How a health practice becomes the one AI recommends

The question changed, quietly

Practices used to compete for a position on a results page. The patient saw ten links, and being third instead of eighth was worth money. Now a large share of patients see one answer, containing two or three practice names, written in a confident voice by something that does not tell them what it left out.

At the end of 2025, 31 percent of surveyed patients said they had used AI to find a new provider. By mid-2026 that number was 47 percent. Among the 465 respondents who had used AI that way, two thirds said they had met incorrect information about a provider, and 60 percent said they trust AI summaries without checking them.²

Read the second half of that again. The errors reported were not exotic. They were addresses, phone numbers, insurance details and opening hours. That is not a model inventing things. That is a model faithfully repeating a stale directory entry because nothing more authoritative contradicted it.

What the research actually supports

The most solid source available is a 2024 KDD paper from a Princeton-led team, which tested content changes across roughly 10,000 queries and measured how often each version was cited in generated answers.¹ Three things worked, consistently:

• Adding relevant statistics with the number stated plainly.
• Adding direct quotations from identifiable people or bodies.
• Citing authoritative sources, close to the claim they support.

Those produced gains in the region of 30 to 40 percent on the paper’s visibility metric. Improving fluency and readability helped less but still helped. Keyword stuffing and stylistic padding did not help at all.

This is a genuinely useful result, and it is worth noticing how boring it is. The things that make an AI assistant quote a page are the same things that make a reference librarian trust it. There is no trick in the list.

The part that is harder to hear

Correlation studies published by search-industry vendors through 2025 and 2026 point at something less flattering: the strongest single predictor of being cited is how established the domain already is.⁴ Sites with very large numbers of referring domains are cited several times more often than sites with very few. That is not something a single practice can fix in a quarter, and any agency implying otherwise is selling you a story.

A separate overlap study, testing 15,000 long-tail queries, found that only around 12 percent of the URLs AI assistants cite also rank in Google’s top ten.³ Search position and AI citation are related, but they are not the same outcome, and a practice that is already ranking well should not assume it is being described accurately.

One more caution before the practical part: those same vendor studies report that between 40 and 60 percent of the sources cited in AI answers change from one month to the next.⁴ Their datasets are not published, so read that as an order of magnitude rather than a measurement. A good month is not a result. A bad month is not a crisis. Judge this over three to six months or do not judge it at all.

The order to do the work in

For a health practice, this is the sequence that produces the most improvement for the least effort, and it is deliberately front-loaded with the unglamorous parts.

1. Make the facts identical everywhere

Practice name, practitioner name, credentials, address, phone, hours, services and service area, matching exactly across the website, Google Business Profile, Bing Places, professional directories, association listings, the NPI registry where relevant, and every social profile. One old profile with a former address is enough to produce the exact error patients reported in that survey. This step is tedious, free, and does more than anything else on this list.

2. Claim and complete the Google Business Profile

Services, service areas, hours, photographs, questions answered, reviews responded to. It feeds Google’s entity understanding directly, and it is the single most-cited source of the operational facts that assistants get wrong.

3. Submit to Bing, not only Google

ChatGPT’s live retrieval leans on Bing’s index. A practice can be comprehensively indexed by Google and close to invisible to ChatGPT because nobody ever opened Bing Webmaster Tools. Verifying a site and submitting a sitemap there is a short job, and it is free.

4. Let the retrieval crawlers in

The crawlers used to answer live questions are separate from the ones used to gather training data, and they use different names. As of 2026 the retrieval agents include OAI-SearchBot, PerplexityBot and Claude-SearchBot, with ChatGPT-User and Claude-User triggered when someone asks an AI assistant to open a specific page. Blocking those removes you from answers. Whether you also allow the training crawlers, such as GPTBot, ClaudeBot and Google-Extended, is a separate and legitimate business decision. Decide each one deliberately, and re-check the names annually, because they change.

5. Answer the question in the first paragraph

Every important page should open with a two to four sentence answer to the question the page exists for, before any story, any history, and any warm-up. An AI assistant summarising your page will use what is near the top. So will a patient in a car park.

6. Put numbers and sources on the page

This is where the research points most clearly. Prices, timelines, session lengths, how many years, how many patients, what is included and what is not. Where a health claim appears, put the citation next to the sentence rather than in a list at the bottom. For a regulated practice this is doubly useful: the FTC expects claims to be substantiated before publication and disclosures to sit close to the claim they qualify.⁵ The compliance requirement and the citation requirement turn out to be the same requirement.

7. Make the structured data agree with the page

Organization, Person for the practitioner with credentials and sameAs links, LocalBusiness or the appropriate medical type where it is factually and legally correct, FAQPage on questions that are actually visible, and BreadcrumbList on nested pages. Two rules matter more than which types you choose. Structured data must never say something the page does not visibly say. And medical schema types belong to licensed providers, so an unlicensed coaching brand using Physician or MedicalClinic markup is not being clever, it is misrepresenting itself in a machine-readable format.

What will not work

• Keyword stuffing. It was measured and it did nothing.¹
• Hidden text aimed at crawlers but not shown to patients. If it is worth telling a machine, it is worth showing a person, and cloaked content is a violation risk in both directions.
• Buying a guarantee of being recommended. Answers vary by phrasing, by user, by day and by model version. Nobody controls that.
• Chasing prompt-volume estimates the way an SEO chases keyword volume. That data is modelled and simulated, not observed. Use real patient language from calls, intake forms and Search Console instead.
• Publishing a testimonial that says something the practice is not allowed to say directly. That is still the practice making the claim.⁵

Who should not bother with this yet

A practice with no reviews, no complete Google Business Profile and a site that is not fully indexed does not have an AI visibility problem. It has a basics problem wearing a more fashionable name, and the first four steps above will do more than any amount of content strategy. Likewise a practice that is at capacity and not taking new patients should spend the money elsewhere.

How to know whether it is working

Write down 20 to 30 questions a real patient would ask, in the words they would actually use. Run them across ChatGPT, Perplexity, Gemini and Claude, and record every answer verbatim, including which sources were cited and every factual error about you. Do it before you change anything, because that baseline cannot be reconstructed later. Then re-run the same list monthly and compare. Track four things: whether you are mentioned, whether the facts are right, who is mentioned instead of you, and which sources the answer leaned on. Read the trend at three and six months.

If that sounds like a lot of manual work, it is, which is why it belongs in a recurring process rather than in a launch checklist. It is also the only honest answer to the question every practice owner asks, which is whether any of this actually made a difference.