What decides which practices ChatGPT and Google's AI recommend

People selling AI visibility tend to talk as if there’s a secret formula. There isn’t one, but there is a reasonable body of evidence now, from large tracking studies, from industry research and from our own tests. It points to a short list of things that matter and a longer list that gets more attention than it deserves.

This is our read of that evidence as of November 2026, written for owners of independent dental, chiropractic, physical therapy, optometry and aesthetics practices. Each claim carries a confidence label, because some of it is solid and some of it is one vendor’s dataset. If you want the step-by-step version, see how to get your practice recommended by ChatGPT.

How we rated the evidence

We gave each claim one of four labels, based on what’s behind it rather than how often it’s repeated online.

• High: an official statement from the engine’s maker, or several independent studies pointing the same way.
• Medium: one large study, or a vendor study with an obvious commercial interest, or data from outside healthcare that fits how the engines work.
• Low: reporting or small tests only, with no large dataset behind them.
• Unknown: widely advised, but nobody has measured it for AI answers.

Most of the research here is observational. It shows which practices get named, not why, so a pattern like “named practices have more reviews” is strong evidence of a link, not proof of cause. Where a study covered other industries, we say so.

How does each AI engine find local practices?

Each engine pulls local business data from a different mix of sources, which is the main reason they recommend different practices. Google’s AI leans on its own Business Profile and Maps data. ChatGPT leans on Yelp data, search results and practice websites. The others are less documented.

Engine	Main sources for local practices	Confidence
Google AI Mode and AI Overviews	Google Business Profile and Maps, then websites and directories	High
Gemini app	Google Maps grounding, Google Search, websites	Medium-high
ChatGPT	Yelp business data, third-party search (Bing widely reported), practice websites, directories	Medium-high
Perplexity	Its own crawler plus a broad mix of directories and websites	Medium-low
Claude	Brave Search results, then the pages behind them	Medium-low
Microsoft Copilot	Bing search and Bing Places	Medium

Google AI Mode, AI Overviews and Gemini

Google’s AI answers for local questions are built mostly on Google’s own business data. A Steady Demand analysis, reported by Driven Metrics, found about 80% of AI Mode’s local-intent citations went to Maps or Business Profile URLs, rising to 88% for dentists. That figure is second-hand, but it fits how Google describes its own tools: the Gemini API documents grounding answers in Google Maps data.

AI Mode isn’t just the local 3-pack in a new format. A Local SEO Data study of 1,120 searches across seven industries found 28.5% of local-pack businesses were missing from AI Mode for the same search. For medical specialists, AI Mode more often skipped business listings altogether and answered with educational content.

ChatGPT

ChatGPT’s local answers draw heavily on Yelp. Yelp licensed its data to OpenAI in July 2026 (Axios), and in Steady Demand’s August 2026 test of 2,880 prompts, Yelp appeared in the business-card data of 95.83% of ChatGPT local responses, while Foursquare appeared in none (again via Driven Metrics).

For the text of the answer and its links, OpenAI says ChatGPT search uses third-party search providers plus partner content. It doesn’t name them on that page; Bing is widely reported to be one. Practice websites carry a lot of weight: a Dental Economics study of 1,405 patient questions in 20 US markets found practice-owned sites were 53.8% of all ChatGPT citations on dental questions, more than three times the next category. 87.2% of the questions cited at least one practice website.

Perplexity, Claude and Copilot

These three are less studied for local health questions, so the confidence is lower.

• Perplexity runs its own crawler, PerplexityBot, and spreads its citations widely. Yext’s citation study found it drew on a broader mix of directories (MapQuest, TripAdvisor) than ChatGPT or Gemini.
• Claude’s web search appears to use Brave Search, based on reporting from 2025. Anthropic hasn’t published a local-business data source.
• Copilot is Microsoft’s, so it runs on Bing. That makes Bing Places the obvious listing to claim, though we haven’t seen a healthcare-specific Copilot study.

Yext’s data also shows the engines leaning on different source types overall: Gemini favoured websites (52.1% of its citations), while OpenAI’s models leaned on listings (48.7%). That study ran in July and August 2025, before the Yelp deal, so ChatGPT’s reliance on listings has likely grown since.

How often each engine recommends any given business also varies a lot. SOCi’s 2026 Local Visibility Index, covering about 350,000 US locations of multi-location brands, found ChatGPT recommended 1.2% of them, Perplexity 7.4% and Gemini 11.0% (via Driven Metrics). Those are chains, not independent practices, but the gap shows how much stricter or narrower one engine can be than another.

Which levers matter most?

Ranked by the strength of the evidence, the levers are: Google Business Profile, then reviews, then claimed listings (Yelp especially), then your own service pages, then health directories, then Bing, then outside mentions, then crawler access. The order reflects how well each is supported, not how much effort it takes.

1. Google Business Profile completeness and primary category

Confidence: high. For Google’s AI surfaces, the Business Profile is most of the source material, per the citation data above. The primary category decides which questions you’re even considered for, and services, hours, photos and Q&A give the AI details to match against specific questions (“open Saturday”, “takes new patients”). BrightLocal’s 200,000-search study found Google Maps named a tracked business in 66% of searches against 32-38% for AI surfaces, so a strong Maps presence is the base, not the finish line.

2. Review count, then rating

Confidence: high for the pattern, medium for exact thresholds. In Cheers’ data, only 3.7% of 9,435 Google AI Mode answers that showed rated businesses named any business under 4.0 stars (home-services questions, not healthcare). A 2026 roundup reports that the median business ChatGPT named had 175 reviews against 97 for those it left out, from 8ight.AI data.

The best healthcare-specific signal comes from the UK. Whito asked ChatGPT, Gemini and Google AI Mode for the best private dentist in six UK cities. Of 59 practices named, five were named by all three engines, and Whito’s read was that review volume was the common thread: the consensus practices had very large review counts.

There’s a wrinkle for healthcare. The Local SEO Data study found that in consumer industries, businesses shown only in AI Mode had 49% more reviews at the median than those shown only in the local pack, but for lawyers and medical specialists the pattern flipped. For those categories, AI Mode was more likely to answer with educational content than with business listings, so reviews alone won’t carry a specialist practice.

All of this is observational. Practices with lots of reviews also tend to be older, busier and better listed, so reviews may partly stand in for those.

3. A claimed, complete Yelp profile

Confidence: medium-high for ChatGPT, low for Google. The Yelp licensing deal and the 95.83% business-card figure make Yelp the most direct data feed into ChatGPT’s local answers that anyone has measured. It matters much less for Google’s AI, which has its own data.

4. Service and practitioner pages on your own site

Confidence: medium-high, and high for dental questions in ChatGPT. The Dental Economics study is the strongest evidence here: practice sites were the largest single source, and when the question named a specialty like orthodontics or periodontics, they made up over 70% of citations. Pages that answer a specific patient question (who’s a candidate, what it costs, how long it takes) give the AI something to quote. Thin “we offer X” pages don’t.

5. Health directory profiles

Confidence: medium. Yext’s study of 6.8 million AI citations found listings were 52.6% of citations in healthcare, with WebMD, Vitals and other industry-specific directories leading. Yext sells listing management and the data dates from mid-2025, so treat the exact split with some caution.

6. Bing Places and Bing indexing

Confidence: medium. Bing powers Copilot directly and is widely reported as one of ChatGPT’s search providers. There’s no published study isolating Bing Places’ effect on AI recommendations for practices, but it’s free and takes 20 minutes.

7. Mentions on best-of lists, local press and Reddit

Confidence: medium in general, low for local health specifically. In Cleareo’s tests (November 2026), 59% of citations for “best agency” questions in Gemini and Perplexity pointed to best-of list pages, and the most-named agency was never cited from its own site; it was named because other pages listed it. That shows how the engines assemble answers from third-party lists. Whether local “best dentist in [town]” lists carry the same weight for practices hasn’t been measured well.

8. AI crawler access

Confidence: high that blocking hurts, none that opening helps beyond access. OpenAI’s crawler docs state that sites blocking OAI-SearchBot won’t appear in ChatGPT search answers. Once crawlers can reach you, there’s nothing more to gain here. It’s a pass-fail check.

What doesn’t matter as much as people say?

Four things get sold as AI fixes without evidence behind them. They aren’t harmful, but they shouldn’t be where your money goes first.

Claimed lever	What the evidence says	Confidence
Schema markup	No meaningful citation lift in a 1,885-page controlled test	Medium-high
llms.txt	No link to citations across 300,000 domains	Medium-high
NAP consistency	Widely advised, never measured for AI answers	Unknown
Keywords inside reviews	No study found supporting it	Unknown, leaning no

On schema, Ahrefs’ controlled test found changes of about +2% in ChatGPT and AI Mode citations, both statistically indistinguishable from zero, and a small decline in AI Overviews. Schema still helps regular search features, so it’s fine as part of a website build. On llms.txt, Search Engine Journal reported on 300,000 domains: about 1 in 10 had the file, and citation rates didn’t differ.

NAP (name, address, phone) consistency is standard local SEO advice, and it’s reasonable housekeeping. We just haven’t found anyone who measured whether it changes AI recommendations. Asking patients to put “best implant dentist in [town]” in their reviews is a different matter: it makes reviews look scripted, and there’s no evidence it helps.

Why do AI answers change from one ask to the next?

AI recommendations are unstable for three reasons: the same engine gives different answers on re-asks, location and wording change the answer, and different engines disagree with each other. Any single check is a sample, not a measurement.

Re-asking the same question

Dr. Bicuspid’s test of ChatGPT, Gemini and Claude across the 50 largest US cities found 38% of answers changed when the question was re-asked. In Chicago, Gemini’s second answer replaced all five of its first picks. BrightLocal found re-asking brought back the same business only about half the time.

Location and wording

Where the question is asked from, and how it’s worded, change the result. The Local SEO Data study found that adding “near me” to a bare query brought business listings back in every case they tested (38 of 38). Patients don’t phrase questions the way marketers do, which is why we test real patient wording from inside the practice’s area.

Engines disagree

Dr. Bicuspid’s three engines named 560 different practices, and only 30 appeared in all three. In 28 cities they didn’t share a single practice. Our own numbers show the same thing at the source level. In Cleareo’s tests (November 2026), 100 answers to 25 buyer questions across ChatGPT, Claude, Gemini and Perplexity cited 616 pages, and 90% of those pages were used by only one engine. The engines disagreed by 20 to 33 percentage points on how often they named mid-tier brands. Those were business-services questions, not practice searches, but they show how differently the engines source their answers. Our 20-city dentist study found the same split with practices: ChatGPT, Claude, Gemini and Perplexity named 489 different practices between them, and only 5 were named by all four.

App versus API

Many tracking tools query AI through the developer API rather than the consumer app patients use. In the same tests, the APIs kept the app’s top pick 84-95% of the time but changed about 40% of the shortlist and most of the sources. Tools built on APIs are useful for trends; they aren’t an exact copy of what a patient sees.

Cited but not named

One more pattern from our tests: being cited isn’t the same as being recommended. One vendor’s pages fed 18 answers, and it was named in only 4 of them. An engine can read your page and still name someone else.

Which health directories count in each country?

In the US the data is reasonably clear; for the UK and Australia it’s thin, and we’d rather say that than guess.

United States

Directory	Evidence it’s cited	Best for
Healthgrades	Large US provider directory; no published AI citation share found	Physicians, dentists, chiropractors, PTs
Zocdoc	Booking channel plus profile; no published AI citation share found	Practices taking online bookings
WebMD Care	Named as a leading healthcare citation source by Yext	Physicians, dentists, specialists
Vitals	Named as a leading healthcare citation source by Yext	Physicians, specialists
RealSelf	Dedicated cosmetic treatment review site; no published AI citation share found	Med spas, cosmetic practices
Yelp	Feeds ChatGPT business cards	All practices

Confidence: medium for WebMD and Vitals, which Yext names; lower for the others, which are large and widely used but haven’t had their AI citation share published. Nobody has published a clean ranking by profession.

United Kingdom and Australia

We haven’t found a large, published study of which directories AI engines cite for UK or Australian health practices, or a survey of how many patients there use AI to choose one. The Whito UK study is the best we’ve seen, and it pointed to review volume and an active awards circuit rather than any single directory. Until better data exists, the safe order is Google Business Profile, then reviews on Google and other third-party sites, then the main national directory for your profession.

Two rules shape what you can do there. In Australia, Ahpra’s advertising guidelines ban clinical testimonials in your own advertising, though reviews on third-party sites you don’t control are generally fine. In the UK, the DMCC Act fake review ban applies from April 2025, and aesthetics clinics can’t advertise prescription-only medicines such as Botox.

What this means for your practice

Put your effort where the evidence is strongest. In practice that means:

1. Get your Google Business Profile complete, with the most specific primary category. This covers most of Google’s AI.
2. Build a steady flow of genuine reviews on Google, with HIPAA-safe replies. This helps every engine.
3. Claim Yelp and Bing Places. This is the cheapest way to close the gap between Google and ChatGPT.
4. Make sure your main services each have a page that answers real patient questions.
5. Check the answers monthly across several questions, engines and runs, and compare yourself with your top competitors rather than chasing a single screenshot.

You can run that check yourself in an afternoon, use one of the AI visibility tools, or request our free AI check, which runs five patient questions for your town in ChatGPT, Google’s AI, Claude and Perplexity, twice each. If you want the work done for you, our AI visibility service covers it month to month. We’re new and don’t have case studies yet, so the free check is the fair place to judge us.