VECTORY · Scoring & Methodology

🎯 NVS — Neuro Visibility Score

A composite 0–100 measure of how well a brand is cited, sentimentally treated, and authority‑backed across the AI engines we track.

Core Formula (0–100)

NVS = 0.50 × CitationCoverage × 100
+ 0.30 × SentimentIndex × 100
+ 0.20 × AuthorityScore × 100

All inputs are normalized to 0.0 – 1.0 before weighting. Final score is clamped to [0, 100] and rounded to 1 decimal.

Component	Weight	What it measures	Data source
Citation Coverage	50%	Fraction of target queries where the brand is mentioned by AI engines	SONAR Observations
Sentiment Index	30%	Tone of mentions (1.0 = positive, 0.5 = neutral, 0.0 = negative)	LLM Classification
Authority Score	20%	Share of authoritative sources (Forbes, .gov, .edu, etc.) among citing sources	Source Influence Graph

✅ Content Quality Score

A 0–100 composite that decides whether a generated page is auto‑deployable, needs a small edit, or must be regenerated. Combines positive signals with deterministic penalties.

Composite Formula (0–100)

Composite = 0.35 × Factual_Accuracy
         + 0.25 × Brand_Voice
         + 0.40 × Citation_Readiness
         − Anti_Pattern_Penalty × 100
         − Freshness_Penalty    × 100

Guard	Weight	What it evaluates	Deductions
Factual Accuracy	35%	Do claims match site ground truth? No hallucinations?	−5…15 per hallucination −3…8 per unsupported claim
Brand Voice	25%	Does content match brand tone? No filler language?	−5…10 per tone mismatch −2…5 per filler phrase
Citation Readiness	40%	RAG retrievability — would AI engines cite this content?	−3…8 per vague/generic section

LLM Rounding Bias Fix: If all three sub‑scores land on multiples of 5 (e.g. 60 / 75 / 70), a deterministic ±1…3 jitter is applied using issue count and verdict length as the seed. This kills the unrealistic “round number” bias common with LLM evaluators.

⚠️ Deterministic Penalties

Applied after the LLM evaluation. Calibrated against analysis of leaked system prompts from GPT‑5.2, Gemini 3, Perplexity Comet, and Claude Opus 4.

Penalty	Range	Triggers	Engine impact
Anti‑Pattern	0.00 – 0.15	Hedge phrases (“It’s important to note”), buzzwords (“leading provider”, “world‑class”), sycophancy patterns	All engines lower citation probability for these patterns
Freshness	0.00 – 0.10	Stale dates (year < 2024) in content body	GPT‑5.2: P>10% instability → triggers web search re‑verification, bypasses content

// Example: content with 3 buzzwords + 2 stale dates
anti_pattern = 0.05 // 5% penalty
freshness = 0.10 // 10% penalty
combined = 0.15 → −15 points from composite

// Composite 72 → 72 − 15 = 57 (Grade: D)

⚔️ Share of Voice (SoV)

How often the brand domain shows up in AI engine responses, measured against every other domain that gets mentioned.

SoV% = brand_mentions / total_mentions × 100

Counts how often the brand domain appears in AI engine responses across all SONAR observations. Competitors are all other domains mentioned, sorted by mention count (top 10 reported).

⚡ SoV = 0% is common for new or unknown brands. It means no AI engine currently mentions the brand domain in response to target queries — the exact gap VECTORY exists to close.

🏛️ Citation Authority Score

Weighted influence of the sources that cite you — not just a count, but how trustworthy they are inside each engine’s ranking model.

authority_score = Σ authority_source_influence / Σ total_influence

// Weighted by influence, not raw count
// authority = source ∈ base authority domains OR niche‑specific domains

Base authority domains (universal)

bloomberg.com forbes.com reuters.com wikipedia.org hbr.org github.com .gov .edu trustpilot.com g2.com

+ niche‑specific domains injected from SONAR enrichment per vertical

💬 Sentiment Index

A 0.0 – 1.0 score for how AI engines emotionally frame the brand. Built from per‑mention classification across every SONAR observation.

sentiment_index = (positive × 1.0 + neutral × 0.5) / total

Index range	Verdict
≥ 0.8	Very positive — brand well‑received by AI models
≥ 0.6	Positive — mostly favorable mentions
≥ 0.4	Neutral — mixed or factual mentions
≥ 0.2	Negative — unfavorable mentions detected
< 0.2	Very negative — brand has reputation issues in AI output

🔬 Quality Gate · Self‑Test Loop

An iterative patcher that re‑tests every generated page against ~10 target queries, patches weak spots, and rolls back if a patch makes things worse.

Iterative Content Patching

Self‑Test

→

Find Weak Queries

→

LLM Patch

→

Re‑Test

→

Compare & Track Best

threshold = 7 // out of ~10 target queries
max_iterations = 5

LOOP:
  if citeable_count < threshold:
    → generate patch sections for weak queries
    → merge + re‑test
    → track best_content (rollback if degradation)
    → check stop conditions
RETURN best_content

Stop conditions

1 ✅ Passed citeable_count ≥ threshold
2 Stall — same weak‑query signature detected. 1 retry with T=0.55, then stop
3 Cost guard — >30k chars patched with <50% of target lift
4 Diminishing returns — last 2 iterations each lifted ≤1 query
5 Max iterations (5) reached

Rollback strategy: The gate always tracks the best‑known content state. If a patch iteration causes degradation, it rolls back to the best state instead of keeping worsened content. Final output is always the best score achieved.

🏆 Grade Scales

Final output grades — the same scale used inside the dashboard, audit reports, and the Quality Gate itself.

NVS Score Grades

Grade	Score	Verdict
A	≥ 80	Excellent — dominates AI visibility
B	≥ 60	Good — strong with room to grow
C	≥ 40	Average — visible but not dominant
D	≥ 20	Weak — significant gaps
F	< 20	Critical — barely visible

Content Quality Grades

Grade	Score	Action
A	≥ 90	Auto‑deploy ready
B	≥ 75	Deploy with minor review
C	≥ 60	Needs editing before deploy
D	≥ 45	Major rewrite required
F	< 45	Regenerate from scratch

Scoring & Methodology