SaliencyLab exists to explain the early perceptual risk while the asset is still movable, before media spend turns a weak opening into a six-figure mistake, and before lift studies measure what was already lost in the first second.
The first-second problem
By the time recall, lift, and sentiment studies report back, the decision has already cost money. Three things go wrong in the first second, and none of them are what downstream research is built to catch.
A weak opening loses the viewer before recall, lift, or sentiment studies ever get the chance to measure an outcome. The opening is worth inspecting before launch; actual viewing and skip behavior depend on the audience and placement.
Lift, recall, and brand-tracking studies can tell a team whether a message landed. They are not designed to explain the first-second perceptual filtering that decided whether the viewer ever got there.
That is the gap SaliencyLab is built to serve: the moment before spend, before validation, when the asset is still cheap to change and a single edit can move the entire campaign trajectory.
Complementary, not competing
SaliencyLab does not replace lift studies, brand trackers, or human panels. It runs earlier, and makes the deeper research more valuable by routing only the surviving routes into it.
Traditional research
Confirms memory, lift, and recall after exposure. Tells the team what happened.
SaliencyLab
Diagnoses the asset before or alongside validation so the team can adjust earlier.
Together
Use SaliencyLab upstream to filter. Escalate into deeper research only when the route deserves it.
| Dimension | Traditional research | SaliencyLab |
|---|---|---|
| Position in workflow | After final cut, before launch, validation | Upstream Before or during creative development |
| What it answers | Did this ad work? (outcome) | Why is it likely to work, or not? (diagnosis) |
| Time per read | 2–6 weeks | 60–90 seconds (images) · typically 3–4 minutes (video), up to 9 minutes |
| Cost per asset | $30k–$150k per study | Included within Pro plan limits |
| Validation target | Sales lift, recall, attributed conversion | Public TikTok CTR and YouTube views, for the platform-native attribute model only |
| Held-out signal | Reported per study | Separate platform-native attribute model: Spearman ρ +0.33–+0.40 (TikTok CTR, YouTube views). These figures do not validate the five-KPI diagnostics. |
How it changes the decision
The creative review meeting today is mostly subjective: gut calls, opinions in a Figma comment, an executive overriding a brand planner. SaliencyLab adds a defensible upstream read, without locking the team out of judgment.
Step 01 · Concept
The team has an asset, a debate, and a launch timeline. Today the decision is made on taste, with two or three voices dominating the room.
Step 02 · Diagnose
Score the asset, read the benchmark frame, and pressure-test the route. The score is a model prediction, not a verdict, but it removes the floor of the debate.
Step 03 · Validate
Escalate into lift, recall, or human-panel work only when the question genuinely requires that depth, on stronger routes, not weak ones.
Truth-first
The validation envelope is the product. Overclaiming kills it. Here is the line, drawn in public.
✓ What is validated, and for which model
✗ What we do not claim
Historical validation snapshot, 2026-08-25, for the separate platform-native attribute model; not five-KPI validation. Meta-Feed cohort data is currently sparse for brand ads, so SaliencyLab returns directional defaults there, not held-out scores. Intervals are brand-clustered: ads from one brand are not independent observations, and treating them as such would make these look tighter than they are. Held-out figures are what to expect on an ad the model has not seen. These describe our platform-native attribute model, not the five-KPI composite shown beside it. Full methodology in /methodology.
We are not trying to measure every possible outcome.
We are trying to explain the creative before the wrong decision gets locked in.
Common objections, answered
If a question is not here, write us, we keep this list honest as the validation envelope grows.