Getting Recommended by ChatGPT and Perplexity: What Actually Works for Clinics

Ask ChatGPT to name a good dermatology clinic in Phoenix and it will name several. Ask Perplexity the same thing and you get a shortlist with citations underneath. Neither assistant knows your practice. Both are doing something more mundane than it looks: running a search, reading a handful of pages, and summarizing what they found. Understanding that process is the whole game, because it tells you where a recommendation can actually be won and where effort is wasted.

This article is about one specific moment: a patient asks an assistant to name a clinic, and your name is either in the response or it is not. For the broader question of how to structure a practice website for AI-driven search, we covered that separately in AI search optimization for medical practices. Here we are narrowing in on earning the mention.

What actually happens when a patient asks for a clinic

The assistants your patients use are not answering from memory. For anything local, current, or commercial — "best knee surgeon near me", "hair transplant clinic reviews Madrid", "who does awake sedation in Austin" — the model triggers a retrieval step first. It writes one or more search queries, pulls back a set of pages, reads them, and generates an answer grounded in what it just read.

Which index it pulls from depends on the assistant. ChatGPT's search and Microsoft Copilot lean heavily on Bing. Perplexity runs its own crawler alongside third-party indexes. Google's AI Overviews and AI Mode work off Google's index. Other assistants license a search provider under the hood. The specifics shift — contracts change, crawlers get rebuilt, features get renamed — but the structure has held steady: retrieve first, summarize second.

The practical consequence is the one most clinics miss. There is no separate "AI channel" sitting beside search that you optimize with different tactics. If your pages are not retrievable in Bing and Google for the questions patients actually type, they are not in the candidate set the model reads, and no amount of clever copywriting changes that. Ordinary medical SEO did not become obsolete; it became the entry fee. Bing in particular deserves far more attention than most practices give it, precisely because it feeds the assistant with the largest consumer audience.

Your own website can only do part of the job

Here is the uncomfortable part. When a model assembles a recommendation, it treats your website as an interested party. Your claim to be the leading clinic in the region carries roughly the weight you would give the same claim from a stranger at a party — some, not much. What moves the needle is corroboration from pages you do not control.

This is why the same names keep resurfacing in AI answers within a given city. Those practices appear on third-party pages the model retrieves alongside their own site: directory listings with complete and consistent profiles, "best clinics in X" roundups published by local media or patient-facing portals, review platforms, professional association member pages, local press, quoted commentary in trade coverage. You do not write any of these. You earn them, or in the case of directories, you simply claim and complete them.

Perplexity makes the pattern easiest to observe, because it shows its sources. It visibly discounts a company's own promotional language and pulls candidate names out of comparison pages and listicles written by other people. If your clinic is absent from every third-party list covering your city and specialty, you are asking an assistant to promote you on your own say-so, and generally it will not.

Sequence matters. Claim and complete the profiles first — free, fast, and they feed the entity signals covered below. Then pursue genuine third-party mentions: local journalists, specialty associations, patient education portals, podcasts, hospital affiliations. That part is slow and nobody can shortcut it.

The entity problem: can an assistant tell you are real?

Before a model recommends you, it has to be reasonably confident that your clinic is a distinct, real organization and not a name it half-recognizes. This is where a surprising number of otherwise well-marketed practices fall down.

Assistants resolve entities by looking for agreement across independent sources. When your name, address, and phone number differ between your website footer, your Google Business Profile, an old directory listing, and the insurance panel page that still shows a suite number you left in 2021, the model has no clean signal to anchor to. Ambiguity gets resolved by dropping you, not by guessing in your favor.

A few things worth auditing this quarter:

• Name, address, phone consistency across every listing you can find, including the ones you forgot you created. Search your phone number in quotes and see what comes back.
• Duplicate and orphaned listings from old locations, former practice names, or a departed partner's profile.
• Structured data on your site: Organization or MedicalClinic and LocalBusiness markup on the home and location pages, Person markup for practitioners, FAQPage where you genuinely answer questions. It does not force a citation, but it removes ambiguity about who and what you are.
• A real, identifiable author behind clinical content. A named practitioner with credentials, a bio page, and a presence elsewhere on the web (association listings, publications, hospital staff pages) is a stronger signal than "the editorial team".
• Profiles that agree with each other on services offered, languages spoken, and hours. Contradiction between your own profiles is a self-inflicted wound.

Write pages an assistant can lift a sentence from

Models extract. They want a passage that answers the question cleanly and can be quoted or paraphrased without the surrounding paragraph collapsing. That preference has practical consequences for how you write.

Lead with the answer. If the page is about recovery time after rhinoplasty, the first two sentences should state the recovery time, with the caveats following rather than preceding. Marketing preambles push the extractable content below the fold of the model's attention, which functions the same way it does for a human skimming.

Be specific. Named procedures, named devices, named conditions, actual numbers, actual price ranges where you are willing to publish them, actual durations. Vague reassurance is unextractable — there is nothing in "we take a personalized approach" that an assistant can hand to a patient.

Use question-shaped headings that match how patients phrase things, and let each page answer one question completely instead of touching six questions shallowly. A page that fully answers "how long does dental implant osseointegration take" is more likely to be retrieved and cited than a services page that mentions implants among eleven other treatments.

Reviews function as a filter, not just social proof

When an assistant is narrowing a list of local providers, review data is one of the few quality signals available to it that is not self-reported. Volume, recency, and the substance of what patients wrote all get read. Reviews that mention specific procedures and specific practitioners are more useful to a model than a wall of five-star ratings with no text, because they connect your entity to the services a patient asked about.

The mechanics of building that volume ethically — timing the ask, making it easy, staying inside platform rules and healthcare privacy constraints — are their own subject, and we wrote about them in how to get more patient reviews. The relevant point here is that review work now pays twice: once with patients reading them directly, once with the systems deciding whether to name you at all.

Check that AI crawlers can actually reach your site

This is the least glamorous item on the list and the one that most often turns out to be the actual problem. A meaningful number of sites block AI crawlers without anyone having decided to.

The user agents that matter are GPTBot and OAI-SearchBot (OpenAI), ClaudeBot, PerplexityBot, and Google-Extended. Check two places, not one. First, your robots.txt — a previous developer or an SEO plugin may have added blanket disallows. Second, and more commonly missed, your CDN or security layer. Cloudflare, Akamai and others ship one-click "block AI bots" toggles and managed bot rules that operate at the network level, entirely independently of what your robots.txt says. A site can serve a perfectly permissive robots.txt while the CDN quietly returns a block page to every AI crawler that shows up.

Worth knowing: these crawlers do different jobs. Some fetch content for model training, others fetch pages in real time to answer a live query. Choosing to block training crawlers is a legitimate editorial decision. Blocking the search-facing ones removes you from answers. If you block, do it deliberately and know which is which.

What does not work

Some of what gets sold as AI-search optimization is not doing anything.

llms.txt is not a switch. The proposed file that tells assistants how to read your site is cheap to publish and harmless, but no major assistant currently reads it as an input to answers. Treat it as an inexpensive bet on a standard that may or may not get adopted, not as a lever.

Keyword stuffing is worse than useless. Models summarize meaning, not term frequency. Repetition makes a passage harder to extract cleanly.

Bulk AI-generated content does not build authority. Publishing sixty thin pages produces sixty pages nothing links to and nobody cites. The corroboration problem described above is not solved by producing more of your own material.

You cannot buy the recommendation. There is no advertising product that places your clinic inside an organic assistant answer. Where paid placements exist on these surfaces, they are labeled and separate from the recommendation. Anyone offering guaranteed inclusion is selling something that does not exist.

How to measure something with no rank tracker

There is no clean equivalent of a keyword position report here, and honest measurement means accepting some fuzziness.

Start with server logs. Filter access logs by the user agents listed above and confirm those crawlers are fetching your pages, which ones, and how often. This is the only hard, first-party evidence you have, and it answers the access question definitively.

Then test the answers. Build a fixed list of twenty to thirty questions a patient would realistically ask — procedure questions, comparison questions, "best clinic in [city]" questions — and run them monthly across ChatGPT, Perplexity, Copilot and Google's AI surfaces. Log whether you are named, who is named instead, and which sources get cited. Those citations are your target list: they tell you exactly which third-party pages the model trusts for your topic.

Watch referral traffic from assistant domains in your analytics, and keep an eye on Bing Webmaster Tools alongside Search Console. Then accept the caveat: these systems are non-deterministic and personalized, results vary by session and location, and nobody — no agency, no tool — can guarantee a citation. What you can do is make yourself the obvious, well-corroborated, easily-verified answer and let the odds work in your favor.

If you want a second pair of eyes on where the gap sits for your practice, that diagnostic — index visibility, entity consistency, third-party presence and crawler access — is the first thing our team at Medical Marketing runs. Most of the time the answer is not that the content is bad. It is that nothing outside the clinic's own website confirms the clinic exists.

Frequently asked questions

Does llms.txt help my clinic appear in AI answers?

Not currently. llms.txt is a proposed standard for telling assistants how to read your site, but no major assistant uses it as an input when generating answers today. It costs almost nothing to publish and does no harm, so treat it as a low-cost bet on future adoption rather than a lever that changes anything now.

Can I pay to appear in ChatGPT or Perplexity recommendations?

No. There is no advertising product that inserts your clinic into an organic assistant answer. Where paid placements exist on these surfaces, they are labeled and kept separate from the recommendation itself. If a vendor offers guaranteed inclusion in AI answers, they are selling something that does not exist, and you should walk away.

How long does it take before an assistant starts naming my clinic?

Fixing crawler access or robots.txt can show up in weeks. Entity consistency and directory profiles take a month or two to propagate. Earning third-party mentions and building review volume is the slow part, typically six months or more before it shifts answers. Anyone promising results in thirty days is describing a timeline these systems do not follow.

Do I need to rank first on Google to be recommended by AI?

No, but you need to be retrievable. Assistants search an index and read the pages that come back, so if you are nowhere on page one or two for the relevant query, you are unlikely to be in the set the model reads. Being cited also depends on third-party pages that name you, which the assistant may retrieve instead of your site.

Should I block GPTBot to protect my clinic's content?

Understand which crawler does what first. Some fetch content for model training, others fetch pages live to answer a patient's question right now. Blocking training crawlers is a defensible editorial choice. Blocking the search-facing ones removes you from the answers patients see. Check your CDN's bot rules too, since they can block independently of robots.txt.