What AI Cites When Health Is at Stake: A 2026 Citation Study, Read for Clinics

What the study measured

The study pulled every domain cited inside AI answers for prompts about 22 consumer topics, using US data, then checked six of them against UK data. Engines covered: ChatGPT, Google AI Overviews, Google AI Mode and Gemini, captured as a single snapshot in June 2026. Eleven topics were “Your Money or Your Life” (YMYL): credit cards, life insurance, mortgages, car insurance, personal loans, cryptocurrency, health insurance, investing, prescription drugs, weight loss and legal services. Eleven were everyday products such as laptops, mattresses and running shoes.

Every cited domain was sorted into five source types:

One caveat matters before any number: the pooled figures lean heavily towards Google AI Overviews, which made up 77 to 88 percent of citations in the topics split by engine. Read the headline numbers as “AI search, led by AI Overviews”. The engine-by-engine findings further down are compared within each engine only.

Finding 1: official bodies take a bigger share when health is involved

Every YMYL topic saw a clear lift in citations from official bodies. More than 11 percent of citations for health insurance and prescription drug questions came from healthcare.gov, CMS, the FDA, the NIH and MedlinePlus. In weight loss, the NIH and CDC alone held 5.1 percent. Health insurance was the most government-heavy topic of all, with official bodies taking 12.5 percent and healthcare.gov the single most-cited source: 14,352 citations from just 582 pages, roughly 25 citations per page.

In everyday topics, nine of eleven categories saw government sources at 0.2 percent or less. The one exception proves the rule: luggage pulled 3.8 percent from tsa.gov because carry-on rules are regulated. The deciding factor is not how serious the category feels. It is whether an authority defines the facts.

What this means for a clinic. You cannot outrank the CDC on “what is osteoarthritis”, and you should stop trying. In my audits, most practice blogs still spend their budget on exactly these definitional pages. They rarely get cited, because the engine already has a better source. The better move is to cite those authorities yourself, accurately, and spend your effort where an authority does not define the answer: your outcomes, your process, your surgeons, your pricing structure, your location and access details.

Finding 2: forums matter less when the stakes rise

Community sources run at about one in seven citations for everyday products (around 14 percent; smartphones 15.2 percent, laptops 14.9 percent). In most YMYL topics that drops to 6 to 9 percent. People debate earbuds on Reddit; they rarely debate prescriptions there in a way an engine will repeat as fact.

Weight loss broke rank at 11.7 percent, because diet and fitness talk is conversational and experience-led. That exception is useful for healthcare. Topics where patients swap lived experience (recovery after ACL reconstruction, life in addiction recovery, ketamine therapy experiences, chronic pain coping) behave more like weight loss than like prescription drugs. In those topics, what patients say about you in public carries real weight in AI answers.

What this means for a clinic. For procedure recovery, rehab and mental health, a community footprint is part of AI visibility: genuine patient reviews, clinician answers in relevant forums under a real name, and surgeon or therapist video explaining recovery honestly. Never astroturf. Engines and moderators both catch it, and in regulated care it creates compliance risk too.

Finding 3: high-stakes topics lean on a few trusted pages

The study used a concentration score: total citations divided by unique cited pages. A higher score means AI keeps returning to the same pages. Credit cards scored highest at 4.04; mortgages 3.77; personal loans 3.48. Legal services scored only 1.47, because it splits into criminal, family, corporate and more, with few sites covering all of it. Furniture, the loosest everyday topic, scored 1.40.

Two forces drive the score: how settled the trusted sources are, and how tightly the topic is drawn. Narrow topics with comparable options settle on a short shelf of sources. Sprawling topics leave room for a long tail.

What this means for a clinic. Healthcare is not one topic. “Knee replacement recovery” is narrow and settled, dominated by hospital systems and national references. “Ketamine clinic for treatment-resistant depression in Denver” or “accredited sober living near Phoenix” is narrow in a different way: local, commercial and thinly covered. That second kind is where a specialist site can become one of the few sources AI returns to. Pick the narrow, under-served questions in your specialty and answer them more completely than anyone else.

Finding 4: intent decides whether your own site is cited

This is the finding I would put on every clinic owner’s wall. Brand-owned sites held 20.6 percent of citations in credit cards, 19.4 percent in life insurance, 17.1 percent in mortgages and 15.8 percent in car insurance. In prescription drugs they held 1.3 percent, weight loss 0.9 percent and legal services 0.6 percent. No drug manufacturer appeared in the top 15 sources for prescription drugs. The first brand in legal services sat at position 60.

The difference is not risk. It is what the person is doing. When they are choosing a named provider, engines cite the provider’s own pages for facts about it. When they want expert advice, brands fade and authorities take over.

Healthcare contains both kinds of prompt, often from the same patient a week apart:

The study shows brands can win even in regulated categories when buying means picking a provider. One health insurer reached position 4 in health insurance despite government sources owning the top of that topic. A credit bureau earned most of its citations through its blog and ask-an-expert guides rather than product pages, reaching the top six in three finance topics as an information provider, not a seller. Both patterns transfer directly to hospitals and specialist groups.

Finding 5: each engine plays a different role

The deeper analysis split five topics by engine. The difference that held everywhere was breadth of sources.

• Google AI Overviews behaves like a librarian. It returns to the same trusted sources far more than any other engine: concentration of 4.76 on credit cards, 3.56 on health insurance and 3.53 on weight loss. Even its loosest topic was roughly double the other engines. Hard to break into, durable once you are in.
• ChatGPT behaves like a reporter. Concentration of only 0.85 to 1.08, with independent media and review sites making up 80 to 88 percent of its citations in each topic. Reddit was its top single source, yet community sources overall were a small share. Easier to enter, but each source is cited less often.
• Gemini and AI Mode sit between the two, closer to ChatGPT’s breadth.

What this means for a clinic. One content investment pays off differently by engine. A small number of authoritative, citable pages is what AI Overviews rewards. A wide footprint across credible third-party sites is what earns ChatGPT mentions. The one large lead my own consultancy won from AI came through ChatGPT naming me, and when I traced why, it was third-party and topical signals, not my homepage.

The UK cross-check: same patterns, local names

Six categories re-run on UK data kept the same splits; only the domains changed. Gov.uk and the NHS stood in for healthcare.gov: UK health insurance pulled 5.2 percent of citations from official sources, against 0.2 percent for running shoes. Comparison sites dominated UK finance the way US ones do in America. Provider-choice topics still favoured brand domains.

For practices marketing across the US, UK, Canada and Australia, this is practical. Your authority layer changes by country (CMS and state boards; NHS, CQC and GMC; provincial colleges; AHPRA), so citing the right national sources on country-relevant pages is not decoration. It is how an engine checks you belong in that market’s answer. My guides to UK, Canadian and Australian advertising rules exist for exactly this reason.

What I add from healthcare audits

The study measures consumer topics. Healthcare buyers add three layers it does not cover, and these are where I see clinics lose AI answers they should win.

1. Entity consistency. AI tools cross-check a practice’s name, address, phone, providers and specialties across the NPI registry, Google Business Profile, Healthgrades, insurer directories and the site itself. Conflicts (an old address, a departed surgeon, a mismatched specialty) make engines hedge or skip you. This is the cheapest fix with the largest effect I find.
2. Credentials in crawlable text. Board certification, fellowship training, procedure volumes and hospital affiliations often live in images, PDFs or nowhere. If a crawler cannot read it, the engine cannot use it to justify recommending you.
3. Restricted-ads sectors depend on this more. Addiction treatment needs LegitScript certification to advertise on Google; psychedelic-adjacent and prescription treatments face tight ad policies. For those businesses, organic and AI answers are not a second channel. They are often the main one. I cover this in LegitScript and rehab SEO and ketamine clinic marketing.

The playbook: seven moves for clinics, labs and treatment centres

1. Sort your prompts by intent. List 40 to 60 real questions patients ask an assistant, grouped as advice, comparison, provider choice and logistics. Track provider-choice and logistics first; they are the citations your own domain can win.
2. Build provider-choice pages that answer like a fact sheet. Procedure pages that state who performs it, their training, where, what the process and recovery look like, insurance accepted, and the next step. Write the answer in the first 100 words, in plain sentences an engine can lift.
3. Stop competing with the CDC. Prune or merge thin definitional posts. Point to authorities with accurate citations, then add what only you know: protocols, outcomes you can document, patient pathways.
4. Become the source behind the sources. Offer clinicians as named experts to health journalists and associations, publish original data (wait times, outcome audits, benchmark studies), and get listed accurately on the directories AI already trusts in your specialty.
5. Fix the entity layer. Reconcile NPI, Google Business Profile, directories and site data; add Physician, MedicalClinic and MedicalProcedure schema that matches visible text. See my medical schema guide.
6. Earn the community layer honestly in experience-led specialties: reviews, clinician answers, recovery video.
7. Measure by engine. AI Overviews and ChatGPT reward different things. Track citation share for your prompt set in each, monthly, and log which page or third-party mention each engine used.

If you want this run against your own practice, my AI search optimisation service starts from exactly this prompt map. For the mechanics behind it (how engines expand one question into many searches and ground their answers), read how AI search chooses healthcare providers.

Limits of this data

It is a single snapshot from June 2026, so it makes no claims about month-to-month movement. Pooled figures are weighted towards AI Overviews by volume. Topics were defined at product level, and healthcare provider prompts (find a surgeon, choose a rehab) were not measured directly; my healthcare reading applies the patterns, it does not re-measure them. Treat the numbers as direction, and test your own prompt set.