Understand the creative. Make the call.
RoastIQ combines five creative diagnostics with source evidence. A separate platform-native attribute model has its own validation record. Use the report to identify what to improve and inform which ad to back; it does not forecast campaign returns.
By the time the data comes back, the budget is already gone.
Most creative testing happens after launch. You spend, you wait, you measure, you cut. RoastIQ moves the verdict to the moment the file is exported.
The five-KPI composite organizes creative diagnostics. Outcome-validation findings belong to the separately identified platform-native attribute model. Keep both distinct from observed campaign results.
- Day 0 · BriefConcept signed offDecision made on taste, not signal.
- Day 7 · Final cutProduction locked$30k–$300k committed.
- Day 8 · Pre-spendRoastIQ scores itMinutes. Verdict before media buy.
- Day 14 · LiveIn-market dataNow you know, too late to change the cut.
From a 60-second file to a scale-or-rebuild call, four layers of analysis.
One pipeline. Four passes. Each layer adds context the next layer needs.
Ingest the creative
Drop in MP4, MOV, JPG, PNG, WebP, up to 60 seconds. We extract frames, audio, transcript, and on-screen text so the same model sees what a viewer sees.
Predict visual attention
A saliency model predicts where eyes go in the first 0–3 seconds. This is a prediction, not measured eye tracking, not proof of where your audience actually looked.
Score the five frozen KPIs
Beat the Skip · Get Noticed · Brand Impact · Sell Proposition · Build Brand. Five dimensions, fixed weights, never reordered. A multimodal model produces diagnostic scores. Any benchmark comparison needs separately checked cohort coverage.
Issue the verdict
The composite maps to one of three calls, Scale, Sharpen, Rebuild, with the evidence trace, confidence label, and benchmark sample size attached. No black box.
Five KPIs. Fixed weights. One number you can act on.
The composite is frozen. We don't let the model choose its own scoring weights, that's how scoreboards become marketing.
Three verdicts. One ladder. No ambiguity.
Apply the rules in order: check Rebuild first, then Scale; all remaining scores are Sharpen. These diagnostic labels guide review, not a media-spend decision.
Boundary examples. Composite 70 with every KPI ≥55: Scale. Composite 70 with one KPI at 54 and the others ≥55: Sharpen. Composite 72 with two KPIs at 44: Rebuild, even though the composite is high. Composite 69 with every KPI ≥55: Sharpen.
Validation belongs to a specific model.
The figures below evaluate the platform-native attribute model against specific public outcomes. They do not validate the five-KPI composite, BuyerLens interviews or attention heatmaps, and they are not evidence of causal campaign lift.
TikTok click-through rate: held-out ρ +0.33. The fitted evaluation reports ρ +0.43 on 691 ads, with a brand-clustered 95% interval [+0.32, +0.53].
YouTube view counts: held-out ρ +0.40. The fitted evaluation reports ρ +0.49 on 397 ads, with a brand-clustered 95% interval [+0.37, +0.58].
These describe our platform-native attribute model, not the five-KPI composite. Cohorts can differ between dated evaluations; do not add their counts into a new validation sample. The historical quintile comparison is not a guaranteed uplift for new ads. Correlation is not causation. Public outcomes can reflect distribution, brand size and other factors alongside creative. Use measured campaign outcomes for business decisions.What RoastIQ is, and what it isn't.
Pre-spend creative diagnostics work because they stay in their lane. We are explicit about ours.
RoastIQ does
- Show a separately evaluated platform-native attribute model for TikTok click-through rate and YouTube view counts. Its validation does not apply to the five-KPI composite.
- Predict visual attention from a saliency model; do not describe it as measured gaze.
- Identify creative attributes and link diagnostic concerns to source evidence for review.
- Benchmark against the calibration pool for your format, platform, and category.
- Apply versioned verdict rules to the stored diagnostic scores. Provider output can vary; a reused result is different from a fresh run.
RoastIQ does not
- Predict in-market sales lift, ROAS, or attributed conversion.
- Measure brand recall, that requires recall-survey methodology.
- Run a real consumer panel (BuyerLens is scenario simulation, not a panel).
- Measure eye tracking, heatmaps are predicted, not observed.
- Replace post-launch measurement. It runs before the spend, not instead of it.
SaliencyLab methodology · Updated
How to use RoastIQ evidence in a creative decision
RoastIQ by SaliencyLab analyzes an image or video in its project context. Its five KPIs organize a creative review; they are modeled diagnostics, not measurements of the people who watched your ad.
What goes into the analysis?
Start with the actual creative, its platform, market, category and campaign objective. Add the brand context and intended audience. The brief describes what you intended to communicate; it is not evidence that the creative successfully communicates it. Images and videos have different evidence: a still image cannot establish pacing, audio clarity or a viewer drop-off time.
What should I read before the score?
Inspect the source frames, visible text and available audio evidence alongside the five KPIs: Beat the Skip, Get Noticed, Brand Impact, Sell Proposition and Build Brand. A heatmap predicts relative visual attention. It is not eye tracking. A suggested edit is a hypothesis tied to that evidence, not proof of an improvement.
How should I interpret benchmark and validation claims?
A comparison is useful only to the extent that its platform, market, category and format fit your creative. Check the stated cohort and coverage before using a percentile. The separate platform-native attribute model has its own validation record; its correlations do not validate the five-KPI composite, BuyerLens answers or your campaign. Meta and unsupported cohorts remain directional.
When should I pause rather than approve the ad?
Pause when key creative evidence is missing, the benchmark is poorly matched, the report contradicts the source, or the recommendation depends on an unsupported claim. Do not turn an unknown into a zero or treat a stable model response as a correct one. Resolve the missing evidence and record the human decision.
How do I turn this into a useful test?
For example, if the product benefit appears only on the closing card, test a version that demonstrates the benefit earlier while preserving the rest of the cut. Record the original, the exact edit, the reasoning and the campaign measure you will compare. A better model score alone is not evidence of better sales, recall or return on spend.
The next ad you ship, scale, sharpen, or rebuild?
Bring a file to a 30-minute call. Get a verdict in minutes. Decide before the media plan goes live.