally: pool every name across both answers and, in the median pair, just 29% of them appear in both.
is all that two identical asks share of their recommendations
Same question, same assistant, minutes apart
of repeat asks returned the identical list
of repeat asks shared no names at all
This is the single number that should change how the industry talks about AI search. Google trained everyone to think in rankings: stable positions you climb one competitor at a time. AI recommendations don’t work that way. Each answer is a fresh draw from a weighted pool, and most of the pool rotates on every draw.
Nobody owns the shelf
If AI answers were a ranking, every market would have fixtures: practices that show up in nearly every answer, the way the same three practices sit in Google’s map pack for months. They almost don’t exist. In 16 of our 19 markets, not a single practice appeared in more than half of the AI’s answers for its own city.
The typical market surfaced about 120 different practice names across its answers. More than half of those names appeared exactly once and never again. The most-recommended name in a typical market captured just 7% of all recommendation slots, and the top three names together held under 18%.
cities have no practice that the AI recommends even half the time
of recommended names appear exactly once, then never again
of a city’s recommendations go to its top three practices, combined
The AI shelf in most markets has no fixtures yet. It’s a rotating carousel, and nobody has claimed a permanent slot.
That cuts two ways. It means nobody can honestly sell you “#1 on ChatGPT.” It also means the window is open: the handful of practices in our sample that broke 80% presence prove stable visibility is achievable, and whatever earns it is being earned right now, before the shelf hardens the way Google’s did.
Even the visible flicker
For each practice we tracked how consistently it appeared when the same question was asked again. The pattern is not “visible practices stay visible.” Where a practice appeared at all, it appeared in every repeat of that question only about 60% of the time. The other 39% of the time it flickered: named in some asks, missing from others.
ChatGPT and Gemini are different worlds
Put ChatGPT’s recommendations for a city next to Gemini’s and you are mostly looking at two different lists. Of all the practices named in a typical city, only about one in five was named by both assistants. The rest belong to one engine’s world or the other’s: of the practices reliably visible on at least one assistant, half were invisible on the other. Being recommended by ChatGPT tells you nothing about Gemini, and the reverse.
And when they do land on the same name, it is rarely a discovery: the practices both assistants agree on are the handful the whole city already knows.
This split is not a quirk of one kind of question. Take any question type, compare the names each assistant recommends for it in the same city, and there is no question where the two agree on even one name in five:
| When the patient asks about… | Names both assistants agree on |
|---|---|
| A specific procedure | 10% |
| Affordable | 11% |
| Best overall | 13% |
| Top rated for a treatment | 14% |
| Accepting new patients | 16% |
| Takes my insurance | 17% |
| Emergency, near me | 17% |
| Open on Saturday | 19% |
| A symptom needing treatment | 19% |
| For my kids or family | 20% |
For a practice, the lesson is blunt. There is no being visible “on AI.” There is being visible on ChatGPT, and being visible on Gemini, and they are separate battles: winning one buys you nothing on the other. Why would two AIs reading the same city disagree this much? Because they read it through different evidence, and show different amounts of it. That is exactly where the next finding goes.
The sources AI cites, and the ones it can’t read
Both assistants ran with live web search enabled. What the patient sees could not be more different. ChatGPT showed at least one source link in 99.3% of its 570 answers, about five citations per answer, 2,621 links in total. Gemini showed a source in just 7%: 95 links across 570 answers. Gemini grounds its answers in search too; it just presents the result as settled fact, with addresses, phone numbers, and “highlights” and no visible evidence trail.
So where do ChatGPT’s citations point? Not where the industry spends its energy. 78% of the citation links go to practice-level websites: the recommended practice’s own site, its offers page, its hosted review page. Only 22% go to platforms and directories of any kind. And the platform list is upside down relative to a decade of local-marketing advice:
| Source | What it is | Answers citing it | Markets |
|---|---|---|---|
| reviews.birdeye.com | Hosted review pages | 81 of 570 | 16 of 19 |
| zocdoc.com | Booking marketplace | 66 of 570 | 13 of 19 |
| deltadental.com | Insurer directory | 45 of 570 | 11 of 19 |
| aae.org | Endodontist society directory | 24 of 570 | 14 of 19 |
| findadentist.ada.org | ADA directory | 22 of 570 | 12 of 19 |
| invisalign.com | Vendor provider locator | 17 of 570 | 10 of 19 |
| healthgrades.com | Reviews directory | 17 of 570 | 11 of 19 |
| yelp.com · facebook.com · instagram.com | Where the reviews live | 0 | 0 of 19 |
Read that last row again. In 570 recommendation answers with 2,621 source links, ChatGPT cited Yelp exactly zero times. Facebook: zero. Instagram: zero. Google reviews: effectively zero. This is not a mystery. Yelp’s robots.txt blocks AI crawlers from its business pages,1 so those reviews cannot serve as evidence in an AI answer. Meanwhile Birdeye’s hosted review pages, which are open to crawlers, have quietly become ChatGPT’s favorite review source, cited in 16 of our 19 markets, ahead of Zocdoc and Healthgrades.
Whoever hosts your reviews on an AI-readable page becomes your review site of record. The platforms blocking AI crawlers have written themselves out of the answer.
A concierge and a brochure
The two assistants have also settled into different characters, which matters for what a patient does next:
| In a typical answer… | ChatGPT | Gemini |
|---|---|---|
| Length | 220 words | 391 words |
| Includes a phone number | 96% | 63% |
| Shows its sources | 99% | 7% |
| Formatted with headers and bullets | 48% | 100% |
| Ends by offering to narrow the choice | 8% | 1% |
ChatGPT behaves like a concierge: shorter answers, nearly always a phone number, and sometimes an offer like “tell me your insurance and I’ll narrow it down.” Gemini behaves like a brochure: longer, heavily structured, more names, less action. Either way, the patient is handed a short list and, most of the time, a number to call. What happens when they call it is a different problem, and the reason our AI Front Desk exists.
One more pattern worth knowing: across all 1,140 answers, the assistants named 2,373 distinct providers in 6,492 recommendation mentions, and 68% of those names are practice brands, not individual doctors. AI recommends institutions more readily than people, which quietly rewards practices with a strong brand name over practices trading on a practitioner’s personal reputation.
What this means for your practice
1. Screenshots are not evidence. Any agency claiming they got you “to #1 on ChatGPT” off a single screenshot is selling a lottery ticket stub. With a 29% median overlap between identical asks, one answer proves nothing in either direction.
2. Ask about your draw rate, not your rank. The honest metric is a presence rate: out of many repeated, varied asks, how many name you? That is a number you can track month over month and actually move.
3. Your website is the evidence AI presents. With 78% of citations pointing at practice-level pages, your site, your offers page, and your AI-readable review page are not just conversion assets. They are the receipts ChatGPT hands the patient. A practice invisible to AI crawlers has nothing to be cited with, and the platforms that block those crawlers can no longer speak for you there.
4. The window is open. No fixtures in 16 of 19 markets means AI assistants haven’t crowned local winners yet. The practices that build citable evidence now are buying tickets in every draw while their competitors are still asking what their rank is.
Where Glace fits
This research comes directly out of how our visibility scan works. We never trusted a single ask, so every Glace scan puts the common patient questions to both assistants and repeats them at intervals, the same way this research was run. The score a practice gets is a presence rate, because that is the only number that means anything in a lottery. The scan also reads the answers themselves: which competitors keep appearing, which sources get cited in your market, and which question types you are absent from.
Start here, free
AI Visibility Check
See your practice’s presence rate across ChatGPT and Gemini, measured the way this research was: repeated real answers, not one screenshot
Then build the evidence
Marketing & AI Visibility
An AI-readable website, service-specific content, and reviews where AI can actually cite them, so you appear in more of the draws
When the patient calls
AI Front Desk
96% of ChatGPT answers hand the patient a phone number. The AI Front Desk makes sure that call becomes an appointment
The honest caveats: 19 markets is a small sample, weighted toward dentistry and the metros we work in. Run-to-run variation partly reflects randomness inherent to these models, which is precisely the point, because the patient’s answer is drawn from that same randomness. And these numbers describe August and September 2026. The models will drift, which is one more reason a one-time check is worth less than a measurement you repeat.
There is no AI ranking to win. There is a lottery, and the only strategy is to hold more tickets: more citable evidence, in more of the places AI actually reads, measured often enough to know it’s working.