How clinics get cited by ChatGPT and Perplexity
A patient asks ChatGPT which clinics in their city do full-arch implants and what it costs. The answer names three practices. If you are not one of them, no amount of Google ranking recovers that conversation — it never reached a search results page.
The mechanics are different from SEO, and less mysterious than the term "generative engine optimization" suggests. Assistants are looking for text they can lift with confidence and a source they can name without embarrassment.
On this page
How the citation actually gets made
Two different pipelines, and it helps to know which one you are trying to enter.
Retrieval-augmented answers — Perplexity, ChatGPT search, Google AI Overviews — run a live query, pull a handful of pages, and compose an answer from passages they extract. Citation here is mostly a document problem: can the passage be lifted cleanly and does the source look credible.
Answers from model memory have no live retrieval. Here you are competing on how often and how consistently your clinic has been described across the training corpus. That is a slower game, won by being mentioned in places that get scraped — directories, industry publications, your own consistently-worded pages.
The first pipeline is the one you can influence this quarter.
Six things that decide it
Ordered by how much difference each makes in practice.
| Factor | What it means concretely |
|---|---|
| Direct answer up top | A 40–55 word answer to the page's question, before any preamble. The single biggest lever — models lift self-contained passages, and a page that buries the answer in paragraph six offers nothing liftable. |
| Question-shaped headings | H2s phrased the way a patient asks, each answered in its first sentence. Not the third. |
| Tables and lists | Structured content gets quoted disproportionately, because the boundaries of the fact are unambiguous. Real <table> markup, never a screenshot. |
| Entity clarity | Schema that says who wrote this, what they do, and where else they exist. Without it your page is an anonymous document. |
| Corroboration | The same claims about you, worded consistently, in more than one place a crawler can reach. |
| Crawler access | GPTBot, ClaudeBot, PerplexityBot and Google-Extended allowed in robots.txt. Blocking them is a decision many sites made by accident. |
Notice what is absent: keyword density, word count, and backlinks. Backlinks still matter indirectly, because they influence which pages get retrieved. They are not the mechanism.
Entity clarity: the part most sites skip
An assistant will not cite a source it cannot identify.
Most clinic sites present as a collection of pages with no stated author, no stated specialization and no link to anything that confirms either. A model reading that has a passage it likes and nobody to attribute it to, so it reaches for a source that is easier to name — often a directory or a competitor with a clearer footprint.
Fixing this is an hour of work on a template. Person or Organization schema with knowsAbout and sameAs pointing at your real profiles. Article with a named author on every page that argues something. FAQPage where you have real questions. Consistent naming — the same clinic name, the same specialization wording, everywhere.
The goal is that a model resolves your clinic to one entity with a definite specialization, rather than to scattered unconnected mentions.
llms.txt, schema and extractability
llms.txt is a plain-text index at your root listing your important pages with a one-line description of each. It is a convention, not a standard, and no assistant is obliged to read it.
Publish it anyway. It costs twenty minutes, it makes your site's structure legible in a single fetch, and the downside is zero. Treat it as a courtesy to crawlers rather than as a ranking factor, and keep the descriptions honest — a page of marketing adjectives helps nobody.
Extractability is the broader point. Content rendered only by JavaScript, key facts trapped in images, prices stated as "call for pricing" — each is a fact a model cannot lift, and it will find one that can be lifted somewhere else.
How to tell whether it is working
Ask the assistants directly, on a schedule, and log the answers.
- Write 20–30 questions a real patient would ask, in their words, including the ones where you expect to lose.
- Run them monthly across ChatGPT, Perplexity, Gemini and Google AI Overviews. Record which sources get named.
- Check your server or CDN logs for GPTBot, ClaudeBot and PerplexityBot — whether they arrive, what they fetch, how often.
- Watch Search Console for queries where you appear but clicks do not follow. That gap is often an AI Overview answering on your behalf.
- Do not expect a dashboard. This is a citation-tracking exercise, and for now most of it is manual.
VERIFIED EXAMPLE · For a luxury US healthcare clinic, the site went from effectively zero AI visibility to appearing across 600+ AI Overview queries within six months, built on direct-answer structure, entity schema and consistent corroboration rather than on new backlinks.
Questions owners ask
Do I need to rank on Google to be cited by ChatGPT?
No, though it helps. Retrieval-augmented answers pull from a search index, so visibility improves your odds of being retrieved. But the citation decision is made on how liftable and attributable your passage is, which is why pages ranking outside the top ten get cited regularly and page-one results sometimes do not.
Should I block AI crawlers to protect my content?
For a clinic, almost never. Blocking GPTBot or PerplexityBot removes you from answers patients are already asking for, and there is no traffic to protect that is worth that trade. The calculation is different for publishers who sell content; it is rarely different for a business selling a service.
Is llms.txt worth doing?
Yes, with realistic expectations. It is a convention rather than a standard and no assistant is required to read it, but it takes twenty minutes, makes your structure legible in one fetch, and carries no downside. Treat it as housekeeping, not as a ranking lever.
How long does it take to start appearing?
Weeks to a few months for retrieval-based assistants, because they query live indexes and pick up changes quickly once pages are crawled. Answers drawn from model memory move on the timescale of training updates, which is far slower and largely outside your control.
Does this replace SEO?
No. It shares infrastructure with SEO — crawlability, structure, authority — but optimizes for a different outcome. A page written to be cited reads differently from a page written to rank: shorter answers, earlier, with the reasoning after rather than before.