How AI assistants choose which doctor to recommend

When a patient types "who is the best hand surgeon in Denver" into an AI assistant, the answer usually names two to five providers. None of them paid to be there. This article breaks down how those names are selected, based on how the systems work and what we see when we test them.

AI assistants do not keep a secret ranked list of doctors. They assemble each answer from what they can find and what they already know, and they are designed to be careful with health information. Understanding the steps makes it clear which parts a practice can influence.

What happens when a patient asks an assistant for a doctor?

For a local or current question, the process usually runs like this:

1. Interpretation. The assistant works out what is being asked: a specialty, a location, sometimes a condition, an insurance plan or a preference such as "takes new patients" or "speaks Spanish".
2. Search. It runs one or more searches behind the scenes. Google's AI features use Google's index. ChatGPT search uses its own crawler plus third-party providers, with Bing widely reported among them. Perplexity runs its own index and retrieval. Copilot uses Bing.
3. Reading. It opens a small number of results: practice websites, directory pages, listings, articles, forum threads.
4. Synthesis. It writes an answer, often naming several providers with a short reason for each, and in many products it shows the sources it used.

Two practical consequences follow. First, if your practice does not appear in the underlying search results for that question, it is unlikely to be named. Second, the reasons the assistant gives are drawn from the text it read, so whatever the sources say about you becomes your description.

Which signals matter most?

No AI company publishes a weighting, so treat this as a practical model rather than a formula. These are the signals that consistently appear in the reasoning and citations of assistant answers about healthcare providers:

Why does consistency matter so much?

Language models are trained to avoid stating things they cannot support. When your website says you are in Suite 210, Healthgrades says Suite 120 and an old directory still lists a partner who retired, the assistant has less reason to name you confidently. When five sources agree that Dr. Patel is a board-certified rheumatologist at a specific address who treats lupus and rheumatoid arthritis, the assistant can repeat that with little risk. This is the entity layer of GEO, and it is often the fastest win for established practices.

Do assistants favor big hospital systems?

For broad questions, often yes. Large systems have strong domains, many pages and extensive press coverage. Private practices compete better on specific questions: a named procedure, a specific population, a neighborhood, a language, a same-week appointment. The more specific the page, the smaller the pool of sources that match it, and the better your odds. That is why a private practice should publish depth on the handful of things it does best rather than a thin page on everything.

How do reviews factor in?

Assistants frequently mention ratings and review counts when recommending a provider, because the directories and listings they read display them prominently. Review content matters too: if many reviews mention short wait times or a physician who explains things clearly, those phrases often surface in the assistant's description. Build a routine that asks every patient for feedback in the same way, keep the process compliant with FTC rules and your state medical board guidance, and respond to reviews without confirming that the reviewer is a patient or disclosing any health information.

Does the assistant check credentials?

Not in the way a credentialing office does. It relies on what sources say. That makes it worth stating credentials precisely and consistently: the full name of the board certification, the fellowship institution, the hospital privileges. It also means that inaccurate claims are risky. If your site says "top-rated" and nothing corroborates it, the claim is unlikely to be repeated and may make the rest of the page look less reliable.

Why do different assistants give different answers?

Each assistant uses different indexes, different retrieval methods and different models, and answers vary even between two runs of the same assistant. Google AI Overviews draw on Google's index and tend to mirror strong local search results. ChatGPT often cites directories and review sites. Perplexity cites a wider spread of sources and shows them all. This is why we test with a fixed set of prompts across several assistants every month and track trends instead of reacting to a single answer.

What can a practice do this month?

• Search your own practice name in ChatGPT, Perplexity, Gemini and Google AI Mode. Note what each says and which sources it cites.
• Ask each assistant the three questions a new patient is most likely to ask about your specialty in your city. Are you named? Who is, and why?
• Fix every factual inconsistency you find across your site, Google, Bing, Apple and the main directories.
• Rewrite your physician bios with verifiable facts and add Physician schema with sameAs links.
• Pick the five procedures or conditions that matter most to your revenue and give each one a strong, dedicated page.

If you want an outside view, our AI visibility audit runs this process for your practice and competitors and returns a prioritized fix list.

What assistants will not do

Responsible assistants are built to avoid giving individual medical advice and to encourage people to consult a clinician. They will not diagnose a patient and then send them to you. They will, however, help a patient shortlist providers for a need the patient has already identified, and that is the moment a practice wants to be visible. Content that respects this line, explaining conditions and treatments accurately without making promises, is the content assistants are comfortable citing.

What does a typical AI answer about doctors look like?

When we test prompts such as "best dermatologist for acne scars in Phoenix", the answers follow a recognizable pattern. The assistant usually opens with a caveat that it cannot personally vouch for providers, then lists two to five practices. Each is followed by one or two reasons: a specialty focus, a notable procedure, a rating or review theme, a hospital affiliation, sometimes a location detail. It often ends by suggesting the patient check insurance and read reviews. Several things stand out when you compare the named practices with those that were left out:

• The named practices almost always have a page that matches the specific need, not only the general specialty.
• Their descriptions across the cited sources match each other closely.
• The reasons given by the assistant are often lifted almost word for word from the practice's own site or a directory profile.
• Practices with sparse or contradictory information are rarely named, even when they are well known locally.

That third point is the most useful for practices. The sentences you publish become the sentences patients read in the assistant's answer. Write them deliberately.

How do assistants handle patient preferences?

Patients add constraints: "who takes Blue Cross", "female OB-GYN who speaks Spanish", "pediatric dentist open Saturdays", "same-week appointment". An assistant can only honor these if the information exists in text somewhere it can find. Many practices mention insurance only inside a PDF, list languages nowhere and bury hours in an image. Put each of these facts on a crawlable page, in a sentence, and mirror them in your Google Business Profile attributes and directory profiles. Preference-based prompts are among the least competitive, because so few practices publish this information clearly.

What role do knowledge bases such as Wikidata play?

Large models learn about well-documented entities from sources such as Wikipedia and Wikidata, and search engines use these to build knowledge graphs. Most private practices will not qualify for a Wikipedia article, and they should not try to create one about themselves. A Wikidata item, where appropriate and properly sourced, or an entry on authoritative professional directories, can help systems connect the practice, its physicians and its location. Treat this as a supporting signal for established practices rather than a first step.

Do assistants read social media?

Sometimes. Public pages on platforms such as LinkedIn, YouTube and Reddit do appear in assistant citations, especially for questions about experiences and opinions. A physician who publishes clear educational videos with accurate titles and descriptions, or who answers questions publicly under their own name, adds to the body of evidence assistants can draw on. Consistency matters here too: the practice name and specialty on social profiles should match the website and listings exactly.

Frequently asked questions

Can a doctor pay to be recommended by ChatGPT or Perplexity?

Organic answers in these assistants are not sold as placements. Some platforms test advertising formats, but those are labelled separately. Being named organically depends on retrieval, clarity and corroboration, not payment.

Why does ChatGPT recommend a competitor instead of me?

Usually because the competitor appears in the search results the assistant retrieves, has clearer information about the specific need and has more consistent third-party mentions and reviews. Check which sources the assistant cites for that answer; they show you where the gap is.

Do AI assistants use Healthgrades and Zocdoc?

They often cite healthcare directories, including Healthgrades, Zocdoc, Vitals and WebMD, for provider questions. Which ones appear varies by assistant and location, so keep all major profiles accurate and consistent.

Does being on a hospital website help AI visibility?

Yes. Affiliation pages on hospital and university sites are strong corroborating sources. Make sure your listing there uses the same name, specialty and practice details as your own site.

How fast can a practice change what AI assistants say about it?

Assistants that search live can reflect updated pages and listings within days or weeks. Knowledge absorbed during model training changes more slowly, over model updates. Focus first on the live sources you control.

Keep reading