This reference is analysis throughout. We have no privileged information about how any specific system selects its sources, and where we describe likely behaviour we say so. That labelling is not modesty. It is the point of this publication.
Why this category is symptom led
Someone notices a line, a hollow, a texture, an asymmetry, a change. They do not know what it is called. They do not know which procedure addresses it, whether any procedure does, or whether it is a cosmetic question at all. They describe it in ordinary language and expect the system to work out what they mean.
That is a translation problem, and translation is exactly what embedding based retrieval is good at. It is also exactly what keyword matching was bad at, which is why the sector learned to build pages around procedure names instead: those were the strings that matched.
The mechanism, plainly
A retrieval based answer system converts the question into a representation of its meaning, finds passages in indexed documents whose meaning is close to it, composes an answer from those passages, and attributes the answer to the sources it used.
Four consequences follow directly from that description, and none of them requires knowing anything proprietary about a specific product.
The unit of retrieval is the passage, not the page. A page can be broadly irrelevant and still contain the one paragraph that answers the question. Conversely, a well ranked page whose relevant content is scattered across it is harder to use.
Specificity beats breadth. A passage that answers a narrow question precisely is a closer match than a passage that covers the topic generally.
Structure is functional. Clear headings, self contained sections and explicit qualifications make a passage extractable. A wall of text is not.
Consistency across sources matters. A system assembling an understanding of an entity from many places is more confident where they agree, which makes coherence a property worth maintaining deliberately.
The practical instruction is unglamorous and it does not need a vendor: write specifically, structure clearly, put the qualification inside the sentence, name the jurisdiction, link the primary source, and say the same true thing everywhere. That has been good practice for twenty years. It is now more mechanically rewarded than it was.
What tends to get cited in a health adjacent category
In categories where the stakes are real, institutional sources are structurally advantaged. Regulators, registers, national health bodies and clinical professional bodies publish material that is authoritative, stable and widely referenced.
That has an uncomfortable implication for commercial publishers here: the best available position is not competing with the institutional source but being the clearest explanation of it. A page that contradicts the regulator is competing against the source the system trusts most in the category, and it loses.
ObservedGenerative answer surfaces attribute sources visibly, which makes the selected set partially inspectable even where the selection logic is not published.AnalysisThe role that adds value without competing with the source is the bridge: accurate explanation of an institutional position, with the institution linked. That is a real editorial job and it is scarce in this sector.SpeculationWe would expect thin, technically optimised pages to lose ground further, because thinness is precisely what passage level retrieval cannot use. That is a mechanism argument, not a measurement.The measurement problem, and why claims about it are unreliable
Attribution from generative answer surfaces is incomplete by construction, because a cited answer frequently satisfies the query without a click. That means referral based measurement systematically undercounts influence and cannot be corrected into precision.
Confident percentage claims about AI visibility should therefore be treated with the same scepticism as market size claims in this sector. Most are derived from hand run prompt samples, which is a legitimate method and a very different thing from a measurement of reach. There is no certifying body, no standard and no audit.
What changes commercially
The likely effect is a change in the shape of the acquisition funnel rather than in its size. If an answer engine satisfies the informational stage, the visits a clinic receives are fewer and later, arriving from people who are further along in their decision.
Fewer visits at higher intent is not a worse outcome. It does break every metric calibrated on session volume, and it makes informational content look like it is failing at exactly the moment it is working. Businesses that cut that content because sessions fell may be removing the thing producing their remaining enquiries.
What to actually do
Reorganise around the question rather than the procedure. Write fewer and longer pages. Make every section self contained. Name the nation. Link the instrument. State what is not known as clearly as what is. Publish verifiable practitioner identity. Keep the description of your organisation consistent everywhere it appears, including in structured data.
None of that is a technique and none of it is proprietary. It is the same list a competent editor would have given you before any of this existed, which is either reassuring or disappointing depending on what you were hoping to buy.