The final issue in this series is about the discovery layer as an institutional question rather than a technical one. What follows is analysis. We have no privileged information about how any specific system selects sources and we are not going to pretend otherwise.
There is a regulatory dimension to this that the sector has not caught up with. If a clinic's own published content is being summarised and reproduced by third party systems, the claims in it travel further and persist longer than the page itself.
Responsibility for a marketing claim sits with the advertiser. A claim reproduced by a summarisation system was still made by the clinic, and the reproduced version may drop the qualifications that made the original defensible.
AnalysisQualifications are the first thing lost in summarisation. A claim written as true in a narrow case will be reproduced as true generally.SpeculationWe are not aware of this having been tested in an advertising complaint in this sector. It looks untested rather than resolved.Write claims that survive being stripped of their context, because they will be. Put the qualification inside the sentence rather than in the paragraph after it.
The mechanical change is that documents are retrieved as passages by semantic similarity rather than as pages by keyword match, and an answer is then composed from several sources with attribution.
The consequences follow directly. Specificity beats breadth, because a specific passage matches a specific question. Structure matters, because a well delimited section is more extractable than a wall of text. Consistency across sources matters, because a system assembling an understanding of an entity from many places is more confident when they agree.
ObservedGenerative answer surfaces attribute sources visibly, which makes the selection partially inspectable even where the selection logic is not disclosed.AnalysisThe practical instruction is unglamorous: write clearly, structure properly, be specific, and say the same true thing everywhere. That has been good advice for twenty years and it is now more mechanically rewarded than it was.SpeculationWe would expect the value of a technically optimised but substantively thin page to fall further, because thinness is exactly what a passage level retriever is bad at using.There is no technique here that substitutes for having something specific and correct to say. That is the least commercially useful conclusion available and we believe it is the right one.
Institutional sources are structurally advantaged in this environment, which has an uncomfortable implication for commercial publishers in health adjacent categories: your best strategy is to be the clearest explanation of the institutional position, not an alternative to it.
That is a narrower role than the sector's content marketing assumes, and it is a more defensible one.
AnalysisExplaining an institution accurately and linking it is a role that adds value without competing with the source. Contradicting it competes with the source and loses.Position your content as the bridge between the primary document and the reader. That is a real editorial job, it is scarce, and it happens to be what retrieval systems are looking for.
Patients arrive with an answer rather than a question, and the answer came with citations they did not read.
The consultation task has changed accordingly: not correcting a misconception but locating where a broadly correct general answer stops applying to this specific person. That is a harder conversation, because the patient's information is not wrong, it is merely not about them.
AnalysisGeneral correctness that is individually inapplicable is a new category of consultation problem and it is more difficult than plain misinformation.SpeculationPractitioners may find they spend more time on individualisation and less on education. That would be a better use of the appointment, if it happens.The sentence that does the work is: that is broadly right, and here is why it does not apply to you. It concedes the source and moves to the assessment, which is the only thing the consultation can offer that the machine cannot.
"Guaranteed AI visibility." Nobody can guarantee inclusion in a system whose selection they do not control. "We are cited by ChatGPT." Possibly true for one prompt on one day, and not a property of the site. "AI SEO certified." No certifying body exists.
Percentage claims about AI referral traffic. Attribution is incomplete by construction. Treat any confident number with the same scepticism you would apply to a market size claim in this sector, which is to say a great deal.
ObservedAnswer surfaces frequently satisfy a query without a click, so referral based measurement systematically undercounts influence and cannot be corrected into precision.The claims being made about AI visibility in this sector right now have the same structure as the claims made about search engine optimisation twenty years ago: real underlying phenomenon, confident vocabulary, no measurement. That resolved eventually. It resolved slowly.
The economics of publishing in this environment favour depth over volume, which reverses the incentive that produced most of the sector's content.
A large number of thin pages was rational when ranking was the mechanism, because each page was a lottery ticket in a different draw. When the mechanism is passage retrieval and entity confidence, thin pages contribute little and dilute the entity's coherence, which is a cost rather than a neutral.
AnalysisContent pruning becomes economically rational for the first time. That is a genuine reversal and most sites have not made it.SpeculationWe would expect the median useful site in this category to get smaller and better over the next few years. That is a preference as much as a prediction and we are labelling it accordingly.Fewer, longer, more specific, better sourced. That is the whole recommendation, and it costs more per page and less in total.
Authority is becoming an entity level property rather than a page level one: who you are, consistently described, verifiably the same across every surface, saying things that agree with the primary sources.
That is slower to build and harder to fake than any previous discovery advantage, which is the best thing about it. It also rewards publications and clinics that were going to behave that way anyway, and offers very little to anyone looking for a technique.
SpeculationEntity coherence looks like the durable advantage in this environment. It is a mechanism argument built on how retrieval systems are constructed, not a claim about any specific product's behaviour.AnalysisWhich means the useful work is editorial rather than technical, and it is the kind of work that does not photograph well in a pitch deck.This is the last issue in the opening series, so we will end where the format started: state what you know, label what you do not, name the source, and let the reader check. Everything else in this issue is downstream of that.