soma
Soma · a brain read-out for ads

Anyone can make a hundred ads. Nobody knows which one wins.

Soma reads how a brain watches each cut, then builds and edits against that read. Scored before you spend a dollar.

Halden SupplySoma 72
AttentionSurprise
Already running ads through the beta
TikTok
Browserbase
Supercell
MrBeast
NextXI
and more
01The science

Attention isn’t one thing. Different regions do different jobs.

Generic eye-tracking tells you where a gaze lands. Soma reads the cortex itself: two networks, measured separately, once a second for the whole runtime.

Meta's TRIBE v2, benchmarked against real fMRI1st of 263 teams, Algonauts 202520,484 cortical points, once a second
Dorsal attention
IPS · FEF · superior parietal

Decides what the viewer looks at, and holds on.

Ventral · surprise
anterior insula · ACC

Fires on the unexpected. That jolt is what a hook actually is.

02Hook scoring

The first three seconds are everything.

Every ad’s first 3 seconds are scored as its hook, separately from the full video, and worth 40% of the total. The ventral surprise signal is what measures it: a hook works by being unexpected, and the salience network is what registers the unexpected.

Ventral salience network · TRIBE z-scored BOLD, baseline-subtracted (0 = black screen)
Product First · rank 1 of 5
hook window averages −0.030 vs baseline
Weak Open · rank 5 of 5
hook window averages −0.075 vs baseline
Hook · Product First
rank 1 · 40% of the total
Hook · Weak Open
rank 5 · 40% of the total
Same product, same length, same batch. The two cuts separate inside the first three seconds — and the hook score is measured on exactly that window, before the rest of the ad gets a chance to recover it.
03The read-out
TRIBE z-scored BOLD, baseline-subtracted (0 = black screen)
Attention (dorsal)
where the ad holds the eye
Surprise (ventral)
the jolt of something unexpected
Comprehension (association)
where meaning gets built
Weak spot · a sustained stretch where attention sits below this cut’s own median — 2 here
04The batch

Compare across the batch.

Dorsal attention network · TRIBE z-scored BOLD, baseline-subtracted (0 = black screen)
05Comprehension

Did the brand actually land?

A held gaze is worthless if the product never registers. Soma reads the ad's own words: on screen via text recognition, out loud via speech recognition. When the campaign is named and the language cortex confirms it registers, the score rises.

On screen · text recognition
Spoken · speech recognition

The product is named at 0:00, on screen and said out loud. Language cortex confirms the words register.

Comprehension
0/100
Named 2×1 spoken, 1 on screen. Language cortex confirms the words register, so the score lifts. Customers submit the campaign name and goal alongside the video, so Soma knows what to listen for.
Spoken · speech recognition8 lines
  1. 0:00

    Japanese style hat versus an American style hat now the reason that I prefer the Japanese style hats over these American ones is that they're

  2. 0:07

    Super thick and trapped in all the heat but with these Japanese style hats

  3. 0:10

    They're super breathable and make them perfect to wear for the summertime

  4. 0:13

    They also have a completely bendable brim so you can make it fit your face shape perfectly and with these Japanese style five panel hats

  5. 0:19

    They're just so much more versatile you can wear them rock climbing

  6. 0:22

    Camping or even just the dinner and the best part is that if you want to rock this same exact style

  7. 0:26

    You don't have to blow your whole paycheck because these are on a huge sale right now with free shipping

  8. 0:30

    So if you still see that link grab them before they sell out again

On-screen type
15%

of the clip’s seconds carry type

Speech is transcribed verbatim. On-screen type is reported as coverage rather than quoted — recognition on small, moving social type is unreliable enough that the strings would misrepresent the ad, while presence does not.
06Generation

Write a brief. Get back the cut that wins.

Describe the ad in plain language. Soma generates against your brand kit and kills the losers before you see them.

The brief

Halden Supply
Welding gear · direct response
Brand kit
Grotesque · medium
Plain-spokenHands-onNo hype
The internal test0 generated·5 survive
Predicted score, one column per candidate

Generating against the brand kit, then scoring every candidate before anyone sees it.

Top candidate72/100
Product first

Open on the helmet in hand, mid-shot, before a single word.

Hook
71
Hold
82
Comp.
59
Why the generator chose it

Ventral surprise peaks when an unexpected object lands in frame, so the object goes first.

Highest surprise of the five, at all 4 samples in the hook window.

The five that clearedSoma

Same brief, five directions. The spread from 52 to 72 is the difference between an ad that works and one that burns the budget, decided before a dollar is spent. Same 0-100, same model, as the real ads further down this page: the best of those reads 53.

07AI ad editing

Change it in a sentence. It re-scores itself.

Edited in plain language. Every candidate edit is scored and the best cut comes back, from your frames, never regenerated.

Attention holds for 59% of the runtime. It drops from 0:01 to 0:07.

Deal first · base score 62 drags it down carries it
0:00
-3
0:01
-4
0:04
-1
0:06
-2
0:12
0
0:14
+2
0:19
+1
0:22
0
0:23
0
0:24
+1
0:26
-1

The opening 0:010:04 shot is dragging it down. Cut it and the score goes up.

Ask for the editReading the cut…

Splice66+4
Splice out the 1.7–4.9s beat

Lands on the helmet at 1.7s.

measured
Splice65+3
Splice out the 0.0–1.7s beat

Opens on the bench, no establishing shot.

measured
Reorder64+2
Move the product beat to the front

Every frame kept; the helmet leads.

predicted
Soma score
62+462

11 shots, each scored for what it costs or earns.

After the splice Before
0:00
0:01
0:04
0:06
0:12
0:14
0:19
0:22
0:23
0:24
0:26
Runtime
0:290:25
Dead air
6.0s2.8s
Shots
1110
Brand named
Static, from this campaign

Same offer, static. Headline top-left.

Add the offer as text at 0:04.Placed under the safe area, held 2.1s · comprehension +6, attention unchanged.
Give me three openings that beat this one.Three cuts generated, scored, ranked. Top cut 68.
08The database

Every ad that comes through makes the model sharper.

The read-out is the product; the database behind it is what a public model cannot buy.

Ads in the demo corpus
0
from report.corpus.ads
Advertiser entries indexed
0
from report.corpus.advertisers
53
51
50
42
40
40
36
soma

Stop guessing which ad wins.

Send us your batch. We rank it, diagnose it, and give you back the cut that wins.

upload mp4, turnaround in hours

usesoma.work