Navigating Generative Search in Healthcare

We spend enough time inside acquisition dashboards to see the problem in boring form. Organic traffic that used to convert quietly is becoming less dependable. Some of it still arrives. Some of it gets answered upstream. When it disappears, the replacement is paid search, call-center work, local reputation spend, and a lot of explaining in Monday pipeline meetings. The issue is not “the future of search.” It is next quarter’s same-store patient volume.

How AI Search Actually Works

The mechanics matter because they are less magical than the category name makes them sound. AI search agents do not browse the web the way a patient does. They break the question into smaller searches, call a traditional search index, read a handful of pages, and then write the answer back with citations.

GEO, or Generative Engine Optimization, builds on top of SEO. It does not replace it. If your page does not rank in the traditional search index, the AI agent never retrieves it. If the AI agent never retrieves it, you cannot be cited. The boring foundation is still technical SEO, page speed, mobile performance, schema, internal links, and enough domain authority to get into the retrieval set. GEO adds a smaller layer on top: writing pages that answer the actual question plainly enough to be quoted.

From Question to Citation

Consider a real scenario. A patient in Los Angeles tears their ACL playing soccer and opens ChatGPT to ask: “Who is the best orthopedic surgeon for ACL reconstruction near me? I want great outcomes and transparent pricing.”

The AI does not answer from memory. It breaks that question into three or four targeted searches, retrieves the top results for each, reads through a small set of pages, and favors the ones with direct answers, local specificity, structured data, and evidence it can point to. Those pages get cited. The rest may still be good pages, but they are not part of the answer the patient sees.

I do not love the phrase “the new Position #1,” but it is directionally right. In the traditional search model, ranking first on Google meant you got a large share of clicks. In the answer model, the patient may never click at all. They see a short list of named options, and your practice is either in that answer or it is not. That dynamic is already visible in local reputation markets. In Independent for 38 Years and 5 Stars on Google, Dr. Robert Layman of Pinnacle Eye Group describes how durable patient trust and review quality compound over time.

The research is useful, if you do not overread it

A peer-reviewed study from Princeton, IIT Delhi, and Georgia Tech (Aggarwal et al., published at KDD 2024) introduced the concept of Generative Engine Optimization and tested specific strategies for improving visibility in AI-generated search results. The headline result was strong: GEO methods improved AI-search visibility by up to 40%, and the Perplexity-specific validation showed gains of up to 37%.

The useful part is that the winning tactics were not tricks. Add real statistics. Quote credible sources. Cite authoritative references. Make the page readable. That is almost disappointingly normal, which is why I trust it more than the usual search-guru vocabulary.

The click-through data gives the risk a shape. Research from Seer Interactive found that Google AI Overviews reduce organic click-through rates by 61% on average. When an AI Overview appears, 83% of searches end with zero clicks. If a provider group has to buy back that missing traffic through paid media, this stops being a content metric very quickly. It becomes a budget conversation.

But brands that are actually cited within AI Overviews saw a 35% increase in organic clicks compared to their baseline. That is the whole game in one awkward sentence: the answer layer can take clicks away, but it can also hand attention to the few sources it names.

The workflow is boring, which is good

The practical work is less glamorous than GEO agencies make it sound. We embed forward-deployed AI engineers into portfolio company operations at Cade, and the content loop that keeps working is straightforward: machines find the gaps, humans write the useful thing, machines measure whether anyone cites it.

In that loop, agents scan competitors, patient questions, local search results, and answer-engine citations. They surface the gaps. Then a credentialed clinician or subject-matter expert writes the actual page. The page gets the boring machinery too: schema, direct answers, sources, service-area detail, and enough structure that a retrieval system can quote it without guessing. After publication, agents test a fixed set of patient questions and track whether the page appears in the answer.

In one regional orthopedic group, pages we reworked this way started showing up more often in AI answers within 90 days. I would not call that proof yet. Attribution is still noisy, and the tools are changing under us. But it was enough to change how we brief writers: the page has to answer the patient’s actual question, carry clinical authority, and leave evidence the answer engine can cite.

The reason this works is not magic. A health page with a named clinician, specific local information, current service details, and real sources is easier for a patient to trust and easier for a retrieval system to use. A page that reads like a generic marketing team paraphrased WebMD has less to grab onto. The commercial consequence shows up in how private practices compete on responsiveness and trust, which Charles Lathram, founder of Cade Partners, discusses in 2025: The Year of Patient Expectations for Private Practices.

Update: May 5, 2026

A week after Buchodi’s Threat Intel published a teardown of ChatGPT ad payloads, OpenAI made the channel explicit. On May 5, 2026, the company announced partner buying, a beta self-serve Ads Manager, and CPC bidding for ChatGPT ads. So the paid layer is no longer just a rumor living in network traffic.

Buchodi’s useful detail was the attribution loop. In observed sessions, ads were delivered as structured single_advertiser_ad_unit objects inside the same stream that carries model output. Each captured ad carried four encrypted references: an integrity payload, oppref, olref, and an ad data token. On click, the merchant page loads OpenAI’s OAIQ SDK, stores oppref in a first-party __oppref cookie, and posts merchant-side events back to bzr.openai.com. That is the outline of a real ad product: insertion, click handling, conversion measurement, and post-click reconciliation.

For healthcare operators, the planning question changes. I would not assume paid conversational discovery is only for giant IOs anymore. OpenAI is clearly moving it toward normal ad buying: partners, self-serve tooling, CPC bidding. The useful mid-market question is whether your content, booking pages, and conversion events are clean enough to measure the channel when it opens up.

Healthcare will attract ad dollars because the queries are uncomfortably high intent. When a patient asks about knee replacement options in their city, they are not casually browsing. They are much closer to choosing a provider than someone reading a generic health article at lunch.

That does not mean every specialty group should start buying ChatGPT ads the minute a sales rep calls. It means the same plumbing that helps organic citations will matter for paid: clear service pages, accurate locations, clean booking flows, conversion events, and a way to tell whether the recommendation became an appointment.

The organic work is still worth doing first. It builds the page the ad can point to, the answer can cite, and the operator can measure. Without that, paid conversational search is just a new place to spend money without knowing what happened.

What I would fix first

If I were sitting with a healthcare operator this week, I would not start with a new content calendar. I would start with the pieces that make the site quotable and measurable.

I would not turn this into a panic project. Most clinics still need ordinary SEO, reviews, clean scheduling pages, and decent local content before they need a formal GEO program. But the answer layer is real now. Teams that learn how to be cited will get a little more free demand. Teams that do not will buy more of it back.