How AI Search Chooses Healthcare Providers: Query Fan-Out, Grounding and Citations Explained

The short version

AI search engines work in four stages: they expand the question into many searches (query fan-out), retrieve pages and passages from an index, ground the answer in those passages, and select which sources and providers to cite. A practice gets named when it supplies clear, consistent, corroborated answers at every stage. Most practice sites fail at stage two or three: the facts that would justify recommending them are missing, buried in images, or contradicted elsewhere on the web.

Step 1: query fan-out, one question becomes twenty

Google describes AI Mode as using a “query fan-out” technique: it issues multiple related searches across subtopics at once, then brings the results together. ChatGPT search and Perplexity behave similarly, rewriting a conversational prompt into several search queries behind the scenes.

Take a realistic patient prompt: “My 58-year-old mum has bone-on-bone knee arthritis. Who are the best surgeons for knee replacement in Tampa who take Medicare, and how long is recovery?” A fan-out might produce sub-queries like these:

The practical lesson: an engine rewards the site that answers the cluster, not the single keyword. If your knee replacement page never states which surgeons perform it, which insurers you accept, or what recovery looks like week by week, several sub-queries come back empty for you, and a competitor who wrote those things down fills the gap.

This is also why I build pages from the full fan-out instead of one head term. On a sector page I map the informational, comparison, provider and logistics questions together. The same map drives my orthopedic sample engagement plan.

Step 2: retrieval, the engine pulls passages, not pages

Retrieval systems split pages into passages and score each passage against each sub-query. That changes how a page should be written:

• Answer first. Put the direct answer in the first one or two sentences under a heading phrased the way a patient asks. “How long is recovery after knee replacement?” followed by “Most patients walk with support on day one, drive at four to six weeks and return to most activities by three months” is retrievable. Three paragraphs of preamble are not.
• One idea per section. A passage that mixes insurance, recovery and a doctor bio scores weakly for all three.
• Facts in text. Credentials in a JPEG, hours in a widget that loads late, prices in a PDF: retrieval often misses them.
• Accessible to crawlers. If robots.txt or a firewall blocks OAI-SearchBot, PerplexityBot or Googlebot, you are absent before ranking starts. I see this on healthcare sites more than anywhere else, usually from an aggressive security plugin.

Traditional ranking still matters here. Google states that pages shown in AI features must be indexed and eligible to appear in Search with a snippet; there is no separate AI index to submit to. Strong classic SEO is the entry ticket. Passage quality decides whether you are quoted once you are in.

Step 3: grounding, why facts you never wrote down cannot help you

Grounding means the model anchors its answer in retrieved sources rather than relying only on what it memorised in training. Google’s Gemini API documents “grounding with Google Search” as connecting the model to real-time web content and returning the sources it used. The aim is fewer invented claims and answers that can be traced.

For a clinic, grounding has two consequences:

1. Unstated means unusable. If no retrievable page says your spine surgeon is fellowship-trained in minimally invasive surgery, a grounded engine cannot honestly say so, so it will not recommend you for that reason.
2. Corroboration wins ties. Engines prefer claims that appear on more than one trusted source. Your site saying “board certified” plus the specialty board’s directory and your hospital profile saying the same is far stronger than your site alone.

Grounding also limits harm. In health topics, engines lean harder on authoritative sources for medical facts. My read of a 2026 AI citation study shows government and medical authorities gaining share in every high-stakes topic measured, while brand sites are cited mainly when someone is choosing a provider. That tells you which part of the answer you can realistically own.

Step 4: selection, which providers get named

After grounding, the engine chooses which entities to mention and which sources to cite. In the provider prompts I test monthly, the named practices share a profile:

• Clear specialisation. “Sports medicine and joint replacement” beats “comprehensive care for the whole family” for a knee prompt.
• Consistent entity data. Name, address, phone, providers and specialties match across the site, Google Business Profile, the NPI registry, Healthgrades, insurer directories and hospital pages.
• Third-party mentions. Hospital affiliation pages, society directories, local news, credible “top doctors” lists, podcast or press quotes.
• Review volume and recency on the platforms the engine reads.
• Specific, citable pages for the procedure and location in the prompt.

ChatGPT tends to draw from a wide range of independent sources; Google AI Overviews returns to a smaller set of trusted pages. Third-party breadth moves the first faster, page authority the second.

GEO, AEO and AI SEO: what the terms actually mean

None of these replace SEO. Indexing, internal links, authority and helpful content are still the foundation every engine draws on. What changes is the unit of competition: from the page to the passage, and from the keyword to the entity.

What I check on a practice site, in order

1. Crawler access: robots.txt, firewall and CDN rules for Googlebot, Bingbot, OAI-SearchBot, PerplexityBot.
2. Index status of every procedure, provider and location page in Search Console.
3. Fan-out coverage: for the 20 most valuable patient prompts, does a page answer each sub-query?
4. Passage quality: answer-first sections, question headings, facts in text.
5. Entity reconciliation: site vs Google Business Profile vs NPI vs directories vs hospital profiles.
6. Structured data that mirrors visible text: MedicalClinic, Physician, MedicalProcedure. See my schema guide.
7. Corroboration: which third-party sources mention the practice and surgeons, and which are missing.
8. Measurement: a fixed prompt set run monthly in ChatGPT, AI Overviews, AI Mode, Gemini and Perplexity, logging who is named and which source was cited.

Mistakes that keep practices out of AI answers

• Publishing dozens of thin symptom posts that compete with national health bodies while procedure pages stay vague.
• Separate near-identical pages for every keyword variant, which splits signals and confuses retrieval. One strong page per intent beats five weak ones.
• Surgeon bios with no procedures, training or affiliations in text.
• Location pages that swap the city name and nothing else.
• Blocking AI crawlers by accident, then wondering why ChatGPT never mentions you.
• Buying fake reviews or planting forum posts. In healthcare this risks platform penalties and regulatory complaints, and engines discount it anyway.

Where to start this month

Write down ten questions your best-fit patients would ask an assistant. Run them in ChatGPT and Google AI Mode with your city named. Note who is named and which pages are cited. Then compare your procedure page with the cited one, section by section. The gap is usually obvious within an hour. If you want me to do this properly across engines and specialties, that is the starting point of my AI search optimisation work, and specialty versions exist for orthopedic practices.