GEO for Medical Publishers

Why has my health website's traffic dropped and what do I do about it?

Most likely because the questions your pages answered are now being answered inside the search result, and because a December 2025 core update coincided with the steepest part of the decline. Independent measurement puts the mechanism beyond reasonable dispute: Pew Research Center's analysis of 68,879 Google searches by 900 US adults found that when an AI summary appeared, users clicked a traditional result in 8% of visits against nearly twice that rate without one, and ended the session on 26% of pages with summaries versus 16% without. Published visibility-index data shows two of the largest US health publishers losing roughly 43% and 52% of index visibility between mid-2024 and January 2026. What to do about it has three parts, in order: change what you measure, because pageviews no longer describe your position; change what your advertisers are buying, because they are buying an audience that no longer arrives the same way; and make your existing editorial quality machine-visible, because the medical review process you already pay for is the one asset in this category that cannot be cheaply replicated. Chasing the lost traffic is the one response that will not work — as one publisher president put it, that traffic is not coming back.

What the visibility data shows

The click-through evidence, ranked by how much you should trust it

Three bodies of evidence circulate on what AI summaries do to clicks. They are not equally reliable and the category treats them as though they are.

The pattern is worth noticing as a discipline point. The most reliable figure on this list is from an independent research organisation that published its sample and method, and it is also the least quoted in the industry conversation. The most quoted figures are the ones produced by companies selling a solution to the problem they describe. That is not an accusation of bad faith; vendor data is frequently the only data. It is a reason to lead with Pew.

The number that matters is not traffic

Traffic is the symptom that gets the headline. The injury appears in the customer relationships, and for the first time there is a public filing that shows it.

Ziff Davis, whose portfolio includes health media properties, recorded a goodwill impairment of $54.8 million entirely against its Health and Wellness segment for the quarter ended 30 June 2026, disclosed in a Form 8-K on 6 August 2026. That figure has been widely reported. The numbers underneath it have not, and they are the more useful ones.

The company did not attribute the impairment to AI search, and this page will not either. Its risk disclosures reference content piracy and developments related to artificial intelligence, including unauthorised use of its content by developers and users of generative AI. That is a legal-risk framing, not a traffic-causation statement. The connection between AI search and the impairment is an inference drawn by observers, this analysis included, and it should be stated as one.

What is not an inference is the shape of the damage. Advertiser count and subscriber count both fell materially, and both fell faster than revenue. For any health publisher reading this, those two numbers belong on the monthly dashboard immediately, because they move before revenue does and they describe the thing that is genuinely hard to rebuild. An impression lost is an impression. An advertiser relationship lost is a sales cycle.

The publisher that did not decline, and why that is not reassuring

One major US health publisher appears on the same visibility chart with no meaningful decline at all, holding a narrow band throughout the period while its peers halved. The obvious reading is that it did something right. The more likely reading is that it had less to lose. A property that was not holding high-visibility positions across a broad condition footprint cannot lose them.

That distinction matters commercially, because the parent company of that property is the one that took the $54.8 million impairment. Flat visibility did not protect the segment. Which is the single most useful observation on this page for anyone currently being reassured by their own visibility chart: visibility index is a measure of search position, and search position is no longer the business.

The corollary is more encouraging than it sounds. If a flat visibility line did not protect a business, a falling one does not condemn one. The determinant is whether the publisher holds relationships — with advertisers, with registered readers, with licensors — that do not route through a search result. That is the actual question, and the rest of this page is about it.

How do health publishers get cited by ChatGPT and Google AI Overviews?

By being the source a model reaches for when a higher-tier source does not cover the question, and by being structurally easy to extract when it does. In health specifically the citation hierarchy is steeper than in any other category: government health bodies and regulators sit at the top, then specialist societies and clinical guidelines, then peer-reviewed literature, then academic medical centres and large health publishers. A publisher is realistically competing for the fourth tier, not the first, and strategies built on displacing a guideline body will fail. What does move the needle is the part of the question space the higher tiers do not cover well — lived experience, practical management, what to expect, how to talk to a clinician, cost and access — combined with the signals that make content extractable: a direct answer near the top, verifiable citations, a named and credentialled medical reviewer with a visible review date, and clean machine-readable structure. The published GEO research found that citing authoritative sources improved a page's own visibility by roughly 27% and that fluency improvements were worth roughly 28%, both of which a health publisher is already set up to do better than almost anyone.

The structural position of a health publisher

It is worth being blunt about where a publisher sits, because a great deal of GEO advice sold to publishers implies a position they do not occupy.

The strategic consequence is that a health publisher's route into a citation runs through the questions the tiers above do not answer well. Guideline bodies do not write about what a scan actually feels like, how to phrase a question to a specialist, what a diagnosis means for work, or what happens if you cannot afford the first-line option. Those are publisher questions, they are high-volume, and they are where the hierarchy flattens.

The full taxonomy — all seven tiers with observed citation frequency per tier, measured rather than asserted — is being built in the how AI engines decide which health publisher to cite analysis. This page states the position; that one measures it.

What makes health content extractable

Four properties, and a health publisher already produces three of them as a matter of editorial standard. The gap is almost never quality. It is that the quality is invisible to a machine.

The medical review point deserves emphasis because it is the cheapest material advantage available in this category. Health publishers spend real money having clinicians review content. That spend was incurred to satisfy a quality standard and a legal posture, and it was never structured so that a retrieval system could see it. Turning an existing cost into a visible signal is a structural project, not an editorial one, and it is the subject of the from E-E-A-T to AI citability analysis. The information-architecture side — how a condition hub should be built so that a model can use it — is covered in the condition hub architecture guide.

The post-search playbook — and why health publishers cannot run most of it

Through 2026 a recognisable playbook has emerged, built and described publicly by news and technology publishers. It is genuinely good work and it is worth studying. This is what the named players reported at a September 2026 industry summit.

Two conclusions from that table before the health-specific argument. First, the reporting from that summit was explicit that AI licensing revenue is not close to offsetting lost referral revenue. Nobody is being made whole by a licence. Second, three of the nine are selling GEO services to other people, which tells you something about where the publishers themselves think the near-term money is.

The structural problem: a health audience does not come back

Look at what the strongest moves in that table have in common. The dynamic paywall, the registration funnel, the subscription conversion, the newsletter business, the homepage investment — every one of them assumes a reader who returns. They are mechanisms for converting a repeat visitor into a known, monetisable relationship.

A health publisher's core audience does not behave that way, and no amount of product work will make it. Someone researching a new diagnosis arrives at the worst week of their year, consumes a great deal very quickly, and leaves when the episode resolves. They do not want a relationship with a website about the condition they hope never to think about again. A person does not subscribe to shingles.

This is not a content-quality problem and it is not fixable by better funnel design. It is the defining economic feature of condition-led publishing and it is why the news-publisher playbook transfers so badly. A general news publisher is converting habit into revenue. A health publisher has no habit to convert, outside a minority chronic-condition and caregiver segment that is genuinely different and should be sized before anything is built for it.

The honest conclusion is that health publishing needs a different answer, and the rest of this page is an attempt at one. It starts from the observation that a health publisher's durable relationship was never with the reader in the first place. It was with the advertiser.

What transfers and what does not

What does work: five moves for a health publisher

These are ordered by how quickly they return something and how little they depend on anyone else's behaviour. None of them is a replacement for search traffic, because nothing is. All five are ways of making the business less dependent on it.

1. Change the dashboard before you change the strategy

If pageviews remain the primary metric, every subsequent decision will be made against a number that no longer describes your position. The replacement is citation share — how often, across a defined and repeated sample of prompts in your subject areas, your property is the source a generative engine draws on and names.

Three things move onto the monthly dashboard alongside it, and all three are leading indicators that most publishers currently review annually or never: advertiser count, registered-user count by segment, and the share of sessions arriving from a named AI referrer. The Ziff Davis disclosure is the argument for the first two — advertisers fell 866 to 797 and subscribers 1.892 million to 1.618 million while revenue fell only 3% in the segment. Revenue was the lagging indicator; the relationships moved first.

Measuring citation share across a library of tens of thousands of pages is a sampling problem rather than a tracking problem, and it has a real method with real sample-size requirements. That method is the subject of the measuring AI visibility across a large content library guide. This page's point is only that the dashboard has to change first, because otherwise you will have no way of knowing whether anything you do next worked.

2. Make the medical review visible to machines

The cheapest unexploited advantage in health publishing. You already pay clinicians to review content. That review is usually represented on the page as a line of text and almost never as structured, machine-readable data linking a credentialled person, their qualifications, a review date and the specific content reviewed.

The published GEO benchmark found that citing authoritative sources improved a page's own visibility by roughly 27% and that fluency improvements were worth roughly 28% — two of the four best-performing methods tested, and both of them things a medically reviewed, professionally edited health property is already structurally good at. The gap between doing it and being seen to do it is a schema and workflow project, not an editorial one, and it is the subject of the medical review workflow and AI citability analysis.

3. Reposition the advertiser proposition from impressions to authority adjacency

This is the most commercially consequential move on the list and the one with the longest lead time, which is why it should start this quarter. A health publisher's durable asset was never the reader relationship — it was the advertiser relationship, built on the ability to reach a condition audience at a defined moment. That ability is weakening, and the advertisers know it, because they are facing the same measurement collapse from the other side.

A pharmaceutical or consumer-health advertiser now has the same problem you do: no rank, no impression log, no referrer, and no idea whether a generative engine mentions their brand when a patient asks about their category. What they want to buy is changing from impressions against a condition audience to presence in the sources that shape what that audience is told. That is a different product, and a publisher with a medically reviewed, heavily cited content estate is unusually well placed to sell it.

One caution that belongs in any conversation of this kind. Co-produced content with a regulated advertiser is subject to promotional rules regardless of where it sits, and the fact that a generative engine paraphrased it does not change whose communication it was. The regulatory position for the advertiser side is set out in the generative engine optimization for healthcare pillar, and a publisher selling into pharma should understand it before building the product.

4. Treat licensing as a distribution negotiation, not a settlement

The evidence from publishers who have done it is consistent and unflattering: AI licensing revenue is not close to offsetting lost referral revenue. One major title has signed few deals and requires fair value, a sustainable partnership beyond a one-time payment, and control over use. Another has signed none and is holding out. A third plans to block the dominant search crawler once it no longer needs the indexing.

What a health publisher has that a news publisher does not is liability exposure as a negotiating asset. Medically reviewed health content is high-risk content for a model developer to get wrong, and the provenance of health information is a live regulatory question in several markets. That is leverage — for attribution terms, for referral commitments and for usage controls, which several publishers now say they value above the cash.

The decision of what to licence, what to block and what to optimise is portfolio-level and differs by content type, and it is not a question this page should answer. The framework that prices the three options against each other is the licence, block or optimise analysis.

5. Commission against prompts, not keywords

The questions people bring to a generative engine are not the queries they typed into a search box. They are longer, they are multi-turn, they narrow toward a decision, and they are frequently about the practical and emotional dimensions of a condition rather than its definition. A commissioning calendar built from a keyword tool is modelling the behaviour that stopped.

The practical starting point is cheap: take your top fifty conditions, write out the ten questions a real person would actually ask an assistant about each, and compare that list against what you have published. The gaps are usually obvious and they are usually in the same places — what to expect, how to cope, what to ask a clinician, what it costs, what happens next. The structured method, with the four-stage and four-persona model behind it, is the subject of the health publisher's prompt map.

Is AI traffic worth having?

A fair question, and one where the published evidence is contradictory enough that it needs stating carefully. The volumes are small and the per-visit behaviour appears better.

The honest synthesis is that AI referral traffic is probably better per visit and far smaller in volume, and that the arithmetic has not yet been published for anyone, let alone for health. A publisher should instrument AI referrers in analytics now, segment them, and measure their own conversion and engagement rather than relying on a retail dataset from a vendor. It takes a week to set up and it replaces an argument with a number.

A twelve-month operating plan

The sequence matters more than the speed. The common failure is to start with content, because content is what a publisher knows how to make, and to discover a year later that there was no measurement in place to tell whether it worked.

What we do not know

A page that only reports findings is incomplete. These are the things this analysis genuinely cannot tell you, and a publisher making a large decision should know which part of the picture is missing.

Conclusion: the relationship, not the ranking

The temptation when a traffic chart falls is to treat it as a ranking problem, because ranking problems have known solutions and fifteen years of institutional muscle behind them. The evidence says it is not a ranking problem. A publisher whose visibility line stayed flat throughout the period sits inside the segment that took a $54.8 million impairment, and the operating numbers that moved first were advertiser count and subscriber count rather than revenue. Position in a search result stopped being the business before most publishers stopped measuring it.

The second temptation is to copy the playbook that news and technology publishers are visibly executing, and that one is more dangerous because the playbook is good. Dynamic paywalls, registration funnels, structured-data CMS investment, AI content feeds, selective licensing — these are serious responses by serious operators, and several of them will work. But the strongest of them convert a returning reader into a known relationship, and a health publisher's core audience does not return. A person who spent four days reading everything you had about a diagnosis is not a lapsed subscriber. They are someone who got better.

What a health publisher actually holds is a different asset than a news publisher holds, and it is more durable than it currently looks. It is a medically reviewed, heavily cited content estate in a subject area where accuracy carries legal weight, serving advertisers who need to reach condition audiences and who are themselves losing the ability to measure whether they do. Three of those four things are strengthening, not weakening. The one that is weakening is the delivery mechanism.

So the work is to make the asset legible and to change what is sold against it. Make medical review machine-visible, because the cost is already incurred and the signal is currently invisible. Move the dashboard to citation share, because otherwise nothing that follows can be evaluated. Size the recurring audience honestly and build relationship products only for it. Take citation share by condition to the advertisers, because they have the same problem and no capability. And treat licensing as a distribution negotiation where attribution and referral terms matter more than the cheque, because on the current evidence the cheque does not close the gap.

None of that restores the traffic. The most useful sentence any publisher said publicly in 2026 was the blunt one: that traffic is not coming back. A strategy built on its return is not a strategy, and the publishers who accept that first will have two years of head start on the ones who do not.

Frequently Asked Questions For GEO for Medical Publishers