The Health Publisher's Prompt Map

How is keyword research different for AI search?

Four differences, three of them measured. First, the input is longer and more personal. One comparison of real against tool generated prompts found 15.1 words against 8.8, and personal pronouns in 52.1% of real prompts against 18.8% of generated ones. A separate analysis reported an average of 23 words when the assistant was not performing a web search and 4.2 words when it was, which tells you the input length depends on what the user thinks they are using. Second, the unit is a short chain rather than a single query. Across 142,808 real conversations the mean was 4.62 turns, but the median was 2 on ChatGPT and 1 on Gemini, Grok and Perplexity. Plan for two moves. Third, much of the space is not a publishing opportunity at all. In 1.1 million classified conversations, 49% of messages were Asking, 40% were Doing and 11% were Expressing, and only the Asking share is something a publisher can win. Fourth, and the practical one, there are no volumes. No equivalent of a search volume figure exists for prompts, which means prioritisation has to come from commercial value and coverage gaps rather than from a sorted list. So the work changes from discovering strings to modelling demand, and the output changes from a keyword map to a commissioning calendar.

The four differences, with what each one changes

The addressable share row is the one that should change a budget conversation. The figures come from a study of 1.1 million conversations classified by intent, and they say that roughly four messages in ten are somebody asking the assistant to produce something: draft this email, summarise this document, write this code. A health publisher cannot compete for any of that. When a board paper cites total assistant usage as the size of the prize, it is overstating the addressable market by roughly a factor of two, and a content strategy built on that number will under deliver for reasons that have nothing to do with the content.

The direction of travel is favourable, though. Over roughly a year in the same dataset, Seeking Information rose from 14% to 24% of conversations while Writing fell from 36% to 24%. The share a publisher can address is growing while the share it cannot is shrinking, which is a better argument for investment than the total usage number and happens to be true.

The multi turn claim, corrected

Every page in this category asserts that AI search is multi turn, and the assertion is doing a lot of work without evidence behind it. The largest study of ChatGPT usage, covering 1.1 million conversations, reports no turn count statistics at all. The best available numbers come from a dataset of 142,808 conversations across five platforms, assembled from publicly shared conversation links.

The limitation matters and runs in a helpful direction. These are conversations people chose to share publicly, and people share the long interesting ones. If anything the real median is lower than two, not higher. So the honest model is an opening prompt and one follow up, with a meaningful minority of single shot questions and a thin tail of long sessions that are not worth planning against.

That correction has a direct consequence for how this map is used. Do not build a four stage prompt chain and assume a reader walks through all four in one conversation. They do not. The stages describe where a reader is when they arrive, not a journey they complete in a session. A reader at the Options stage opens with an Options question. They did not pass through Sensation and Naming on the way, because that happened three weeks ago on a different device.

What questions do people ask AI about health conditions?

The best answer available comes from a peer reviewed study of 617,827 health conversations drawn from a generalist assistant's logs over one month in January 2026, classified into twelve health intent categories. The largest block by a distance is health information and education at about 41%, followed by personal symptoms, test results and conditions at nearly one fifth. The remainder spreads across academic and research support, medical paperwork, care navigation and emotional wellbeing. Two findings in that dataset matter more to an editorial plan than the headline split. The first is that roughly one in seven symptom and condition questions was asked on behalf of somebody else, a partner, a child or a parent, which is a reader almost no health library is written for. The second is that intent separates sharply by device: academic support and research accounted for 16.9% of desktop conversations against 5.3% of mobile, while symptom questions and health concerns were 15.9% of mobile against 6.9% of desktop. Personal health questions also rise in the evening and at night. So the honest answer is that people ask three broadly different things, from three different situations, and most health content is written for only one of them.

What the health conversation data shows

The study is the strongest evidence currently available on this question and its limits should be stated alongside it. One platform, one month, queries only with no outcome data, and the full twelve category breakdown appears in a figure rather than in reportable numbers. Within those limits it is a very large sample of real behaviour rather than an inference from keyword tools.

The caregiver finding deserves its own paragraph because it is the commercially useful one. One in seven is not a rounding error, it is a substantial minority of the highest intent queries in health, and the person asking has a different problem from the patient. They are describing somebody else's symptoms with imperfect information, often cannot answer the follow up questions an assistant asks, carry a different kind of anxiety, and are frequently the person who will actually make the appointment, choose the provider or buy the product. They are, in other words, the decision maker, and almost nothing in the category is written for them.

The two axes, and why these two

A map is only useful if its axes explain most of the variation and can be observed from outside. These two do. One is where the reader has got to, which determines what they need. The other is who is asking, which determines how they phrase it and what they will do next. Everything else, including condition, specialty and format, is detail inside a cell rather than an axis.

Axis one: the four stages

These are not funnel stages in the marketing sense and they are not a journey the reader completes in one session. They describe the state a reader is in when they open the conversation, and the correction in the previous section applies: a reader arrives at a stage, they do not walk through them.

The two stages that matter commercially are three and four, and they are the two that most health libraries serve worst. A large publisher usually has excellent Naming content, because that is what a decade of keyword research optimised for, and thin Options content and almost no Action content. That is also exactly the split the citation evidence predicts will hurt, because the commercial question panel in our citation research was dominated by platforms, retailers and forums rather than by publishers, and Action is where those competitors live.

Axis two: the four personas

Three of these four are grounded in the conversation data rather than invented in a workshop, which is unusual for a persona model and is the reason to prefer this one over the segments a brand team would produce.

The professional persona is the one most publishers already serve accidentally and never deliberately. A consumer health library with genuine clinical depth gets read by nurses, pharmacists, students and allied health professionals who arrived looking for a quick, reliable answer. At 16.9% of desktop health conversations this is not a fringe audience. It is also the one cell where the provenance work described in the medical review workflow pays off most directly, because this reader actually checks.

The map: sixteen cells, published in full

Three of these four are grounded in the conversation data rather than invented in a workshop, which is unusual for a persona model and is the reason to prefer this one over the segments a brand team would produce.

The professional persona is the one most publishers already serve accidentally and never deliberately. A consumer health library with genuine clinical depth gets read by nurses, pharmacists, students and allied health professionals who arrived looking for a quick, reliable answer. At 16.9% of desktop health conversations this is not a fringe audience. It is also the one cell where the provenance work described in the medical review workflow pays off most directly, because this reader actually checks.

The map: sixteen cells, published in full

Four stages by four personas. The worked example below runs type 2 diabetes through every cell, because it is a condition with a real caregiver population, a large diagnosed population and a professional audience, so nothing has to be invented to fill a box. The structure transfers to any condition. The prompts do not, and should be rewritten per condition using the method in the next section.

Each cell carries an opening prompt, the follow up that actually decides the outcome, and an honest verdict on whether a publisher can compete there. The verdict column is the one to read first, because four of the sixteen cells are not worth commissioning for.

Stage 1. Sensation: no name for it yet

The caregiver row in this stage is the single clearest content gap in the whole map. A person describing a parent's symptoms, who cannot persuade that parent to see a doctor, has a question that is half clinical and half practical, and the practical half is unanswered everywhere. That is a page nobody has written and it would be cited, because there is nothing else to cite.

Stage 2. Naming: is this the thing

Stage 2 is where the entire category already competes, and the opening prompt in the first row is the most contested string in health publishing. The citation evidence says the incumbents on questions like that one are extremely entrenched, with ten organisations taking more than half of all citations in a clinical question panel. Displacing them on what are the symptoms of is not a realistic objective. Owning the follow up is.

Stage 3. Options: what can be done

Stage 3 is where a health publisher has the clearest structural advantage and usually the thinnest library. A clinical source will not tell a reader what reversing it realistically involves week by week, because that is not what a guideline is for. A commercial site will tell them, badly, because it is selling something. A publisher with editorial independence and clinical review is the only party with both the permission and the incentive to answer it properly, and the follow up question in the first row is as close to a free commercial opportunity as this map contains.

Stage 4. Action: a decision is imminent

Read the first row and the second row together, because the contrast is the lesson of the whole map. The product choice question is one a publisher will lose, and the citation evidence is unambiguous about it: in a panel of commercial health decisions, the top cited domains were platforms, retailers and forums, and no peer reviewed journal appeared in the top hundred. The appointment question sitting directly underneath it is one a publisher can own outright, costs almost nothing to produce, and leads to the same commercial outcome by a different route.

The map at a glance, scored

Four of the sixteen cells are not worth commissioning for, four are crowded, and four are close to empty. That distribution is typical and is the reason to audit before planning.

Count the caregiver row. Three of its four cells are described as empty and the fourth as thin, and the persona accounts for roughly one in seven of the highest intent health queries measured. If this page produces one commissioning decision, it should be that one.

The two turn chain, and why the second turn is the one to brief for

The follow up column in those four tables is the part of this page that will change what gets written, so it deserves its own section. The median conversation is two moves. The first move establishes the topic. The second move is where the reader's actual problem surfaces, and it is almost always narrower, more personal and more decision shaped than the opening.

Most health content is written as an answer to the first move. Comprehensive, encyclopedic, covering the topic. Then the reader asks the second question and the page has nothing to say, because the second question was not a topic question. It was a circumstance question.

The pattern across all four is the same. The opening move asks for information and the second move asks for judgement. Information is a commodity that every source in the citation study already supplies, and judgement is the thing an experienced editorial team can produce and a guideline body structurally cannot. That is the whole argument for a health publisher's continued relevance, compressed into one column of a table.

Building your own prompt set from the map

The map is the structure. The prompts inside it are yours and have to be built, because a generic list will describe a generic publisher's demand. This takes a few days rather than a quarter, and the sequence matters more than the tooling.

1. Start with what readers already asked you. Internal site search logs are the single best source and almost always ignored. They are real language, from your actual audience, already in question form a surprising amount of the time. Pull twelve months, strip navigational queries, and keep anything phrased as a question.
2. Add what your commercial and editorial teams get asked. Customer service tickets, comment threads, the questions that come in through contact forms, and what the sales team hears. This is the richest source of second turn questions, because people only escalate to a human once the obvious answer failed them.
3. Add the classical demand you already know. Your existing keyword research is not worthless, it is incomplete. Convert the head terms into the sentences a person would actually type, and keep both forms so you can compare later.
4. Now fill the map and look at the holes. Place every prompt into one of the sixteen cells. The distribution will be badly skewed toward Naming for the person themselves, and the empty cells are the finding. Do not fill them from imagination yet.
5. Fill the empty cells by asking the right humans. For the caregiver cells, ask people who have cared for somebody with the condition. For the professional cells, ask the clinicians on your review panel what they get asked. Thirty minutes with three people produces better prompts than a day with any tool.
6. Write each prompt as a sentence with a pronoun in it. Real prompts carry personal pronouns in roughly half of cases against under a fifth for tool generated equivalents. A prompt set written in keyword voice will not behave like real demand when you run it.
7. Pair every prompt with its likely follow up. This doubles the set and is the step that makes it useful for commissioning rather than only for measurement.
8. Stop when the cells are balanced, not when you hit a number. Then hand it to the measurement programme, which decides how many of them you can afford to run and how to allocate across the cells. That decision is specified in measuring AI visibility across a content library and is deliberately not made here.

What not to do

Turning the map into a commissioning calendar

If the map does not change what gets written, it was an analysis exercise and the time was wasted. The output is a brief list, ordered, with a named cell against every brief. Here is the shape of one quarter for a publisher who has run the audit and found the typical distribution.

Note what is not on that list. No new Naming content for the person themselves, which is where most health publishers would instinctively invest because it is where their traffic currently is. That cell is crowded, the incumbents are entrenched, and adding a fourteenth symptoms of page to the web is not a strategy. Defend what you have there and spend the commissioning budget on the four cells where there is nothing to displace.

Once the briefs exist, how the pages are actually structured, what the answer block contains and what markup they carry belongs to condition hub architecture for AI answers. This page stops at the brief, deliberately.

Limitations, and what would change our mind

The map is a model and the evidence under it is uneven. Here is the full accounting, source by source.

What would change our mind. Turn count data from a platform's own logs rather than from shared conversations would settle whether the two turn model holds, and it is the single most useful thing an assistant provider could publish. A health specific Asking and Doing split would sharpen the addressable share argument considerably, which currently relies on a general usage figure. A replication of the caregiver finding on a second platform would move it from one study to a pattern, and given that it is the basis for the strongest recommendation on this page, that replication matters more than anything else in this list.

Conclusion: model the demand, do not buy the list

The keyword list worked because the world it described was finite, and that world is gone. What replaces it is not a longer list. It is a model with two axes that explain most of the variation: where the reader has got to, and who is actually asking. Sixteen cells, of which most publishers serve one well.

Three things in the evidence should change a plan this quarter. Roughly half of assistant usage is task completion a publisher cannot compete for, so the opportunity is smaller than the headline numbers suggest and should be sized honestly. The conversation is two moves, not ten, so the brief needs a second question rather than a longer outline. And roughly one in seven of the highest intent health questions is asked on behalf of somebody else, which is a reader almost nobody writes for and the clearest commissioning opportunity in the category.

None of that requires a tool. It requires a few days with your own site search logs, your support tickets and three people who have actually cared for somebody with the condition you publish about. The vendors selling prompt volumes derived them from keyword data, because no platform publishes prompt volumes, and a model you built from your own demand will beat a list you bought from somebody else's.

For the citation evidence that makes query intent the dominant variable, see how AI engines decide which health publisher to cite. For how many of these prompts to run and how to allocate them, see measuring AI visibility across a content library. For the provenance that the professional persona checks, see the medical review workflow. For how to build the pages once commissioned, see condition hub architecture for AI answers. The framework is in the generative engine optimisation pillar, the commercial case in GEO for medical publishers, and the crawler access decision in licence, block or optimise.

Frequently Asked Questions ForPrompt Mapping for Health Content