Cortex scores paid-social video creative by comparing a candidate ad with your brand's own winners and losers.
Cortex analyzes each ad in two layers: Gemini extracts structured creative moments and psychology, while the fMRI-trained encoder predicts brain response. The score combines whole-video neural fit, moment-level neural fit, anchor-window fit, and calibrated creative-profile fit against the brand's winners and losers.
Predicted response, not live measurement
Cortex uses model-predicted brain-response patterns. A scoring run is not live viewer measurement and does not recruit a new panel for each candidate. The system predicts response patterns from the video and compares those patterns with your brand's reference bank.
This matters for interpretation: Cortex is a creative decision tool, not a claim that a specific audience was measured during the scoring run.
Reference-bank contrast
The reference library defines the brand standard. Winners and losers are both required because Cortex needs contrast:
- Winners define what strong brand-specific response looks like.
- Losers define what weak or off-pattern response looks like.
The signature is built from eligible references. Labels, review status, and weights influence what the next signature learns.
Whole-video and element-level comparison
Cortex looks at the whole video and, when timeline coverage is available, comparable moments inside the video.
Whole-video comparison asks whether the overall candidate pattern resembles winners or losers.
Element-level comparison asks whether a specific kind of moment behaves like equivalent moments in winners. For example, the candidate hook is compared with winner hooks, the CTA with winner CTAs, and transformation moments with winner transformation moments when enough examples exist.
Element-level evidence is often more actionable because creative teams can edit the exact moment that underperformed.
Product scores
The result page separates the product decision score from diagnostics:
- Final score: the headline decision score.
- Brand winner-likeness: how much the candidate resembles the brand's winners rather than losers.
- General attention: how strongly the candidate pulls attention, independent of the brand bank.
- Video score: the whole-video brand match.
- Element score: the scored-moment brand match.
- Confidence: how much trust to place in the result.
The final score blends brand fit and attention. Diagnostics explain why.
Creative analysis and recommendations
Gemini creative analysis turns scores, structured moments, psychology, and reference comparisons into recommendations. It should be read as an explanation of the score payload, not as a separate model overriding the numeric result.
If recommendations are unavailable but numeric scores are present, the scoring can still be used.
What improves accuracy
The strongest inputs are not secret settings. They are good customer data:
- Clean winner and loser labels.
- Enough examples on both sides.
- Balanced reference counts.
- Reference weights that downweight noisy examples.
- Rebuilt signatures after reference review.
- Calibration and diagnostics reviewed before high-volume scoring.
The app surfaces these checks so teams can improve the bank over time.