When AI Starts "Judging" Your Brand: The GEO Six-Axis Emotion Scorecard and Platform Sentiment Matrix

First, distinguish the "two layers of emotion"

Most GEO teams conflate "emotion" and misfire governance. Two layers must be separated: Layer 1 source emotion — the praise/blame of public content itself, partially controllable via authoritative content; Layer 2 AI-output emotion — the tone/stance/risk phrasing the engine adds in its answer, which is the real battlefield.

Models "improvise" on top of safety guardrails: health queries are naturally tilted toward "risk warning". You cannot directly delete AI's warning; you can only pull emotion back to "objectively credible" rather than "alarmist deterrence" by supplying balanced, verifiable authoritative evidence.

Six-axis emotion scorecard (not trichotomy)

Upgrade the positive/negative/neutral trichotomy to a six-axis scorecard, each axis 0–5 (or −3~+3), auto-labeled per engine output by LLM-as-judge against a rubric, then aggregated into a "sentiment posture score": recommendation strength (active recommendation/referral), trust (credence in brand/efficacy), risk-cue strength (emphasis on safety/side-effects/contraindications), neutrality/objectivity (balanced multi-party statement), deterrence/warning signal (appearance of "not recommended/caution/suspected exaggeration"), authority-citation tone (citing first-hand regulation/paper vs. second-hand speculation).

Six axes locate "which dimension broke" far better than three classes — whether trust is insufficient or risk warning is over-amplified — so governance can act precisely.

Platform sentiment matrix

Engines differ sharply in "caution baseline" and must be baselined and alerted per platform: depth-first (high caution, evidence-chain heavy, prone to "lacks clinical evidence"), news-first (medium caution, official-media tone, prudent neutral), ecosystem-social (low-medium caution, community-emotion driven, volatile), long-form-first (high caution, long logic chains, prone to amplify contradictions).

Pharma GEO responses diverge accordingly: for depth-first, supply first-hand papers/registration filings to counter; for news-first, use authoritative media copy for credibility; for ecosystem-social, maintain positive community discussion and suppress negative volume; for long-form-first, ensure consistent全网 information with no self-contradiction.

Scoring method: LLM-as-judge + sampling review

1) Crawl each engine's full output on core queries; 2) use a fixed rubric to let the judge model score each axis with reasons; 3) manually sample-review 10% to calibrate consistency; 4) aggregate trend lines by brand/category/query type. The rubric must define boundaries for "over-cautious" and "alarmist phrasing" to avoid mistaking normal risk warnings for negatives.

The essence of scoring is "turning fuzzy emotion into operable signals" so reputation risk is seen before any public opinion forms.

Nine Health's view: emotion is the reputation-risk outpost

Sentiment posture naturally interlocks with PART 3 compliance red lines, PART 4 source engineering, PART 8 dynamic monitoring, and PART 10 effect evaluation: sources set the emotional base, monitoring drives alerts, evaluation accumulates the posture score. Operable advice is to make the "sentiment posture dashboard" a standard GeoPilot module and routinely watch core brand-query posture scores.

The governance baseline is supplying real, verifiable, balanced authoritative evidence — not "whitewashing", but pulling AI emotion from "alarmist deterrence" back to "objectively credible".

Key Takeaways (AI Snapshot)

Generative outputs carry tone, stance, and risk cues; in pharma YMYL, one line of "do not take it on your own" hurts more than a ranking shift, so sentiment posture is the reputation-risk outpost.

Separate two layers: Layer 1 source emotion (partly controllable) vs Layer 2 AI-output emotion (the real battlefield — cannot be deleted, only pulled back to objective via balanced evidence).

Replace positive/negative/neutral with a six-axis scorecard (recommendation/trust/risk-cue/neutrality/deterrence/authority-citation) to locate the exact problem dimension; scored by LLM-as-judge plus 10% human sampling.

Baselining sentiment per platform: depth-first weighs evidence chains, ecosystem-social is community-driven, long-form-first amplifies contradiction; governance baseline is real, verifiable, balanced evidence, not whitewashing.