Skip to main content
RoastIQBuyerLensHugoPricingBlogAbout
Client accessDiscuss your campaign
Home/ Hubs/ Creative Analysis
Hub · 6 guides + case gallery · updated 18 May 2026

Creative analysis,
written down.

Six guides. One methodology. Read alongside archived ad diagnostics from TikTok, Meta and YouTube. Read them in order to learn the system, or jump to the chapter you need before your next edit.

5 diagnostic KPIs Separate model-validation cohorts ρ +0.40 YouTube views · held-out OOS Directional outside supported cohorts
Open the field guide → Score an ad first
Oussama Nakhil portrait
Oussama Nakhil, Founder
Multiple years buyer-side: NielsenIQ insights, then L'Oréal Groupe in global marketing insights, where I managed Kantar as a vendor. Built SaliencyLab to replace the slow, expensive consumer research loop with something pre-spend, public-data, and honest about its ceiling.
5
KPIs frozen · never renamed
Minutes
to a verdict, per ad
3
platforms discussed; Meta remains directional
// Read this first

"Most creative advice is rumour. This field guide is the opposite, every rule is anchored to a number, every number is anchored to an ad in a benchmark we keep growing. Read it once and you will leave with a methodology you can defend in a room of skeptical buyers."

The field guide

Six guides. One methodology.

Featured · Case gallery

Nine real ads, nine archived diagnostics.

Nine real ads across eight brands: Nike, Coca-Cola, BMW, L'Oréal Paris, Netflix, CeraVe, Nivea and Booking.com. Original model outputs are preserved as directional examples, with their evidence limits.

9 ads 8 brands 3 platforms 2 archived label bands
Open the gallery →
01 / 06
Beat the Skip
Opening creative review worksheet
Inspect the first creative moments and record what becomes understandable. Turn the review into a testable edit.
14 min · CraftRead →
Opening frame illustration
02 / 06
Hook timing
Hook rate vs thumb-stop: check the metric definition
Record the exact numerator, denominator and viewing threshold before comparing reports. Creative timing and model diagnostics are not viewer retention.
8 min · Beat the SkipRead →
Hook timing illustration
03 / 06
Brand impact
Brand cue placement: review the full sequence
The single best predictor of Brand Impact in our 2,000+ sample is not how often the logo appears, it is when the first distinctive asset hits the frame. The curve, the audit, and the category exceptions.
6 min · Brand ImpactRead →
Brand cue placement illustration
04 / 06
Scoring
Reading a RoastIQ verdict end to end
An authored example of diagnostic scores, directional gates and attributed human choices. It is not a saved client analysis or measured outcome.
12 min · ScoringRead →
Score interpretation illustration
05 / 06
Methodology
Synthetic panels vs live audience tests: a decision tree
What a 6–12-persona simulation can explore, and when you need a real audience. The decision tree, the Kantar comparison, and the honest limits of LLM-based personas.
10 min · ValidationRead →
Synthetic vs live test illustration
06 / 06
Methodology
What "benchmarked" really means
Review cohort scope, source provenance and sample limits before interpreting a diagnostic score. Validation and repeated-run stability remain separate.
9 min · ValidationRead →
Benchmark credibility illustration
// The methodology, in one page

If you only read this section.

RoastIQ combines five model diagnostics into a composite. The evidence gate determines whether a recommendation is supported; Meta and unsupported cohorts remain Directional only. Validation of the separate platform-native attribute model does not establish validation of this composite.

The five KPIs (frozen).

Composite is a weighted sum, not an average. Beat the Skip carries the most weight because losing the first 2 seconds nullifies everything downstream.

25%
Beat the Skip
20%
Get Noticed
20%
Brand Impact
20%
Sell Proposition
15%
Build Brand

Read the verdict with its evidence.

Check the saved evidence gate first. Meta, unsupported cohorts and missing lineage remain Directional only. An eligible recommendation is a review input, not a spending instruction.

≥70
SCALEReview the supported strengths and define a proportionate test.
55-69
SHARPENInspect the specific weakness and propose an edit.
<55
REBUILDRevisit the brief where the evidence supports a structural change.
// Honesty pact

What we can claim. What we will not.

Half of being trustworthy is being clear about your ceiling. The full version of this is in "What benchmarked really means". The short version is here.

// Separate platform-native attribute model

  • The platform-native attribute model has a separate TikTok CTR evaluation.
  • The platform-native attribute model has a separate YouTube view-count evaluation; format coverage needs its own check.
  • Held-out Spearman ρ +0.33 for TikTok CTR and +0.40 for YouTube views, for the platform-native attribute model. These are not five-KPI composite validation estimates.
  • The platform-native attribute model recorded 6.5× pool-wide top-versus-bottom quintile lift. This is a pool association, not a promised campaign lift or five-KPI validation.
  • Actual processing time and cost depend on the run. Missing model, video or review costs remain unknown until recorded.

// What SaliencyLab will not claim

  • Does not predict sales lift, ROAS, attributed conversion, or in-market brand recall.
  • Does not reproduce survey-based recall methodologies, we use behavioral proxies, not consumer interviews.
  • Meta-Feed scores are directional defaults, brand-ad outcome data is sparse, no held-out validation yet.
  • Heatmaps are predicted visual attention, not measured eye tracking.
  • Attribute detections can be wrong. A completed human-coder reliability study is not established; inspect the source evidence and review uncertain labels.