Methodology
How SaliencyLab Produces Enterprise-Grade Creative Signals
SaliencyLab combines multimodal model outputs, structured scoring rules, and benchmark confidence metadata. Results are designed for decision support before media spend, not post-campaign attribution.
1. Analysis Pipeline
- Input ingestion for image/video creatives and optional transcript context.
- Video default path uses hybrid FFmpeg frame contract, Google Video Intelligence shot/label extraction, and Gemini semantic synthesis.
- Visual and language analysis produces core metric primitives (attention, clarity, branding, emotion, CTA).
- Perception layer generates diagnostics, attention decay, and drop-point explanations.
- Enterprise layer enriches payload with pillar scores, skip prediction, KPI families, and matrix classification.
2. Scoring Framework
RoastIQ
Composite score: Attention 25% + Clarity 20% + Branding 20% + Emotion 20% + CTA 15%.
Enterprise Pillars
Brand, Creative, and Behavioral pillar scores provide executive-level decomposition for faster decision-making.
Skip x Impact Matrix
Beat the Skip and Brand Impact scores map each creative into goal/missed/wasted/avoid opportunity zones.
3. Benchmark Confidence
Benchmark cards include sample count, source type, platform norm version, and confidence level. This prevents over-confidence when data density is low or heuristic estimates are used.
4. Validation and Governance
- Model responses are schema-normalized before persistence.
- Analysis payloads include model version, confidence estimate, and generation timestamp.
- Benchmark metadata is versioned to support reproducibility and audits.
5. Known Limitations
Forecasts are decision-support signals, not guaranteed outcomes. Performance still depends on media buying, audience selection, offer quality, and competitive dynamics.