The five factors, in order of observed weight
Every AI Overview panel contains between three and six cited sources. Not all sources compete equally for those slots. Five factors dominate which pages the model picks, and in the tracked data these five factors are strongly additive — a page that scores well on one or two but poorly on the others usually gets displaced within one or two render cycles when a stronger competitor page enters the query cluster.
1. Direct-answer clarity in the first 100 words
The AI Overview model has to synthesise a two-to-four sentence answer from its citation set. It heavily prefers pages that themselves open with a direct answer to the query — a bolded paragraph immediately after the H1, a TL;DR block, an FAQ with the query as the question. Pages that bury the answer under a "let us first understand what is X" preamble are cited far less reliably, even when they eventually contain the exact same substantive answer.
In practical terms this means the first 100 words of any healthcare page targeting an AIO-eligible query need to contain the answer, in prose that could stand alone as the AIO answer itself. A 2000-word LASIK explainer whose first 100 words describe "what is LASIK" as history rather than "what does LASIK recovery look like" as answer will not earn citation on recovery-time queries no matter how thorough the later sections are.
2. Extractable structure
Overviews cite pages the model can lift clean self-contained blocks from. FAQPage schema with answers under 300 characters. Q&A pairs that read as complete units. Comparison tables that pull out cleanly on decision-stage queries ("FUE vs FUT", "root canal vs extraction", "SMILE vs LASIK"). Long-form pages with no extractable structure — a wall of prose with no headings, no FAQ, no tables — are systematically underrepresented in the citation set even when they rank well on classical signals.
The mechanism is not mysterious. The model is asked to produce a compact answer with citations; pages that already offer compact citable units make the model's job easier. Extractable structure is not decoration; it is retrieval infrastructure.
3. Entity strength
Named-expert authorship, Physician schema, MedicalBusiness and MedicalProcedure entities, cross-referenced consistently across the site. The overview model favours pages whose entities it can disambiguate cleanly — a Person entity tied to a specialty tied to a real clinic in a real city carries more weight than an unattributed brand-only page. In tracked data, adding Physician schema and credentialled bylines is often the single highest-leverage lift for a brand that already has FAQ and direct-answer content in place.
Entity strength also compounds. Once the Google Knowledge Graph is confident about a brand's entities, the same brand gets cited more readily on adjacent queries in the same cluster — the retrieval step has to do less disambiguation work each time.
4. Freshness
Overview citations skew toward pages with a recent dateModified. In healthcare specifically, "recent" appears to mean within 90 days for evolving topics (treatment cost, protocol updates, regulatory guidance) and looser tolerances for stable anatomy or generic condition-explainer content. Pages that were published in 2022 and never updated get cited far less than pages that received substantive edits and a fresh dateModified within the last quarter, even when the underlying content is largely equivalent.
This does not mean rewriting content quarterly for no reason. It means updating dateModified when a substantive edit actually happens, timestamping the changed sections in the visible body copy, and running a genuine freshness pass across the tracked cluster at least twice a year.
5. Source diversity
An AI Overview typically wants a mix of source types — a brand page, an editorial or educational source, and often a video. The model rarely cites five brand pages from the same domain even when they are the strongest content available; it wants breadth. The practical implication for a healthcare brand is that YouTube assets targeted at the same AIO-triggering queries (video answers to "is LASIK painful", "what is IVF success rate by age") often earn citation slots that text-only pages cannot. This is why ICG's video workstream via YODA is not a separate marketing programme; it is a citation-earning workstream aligned to the same cluster set the text AEO work targets.
