Inside the machine: how generative engines select evidence

Three layers, three levers

• Large language models — the reasoning layer. LLMs are trained on vast text corpora and generate answers by predicting language. They do not 'know' facts in a database sense; they reproduce patterns. Content that appears consistently, unambiguously, and in machine-parsable form across the training corpus becomes the pattern the model reproduces. Content locked in images, behind logins, or phrased in ambiguous promotional language does not.
• Semantic search — the meaning layer. Modern retrieval converts queries and documents into vector representations of meaning, not keywords. A clinician's question about 'skin clearance' must connect to a manufacturer's PASI-90 data even though the words never co-occur. This is why clinical synonym coverage, plain-language companion summaries, and consistent entity naming (one canonical product name, one canonical indication phrasing) materially affect retrievability.
• Retrieval-augmented generation — the evidence layer. RAG systems fetch discrete text 'chunks' at answer time and hand them to the model as grounding context. The chunk — not the page, not the site — is the unit of retrieval. A chunk that contains an efficacy figure without its population definition, or a dosing instruction without its monitoring requirement, will be used exactly as incompletely as it was written. Co-location of related facts within the same retrievable unit is therefore a safety property, not a stylistic preference.^15,24

What the peer-reviewed evidence says works

The foundational academic study of GEO — presented at ACM SIGKDD 2024 by researchers from Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi — tested nine content-modification strategies across 10,000 queries and multiple generative engines.¹⁰ The findings translate directly into pharmaceutical content practice: adding statistics raised AI-answer visibility by roughly 40 percent; adding quotable, well-attributed statements raised it by roughly 28–40 percent; and citing authoritative external sources raised visibility for lower-ranked content by up to 115 percent. Legacy keyword stuffing reduced visibility by around 10 percent.¹⁰ In short, the tactics that work are precisely the disciplines pharmaceutical medical affairs teams already practice — numeric precision, attribution, source citation — applied to public digital content rather than only to regulatory dossiers.

Exhibit 3

Evidence-based GEO: what moved visibility in controlled testing

Two further findings deserve board-level attention. First, visibility gains concentrate in lower-ranked content: the Princeton team found pages around position five gained most (up to 115 percent), while position-one pages barely moved — meaning GEO disproportionately rewards challengers and disproportionately punishes incumbents who assume their search dominance carries over.¹⁰ Second, AI citation and search ranking are only loosely coupled outside Google's own surfaces: roughly 28 percent of the pages most cited by one leading assistant have no organic search visibility at all, and only around 43 percent of Google position-one pages earn citations from it.¹³ A strong SEO position is neither necessary nor sufficient for AI visibility.