How AdZhi scores your ads
Every score is grounded in signal processing on the actual audio and video, AI linguistic evaluation calibrated against published research, and structural analysis mapped to proven persuasion frameworks. Here's exactly what feeds each score, how it becomes a forecast we grade against your real results, and where we're still building the evidence base.
Earned accuracy, or none at all
A forecast is only worth the proof behind it. AdZhi holds itself to one rule: accuracy is measured automatically, and shown on every report.
AdZhi only claims predictive power when it has earned it: measured out-of-sample, per account.
- Graded against reality. Each forecast is frozen at analysis time and later compared to the outcome you record: the number it committed to, never one revised after the fact.
- Skill, not just error. Accuracy is reported as how far the forecast beats the naive baseline of guessing your account's average. A forecast no better than that guess claims nothing.
- Calibrated to you. The correction learns from your own realised results; an account-specific calibrator beats the pooled benchmark, and the report names which is in force.
- Your score, re-weighted to you. Beyond the forecast, the AdZhi Score itself re-weights its signals toward what drives realised outcomes. Climbing from the global model to your industry benchmark to your own workspace as the data warrants, and applied only once it beats the more general weighting on your own data, so personalisation can never make your scores less predictive.
- Honest until proven. Until a metric clears the bar on real outcome pairs, it reads “directional”, never “calibrated.”
Three layers of analysis, working together
Not a single model and a number. Each analysis runs acoustic measurement, linguistic evaluation and structural mapping, then combines them.
Acoustic
Measured from the waveform.
- Pitch (F0) and variation over time
- RMS energy envelope + CTA Momentum
- Voice clarity proxy (credibility)
- Speech rate, disfluency detection
- Spectral shape (centroid, rolloff), attention decay model
Linguistic
Evaluated against research.
- Hook strength (LLM-scored vs a persuasion rubric)
- VADER sentiment (peer-reviewed NLP)
- CTA identification + urgency
- Value-proposition clarity
- Self-critique loop for consistency
Structural
Mapped to frameworks.
- Narrative arc → AIDA / PAS
- Act completeness: hook, build, urgency, close
- Timestamped gap detection
- Section-level energy alignment
- Opening pattern classification
From raw signal to score
Every number AdZhi shows sits at one of three levels. Raw measurements feed the proprietary signals; the signals and core components roll up into the three headline scores.
Raw features
Measured directly from the waveform, the frames and the transcript: deterministic and reproducible. A representative core — the pipeline computes more still.
Show all 52 raw featuresHide raw features
Proprietary signals
Coined signals that combine the raw features into a read you can act on. Hover any signal to see the features it’s built from. Composite + heuristic today, graded against your outcomes over time.
Headline scores
The top-line numbers you compare and act on.
Voice IQ vs Trust Index: Voice IQ is the overall delivery score; Trust Index is the specific credibility dimension within it. The AdZhi Score is a weighted composite of six components; everything runs across an 11-stage pipeline, and the 20-dimension Persuasion DNA fingerprint is a separate library-matching vector, not a score.
What each number means
Overall creative effectiveness
A composite across six components: a single-number summary for comparing variants, creators and campaigns. Above 75 is strong; 50-75 points to specific weaknesses; below 50 usually signals a hook or CTA structural issue.
Acoustic delivery quality
Not subjective: signal processing on the actual audio. Credibility from voice clarity (a spectral-flatness proxy); expressiveness from pitch variation (Zuckerman & Driver, 1989, as a proxy for vocal expressiveness). By construction, scores spread across roughly 35-85.
Opening hook strength
LLM analysis of the opening seconds: the mechanism (curiosity, pattern interrupt, bold claim) and how effectively it deploys it. Reports an interval (e.g. 7.2 ± 0.6): the ± is the spread across the self-critique loop’s passes, so it widens when the passes disagree and narrows when they converge — a measure of the model’s own certainty, not a claim about your ROAS.
Narrative arc completeness
Maps the script to AIDA and PAS, identifying which elements are present, weak or missing: “no urgency signal between 00:22 and 00:38”, not “improve the structure.”
How the composite is calculated
A weighted sum of six normalised components (audio path, shown below). For video, the weights rebalance to introduce a ~7% visual-alignment term. Reflects the real implementation in backend/metrics.py.
AdZhi = 100 × (
0.25 × Hook/10
+ 0.20 × CTAQuality/100
+ 0.15 × AttentionRetention
+ 0.15 × Structure/10
+ 0.15 × VoiceIQ/100
+ 0.10 × MisalignmentScore
)Weights are research-informed heuristics (Hook 25%, informed by Meta's published creative guidance; Attention 15% from Binet & Field; Voice IQ 15% from research linking vocal clarity to perceived credibility — measured here as a spectral-flatness proxy, not a true harmonics-to-noise ratio). The global default, not regression outputs. As your account logs real outcomes, AdZhi re-weights them toward the signals that actually drive your results, and applies that only once it beats the global weighting on your own data.
From scores to a forecast
The scores aren't the end of the pipeline. AdZhi maps the signal profile to the outcomes that actually decide spend efficiency, then grades each forecast against your realised results and recalibrates to your account.
Will the first frames hold the scroll
From hook decay, opening energy and sound-off readiness: the ceiling on how many people stay past the first few seconds.
How far the watch lasts
From the attention-stability and emotional-trajectory curves: where energy dips and viewers tend to drop off.
Likely click-through range
A banded estimate (not a false-precision number) from CTA momentum, value clarity and persuasion half-life.
How fast it will wear out
From creative entropy and structural predictability: how quickly the audience is likely to tune it out.
What is, and isn't, proven
We publish what we haven't proven yet. A vendor that hides its gaps is hiding something.
Currently validated
- Acoustic feature extraction: deterministic, reproducible, open-source (librosa).
- Statistical-significance gating: a correlation is only surfaced once it clears p<0.05 on at least 8 matched outcome pairs (and, across many signals, a Benjamini-Hochberg adjustment for multiple comparisons). Below that bar it reads "directional", never "validated".
- Score reproducibility: same file produces identical scores; version-pinned.
- Sentiment (VADER): peer-reviewed (Hutto & Gilbert, AAAI 2014), open-source.
In progress
- Score-to-ROAS correlation: needs per-account performance matched to signals.
- Cross-industry benchmark generalisation: percentiles are directional at current n.
- Pooled global weight regression: the default weights remain research-informed heuristics, not yet regressed across all accounts (per-account re-weighting from your own outcomes is live).
- Hook-type classification accuracy: not yet evaluated against human-labelled ground truth.
- Creative segmentation (k-means): clusters are deterministic; ranking segments by ROAS strengthens as you log outcomes.
We will not claim statistical validation before we have it, or cite correlations we haven't measured. When a component moves from research-calibrated to empirically validated, we'll say so, and name the dataset that validated it.
Start analysing your ads
The first analysis is free: the full breakdown across acoustic, linguistic and structural layers, with timestamped direction.