If you still build your search strategy solely around keyword research, link building, and pushing pages into the classic Google and Yandex top 10, I have sobering news: traditional organic search is rapidly losing its monopoly on buyer decisions.
User behavior has shifted at the root. Today, someone facing a high-stakes problem—whether it's a suspicious mole, a stubborn chronic symptom, hiring a corporate lawyer, or finding an architect for high-load IT systems—is far less likely to wade through pages of ad-packed search results. Instead, they open ChatGPT, switch on Yandex's Alice in generative mode, or query Google AI: "Recommend the best melanoma oncologist in Moscow," "Who is the leading liver specialist?", or "Who should I see for a high-risk pregnancy?"
In that single second, a fundamental shift takes place. The search engine no longer hands the user ten blue links to clinic websites. It writes a cohesive answer naming specific experts, summarizing their credentials, and telling the user directly where to spend their money. Welcome to Generative Engine Optimization (GEO)—optimizing for generative AI.
But how does an AI actually decide which doctor to recommend? Does it even know which clinic that specialist works at right now? And does the commercial traffic actually reach the medical center footing the bill for their salaries and PR?
To answer those questions with hard numbers rather than guesswork, agency TexTerra ran a large-scale empirical study. The team tested recommendations from three leading AI platforms across a real sample of Moscow clinics and physicians—and uncovered a massive gap between an expert's personal fame and their employer's actual sales.
Anatomy of the Experiment: 360 Answers, 3 AIs, and 5 Prompt Tiers
To rule out random model hallucinations, researchers designed a strict matrix methodology. They chose private healthcare as the testbed—an industry where trust in the individual practitioner and the cost of a mistake are sky-high, and patients scrutinize a doctor's hands just as closely as the clinic's sign on the door.
The sample covered five respected Moscow medical centers of varying sizes and specialties. From each clinic, analysts picked two doctors: one with broad public fame and regular federal media appearances, and a second with strong peer reputation inside the medical community but a quieter media footprint:
- Dr. Laskov's Clinic: Mikhail Laskov (oncologist, clinic head) and Alexander Abolmasov (oncologist, deputy head);
- European Medical Center (EMC): Tatiana Polunina (gastroenterologist, hepatologist, department head) and Marina Berezhnaya (head of emergency care);
- Mother and Child Group: Mark Kurtser (RAS academician, founder and CEO) and Nelly Agaronyan (executive director and chief physician, Lefortovo branch);
- Medsi Clinic Network: Gennady Konovalov (chair of the Academic Council, scientific director) and Alexey Barinov (neurologist);
- Systemic Medicine Clinic: Zukhra Pavlova (endocrinologist, Doctor of Medical Sciences, scientific director) and Anna Terekhova (endocrinologist).
The test prompt set mirrored a real patient's journey from branded lookup to cold discovery in an unfamiliar specialty. In total, the team built 40 query scenarios split into five functional blocks:
- "Who is this" (10 queries): direct identification of a specific specialist ("Who is doctor Mikhail Laskov?");
- "Where do they practice" (10 queries): navigational lookup for booking ("Which clinic does endocrinologist Zukhra Pavlova practice at?");
- "Clinic doctors" (5 queries): testing knowledge of an institution's roster ("Name well-known doctors at the European Medical Center");
- "Who to read" (10 queries): asking for authority and media presence ("Which doctors should I read on hepatology and liver disease?");
- "Recommend a doctor" (5 queries): commercial category queries for an urgent need with no brand or name mentioned ("Recommend a good chemotherapy oncologist in Moscow").
Data collection ran October 26–28, 2026. Every single one of the 40 prompts was fed to three systems: ChatGPT (with web search enabled and memory cleared), Yandex Alice (fresh chat mode geo-locked to Moscow), and Google AI (AI Overviews / generative search in incognito mode).
Crucially, every question was asked three times in completely independent sessions. In the end, researchers collected and manually labeled 360 full AI responses, cross-checking every detail against official clinic websites as of October 29–30.
The Visibility Paradox: Why 96% Name Accuracy Collapses to 16% in Category Search
The first finding looks encouraging on the surface—until you spot the trap.
Split-screen infographic of AI search results: 100% doctor recognition by last name vs. losing the clinic in general category search.
When a user includes the doctor's exact last name (the "Who is this" and "Where do they practice" blocks), modern AIs show near-flawless awareness. In 173 out of 180 responses (96.1%), algorithms correctly named the specialist's current workplace. If the patient already knows who they want, AI won't get in the way of them finding your clinic.
The moment you switch to commercial category queries—where a patient has a symptom or need, not a specific name—the system suffers a catastrophic breakdown.
Across the "Who to read" and "Recommend a doctor" blocks, the selected experts appeared in only 32 out of 135 possible responses. Worse still for business, the doctor was named alongside their current workplace in just 22 out of 135 responses—a dismal 16.3%.
Breaking down those 22 successful mentions by platform exposed stark algorithmic differences:
- ChatGPT delivered 15 accurate "doctor + clinic" pairings;
- Google AI produced 7 mentions that included the clinic;
- Alice scored an absolute zero—not once in any category response did Yandex's AI link the expert to their employer.
That creates a fundamental publicity paradox. Investing in a doctor's personal brand mostly builds the doctor as an independent media figure. The patient reads the expert's name in the AI answer, while the clinic paying for their office and equipment stays invisible.
Three structural ranking patterns stood out clearly:
- Name-match and heavy media-presence advantage: The runaway leaders in mentions were Mikhail Laskov (whose surname matches the clinic's brand), Mark Kurtser (founder of a massive public healthcare holding), and Zukhra Pavlova (author of popular books and a regular national TV guest).
- The insider-reputation barrier: Specialists whose authority lives mainly inside medical circles are virtually absent from generative results. Medsi's scientific director Gennady Konovalov was mentioned only once across the entire thematic test pool. Top neurologist Alexey Barinov wasn't named by a single model across nine runs about the network's doctors.
- Non-transferable status: Even when an AI knows a doctor belongs to a clinic, it fails to carry that knowledge into category search. All three systems effortlessly listed endocrinologist Anna Terekhova under Systemic Medicine Clinic, yet in endocrinology category queries she didn't appear once in 27 runs.
For a deeper look at how search crawlers and citation engines work under the hood, read promoting a business in generative AI: citation mechanics in ChatGPT and Alice.
Platform Showdown: Output Volatility, Answer Length, and Hidden Sources
The experiment proved one thing beyond doubt: generative search is wildly stochastic. Assuming a prompt you checked once will keep returning the same answer is a dangerous marketer's delusion.
In the "Recommend a doctor" category, not a single surname was named by all three competing systems at once. Each platform builds its output from its own training corpus and distinct search indexes.
Even inside a single platform, answers swing dramatically from session to session. Analysts compared recommended doctor lists across repeat runs:
- ChatGPT averaged just 24% overlap in names between two answers to the identical prompt;
- Alice logged a 25% repeatability rate;
- Google AI proved slightly steadier at around 30%.
That means two-thirds of AI recommendations churn from one run to the next. Checking your brand's AI visibility with a single spot query is meaningless—only regular runs of matrix prompt sets reveal the real picture.
Answer anatomy varies sharply across platforms as well:
| Platform metric | ChatGPT (OpenAI) | Alice (Yandex) | Google AI (Overviews) |
|---|---|---|---|
| Median response length | ~5,000 characters | ~6,300 characters | ~2,200 characters |
| Link policy | 626 links (366 unique domains) | Links hidden in 27 of 30 answers | Concise cards with links |
| Category behavior | Generously cites blogs and media | Generates long cohesive essays | Compact summary with vetted clinics |
Alice writes the longest, most detailed essays, yet hides primary sources almost completely—giving readers no one-click path to the clinic's site. ChatGPT acts like a content aggregator, generating 626 links across 366 unique domains in just 30 "Who is this" answers. Meanwhile, Google AI's tight length cap (~2,200 characters) makes earning a spot there the hardest prize of all, with roughly a third as many mention slots as its rivals.
The Aggregator Trap: How Middlemen Intercept Hot Traffic
If you assume getting your top specialist mentioned in an AI answer automatically rings your call center, brace yourself. The study uncovered a massive parasitic aggregator funnel.
In the "Where does the doctor practice" block, a hefty share of answers didn't link to official clinic websites at all:
- ChatGPT: 6 out of 30 answers pointed exclusively to third-party aggregators;
- Alice: 7 out of 30;
- Google AI: 3 out of 30.
The most damaging leak showed up in phone numbers. In 31 AI responses, the booking number listed for the doctor was a call-tracking proxy belonging to aggregators like NaPopravku, ProDoctorov, or SberHealth.
Some cases bordered on absurd:
- For appointments with academician Mark Kurtser, seven out of nine AI responses gave a single NaPopravku proxy number. The patient calls a third-party service, triggers an aggregator commission, or gets routed by an operator to a different partner facility altogether.
- For Alexander Abolmasov, AI generated seven different phone numbers across nine answers—a jumble of old branches and partner call centers.
What does that mean for clinic unit economics? The business pays the doctor's salary and buys the equipment, while an aggregator hijacks the generative-search lead and bills the clinic 1,500 to 5,000 rubles per confirmed booking.
To see the full methodology for spotting these leaks, check out fact matrix vs. hallucinations: TexTerra's methodology for brand GEO audits in AI search.
Digital Litter Syndrome: Why 70% of Experts Are Surrounded by Misinformation
Where do AIs get this bad data? Are they hallucinating?
Manual fact-checking of source pages showed that models rarely invent errors out of thin air. They faithfully mirror the mess already sitting on public websites.
Seven out of ten surveyed doctors had conflicting, incomplete, or badly outdated information online. And those exact errors surfaced word-for-word in AI answers:
- Phantom workplaces: A doctor left a state research institute or rival clinic three years ago, yet their old aggregator profile stayed live. The AI reads both pages and tells patients the doctor practices at two competing institutions.
- Muddled titles and credentials: Mismatches between a clinic's official site and directory listings cause models to confuse department heads with junior residents.
- Misattributed photo captions: Crawlers index images from news articles. If a lead surgeon's photo sits next to another specialist's quote, the AI pins the wrong quote and specialty onto your doctor.
- Same-name collisions: When a respected expert shares a name with someone in an adjacent field, language models happily splice their bios into one composite—crediting an oncologist with chiropractic skills.
If a clinic neglects basic hygiene across its digital footprint, search bots feed AIs informational mush—damaging your brand at the very first touchpoint.
Next-Gen E-E-A-T: Turning an Expert Into a Digital Entity
To make generative models reliably recommend your people and your business, you need to grasp a core principle of modern LLMs: the shift from keyword-indexed pages to Knowledge Graphs built on semantic Entities.
Search engines no longer read text as a bag of keywords. To them, a clinic is a MedicalOrganization entity, and a doctor is a Physician or Person entity. Between them, you need an explicit, repeatedly corroborated predicate link: worksFor, alumniOf, and hasCredential.
This shift rests squarely on Google's E-E-A-T framework (Experience, Expertise, Authoritativeness, Trustworthiness), updated in summer 2026 for generative search. Algorithms demand hard proof of an author's firsthand practical experience.
International data from Profound shows that on professional queries, up to 35% of ChatGPT citations point to LinkedIn posts. Meltwater analysts also found that 75% of those citations lead to individual author profiles—and more than half have modest followings under 10,000. AIs aren't hunting for bloated million-follower influencers; they're looking for recognized practitioners with a verifiable track record.
In Russia, where access to some foreign platforms is restricted, other venues act as authority engines:
- Expert blogs hosted on corporate domains;
- Major business and tech media (RBC, Vedomosti, Forbes, Habr, VC);
- Professional registries and academic databases (eLibrary, CyberLeninka);
- Verified independent aggregators with strict review moderation.
For a step-by-step playbook on adapting content for these algorithms, read generative engine optimization (GEO): how brands get into ChatGPT, Perplexity, and Yandex Neuro answers.
The GEO Optimization Playbook: 5 Steps for Clinics and B2B Teams
Based on the study, TexTerra's experts put together a practical checklist to bind an expert's personal brand to the company brand—and make AIs send traffic to the source instead of middlemen:
- Step 1. Deploy Schema.org markup on every key landing page.
Every staff profile on your site should use
PhysicianorPersonschema withworksFor(pointing directly to your company's root domain),jobTitle,alumniOf, and asameAsarray linking to Wikipedia, directories, and social profiles. That locks an unambiguous record into the global Knowledge Graph. - Step 2. Enforce an unbreakable-link rule in content policy. Ban publishing any expert column, press comment, or conference bio without current title and company. The rule is simple: the expert never exists as a standalone entity—every metadata field, photo caption, video lower-third, and description ties them to your brand.
- Step 3. Deep-clean third-party aggregators. Marketing needs to own doctor cards across every middleman platform. Delete profiles of former staff, update schedules, scrub old branches, and replace proxy tracking numbers with your clinic's direct line wherever the platform plan allows.
- Step 4. Publish content with unique clinical substance. Generic Wikipedia rewrites get ignored. To enter training sets or RAG search context, articles need real frontline detail: atypical case breakdowns, treatment protocols, documented progress photos (with patient consent), and the specialist's own takeaways.
- Step 5. Run quarterly GEO monitoring. Build a matrix prompt set of 30–50 queries covering branded, navigational, and problem-based searches. Run it once a quarter across at least three independent sessions in ChatGPT, Alice, and Google. Track your Share of Voice and squash hallucinations before they cost you patients.

