Methodology
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 established 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 towards what tracks realised outcomes. It climbs from the global model to your industry benchmark to your own workspace as the data warrants, and a new weighting applies only once it beats the more general one 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 measurement to score
One chain, read top to bottom: 52 measurements are taken from the file, each belongs to one of 8 sections, and the sections add up to the one AdZhi Score. Open any level to see what is in it.
Measurements
Measured directly from the waveform, the frames and the transcript: deterministic and reproducible. Every feature has one home section, where the report shows it. 22 of them move a section’s value (18 computed into it, the rest as context for an AI rating); the others are evidence shown under their section. 47 feed the learning model, which tests each one against the results you log and learns which matter for your ads; the other 5 are held out, and each says why.
Show all 52 measurements and their sectionsHide measurements
- Pitch (F0)Voice IQ: context for Hook & Structure rating; learning model
- Pitch variationVoice IQ: Voice IQ · Attention · Alignment; context for Hook & Structure rating; learning model
- Voice clarityVoice IQ: Voice IQ · Alignment; context for Structure rating; learning model
- RMS energyVoice IQ: Voice IQ · Call to action · Attention · Alignment; context for Hook rating; learning model
- Energy dynamic rangeStructure: context for Structure rating; learning model
- Loudness (dBFS)Visuals & sound: evidence · learning model
- Spectral centroidVoice IQ: Voice IQ; context for Structure rating; learning model
- Zero-crossing rateVoice IQ: evidence
- MFCCVoice IQ: evidence
- ShimmerVoice IQ: Voice IQ; learning model
- Tempo & beatVisuals & sound: evidence · learning model
- Musical key & modeVisuals & sound: evidence · learning model
- Music / voice balanceVisuals & sound: evidence · learning model
- Pause & silence ratioAttention: evidence · learning model · Persuasion DNA
- Dramatic pausesAttention: evidence · learning model
- Peak-energy timingStructure: evidence · learning model
- Scene-cut rateVisuals & sound: evidence · learning model
- Average shot lengthVisuals & sound: evidence · learning model
- BrightnessVisuals & sound: evidence · learning model · Visual quality
- ContrastVisuals & sound: evidence · learning model · Visual quality
- SaturationVisuals & sound: evidence · learning model
- Colour temperatureVisuals & sound: evidence · learning model
- Dominant paletteVisuals & sound: evidence
- SharpnessVisuals & sound: evidence · learning model · Visual quality
- Motion intensityVisuals & sound: evidence · learning model
- Camera stabilityVisuals & sound: evidence · learning model
- Visual entropyVisuals & sound: evidence · learning model
- Face visibility %Visuals & sound: Visuals & sound; learning model
- Eye-contact %Visuals & sound: Visuals & sound; learning model
- Speaker countVisuals & sound: evidence · learning model
- Caption coverageVisuals & sound: Visuals & sound; learning model
- On-screen text & CTAVisuals & sound: Visuals & sound; learning model
- Camera & edit styleVisuals & sound: evidence
- Aspect ratioVisuals & sound: evidence · learning model
- Speech rate (wpm)Alignment: Attention · Alignment; learning model
- Pacing varianceAttention: evidence · learning model
- Disfluency rateVoice IQ: Voice IQ · Attention · Alignment; context for Hook & Structure rating; learning model
- Hook delivery timeHook: context for Hook rating; learning model
- CTA timingsCall to action: evidence · learning model · CTA Momentum
- Sentiment (VADER)Structure: evidence · learning model · Emotional Compression
- Sentiment arcStructure: evidence · learning model · Persuasion DNA
- Primary toneHook: context for Hook rating
- Power wordsCall to action: Call to action · Offer; learning model
- Urgency cuesOffer: Call to action · Offer; learning model
- Social proofOffer: Call to action; learning model
- Cialdini principlesOffer: Offer; learning model
- QuestionsAttention: Attention; learning model
- Benefit vs featureCall to action: Call to action; learning model
- Claims & compliance riskOffer: evidence · learning model
- Brand-safety flagsOffer: evidence · learning model
- Named entitiesAttention: Attention; learning model
- ReadabilityAttention: evidence · learning model · Cognitive Load Score
Sections
Each section answers one question about the ad and is scored 0–100 from its inputs (7 for an audio-only ad, which has no picture). Open a section for exactly how. Every input is marked Measured (computed directly from the audio, frames or transcript) or AI read (a model’s judgement against a fixed rubric).
HookDoes the opening stop the scroll?21% · audio 23%
- Hook rating 0–10 AI read
How strongly the first seconds earn attention, rated against a fixed rubric from the transcript, timing and opening delivery. The section’s base value (rating ÷ 10). - Hook type category AI read
The kind of opening: question, bold claim, story, shock, relatable problem, curiosity gap, social proof, product demo or other. Bold claim, shock, curiosity gap or question adds 8%.
Section score: Hook rating ÷ 10, then ×1.08 for a bold-claim, shock, curiosity-gap or question opening (capped at 100%). If it can’t be read: 40%, marked as estimated.
Call to actionIs the ask clear and delivered with intent?17% · audio 18%
- CTA intensity 0–100 Measured
Persuasion intensity at the moment of the ask: vocal emphasis, energy against the ad’s average and urgency language. The section’s base value (intensity ÷ 100). - Calls to action count Measured
Explicit calls to action found in the script. +5 points each, up to +15.
Section score: CTA intensity ÷ 100, plus 5 points per explicit call to action (up to 15), capped at 100%. If it can’t be read: 30%, marked as estimated.
OfferDoes the ad give people a reason to act now?8% · audio 8%
- High-impact levers 0–3 Measured
Whether the script or the on-screen text offers a discount or price incentive, urgency or scarcity, and risk-reversal (a guarantee, refund or free trial). 25 points each. - Supporting levers 0–2 Measured
Social proof (reviews, ratings, authority or community) and a free element or bonus, said or shown on screen. 12.5 points each.
Section score: Discount, urgency and risk-reversal (a guarantee or free trial) add 25 points each; social proof and a free bonus add 12.5 each (maximum 100%). Said or shown on screen counts the same. The clear ask itself is scored in Call to action, not here. If it can’t be read: 40%, marked as estimated.
AttentionHow long does the ad hold a viewer?13% · audio 14%
- Attention half-life seconds Measured
The second at which modelled viewer attention first falls below 50%. The model decays over time, faster for a monotone, fast or hesitant read, and recovers at pitch changes, energy spikes, questions and named brands or products. Sets the drop-off risk, and the bonus for holding late. - Drop-off risk low · medium · high · critical Measured
Derived from the half-life: under 5s critical, under 8s high, under 14s medium, otherwise low. The section’s base value.
Section score: Drop-off risk sets the base (low 100%, medium 65%, high 35%, critical 10%); ×1.1 if attention never falls below half, ×1.05 if it holds past 70% of the ad. If it can’t be read: 50%, marked as estimated.
StructureDoes the ad build to its point?13% · audio 14%
- Structure rating 0–10 AI read
How well the script moves through its acts (hook, problem, solution, proof, ask), rated against a fixed rubric. The section’s base value (rating ÷ 10). - Missing elements list AI read
Which essential acts are absent. −10 points for each missing hook, solution or call to action.
Section score: Structure rating ÷ 10, minus 10 points for each missing hook, solution or call to action (floor 0). If it can’t be read: 40%, marked as estimated.
Voice IQDoes the delivery sound credible and engaged?13% · audio 14%
- Credibility 0–100 Measured
Vocal clarity, a spectral-flatness proxy linked in research to perceived credibility. It measures the sound, not trustworthiness itself. 30% of Voice IQ. - Expressiveness 0–100 Measured
Pitch range and inflection: how far the delivery is from monotone. 25% of Voice IQ. - Hook energy 0–100 Measured
Vocal energy in the first three seconds, against the ad’s average. 20% of Voice IQ. - Fluency 0–100 Measured
The absence of filler words and stumbles. 25% of Voice IQ.
Section score: The weighted average of its four signals (30 / 25 / 20 / 25), over those that could be measured, ÷ 100. If it can’t be read: 50%, marked as estimated.
AlignmentDoes the voice match what the script says?9% · audio 9%
- Critical mismatches count Measured
Moments where the delivery contradicts the words, such as a flat voice on the key claim. −20 points each. - Major mismatches count Measured
Smaller voice–script mismatches. −8 points each.
Section score: 100%, minus 20 points per critical and 8 per major mismatch (floor 0). If it can’t be read: 50%, marked as estimated.
Visuals & soundDo the picture and sound carry the message, including with the sound off?6% · video only
- Visual–audio alignment 0–100 Measured
Whether what is on screen supports what is being said at the same moment. 70% of the section. - Sound-off readiness 0–100 Measured
How much of the message survives with the sound off: captions and on-screen text. 30% of the section.
Section score: 70% visual–audio alignment plus 30% sound-off readiness, ÷ 100. Not scored for audio-only ads. If it can’t be read: 0%, marked as estimated.
AdZhi Score · 0–100
The sum of each section’s value (0–100%) times its weight. Weights sum to 100%; an audio-only ad has no visuals & sound section and uses its own weights. As an account logs results, the weights can be re-fitted to it, but only once that beats the global weights on its own held-out outcomes. Graded A from 80, B from 68, C from 55, D from 40, F below.
Diagnostics: shown in reports, not in the score
Named readings that explain an ad from other angles. None of them is added into the AdZhi Score, and none is a second score to chase. Each sits beside the section it helps explain, so you can see why that section scored as it did.
- Persuasion Half-Life seconds
- Where persuasion intensity falls below half its peak. Helps explain the Attention section.
- Trust Index 0–100
- A credibility read from voice clarity, pitch monotony, disfluency and vocal warmth. Computed separately from Voice IQ, from overlapping inputs. Helps explain the Voice IQ section.
- CTA Momentum 0–100
- Persuasion intensity and vocal energy at the call to action, against the ad’s peak. Helps explain the Call to action section.
- Hook Decay Rate rate
- How fast persuasion and modelled attention fall across the opening. Helps explain the Hook section.
- Emotional Compression 0–100
- How much emotional range the script packs in, and how fast it moves. Helps explain the Structure section.
- Attention Stability 0–100
- How steady the modelled attention curve is. Helps explain the Attention section.
- Cognitive Load Score 0–100
- How hard the ad is to follow: readability, sentence length, speech rate, specificity, background noise and disfluency. Helps explain the Attention section.
- Creative Entropy 0–100
- How varied the language is: vocabulary diversity, phrase repetition and sentence-length variety. Helps explain the Structure section.
- Persuasion DNA 20-dimension profile
- A fingerprint of the ad’s persuasion pattern, used to find similar ads. A profile, not a score.
What each number means
AdZhi Score: overall creative effectiveness
The weighted sum of 8 section scores (7 for an audio-only ad), built from 18 signals: one number for comparing variants, creators and campaigns. Graded A from 80, B from 68, C from 55, D from 40, F below. The grade tells you where an ad sits; the section with the most headroom tells you what to fix.
Voice IQ: vocal delivery
Signal processing on the actual audio, from four parts: credibility 30% (voice clarity, a spectral-flatness proxy, with warmth and steadiness), expressiveness 25% (pitch range and inflection), hook energy 20% (vocal energy in the first three seconds) and fluency 25% (the absence of fillers and stumbles). Parts that cannot be measured are left out and the rest re-weighted. One of the eight sections of the AdZhi Score.
Hook: opening strength
One of the sections of the AdZhi Score. The 0–10 rating described here is its main input; the report shows the section itself out of 100. An AI rating of the opening against a fixed persuasion rubric: the mechanism (curiosity, pattern interrupt, bold claim) and how well it lands, given the opening words, timing, tone, pitch and energy. Reports an interval (e.g. 7.2 ± 0.6): the ± is the spread across the self-critique passes, a measure of the model’s own certainty, not a claim about your ROAS. The base of the Hook section.
Structure: narrative arc
One of the sections of the AdZhi Score, not a separate score. Maps the script to AIDA and PAS, identifying which acts are present, weak or missing: “no urgency signal between 0:22 and 0:38”, not “improve the structure”. The base of the Structure section; each missing hook, solution or call to action costs 10 points.
How well the score predicts real results
Every day we check the score against the results advertisers have logged or linked, and publish what we find here, including when there isn’t enough data to say.
Loading today’s figures…
How the score is calculated
Each section scores 0–100; the AdZhi Score is their weighted sum. An ad with visuals has 8 sections; an audio-only ad has no visuals & sound section and uses its own weights. These are the weights in the code.
| Section | With visuals | Audio only |
|---|---|---|
| Hook | 21% | 23% |
| Call to action | 17% | 18% |
| Offer | 8% | 8% |
| Attention | 13% | 14% |
| Structure | 13% | 14% |
| Voice IQ | 13% | 14% |
| Alignment | 9% | 9% |
| Visuals & sound | 6% | — |
Weights are research-informed heuristics (the hook weighted highest, informed by Meta’s published creative guidance; attention from Binet & Field; voice from research linking vocal clarity to perceived credibility, measured here as a spectral-flatness proxy, not a true harmonics-to-noise ratio). They are the global default, not regression outputs. As your account logs real outcomes, AdZhi can re-weight the sections towards what tracks your results, and applies that only once it beats the global weighting on your own held-out data.
When a section can’t be read (no audio, no transcript, a rating call that failed) it takes the neutral value shown in its definition rather than being dropped, so scores stay comparable across ads. Your report names which sections were actually measured, an estimated section is never the advice, and the ± beside the score pays for the rest: every filled section widens the interval by the spread of a flat prior over its whole range, weighted by how much of the score it carries. A score built from fewer real readings is never shown as more certain than one built from more. Below two real core signals the score is marked provisional and is never used to train anything.
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: the 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.