Soma · live result

We can watch a brain watch your ad.

Upload a cut. We read the viewer's second-by-second cortical response with a brain-encoding model — then tell you which edit holds attention, and the exact second each one loses the room. Below is a real run, on a real ad. No panel. No survey. From the file.

Act I · one ad, second by second

The brain doesn't watch evenly.

Here is one 30-second cut of a welding-gear ad. The line is its attention — dorsal-attention cortex, per second. Watch where it sags: at 0:12, right where the hook hands off to the offer, attention drops 3.3 SD below the cut's own baseline. That's a recut, flagged automatically.

0:00
weak spot, flagged
Act II · which cut wins?

Five edits of the same ad. One timeline. A ranking.

Same creative, five ways to cut it. The model ranks them by attention-hold. Two independent methods agree on the ends: lead with urgency, and the slow story-hook cut loses.

attention head · reads how the ad earns focus, second by second
Act III · where the ad lights up

The same tensor shows us where the ad lands in the brain.

One read gives the whole cortical profile — which brain systems the ad actually drives. This cut leads in the attention and language networks: it earns focus and it carries a message. The profile is the map of where a cut is working, second for second.

Cortical network profile · mean activation
│ = whole-cortex baseline · slate = where the ad lights up

A digital fMRI of your ad.

attention · how the ad holds focus comprehension · language-load lane cortical profile · where it lights up

The model is public. The moat is the outcome flywheel — every ad you run makes us sharper, and no one selling an unverifiable "90% accurate" number can follow.

Real run — TRIBE v2 (Algonauts-2025-winning) → fMRI, fsaverage5 · attention = trained attention head, pooled across the 5 cuts · profile = per-network mean activation.