I'm a Computer Scientist. Here's How AI Search Actually Picks Which Practice to Recommend

Most articles about ranking in AI search are written by marketers guessing at a black box. I am not guessing. I spent fifteen years building software, including AI and machine-learning work going back to a Bayesian inference project I published on in college, before I ever touched healthcare marketing. So when a patient types “best dermatologist near me” into ChatGPT and your practice does not appear, I can tell you what actually happened, not what a content calendar says happened.

Here is the part nobody wants to admit. The model did not “decide” you were worse than the practice it named. In most cases it never saw you at all.

The answer is assembled, not recalled

When someone asks an AI tool for a local practice recommendation, the system is not reaching into a memory of every clinic in your city. Large language models do not store a clean directory. They store statistical patterns about language. Ask one for a local business by name and you will often get a confident, wrong answer, because the model is predicting plausible text, not reading a database.

So the good tools cheat. ChatGPT with browsing, Perplexity, and Google’s AI Overview do not answer your question from memory. They run a real-time retrieval step first. They issue searches, pull back a handful of sources, and then write an answer grounded in those sources. The language model is the writer. The retrieval layer is the bouncer at the door, and it decides who gets into the room before the model writes a single word.

That distinction is everything. If you are not in the retrieved set, no amount of “quality” gets you into the answer. You lost before the model started typing.

Four things actually move you into the retrieved set

Once you see it as a retrieval problem, the levers stop being mysterious. Here is where I focus, in order of impact.

1. Machine-readable identity. The retrieval layer needs to parse what you are, where you are, and what you treat, in a structured form it does not have to interpret. That means schema markup. MedicalBusiness, Physician, and MedicalSpecialty schema on your site, with consistent name, address, and phone, is the difference between a crawler understanding “board-certified dermatologist in Tulsa treating psoriasis” and seeing an unlabeled blob of marketing copy. Most practice sites I audit have none of this, or have it broken. Your competitors who show up usually have it clean.

2. Corroboration across independent sources. Retrieval systems weight agreement. When your name, specialty, and location say the same thing on your site, your Google Business Profile, Healthgrades, your hospital affiliation page, and a few directories, the system gains confidence that you are a real, specific entity worth surfacing. When those sources disagree, an old address here, a former practice name there, the system discounts you to avoid recommending something wrong. Consistency is not housekeeping. It is a ranking signal.

3. Content that answers the actual question. The queries that matter are not “dermatologist Tulsa.” They are “how long does Mohs surgery recovery take” and “is a mole that changed color dangerous.” Those are the questions patients ask the AI, and the AI pulls answers from pages that address them directly. A practice site with no blog feeds nothing into that pipeline. A site with clear, specific, clinically honest answers becomes a source the model quotes. You want to be the page it cites, because the citation is the new front door.

4. Freshness and crawlability. Retrieval favors pages it can reach and pages that look maintained. If your site blocks crawlers, hides content behind scripts the bots do not execute, or has not published anything since 2022, you are a stale source. Stale sources get passed over for active ones.

What does not move the needle as much as you think

Your star rating matters to humans. It matters much less to the retrieval step, because the model is not running a popularity contest. I have watched 5.0-rated practices with a thousand reviews stay invisible while a smaller practice with clean schema and a real content library got named in the answer. Reviews help you convert the patient once they find you. They do very little to get you found in the first place.

Ad spend does not help here either. The AI answer is not an auction. You cannot buy your way into the cited sources. You earn it by being the most parseable, corroborated, relevant source the retrieval layer can find.

Why this is moving fast

The share of patients who open an AI tool instead of a search engine is climbing every quarter, and it climbs fastest among exactly the people you want, younger patients, higher-income patients, the ones who research before they book. The franchise groups and private-equity-backed platforms already know this. They have the engineering budget to fix their schema and publish at volume. The independent practice that treats AI search as next year’s problem is handing those groups a head start that compounds.

The good news for an independent practice is that this is a technical problem, and technical problems have concrete fixes. You do not need a bigger ad budget. You need your site to be legible to a machine, your identity to agree with itself across the web, and a library of pages that answer what patients actually ask. That is buildable, and it is buildable faster than rebuilding a reputation.

How to check where you stand

Open ChatGPT or Perplexity and ask it, as a patient would, for the best practice in your specialty and city. Then ask the follow-ups a real patient asks. See whether you appear, see who does, and look at what those practices have that you do not. Nine times out of ten the ones who show up have cleaner structured data and more on-topic content, not better medicine.

I run this audit for practices and hand over exactly what is missing and what it would take to fix, with no pitch attached. If you want to know why the AI cannot find you, that is the place to start.

William Hunt HuntGrowth | ryan@huntgrowth.net