Randomise the eligible, not the arrivals
Control and treatment have passed the same evidence and eligibility gates.
PROOF
Built so the number survives the CFO.
Randomise only the eligible population. Preserve the control. Report gross profit after response and incentive costs—with experience guardrails beside it.
treatment − holdout − cost
A permanent holdout, assigned before anyone is helped.
Margin, net of what the help cost
Discounts, service cost and intervention cost are charged before value is reported.
Interruption is charged on the same page
Dismissal, bounce, frustration and complaints sit beside economic value.
Losing arms retire themselves
Underperforming arms return to silence after a pre-agreed learning window.
ILLUSTRATIVE RESULT
One journey, fourteen days, two groups.
An example of the reporting contract—not a lift claim.
Holdout
5.8% conversion
277 ELIGIBLE VISITORSTreated
9.7% conversion
1,109 ELIGIBLE VISITORSIncremental gross profit
€11,240
95% CI €6.8K–€15.7KReturn on total cost
5.9×
ILLUSTRATIVE DATASix questions, in the order they have to be answered.
01 · Technical validity
Did the system receive, decide, render and attribute correctly?
02 · Decision quality
Were the selected moments commercially plausible?
03 · Controlled outcome
Did treatment outperform the preserved control?
04 · Economic value
Was the gain valuable after costs?
05 · Experience safety
Did the visitor guardrails hold?
06 · Repeatability
Does the pattern hold across journeys or customers?
What we are allowed to say, at each stage of evidence.
Prototype
Functional rules engine using illustrative deterministic data.
NOT PRODUCTION INTELLIGENCE OR PROVEN LIFTShadow
Observed eligible moments and recorded would-be decisions.
NOT RECOVERED REVENUEControlled pilot
Treatment and control results for the agreed scope and window.
NOT A UNIVERSAL CAUSAL EFFECTReplicated proof
Repeated incremental value under stated conditions.
NOT A CLAIM FOR EVERY SITE OR VISITORSHADOW BEFORE SIGNAL
See the maths on your own journey.
Send us one signup, checkout, application, or upgrade flow. We’ll identify where customers hesitate or drop off, then show what to test.