Healthcare SEO for ChatGPT, Perplexity and Claude: What Changed

Healthcare SEO for ChatGPT works on retrieval rather than ranking. What changed is mechanical: crawler access tokens, passage-level extractability, entity consistency, and off-site citations. What did not change is the trust standard. E-E-A-T, YMYL scrutiny, named clinical authorship, and primary sourcing still decide which sources AI engines repeat.

TL;DR

• Retrieval replaced ranking as the thing to optimize. Ranking orders results a patient scans. Retrieval decides which pages a model reads before writing one answer.
• Four mechanics genuinely changed: crawler tokens, passage extractability, entity consistency, and the off-site pages that get cited instead of yours.
• Six fundamentals did not change, and five were already mandatory in healthcare because health content sits in the strictest quality category.
• Googlebot access underwrites four of the five major engines, including Claude, whose web search runs on Brave, and Brave will not crawl what Googlebot cannot.
• FAQ rich results ended on May 7, 2026. Healthcare held eligibility longest and has now lost it. Keep the markup for machine readability.
• The widely quoted 40% GEO lift is a benchmark ceiling from a synthetic engine, not an average from live ChatGPT or Perplexity.
• Perplexity is the one engine that needs a separate workstream, because it runs its own crawler and index that no Google or Bing tool reaches.

One liner: The plumbing changed, the trust standard did not, and most of the “AI SEO” invoice is work you were already buying.

Key Numbers

Healthcare SEO for ChatGPT Starts With One Distinction

In short: Healthcare SEO for ChatGPT is about retrieval, not ranking. Ranking sets the order of ten results a patient scans and chooses from. Retrieval decides which pages a model reads before it writes the one answer that patient actually sees.

Hold onto that difference, because most of the confusion in this category comes from collapsing it.

You have probably been sold this work under four names already. Generative Engine Optimization. Answer Engine Optimization. LLM SEO. AI SEO. They describe overlapping activity, and the acronym you hear depends mostly on which agency you spoke to last. None of them is a separate discipline with its own rulebook, and treating them that way is how healthcare marketing budgets end up funding files nobody reads.

The stakes are not theoretical. OpenAI reports that more than 5% of all ChatGPT messages worldwide are about healthcare, that roughly one in four of its 800 million-plus regular users sends a health prompt in a given week, and that about seven in ten of those health conversations happen outside typical clinical hours [12]. That figure is OpenAI’s own data about OpenAI’s own product, so treat it as directional rather than audited. A separate OpenAI and NBER study of ChatGPT use found that asking, the search-like behavior of putting a question to the model, accounts for roughly half of all messages [13]. Even discounted, it describes a patient population asking questions at 11pm that your phone line cannot answer.

Here is what a retrieval failure looks like in practice. A practice manager asks ChatGPT to recommend a clinic in her own city and watches a competitor’s name come back. Her site ranks fine. It was never read.

The rest of this piece splits the work in two: four mechanics that genuinely changed, and six that did not. If you want the fuller strategic picture first, start with our overview of generative engine optimization for healthcare.

What Actually Changed for LLM SEO in Healthcare

In short: Four things genuinely changed: crawler access became several separate decisions instead of one, retrieval happens at the passage level rather than the page level, entity consistency across the open web became a ranking input, and the page that gets cited is often not yours.

1. Crawler access is four decisions, not one

This is the change most healthcare organizations get wrong, and it costs the most.

OpenAI runs separate robots.txt tokens for separate purposes. GPTBot collects content that may train future models. OAI-SearchBot handles inclusion in ChatGPT’s search features, and OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. ChatGPT-User fetches a page when a person asks ChatGPT to visit it. Each setting is independent of the others, so a site can allow OAI-SearchBot to stay visible in search while disallowing GPTBot to keep its content out of training [6][7]. Anthropic uses the same three-way split: ClaudeBot for training, Claude-SearchBot for search indexing, and Claude-User for user-requested fetches, and warns that blocking Claude-SearchBot may reduce visibility for user-directed web search [8].

You will find pages claiming that GPTBot handles training and live search retrieval together, so blocking it removes you from both. OpenAI’s own documentation says the opposite. The distinction decides whether a health system that blocked “the AI bots” on legal advice in 2024 is invisible inside ChatGPT search today.

A working starting point:

# OpenAI
  User-agent: GPTBot
  Allow: /
  User-agent: OAI-SearchBot
  Allow: /
  User-agent: ChatGPT-User
  Allow: /
  
  # Anthropic
  User-agent: ClaudeBot
  Allow: /
  User-agent: Claude-SearchBot
  Allow: /
  User-agent: Claude-User
  Allow: /
  
  # Perplexity
  User-agent: PerplexityBot
  Allow: /
  User-agent: Perplexity-User
  Allow: /

One compliance note before anyone forwards this to legal. Allowing a search crawler on public marketing pages is a different decision from anything touching protected health information. Patient portals, forms and anything behind authentication are a separate conversation and should stay blocked.

Compliance with those directives also varies by vendor, which the generic advice never mentions:

Sources: OpenAI and Anthropic crawler documentation, plus independent user-agent audits [6][8][9].

2. Retrieval happens at the passage level

Models pull passages, not pages. A self-contained 45-word answer sitting under a question-shaped heading can be lifted whole. The same clinical fact, buried in the sixth paragraph of a 3,000-word guide and dependent on two earlier paragraphs for context, cannot. Page authority still helps you get read. Passage structure decides what survives the read.

3. Entity consistency became an input

Provider name, credentials, specialty, locations and hospital affiliations have to agree across your site, your directory listings, insurer provider finders and professional registries. When sources conflict, a model reconciles by weight of consensus rather than by whose site it is. A physician listed under two spellings across four directories is an entity problem, and no amount of on-page work fixes it.

4. The cited page is often not yours

In a synthesized answer, the source list frequently points at a third-party page about your organization rather than a page you control. Directory profiles, local news coverage, association listings and review platforms all compete to be the thing a model reads about you. That reallocates a meaningful share of the work off your own domain.

If you want the full build sequence for these four, see our AI search optimization for healthcare service overview.

The AEO vs SEO Healthcare Question, Answered Plainly

In short: AEO and SEO are not competing disciplines in healthcare. Six fundamentals carry over unchanged, and five of the six were already mandatory for medical content because health sits in the strictest quality category search engines have.

That is the part the urgency-sellers skip. If your organization has been doing healthcare SEO properly, you have been building for AI retrieval since before anyone coined a name for it.

Google’s own guidance on its AI features is blunt about this. Its documentation, updated in June 2026, tells site owners they “don’t need to create new machine readable files” or special markup to appear in Search including its generative AI capabilities, because Search does not use them [10]. Ranking in AI Overviews and AI Mode runs on the same fundamentals as ranking anywhere else on Google.

Here is what carries over, and what each item does on both sides:

One of the six did intensify. Authorship verification carries more weight now, because a model reconciles conflicting information rather than displaying two versions side by side and letting the reader choose. When your provider’s credentials appear three ways across four directories, a search engine shows the mess. A generative engine picks one and states it with confidence, and it will not always pick yours.

Why local presence survived intact

Constant six deserves an explanation rather than an assertion. Clinical informational queries and local provider queries behave differently. “Is this condition serious” is answered by a generated summary. “Pediatric dentist near me” is still answered by a map pack and a set of links, and Google has adjusted health coverage in that direction, removing AI Overviews from some medical queries following reporting on misleading medical summaries [16]. Two ecosystems, not one. Your local acquisition work was never the thing under threat.

The citation quality gap is your opening

This is the most encouraging finding in the category, and almost nobody in healthcare marketing is acting on it.

An SE Ranking analysis of health-related AI Overviews found that roughly 34% of citations came from sources it classified as trusted medical, while government health bodies and academic journals together accounted for under 1%. The single most-cited domain was YouTube, at 4.43% of all citations in the dataset [15]. Read that carefully before you use it: the study covered German-language queries, and the sourcing caveats matter, which the next section takes up directly.

The structural point survives the caveat. AI citation in health is not yet dominated by institutional authority. A credentialed organization publishing well-structured, well-sourced content is competing against a field that is thinner than its reputation suggests.

None of this means your current execution is adequate. Unchanged fundamentals and strong fundamentals are different claims. Most healthcare sites fail at least two of the six, usually thin authorship and unsourced clinical claims lifted from competitors. The point is narrower and more useful than “nothing changed”: the fix list is the one you already had, not a new one you need to buy.

So when a proposal lands with these six itemized as new AI line items, read it as a re-invoice.

Engine by Engine: ChatGPT, Perplexity, Claude, AI Overviews and Gemini

In short: Five engines do not mean five projects. Three of them read Google’s index, a fourth depends on Googlebot access indirectly, and only Perplexity runs a retrieval stack that needs separate attention.

Googlebot access underwrites more of this than the acronym sellers admit

Here is the finding that should reorder your priorities. Brave’s crawler documentation states that its bot uses no differentiated user agent, and that if a domain or page is not crawlable by Googlebot, Brave’s bot will not crawl it either [19]. Claude’s web search runs on Brave [18]. So Googlebot access is a prerequisite for AI Overviews, for AI Mode, for Gemini, and, one step removed, for Claude.

Any advice that tells a healthcare organization to shift budget away from Google because patients have moved to chatbots has the plumbing backwards. Four of these five engines lead back to a Googlebot crawl.

Access is a prerequisite. Ranking is no longer a proxy

Two different claims get confused here, and keeping them apart matters.

Being crawlable is still required. Being in the top ten is no longer a reliable predictor of being cited. Ahrefs found in July 2025 that about 76% of AI Overview-cited pages also ranked in the top ten, with a median cited page at position three [20]. A later Ahrefs analysis of 863,000 keywords and 4 million AI Overview URLs put that overlap near 38%, and a BrightEdge analysis on different methodology put it lower still [21]. Any page still quoting the 76% as the current state of play is citing a real study a year past its shelf life.

The practical reading: crawl access first, because without it nothing else runs. Then structure, because rank alone will not carry you into a citation.

Perplexity healthcare visibility is the one separate workstream

Perplexity is the outlier worth its own line item. It pulled its search function in-house and rebuilt it on its own crawler and index rather than renting one [17], so neither Search Console nor Bing Webmaster Tools reports on it and neither one reaches it. If PerplexityBot is blocked, no amount of Google work recovers the visibility. For healthcare specifically, Perplexity skews toward research-intent users comparing treatments and providers, which is a higher-value read than raw volume suggests.

Citation display decides whether visibility becomes traffic

Notice the fourth column. Some of these engines hand the patient a clickable source list. Others answer and move on. That single difference determines whether AI visibility shows up in your analytics at all, which is the measurement problem a later section has to solve.

Four AI Search Claims Healthcare Marketers Keep Hearing, Traced to Source

In short: Three questions dissolve most AI search statistics: who collected it, over what sample, and can you reach the original document in one click. Claims that fail all three are circulating widely in healthcare marketing right now.

Claim 1: “GEO delivers a 40% lift”

There is a real paper behind this number, which is why it travels so well. Aggarwal and colleagues published “GEO: Generative Engine Optimization” at KDD 2024, and it reports that their methods can boost visibility by up to 40% in generative engine responses [4].

Read the qualifiers. The 40% is a ceiling, not an average. The three strongest tactics produced a 30 to 40 percent relative improvement on a metric called Position-Adjusted Word Count, and low-ranked sources gained disproportionately, meaning the headline gain came largely from pages that started with nothing to lose. The tests ran on roughly 10,000 queries against a system built to mimic Bing Chat, then were validated on Perplexity. That is a controlled benchmark, not your practice website inside ChatGPT.

A later benchmark points the other way. C-SEO Bench found that most conversational-SEO tactics do not help, that several actively hurt, and that plain source relevance keeps working [5]. Both findings are legitimate. Only one of them makes it into sales decks.

Any performance claim built on an “up to” should show its average next to it. When it does not, the average is the part you were not meant to see.

Claim 2: “You need an llms.txt file”

Google’s guidance on its AI features, updated in June 2026, tells site owners they do not need new machine readable files, AI text files, markup or Markdown to appear in Search including its generative AI capabilities, because Search does not use them [10]. Gary Illyes said at Search Central Live in July 2025 that Google does not support llms.txt and has no plans to, and John Mueller compared the idea to the keywords meta tag, a file where the site owner declares what the site is about when the engine could read the site instead [11].

The honest nuance: llms.txt has a small real niche with AI coding assistants reading developer documentation. A dermatology group’s website is not that niche. If it appears as a line item on a healthcare proposal, ask which engine consumes it.

Claim 3: “82% of health searches trigger AI Overviews”

This one is being quoted at US healthcare marketers as a general fact. It comes from an SE Ranking study, and the study is real. What travels with the number is not.

The analysis covered 50,807 German-language health queries, captured from Berlin, in a one-time snapshot taken in December [15]. It is a legitimate finding about German-language health search on one date. It is not a US benchmark, and it is not a trend line. The researchers themselves framed the German regulatory environment as the reason the result is notable rather than as a proxy for everywhere else.

Use it if you like, with its sample attached. Quoting it bare is how a national figure gets invented.

Claim 4: The round numbers with no publisher

Ranking pages in this category carry benchmark tables reporting things like a 112% citation lift from proprietary data, or a 27% conversion rate against 2.1% for traditional organic, described as a 13x improvement. No sample size, no collection window, no independent publisher, no link to an original document. Trace one and you usually land on another agency page citing a third, which cites a press summary of a study nobody opened.

Apply the test to this article. The ChatGPT usage figures quoted earlier come from OpenAI, about OpenAI’s own product, published without independent audit. They are labeled that way in the body because the test applies here too. A figure you can source and qualify is worth more than a rounder one you cannot.

What FAQ Schema Is For Now That the Rich Result Is Gone

In short: FAQ rich results stopped appearing in Google Search on May 7, 2026, and healthcare was one of the last verticals holding eligibility. Keep the markup, because it stays valid and machine readable. Stop expecting a SERP feature from it.

The healthcare schema types that still earn their place

Before the deprecation, the wider point. FAQPage is not the only structured data that matters for a medical site, and for most healthcare organizations it is not the most valuable one.

Physician and MedicalOrganization markup are the entity layer that section three called load-bearing. They are how a model confirms a credential rather than inferring it.

The FAQ timeline, stated precisely

If you work in healthcare marketing, this change hit your vertical differently than everyone else’s.

Google deprecated the rich result feature, not the vocabulary. FAQPage remains a valid Schema.org type, and Google’s documentation notes there is no need to proactively remove unused FAQ structured data. So leave it in place.

Whether FAQ markup improves your odds of being cited in an AI answer is a separate question, and it is not settled. You will find pages asserting a specific multiplier, one of them claiming FAQPage schema makes a page 3.2 times more likely to appear in AI Overviews, with no study, sample or publisher attached. Apply the three questions from the previous section and that number dissolves.

The format is what earns the retrieval

What does hold up is the question-and-answer shape itself, and it works with or without the markup. Compare two versions of the same fact:

Buried: “…and while several factors influence this, most patients find that recovery time depends heavily on the procedure performed, with arthroscopic approaches generally allowing a faster return to activity than open techniques, though individual healing varies considerably.”

Retrievable: “How long is recovery after arthroscopic knee surgery? Most patients return to desk work within one to two weeks and to low-impact activity within four to six weeks. Open procedures take longer. Your surgeon’s timeline depends on the specific repair and your baseline activity level.”

Same information, illustrative rather than clinical. The second one can be lifted whole by every engine in the table above, because it stands alone under a question a patient would type.

For the record, this article ships FAQPage markup. Not for a rich result, which no longer exists, but because it stays machine readable for systems that parse it. The schema and the argument agree.

What You Can Measure, What You Can Estimate, and What Nobody Can Track

In short: AI visibility splits into three tiers. Crawler hits and referral sessions are measurable from your own logs. Share of voice is estimable if you disclose your method. The patient who reads an answer, never clicks, and books by phone is not trackable by anyone, including the vendor selling you a dashboard.

If OAI-SearchBot has never touched your service pages, no content strategy fixes that. You have found a robots.txt or crawl-budget problem rather than a content problem, and the answer is already sitting in logs you are paying to store.

The size of the untrackable tier is worth stating to leadership before someone else promises them a number. Pew analyzed 68,879 Google searches from 900 US adults and found that when an AI summary appeared, users clicked a traditional result in 8% of visits, against 15% when no summary appeared. They clicked a link inside the summary in 1% of visits, and browsing sessions ended on 26% of AI-summary pages against 16% without one [14]. Pew’s own caveats apply: this is Google, in the US, in one month, and association rather than proof of cause.

A patient asks at 11pm, reads the answer, clicks nothing, and calls your office on Thursday. That path has always existed in healthcare. AI answers widened it.

Which brings up what to refuse. If a vendor reports an “AI visibility score” without disclosing the prompt set, the engines queried and the collection window, that is a number, not a measurement. The same three questions work on tools as well as on statistics.

Your First 30 Days of Healthcare SEO for ChatGPT

In short: Work in this order: access, then structure, then measurement, then spend. Most guidance in this category leads with content tactics, which accomplishes nothing if the crawler never reached the page.

Three of the first three items cost nothing beyond someone’s afternoon, and they are the ones that decide whether any later work is visible at all.

One caution about the time box. Thirty days sets the work up. It does not deliver the outcome. Crawl cycles, index refreshes and the slow reconciliation of your entity record across a dozen third-party sources run on their own schedule, and anyone promising AI citations inside a month is selling the same “up to” arithmetic this article spent a section taking apart.

What your website is for now

One strategic note before the close. If engines answer the informational questions before a patient reaches you, the job of your website shifts. It becomes less of a discovery surface and more of a destination: service-line depth, provider credentials a patient can verify, scheduling that works on a phone, and the specific proof no generated summary can reproduce. Discovery moves upstream. Conversion stays with you.

Here is the honest close. Almost everything above is doable in-house. None of it requires a proprietary framework, a new acronym, or a file format invented last year. What it requires is a person whose job it is, working through the list in order and rechecking it as the engines change, and most healthcare marketing teams do not have that person sitting idle.

That is the entire value proposition, and it is the one we would rather be judged on. If you want it handled, Healthcare SEO Services does this work for medical practices and health systems. If you want to run it yourself, the list above is the list.

Frequently Asked Questions

Does ChatGPT use my website?

Only if it can reach it. OpenAI runs OAI-SearchBot for its search index and GPTBot for training, controlled by separate robots.txt tokens. Sites that disallow OAI-SearchBot will not appear in ChatGPT search answers. Check your robots.txt and your server logs before assuming a content problem.

Should healthcare sites block GPTBot?

That depends on whether you object to training use, and it is a separate decision from search visibility. Blocking GPTBot keeps your content out of OpenAI model training. It does not remove you from ChatGPT search, which OAI-SearchBot controls. Block one and allow the other if that fits your policy.

Is FAQ schema still worth it in 2026?

Yes, for machine readability rather than a rich result. Google stopped showing FAQ rich results on May 7, 2026, and removes the reporting surfaces through August. The FAQPage type stays valid, and Google says unused FAQ markup causes no problems. Keep it and clean up your dashboards.

What is the difference between AEO and SEO for healthcare?

Less than the acronyms suggest. Answer engine optimization emphasizes self-contained passages a model can lift. Healthcare SEO already required that, plus E-E-A-T, clinical authorship and primary sourcing, because health content faces the strictest standards. Treat AEO as an emphasis inside SEO, not a replacement for it.

Does Claude use Google?

Indirectly. Anthropic’s documentation says Claude’s web search calls the Brave Search API. Brave’s crawler documentation states that if a page is not crawlable by Googlebot, Brave’s bot will not crawl it either. Googlebot access is therefore a prerequisite for appearing in Claude’s cited results.

How do I get cited by Perplexity?

Allow PerplexityBot in robots.txt first, because Perplexity runs its own crawler and index that neither Search Console nor Bing Webmaster Tools reaches. Then write pages where a complete answer sits near the top in plain prose. Perplexity cites passages it can quote without stitching them together.

Do I need an llms.txt file?

No. Google’s AI guidance, updated June 2026, says you do not need new machine readable files or markup to appear in Search including its AI features. Gary Illyes confirmed Google does not support llms.txt. Its real niche is coding assistants reading developer documentation, not healthcare marketing sites.

How do I measure AI search visibility?

In three tiers. Crawler hits by user agent in server logs and referral sessions from AI domains are measurable. Share of voice against a fixed prompt set is estimable if you disclose the method. The patient who reads an answer and calls without clicking is not trackable.

Did AI Overviews replace healthcare SEO?

No. Google’s guidance is that appearing in AI Overviews and AI Mode runs on the same fundamentals as ranking in Search. The surfaces multiplied and click behavior changed. The underlying requirements, crawlability, demonstrated expertise, accuracy and local presence, are the ones you already had.

How long does this take?

Thirty days sets the work up. It does not deliver the outcome. Crawl cycles, index refreshes and entity reconciliation across directories and registries run on their own schedule. Anyone promising AI citations inside a month is quoting a ceiling as though it were an average.

Definition Bank

Sources

1. Google Search Central, FAQ structured data documentation, deprecation notice, 2026.
2. Search Engine Journal, Google Drops FAQ Rich Results From Search, 2026.
3. Google Search Central announcement of August 8, 2023, restricting FAQ rich results to well-known, authoritative government and health websites, as reported by Portent.
4. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, GEO: Generative Engine Optimization, KDD 2024, arXiv 2311.09735.
5. Puerto et al., C-SEO Bench: a benchmark of conversational SEO tactics, 2025.
6. OpenAI, Overview of OpenAI Crawlers, developers.openai.com, accessed 2026.
7. OpenAI platform bot documentation, OAI-SearchBot entry, accessed 2026.
8. Anthropic crawler documentation covering ClaudeBot, Claude-SearchBot and Claude-User, as reported by Search Engine Journal, 2026.
9. Independent AI user-agent audit comparing vendor robots.txt compliance, 2026.
10. Google Search Central, AI features and your website, updated June 15, 2026.
11. Gary Illyes, Google Search Central Live, July 2025, and John Mueller, public statements on llms.txt.
12. OpenAI, AI as a Healthcare Ally, January 2026, reported by Fierce Healthcare and Healthcare Dive.
13. Chatterji et al., How People Use ChatGPT, NBER working paper w34255, 2025.
14. Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, July 22, 2025.
15. SE Ranking, health-related AI Overviews and source quality, analysis of 50,807 German-language health queries, January 2026.
16. The Guardian investigation into misleading medical AI summaries and Google’s subsequent removal of AI Overviews for some medical queries, January 2026.
17. Perplexity search architecture, in-house crawler and index, engineering coverage, 2026.
18. Anthropic support documentation, MCP Web Search, confirming the Brave Search API, 2026.
19. Brave Search Help, Brave Search Crawler, accessed 2026.
20. Ahrefs, AI Overview citations and search rankings, analysis of 1.9 million citations, July 2025.
21. Ahrefs analysis of 863,000 keywords and 4 million AI Overview URLs, plus BrightEdge citation overlap analysis, 2026.