SEISMO deploys short-form video into real markets and learns cohort behavior dynamics from what comes back. Not synthetic data, not panel surveys: deployment results are the only training signal.
Each round the model records the expected reactions first, then checks them after deployment. The weights update by exactly how much it was wrong.
Creatives tagged by hook, angle and format are deployed by market, platform and time slot. Every condition is logged.
The signal is the order and ratio in which views, completions, likes, comments, saves and shares occur.
Judge which of Propagation, Trust, Habit or Exploration moved this cohort, and assign a confidence level.
We record and seal the next creatives’ reaction ranking first. The gap to reality feeds the next round.
Budgets make views. Cohorts make dynamics. SEISMO estimates the mix of four dynamics from the order and ratio of reactions.
Shares and reposts grow before views do
Share/view ratio · second-degree reach · re-uploads
“Who passes it to whom?”
Saves and rewatches rise before shares
Save/like ratio · profile visits after completion · ingredient and evidence questions
“What do they trust enough to buy?”
The same cohort reacts repeatedly to the same format
Reaction retention on re-exposure · variance by format · time-slot stability
“What refuses to change?”
Unfamiliar angles draw question comments first
Comment/like ratio · share of question comments · comparison requests
“Where does new demand open?”
Repeating rounds in one market optimizes the dynamics weights for that market. The closer we get to the baseline (natural redeploy variance), the less room there is to improve.
What is unverified stays blank. The following is outside the model’s scope.
Every judgment is at cohort level. We do not track or store individual accounts.
Without off-platform sales data, results stop at reaction dynamics. Conversion is reported only when joined with client data.
Dynamics learned in Japan apply to Japan only. Another market is learned again through redeployment.
Differences within redeploy variance are not counted as hits. The first judgment is reported after at least two rounds.
One market · about 300 creatives · 14 days · sealed forecast included. Results delivered in the D+15 report.