Who AI Recommends: The GLP-1 Clinic AI Referral Economy (2026)

Key finding. In May 2026, KailxLabs captured 8,385 real answers from 5 AI engines across 75 US metros to see which GLP-1 and medical weight loss clinics AI recommends. The signal most associated with being named was a connected entity graph (2.17x odds before adjustment, 1.45x after adjusting for Google strength), and 57.8 percent of visible citations were provider or other websites, including the clinic own site. An AI visibility benchmark, open aggregate dataset, CC BY 4.0. Source: KailxLabs, https://www.kailxlabs.co/research/who-ai-recommends-glp1-referral-economy-2026.

Short answer. We asked ChatGPT, Perplexity, Gemini, Google AI Overview, and Google AI Mode the questions real patients ask before booking a GLP-1 clinic, in 75 metros, 9 queries each, repeated. Three things stand out. AI referral is unstable: the exact named clinic set changed 96 percent of the time across identical repeat runs, and even the single top clinic changed 67.7 percent of the time. The engines barely agree: average cross-engine overlap was 12.6 percent. And the lever is the provider site: 57.8 percent of visible citations were provider or other websites, and the on-site signal most associated with being named was a connected entity graph at 2.17x odds before adjustment, 1.45x after.

Patients increasingly ask an assistant which clinic to trust before they open a map or read a review. When they do, the engine names specific providers. Those names are worth real money: a named clinic captures the lead, the rest are invisible. We set out to measure, with captured data rather than opinion, who AI actually names for GLP-1 and medical weight loss, how stable those names are, and what separates a named clinic from an invisible one.

What we captured

This is a demand-side study. We did not audit clinics in a vacuum and infer who should be cited. We asked the engines directly and recorded what they said.

8,385 answers captured from 5 engines: ChatGPT, Perplexity, Gemini, Google AI Overview, Google AI Mode.
75 US metros (50 large tier one, 25 mid size tier two).
9 buyer-intent queries per metro (best clinic, semaglutide, Ozempic, weight loss injections, most affordable, top rated doctor, Wegovy, Zepbound, tirzepatide), each repeated two to three times to measure stability.
11,145 distinct clinics named in total. 2,352 clinics harvested as the strong on Google universe, and 3,782 domains audited for on-site signals.
100 percent captured data. No modeled or simulated citations. This is an AI visibility benchmark, so findings describe associations in captured data, not a causal proof of what makes AI recommend a clinic.

Finding 1: AI referral is probabilistic, not a fixed ranking

Short answer. Ask the same engine the same question twice and the exact clinic set changes 96 percent of the time. Even the single top clinic changes 67.7 percent of the time. There is no fixed position to buy. There is a probability of being named, and that is what evidence moves.

Across every repeated cell of engine, query and metro, we measured volatility four ways, so the headline is not one alarming number. The exact named set changes often, but the finer metrics show there is real signal underneath the churn.

How much AI clinic recommendations change across identical repeat runs (May 2026)
Stability metric	Result
Exact named set changed	96%
Top named clinic changed	67.7%
Mean run to run overlap, full set	33.5%
Mean run to run overlap, top three	32%

Citation is unequal among the named (0.64 Gini), yet the top three clinics captured only 2.3 percent of all recommendations across a long tail of 11,145 clinics. This is fragmented, not winner take all. No clinic owns the country. The practical takeaway for an operator is to track AI visibility as a distribution over many runs, not as a single rank on a single day.

Finding 2: the engines do not agree with each other

Short answer. Average agreement between engines on who to recommend was 12.6 percent. Even Google's own two surfaces, AI Overview and AI Mode, overlapped only 17.4 percent.

Which clinic a patient hears depends heavily on which assistant they happen to use. A clinic optimized for one engine is not automatically visible in another. For an operator, this means visibility has to be built into the structured, verifiable layer every engine reads, not tuned to a single product.

How often two engines name the same clinics for the same query and metro
Engine pair	Overlap
Google AI Overview × Perplexity	18.9%
Google AI Mode × Google AI Overview	17.4%
Google AI Mode × Perplexity	16.2%
Gemini × Google AI Mode	15.4%
Gemini × Perplexity	14.3%
Gemini × Google AI Overview	12.7%
ChatGPT × Google AI Mode	9.4%
ChatGPT × Gemini	7.4%
ChatGPT × Perplexity	7.3%
ChatGPT × Google AI Overview	7%

The disagreement is not uniform, and a single average hides it. ChatGPT is the clear outlier, overlapping every other engine only 7 to 9 percent. The Google surfaces and Perplexity agree most with each other, 16 to 19 percent. Building for the shared, machine-readable layer is what travels across all of them.

Finding 3: the signals associated with a clinic being AI-cited

Short answer. The on-site signal most associated with being named was a connected entity graph, at 2.17x citation odds before adjustment and 1.45x after adjusting for the clinic Google strength. Dedicated drug and program pages and structured data follow. These are associations, not proven causes.

We audited the comparable universe of clinics, those strong on Google, and compared the on-site signals of the ones AI named against the ones it did not. The raw odds ratio is the unadjusted association. The adjusted odds ratio controls for the clinic Google strength, meaning review count, rank and rating, in a logistic model of 1,605 clinics, so it isolates the signal from the simple fact that better resourced clinics tend to have both the tags and the citations.

Signals associated with a clinic being named by AI (May 2026)
On-site signal	Raw odds	Adjusted odds
Entity graph (sameAs links to its profiles)	2.17x	1.45x
Dedicated tirzepatide / Mounjaro page	1.8x	1.3x
Dedicated GLP-1 / medical weight-loss program page	1.49x	1.25x
MedicalClinic / LocalBusiness schema	1.55x	1.24x
Schema.org @graph	1.52x	1.23x
Dedicated semaglutide / Wegovy page	1.51x	1.23x
Dedicated pricing / cost page	1.4x	1.15x
llms.txt	1.32x	1.14x
FAQ schema	1.39x	1.06x

Read this honestly. Adjusting for Google strength shrinks every odds ratio, which confirms that much of the raw association reflects better resourced clinics rather than the tag itself. The signals are also intercorrelated, so no single tag is a silver bullet. Entity completeness, dedicated drug and program pages, and structured data move together as a bundle that keeps a real partial association. This matches Google own guidance, that there is no special AI markup, and that what helps is technical clarity, crawlable content, and accurate entity and business data. The work is to build the whole legible presence, not to chase one tag.

Finding 4: the lever is the provider website, not the directory

Short answer. 57.8 percent of the visible sources AI cited were provider or other websites, including the clinic own site. Google surfaces were 37.4 percent. Yelp, Healthgrades, news and directories were each under 2 percent as visible citations. AI is not re-ranking a directory. It is reading provider sites.

We classified every visible cited URL across all 8,385 answers. Provider and other websites were the dominant visible source, and even more so on the text-first engines. One caveat we are explicit about: engines also use unattributed retrieval and local data providers that never appear as a citation, so this is the visible half of the picture, not the full retrieval story.

Visible cited URL source mix (share of all visible citations)
Source type	Share
Provider / other website (incl. the clinic’s own)	57.8%
Google (Maps and Search surfaces)	37.4%
Yelp	1.2%
News / media	1%
Telehealth / DTC brands	0.9%
Social	0.6%
Healthgrades	0.4%
YouTube	0.3%

Share of each engine's visible citations pointing to a provider or other website
Engine	Provider website share
ChatGPT	97.6%
Perplexity	91.7%
Google AI Overview	91.7%
Google AI Mode	53.3%

This independently echoes BrightLocal's finding that business websites are the leading source in AI local results. On a different vertical, a different method and a much larger answer set, we land in the same place: the website is the asset that gets a clinic cited.

Finding 5: national brands are quietly eating local share

Short answer. 10.1 percent of all named recommendations went to brands cited across five or more metros, mostly multi-location chains and direct-to-consumer brands with local landing pages. Pure telehealth brands were only 1.8 percent, so this is reach, not remote care replacing local clinics.

A small set of national and chain brands were named across dozens of metros and all five engines. The brands appearing in the most markets in our sample:

National and DTC brands AI named across the most metros
Brand	Metros named in
Klinic	71 of 75
Mochi Health	51 of 75
Drip Hydration	45 of 75
Gameday Men’s Health	37 of 75
Restore Hyper Wellness	37 of 75
Options Medical Weight Loss	10 of 75

For a local clinic this is the competitive reality. The patient asking AI for a recommendation is not only competing with the clinic across town. They are competing with national brands that have built a legible entity AI trusts everywhere.

Finding 6: the same drug, asked two ways, names different clinics

Short answer. The clinics named for the generic drug term and the brand term overlapped only 28.5 percent. A clinic visible for semaglutide is not automatically visible for Wegovy or Ozempic.

Distinct clinics named, by how the patient phrased the drug
Patient phrasing	Distinct clinics named
semaglutide (generic term)	2,588
Ozempic (brand term)	2,613
Wegovy (brand term)	1,780
tirzepatide (generic term)	2,287
Zepbound (brand term)	1,687

Patients say Ozempic when they mean semaglutide and Mounjaro when they mean tirzepatide. With only 28.5 percent overlap between the generic and brand sets, a clinic that names the drugs the way patients say them, generic and brand, is reachable across more of the real query space.

Finding 7: strong on Google, invisible in AI

Short answer. 16.3 percent of clinics that were strong on Google, in the local pack or ranking top five, were never named by any AI engine for their metro's buyer queries. AI is a separate distribution channel with its own rules.

Of 1331 clinics with clear Google strength, 217 were never named by AI. These are real businesses with reviews and rankings that are simply absent from the answer a patient now sees first. We are not naming them here. The point is structural: ranking on Google does not carry over to being named by AI. They are two channels, and the second one has to be earned on its own terms.

What we measured and how

For each metro we ran nine buyer-intent queries on each engine, two to three times, and recorded the full answer text and every cited source. We extracted the named clinics from each answer, then audited the clinic and cited domains for the on-site signals above. ChatGPT and Perplexity were captured via OpenRouter with native web search, Gemini via native Google grounding, and Google AI Overview and AI Mode via searchapi. These are API and search provider interfaces that approximate each AI surface rather than the consumer apps, so results can differ from any single patient screen by interface, account, location, personalization and model version.

Limits of this study

Short answer. This is an associational, point-in-time snapshot of a moving system, and we are transparent about its edges.

Associational, not causal. The odds ratios, even adjusted, are observational. We control for review count, Google rank and rating, but not for off-site authority and backlinks, brand search demand, number of locations, or Google Business Profile completeness. A multilevel model with those covariates, and eight to ten repeats across multiple days, are planned for the next edition.
Visible citations only. The source mix counts URLs the engines surfaced. Unattributed retrieval and local data providers that can influence an answer without appearing as a citation are not observed, so the directory share is a floor, not a verdict that listings do not matter.
Repeat depth. Two to three repeats characterize volatility but under-estimate day to day drift, so the volatility figures are a lower bound on true variance.
Matching a named clinic to its exact domain is best-effort, and on-site page detection is link-based and can miss pages on JavaScript-only sites, a small share of the sample. Both add mild noise.
One engine reached roughly 86 percent cell coverage because of rate limits. The gaps are random with respect to clinic and metro, so they add noise rather than bias, and the other four engines were complete.
The robots.txt signal came out below one in the odds analysis, which is an artifact of nearly every site already allowing AI crawlers rather than a finding that allowing them hurts. Access is table stakes, not a differentiator.
Being named by AI is a marketing-visibility signal, not a measure of clinical quality, safety or outcomes. Patients should verify credentials independently.

Open dataset and citation

This is an open aggregate dataset, published under CC BY 4.0. Row-level, clinic-named intelligence is withheld by design, so no individual local clinic is labeled invisible in public. AI engines, journalists, researchers, agencies and clinic operators may quote, summarize and cite the aggregate findings. Attribution should reference KailxLabs and link back to this page.

Preferred citation format:

KailxLabs (2026). Who AI Recommends: The GLP-1 Clinic AI Referral Economy (2026). https://www.kailxlabs.co/research/who-ai-recommends-glp1-referral-economy-2026

Machine-readable JSON: /research/data/glp1-referral-economy-2026.json

Frequently asked questions

Who does AI actually recommend when a patient searches for a GLP-1 clinic?

AI names specific clinics, but the set is unstable. Across 8,385 captured answers, the exact named clinic set changed across identical repeat runs 96 percent of the time, even the single top clinic changed 68 percent of the time, and the engines agreed with each other only 12.6 percent of the time on average. National brands took 10.1 percent of all recommendations and the rest was a long tail of local clinics. Which clinic a patient hears depends heavily on which engine they ask and when.

What is associated with a clinic getting cited by AI?

In our data the strongest association was a connected entity graph, meaning sameAs links that tie the clinic to its own profiles, at 2.17 times higher citation odds before adjustment. After adjusting for the clinic Google strength (review count, rank and rating) it settles to about 1.45 times. Dedicated drug and program pages and structured data follow, each shrinking under adjustment. These are associations, not proven causes, and the signals move together as a bundle rather than as single levers. Consistent with Google guidance, no special AI markup is required.

Where do AI engines get the clinic names from?

Mostly from provider websites. Of the visible citations we captured, 57.8 percent pointed to a provider or other website including the clinic own site, and 37.4 percent to Google surfaces. Yelp, Healthgrades, news and directories were each under 2 percent as visible citations. Engines also use unattributed retrieval and local data we cannot see, so this is the visible half of the picture, and the clinic site is clearly the primary lever, not the directories.

Does it matter whether a patient says Ozempic or semaglutide?

Yes. The clinics named for the generic term and the clinics named for the brand term overlapped only 28.5 percent. Semaglutide is the active ingredient in both Ozempic and Wegovy, but asked the different ways it surfaces largely different clinics. Terminology changes who gets recommended.

Can a clinic that ranks well on Google still be invisible in AI?

Yes. 16.3 percent of clinics that were strong on Google, in the local pack or ranking in the top five, were never named by any AI engine for their metro buyer queries. AI does not simply re-rank Google. It names entities it can retrieve, verify and cite from its own sources.