Content & Answer Pages · 9 min read · July 15, 2026
What patients actually ask the AI, and how to collect those questions
A patient with a headache, a lump in the neck or doubts about a recommended operation increasingly types the whole story into ChatGPT or Gemini instead of two words into Google. Those questions are the least filtered patient research you will ever read: they show you word for word what people are afraid of, and whether your practice gets named in the answer they get back.
Why AI questions look nothing like Google searches
In a Google box a patient types "dermatologist Munich" or "mole dangerous". Two or three words, because everyone learned years ago that a search bar only rewards keywords. In ChatGPT, Gemini or Perplexity they write the way they would talk in your waiting room. "For three weeks I've had a mole on my back that has changed and sometimes itches. Should I be worried, and which doctor do I go to?" That is not a keyword. That is a history, a fear and a request for triage in one paragraph, and it tells you more than a click statistic ever did.
For a specialist that is a new kind of data. Every question carries the symptom combination, how long it has been going on, the fear underneath it and the decision the patient is trying to make. Read enough of them and you know your catchment area better than any patient survey would tell you. You also learn the harder thing: whether the AI names your practice as the place to go, or names somebody else.
This is where the gap between classic SEO and Generative Engine Optimization, GEO for short, opens up. Ranking first on Google no longer means you are in the answer. An Ahrefs analysis found only around 6 to 8 percent of the URLs ChatGPT cites also sit in Google's top ten for the same query, and roughly 80 percent of them do not rank in the top 100 at all. Getting cited is a separate selection process, and it is the one that decides whether a patient in your region hears your name.
What patients actually ask, in their own words
The questions fall into a few repeating shapes. Symptom interpretation: "I get heart flutters as I fall asleep, is that dangerous?" Specialist assignment: "Which doctor do I see for constant ringing in one ear?" The second opinion: "My orthopedist wants to operate on a disc, do I really need that?" And preparation: "What should I ask my cardiologist at the first appointment?" Each shape is a page you either have on your site or you do not.
The fear-driven ones tell you most. "Can a PSA value of 4.5 mean cancer?" "Is a lump in the breast always malignant?" That patient is not collecting information, they want to know whether to panic and what to do on Monday morning. If your site answers exactly that, with the real warning signs and a clear line about when to come in, you have written the passage a model can lift, with your specialty attached to it.
Then the logistics: "Which dermatologist in Regensburg is taking new statutory-insurance patients?" "Where do I get a short-notice appointment with a neurologist?" These are local and close to a decision, and they are the questions practice websites answer worst. Consultation hours, which insurance you accept, how appointments are booked, whether you are taking new patients at all: written out in plain text on your own page, that can be quoted. Buried in a PDF or living only in your receptionist's head, it gets guessed at, or the practice down the road gets named instead.
Where to get this data without a dashboard
The awkward part first: OpenAI and Google publish nothing resembling Search Console for AI answers. No report lists the prompts your practice appeared in. So you build the list yourself, and the cheapest source is already in the building: for four weeks, have everyone write down the questions that actually get asked at the desk, on the phone and in the room. Those sentences are close to word for word what the same people type into a chatbot at home.
The second source is testing it yourself. Take your twenty most common patient questions and put them to ChatGPT, Gemini and Perplexity in turn, then write down who gets named. A professional association? A clinic portal? A rating site? A colleague two streets away? Nobody from your region at all? One afternoon of this shows you where the gap is and which sources you are really competing with, which is not the same list as your Google competitors.
Public question sources help with the wording: Answer the Public, Reddit threads, gutefrage.net, Google's own autocomplete. People phrase a worry there as loosely as they do to a chatbot. Read them for phrasing rather than for placement, though. In an Ahrefs study of 1.4 million real ChatGPT prompts, only about 2 percent of the Reddit URLs the model retrieved ended up cited, against roughly 88 percent from its general search channel. Collect the phrasings in one table, sorted by symptom, by fear and by what the person wants to do next. That table is the raw material for everything downstream: your content and your AI visibility both come out of it.
Turning those questions into pages the AI can cite
A model quotes you when your page answers the question more usably than the other pages it retrieved. In practice that means one clearly structured section per recurring question, on your own site, written as an answer rather than as a paragraph about how caring your team is. "Are heart flutters dangerous?" gets: what it usually is, the warning signs that change the picture, what to do at home, and the sentence that says when a cardiologist should look at it. Write it once, properly, and it serves the patient reading it and the system retrieving it.
Form does much of the work. Put the question itself in the heading, the answer in the first sentence, the detail underneath. Stay clinically correct but use the patient's words, because retrieval runs on the overlap between their phrasing and yours: "Tinnitus: when ringing in the ear needs a doctor" is reachable in a way that "Otological differential diagnostics" is not. One page for both readers. Google's own guidance is explicit that you should not write a separate version "for AI", and it treats mass-produced AI-facing pages as scaled content abuse.
Then the accountability layer. Sign the text with the doctor's name and specialty, cite the guideline or study behind each medical claim, and show when it was last reviewed. The academic GEO paper presented at SIGKDD 2024 found that adding quotations, statistics and explicit source citations did most to improve how often a source was used, though it measured that on a simulated engine, so treat it as a direction rather than a promise. It fits what you would expect on medical topics: a page that names Dr. med. Weber, specialist in cardiology, and carries a review date gives a model something an anonymous advice blog cannot.
E-E-A-T: why a named doctor beats an anonymous blog
Health sits in the category Google's quality guidelines call "Your Money or Your Life", where bad information can do real harm, and it is judged accordingly. The keyword to know is E-E-A-T: experience, expertise, authoritativeness and trustworthiness. For a named specialist that framework is a gift, because you can evidence all four in a way an anonymous advice site structurally cannot.
Concretely: a named medical author with their qualification on every clinical text, a complete imprint, a real address, references to guidelines or your professional association, and a visible review date. What happens off your own site counts at least as much. In an Ahrefs analysis of roughly 75,000 brands, how often a brand was mentioned across the web correlated with its AI citation rate at about 0.66, some three times more strongly than backlinks did. For a practice, that means being named on clinic pages, in specialist directories, in local press and by professional bodies is worth more than any on-page trick.
Most practice sites work against themselves here without noticing. "Our team" instead of a name. No title, no date, no source. Patients may read that as modesty; a retrieval system reads a page with nothing to verify. Put the name back in: who wrote it, what they are qualified in, where the knowledge comes from, when it was last checked. Anonymity is the cheapest visibility problem you have and the fastest to fix.
Where advertising law and confidentiality draw the line
You work inside a tighter frame than a shop does. The German Medicinal Products Advertising Act limits how medical services may be advertised, and the professional code rules out misleading claims and promises of success. That constraint is not an obstacle here, it is the format. "We cure your tinnitus" is out. "How tinnitus diagnostics work in our practice, step by step" is clean, and it is the more quotable sentence anyway, because it describes something specific instead of claiming an outcome.
Collecting questions has a hard boundary. No names, no case details that could identify anyone, nothing recognizable in your texts and nothing recognizable pasted into a chatbot. Record the pattern, not the patient: "mole, changed, three weeks, afraid of melanoma" is the useful form. Confidentiality does not end at the keyboard, and no amount of visibility is worth a data-protection incident.
Say the limits out loud on the page. A line stating that the text was reviewed by a named doctor and does not replace an examination reassures the patient and helps a system judging whether a health page is responsible. There is a defensive reason to be precise, too. Columbia's Tow Center ran 1,600 queries through eight AI search tools and found more than 60 percent of the answers got the source wrong, usually stated with full confidence. These systems will say things about your practice either way. The clearer your own page is, the less they have to invent.
A three-week start, then a quarterly rhythm
Weeks one and two: a sheet at reception and by the phone, and every question that gets asked goes on it. No analysis yet, just collection. Week three: you sit down for an afternoon, put the twenty most frequent of those questions to ChatGPT and Gemini, and record who gets named in each answer. That is your baseline. It costs nothing, and it is more than most practices in your specialty have ever measured.
Then you choose. Take the ten questions closest to a decision, which will almost always be the frightened ones and the "which doctor do I even need" ones. Each gets one section on your site: the question as the heading, the answer in the first sentence, the detail below, a named author, a review date. Ten sections a colleague would put their name to beat fifty thin ones, and they are far easier to keep current.
Then it becomes a quarterly half hour. Re-run the same prompts, note what changed, add the questions your team has collected since, and refresh anything stale, starting with hours and insurance details. Which sources these engines lean on moves around a great deal from month to month, so the value is in the repetition rather than in the first pass. A practice that has kept this up for a year is not one project ahead of the one starting now, it is a habit ahead.
Conclusion: the questions are the strategy
The questions patients put to an AI are the least filtered research you will ever get about your own catchment area. Nobody phrases a search box the way they phrase a worry, and nobody tells a receptionist everything they will type at eleven at night. Collect those sentences and you learn what your patients are actually deciding between. That helps your visibility, and it helps the conversation in the room.
None of this is a trick, and there is no file you can upload to be preferred. Google states plainly that its AI features run on its ordinary search ranking systems and that no AI-specific markup or llms.txt is required. The work is the unglamorous kind: take the knowledge already in your head, write it where a crawler can reach it, and attach a name, a date and a source to it. That is the whole bridge between your expertise and the answer a patient reads.
The first step takes five minutes: ask your team which three questions came up most this week. Write those three down and answer them properly on your own site. Do that well enough, often enough, and the next patient who asks a chatbot which specialist to see hears the name of your practice in the reply.
Common questions
Does writing for AI visibility breach the medical advertising rules?
No, as long as you are informing rather than promising. The Medicinal Products Advertising Act and the professional code prohibit misleading claims and guarantees of cure; they do not prohibit clear patient education. Drop the superlatives, describe the procedure instead of the outcome, and add a line saying the text does not replace an examination. None of that touches the parts a model actually quotes.
How do I find out whether my practice shows up in ChatGPT or Gemini?
Ask them. Put your patients' real questions to each engine, including the blunt one: a specialist in your field, in your city, taking new patients. Write down every source each answer names, and vary the wording, because different phrasings pull different sources. Where you are absent, you can see exactly who was chosen instead of you. Repeat it once a quarter, because the answers move.
Is this worth the effort for a small specialist practice?
It suits a small practice better than a large group. A clinic chain needs a committee to change a sentence; you can publish ten precise answers in two weeks and sign them yourself. And you are not competing on budget: a named specialist with a real address and a review date is exactly what these systems look for on health topics, and it is what anonymous lead-generation portals cannot produce. The advantage lies in starting while your local colleagues have not.
Read on