One in four U.S. adults have already used AI for health information or advice, often before they ever open a search engine. So how does AI search work when the question is about your health? The assistant runs a live web search, reads the pages it finds, and writes one answer with citations. We read 27,812 of those answers to see how they get built. No single brand wins the answer. AI stitches it together from a spread of sources, and your job is to earn a slot in that spread.
This is the pillar for that finding. If you want to understand how AI search works in healthcare end to end, start here, then follow the links into the deeper breakdowns.
How do we know this? The study behind the numbers
Every figure here comes from one dataset. We analyzed 27,812 AI answers generated from 6,953 unbranded consumer-health prompts, US-based, across four engines: ChatGPT, Gemini, Perplexity, and Claude (June 10 to 17, 2026). Those answers carried 675,425 individual citations spanning 33,141 distinct domains across eight healthcare verticals.
"Unbranded discovery" is what matters here. We asked the questions a patient asks before they know which brand to pick: how do I treat this, who should I see, what are my options. That is the moment a brand gets discovered or skipped. We did not score sentiment or audit accuracy. This study is about the mechanics: what AI reads and how it puts the answer together.
How does AI actually assemble a healthcare answer?
AI does not reach for one authoritative site and repeat it. It runs a live search, pulls passages from many pages, weighs them, and writes a single answer with citations. The pattern behind most AI search is retrieval-augmented generation: the system retrieves relevant content, adds it to the prompt, then generates a grounded response that points back to its sources.
So an AI answer is a committee, not a verdict. One definition before we go further. A brand is mentioned (we call this Discoverability) when an AI names it in the answer text. A page is cited when the assistant attributes part of its answer to a URL on that brand's domain. Those are different events, and the gap between them matters more than most marketers expect. A brand can be named constantly while none of its own pages ever get cited.
What sources does AI pull from?
AI builds each healthcare answer from five source types working together. In our dataset, the 675,425 citations split across them like this, and the table below is the one worth quoting.
Source type | What it is | Role in the answer |
|---|---|---|
Social and community | Reddit and other forums where patients compare notes | The consensus backbone. Reddit was the single most-cited domain in every vertical, near 12% of all citations, in about one of every three answers. |
Provider sites | The clinic, brand, or practice's own domain | The layer you fully control, and usually the most under-built for AI. |
Aggregators, directories, and news | Listing sites and health news | The gatekeeper shortlist AI trusts to name specific options. |
Government and academic | Public-health bodies and research institutions | The trust anchor for clinical claims. Rarely a brand slot. |
Reference | Wikipedia and general encyclopedic sources | Definitions and background that frame your category. |
No brand owns the whole answer. We break each layer down on its own: the five source types in full, why Reddit leads the community layer, and the directories AI trusts. Inside specific verticals, individual brands did earn outsized Discoverability: in our dataset, the top-named brand appeared in 29% of at-home lab testing answers, 36% of hormone-therapy answers, and 65% of mental health answers. Every other provider was fighting for the slots that were left.
Do all four AI engines read the web the same way?
No, and this is the finding most marketers miss. "AI" is not one system. In our dataset the four engines read the web at very different depths and named very different numbers of brands, and they agreed on who to recommend only about 7% of the time.
Engine | Sources read per answer | Brands shortlisted per answer | Reddit in the mix |
|---|---|---|---|
ChatGPT | ~32 | ~3-4 | Yes |
Gemini | Not separately reported | ~9 | Yes |
Perplexity | Not separately reported | ~3-7 | Yes |
Claude | ~7 | ~3-4 | ~0% |
That table is why a single scorecard falls apart. ChatGPT read roughly 32 sources per answer and still shortlisted only three or four brands. Claude read about seven and referenced Reddit close to 0% of the time, so a Reddit-driven strategy that wins on one engine can be invisible on another. Gemini named a wider field, around nine brands, so making its list is easier but each slot counts for less. You can rank first on one engine and go unnamed on the next. Each engine should be tracked as its own channel.
Does it matter whether AI searches live or answers from memory?
Yes. An assistant answering from its training memory behaves differently from one running a live search. In Yolando Research (June 2026), answers produced without a live web search named about 30% fewer brands, 4.0 versus 5.8, and skewed toward the largest incumbents. Live retrieval widens the field and gives challengers a route in.
You have to live in both places. Your brand needs to be established enough to sit in the model's training data, and present enough on the live web that the assistant reads you when it searches at the moment of the question. Neglect either and you drop out of a share of answers.
Why does AI add a medical caveat to some categories?
Assistants hedge more as the stakes rise. Across our eight verticals, the share of answers carrying a medical caveat, a line telling the reader to consult a clinician, tracked the risk of the category.
Category | Answers with a medical caveat |
|---|---|
Lab testing | 70% |
Weight loss | 60% |
Hormone / TRT | 56% |
Med spa | 43% |
Psychedelics | 42% |
Fertility | 32% |
Dermatology | 30% |
Mental health | 20% |
The caveat does not push brands out of the answer. It wraps them in a reminder to talk to a professional. In higher-stakes categories, the surrounding context AI trusts (government, academic, and clinical-body sources) does more of the framing, so accuracy and presence in those layers matter more.
Why does Reddit keep winning the community slot?
AI treats community consensus as evidence of what real patients actually think, and Reddit is where that consensus lives. Across all eight verticals it was the most-cited domain, near 12% of citations and present in roughly a third of answers. For treatments that carry fear, cost, or stigma, the honest questions show up there: does this actually work, what did recovery feel like, was it worth it.
You cannot control Reddit. It is tens of thousands of independent threads. But you can steer it: earn authentic mentions, show up honestly under a real identity, and watch the threads that name your brand or your treatments. We go deeper in the Reddit citation analysis.
What does this mean for your brand?
You are competing for a slot in a portfolio, not for the whole answer. That reframe changes the work: stop trying to win the entire response and start earning a defensible place in each source type that feeds it.
Three moves follow directly from the data:
Build your provider site for AI extraction. It is the one layer you fully control, and the one most brands leave under-built. Clear, structured service and condition pages give the assistant something clean to read and cite.
Earn your place in the community and directory layers on purpose. They feed the answer whether you tend them or not, so treat Reddit presence and directory listings as real work, not afterthoughts.
Measure Discoverability across all four engines separately. They agree only about 7% of the time. A page that earns a citation on ChatGPT may do nothing on Claude, and one lucky mention is not a position: in our dataset, 71% of brands AI named appeared exactly once. Consistency is the whole game.
See where your brand sits in the portfolio
You cannot fix a portfolio you cannot see. Yolando's healthcare AI search platform tracks your Discoverability across ChatGPT, Gemini, Perplexity, and Claude, shows which source types feed the answers in your category, and generates the on-brand pages that earn you a slot. Claim your place in the answer.
