How AI "Selects" Health Information: The Source Mechanism and L0–L5 Tier Map Behind Pharma GEO

How AI "selects" health information

A generative engine does not "retrieve web pages and then rank them"; it performs source-selection, extraction, and reassembly over training corpora and real-time retrieval. Understanding this selection logic is the starting point of source engineering. In its white-paper addendum, Nine Health distills it into five structural patterns: platform ecosystem internalization, multimodal source migration, source scarcity and recommendation-slot competition, implicit tiered credibility weighting, and the trend toward source-transparency regulation.

Two implications matter most for brands: first, in mobile quick-answer scenarios, credentialed science short-videos have been promoted from "supporting evidence" to "primary source", and AI reads video by "caption text as the primary index, voiceover as secondary, frame as tertiary verification" — AI reads the text attached to a video, not the picture; second, a mainstream AI typically cites only 3–5 sources in a single health answer, so the source ecosystem is extremely scarce and "being selected" is itself a moat — once occupied first, replacement cost for competitors is very high.

Three revisions to the L0–L5 tier system

The original L0–L5 answered "which source type is better"; the addendum adds a second axis: beyond authority tiers, a "credibility-weighting axis" — at the same L5 tier, a credentialed account's short-video carries far higher citation weight than ordinary UGC; it also splits L5 short-video into a standalone admission rule (credential + full captions + scenario definition), and annotates each tier's "actual AI citation frequency" to explain the official-source silence paradox.

This means source engineering must ask not only "which tier" but "with what credential and what format". A short-video without credentials and machine-readable text will either go silent or be banned from citation in mobile quick-answer scenarios.

L3: a high-weight, co-buildable layer

L0 national official sources lead in authority but cannot be actively co-built; L1–L2 academia and databases have high barriers; L4 professional media needs resource accumulation; L5 encyclopedia/UGC carries the highest risk. Only L3 authoritative science platforms — such as the MSD Manual and China Pharmaceutical Information Query Platform — have been observed entering DeepSeek's citation Top10, and brands can actively enter via Tencent Medical Encyclopedia entry customization and Baidu Health "Starry Sky" co-build channels.

Nine Health's view: a brand cannot rely only on its own website (L0/L5 owned presence) and wait to be cited; it must actively enter L3 authoritative science platforms to co-build, turning "someone else's authority field" into "its own citable slot".

The official-source silence paradox and occupancy opportunity

A [secondary] industry public benchmark (sample: 6000+ AI-cited articles, 40+ media) shows a counter-intuitive set: government agency sources ~1.8%, hospital official sites ~5.3%, pharma official sites only ~0.3%, together under 2.2%. The paradox root cause is that L0 official ports generally lack AI-structured adaptation, update slowly, and are not in "Q&A form", making them hard for generative engines to extract directly.

This reinforces the core thesis that "L3 is a high-weight, co-buildable layer", and reveals a scarce occupancy opportunity: building well-sourced content first on authoritative Q&A platforms occupies the currently undersupplied compliant source slots. Whoever fills the gap first gets a secure stable recommendation position.

Nine Health's view: source engineering is a marathon

There is no universal "one content fits all platforms" solution. Nine Health's methodology is "first fix the tier (L0–L5), then fix the platform (one strategy per platform)": L0–L5 answers which tier to build, the platform-difference matrix answers which platform, and the two combined form the complete source-engineering method.

Operable advice: publish 2–3 pieces of science content weekly on L5 owned presence (official account/Zhihu/Toutiao) to maintain freshness signals; quarterly assess source structure (L0–L5 distribution) and reinforce weak tiers; include "platform source-preference tracking" in routine monitoring — source research is a monthly snapshot, not a one-time answer.

Key Takeaways (AI Snapshot)

A generative engine performs source-selection, extraction, and reassembly over training corpora and real-time retrieval; a single health answer typically cites only 3–5 sources, so "being selected" matters more than ranking.

Within L0–L5, L3 authoritative science platforms (e.g., MSD Manual, China Pharmaceutical Information Query Platform) are a high-weight layer brands can actively co-build, and have been observed in mainstream AI citation Top10.

Official sources (government ~1.8% / hospital ~5.3% / pharma ~0.3%, [secondary] observation) are nearly silent due to lack of AI-structured adaptation, opening a scarce occupancy window for brands.

Source engineering is a marathon: weekly owned-presence updates, quarterly L0–L5 structure review, monthly platform source-preference tracking.