The four-step decision loop
Every creative goes through the same decision process. The question is whether the process happens with structured data or without it. Running it with structured data, before media budget is committed, is what pre-spend creative intelligence means in practice.
- Score the creative. How does this ad perform against market benchmarks?
- Understand resistance. If a KPI is weak, which buyer segments object and why?
- Apply the fix. Change the specific element that is causing the problem.
- Rerun the diagnostic. Confirm the fix landed and the score improved.
In SaliencyLab, all four steps happen inside one study. The context flows automatically between them.
Step 1: RoastIQ scores the creative
Upload the video. RoastIQ returns the 5-KPI diagnostic in under three minutes:
- Beat the Skip: 81
- Get Noticed: 76
- Brand Impact: 68
- Sell Proposition: 53
- Build Brand: 76
- Composite: 71, verdict: Sharpen
That verdict is worth pausing on, because the composite is above 70 and the creative still did not earn a Scale. Scale requires a composite of 70 or above and no individual KPI below 55. Sell Proposition at 53 trips the floor rule, so the creative lands in Sharpen instead. A strong average with one structural hole is not a creative that is ready to run.
The evidence timeline shows the product pour arrives at 18 seconds. The visual attention data shows high predicted engagement in the first 3 seconds, then a drop at second 8.
The weakest KPI is Sell Proposition at 53. The next step is to understand why.
Step 2: BuyerLens explains the resistance
Click "Open BuyerLens" from the RoastIQ tab. The study planner opens with three nodes, Audience, Participants, and Goal, all pre-filled from the RoastIQ context.
The audience is already set to FMCG buyers on Instagram Reels, UK market. The goal reads: understand why Sell Proposition scored 53.
Select 3 buyer personas. Click "Run interviews".
Three synthetic buyers evaluate the creative. The objection map reveals a shared tension: the value proposition arrives at 18 seconds, well after the skip decision window. All three personas flag the same structural timing issue.
Convergence across all three is the signal to pay attention to here. One persona raising a timing objection is noise. Three independently landing on the same one points at something structural in the cut.
Hold the output at the right altitude: synthetic interviews are scenario simulation, not consumer evidence. Convergence gives you a strong edit hypothesis, it does not replace what real buyers and live campaign data will tell you. For when this second layer is worth running at all, see when to run synthetic buyers after a score.
Step 3: The brief directs the fix
The Brief tab synthesizes both layers, RoastIQ market data and BuyerLens objection data, into an actionable edit direction:
What to fix:
- Move product demonstration to seconds 3 to 5
- Simplify message hierarchy, one ask per scene
What to keep:
- Opening hook (Beat the Skip scored 81)
- Brand cue timing at 3 seconds
- Community celebration sequence (Build Brand scored 76)
Priority edit order:
- P0: Move product pour to seconds 3 to 5 (Sell Proposition +20 to 25 expected)
- P1: One message per scene (Brand Impact +5 to 10 expected)
Step 4: Rerun confirms the fix
Upload the new cut as a new study in the same project. Run RoastIQ again. Compare the scores side by side.
If Sell Proposition moves from 53 into the mid 70s, it clears the 55 floor, the composite rises to roughly 75, and the verdict changes from Sharpen to Scale. If it does not move, the study planner shows what to try next.
Note what the rerun does and does not tell you. It confirms the creative now scores better on the predicted engagement and click-intent signals the model was trained on. Those signals carry held-out, out-of-sample validation of Spearman +0.30 to +0.32 against public outcomes as of May 2026; the methodology page has the detail. It is not a measurement of what the ad will earn in market. That still comes from your performance data after launch.
Why one workspace matters
The alternative is scattered: run an AI score in one tool, do interviews in another, write the brief in a doc, upload the new cut somewhere else. Context is lost at every handoff.
In SaliencyLab, the context is preserved. The RoastIQ evidence feeds BuyerLens. BuyerLens objections feed the brief. The brief directs the edit. The rerun confirms the fix. One study, one project, one decision loop.