Synthetic data. Fictional product. Simulated outcomes.
Trial-to-buyer/Overview
Snapshot · 22 Nov 2026Browser-only demo

SIMULATION ON SYNTHETIC DATA / CMS-2

Cost per incremental buyer, with the sample behind it.

An adaptive action × timing × touchpoint, measured against people who receive no follow-up.

Adaptive · cost / incremental buyer

45,198₫

Ratio of all 24-run totals · inspect both components

Adaptive conversion

26.9%

Fixed conversion: 28.3% · primary seed

Synthetic trialists

12,000

11697 eligible · 3 disjoint experiment arms

Outreach suppressed

2,401

Consent, purpose or frequency gate · existing buyers counted separately

Conversion and cost trade-off

Fixed: 28.3% conversion, 3,185,282₫ follow-up / 1,000. Adaptive: 26.9% conversion, 809,918₫ follow-up / 1,000.

Adaptive conversion is lower across the reported repeated-seed range. Adaptive uses less follow-up cost per 1,000 trialists across the reported range.

Sample size decides whether a 10–12 week pilot can prove anything.

A programme may fall back to redemptions because it never recruited enough trialists to measure incremental buyers.

Main read · planned sample

3,897+ trialists per arm

Planner requirement: 3,463 / arm, including 15% attrition.

Primary seed · adaptive versus hold-out
26.9% vs 24.8%
Estimated additional buyers · primary seed
79.7 · 95% interval 4 to 155.5
All 24 seeds · additional buyers / 1,000
18 · range -2.2 to 38.5

This primary seed resolves positive adaptive lift over hold-out; compare conversion and cost directly against fixed before deciding.

Underpowered read · same assumptions

972+ trialists per arm

Planner requirement: 3,463 / arm, including 15% attrition.

Primary seed · adaptive versus hold-out
26.1% vs 23.3%
Estimated additional buyers · primary seed
27.7 · 95% interval -9.5 to 64.9
All 24 seeds · additional buyers / 1,000
18.3 · range -19.3 to 41.8

This primary seed does not resolve positive adaptive lift over hold-out; more sample alone does not fix a weak policy or small effect.

Both sizes use the same response model and unit costs, equal 10,000₫ redemption budget per assigned trialist, overlapping intent signals and weak equal priors. The plan sizes conversion against hold-out, not a difference between two unstable cost ratios.

Did each programme actually run?

Equal 10,000₫ coupon budget per assigned trialist, charged on redemption. No simulator reservations.

ArmTrialistsSurveys sentCoupons issued / redeemedNothing sentCoupon budget used / allocated
holdout390000 / 03900Budget-suppressed: 00₫ / 39,000,000₫
fixed39008703523 / 1012377Budget-suppressed: 010,120,000₫ / 39,000,000₫
adaptive38971551144 / 299857Budget-suppressed: 01,820,000₫ / 38,970,000₫

Fixed: survey when incomplete and a survey-capable channel is consented, followed by a coupon. Voucher-only permission permits the coupon but cannot carry the questionnaire. Recent frequency caps delay this sequence within the 30-day window; withdrawal and missing purpose/channel permission suppress it. Adaptive still chooses one next step, not a full repeated journey.

Stop treating every trialist the same.

A coupon is useful only when it changes a purchase.

◌Trial signals

Pack size, intent,
channel permission

◎One next step

Action × timing
× touchpoint

▣Message gate

Rules + a second
model review

The useful insight

The engine must learn which action and timing are worth their cost. Intent is noisy, some surveys are absent, and cold-start exploration can lose buyers. Keep a random hold-out and compare conversion as well as cost.

The same question, compared fairly.

Seeded simulator · six feedback waves

Hold-out24.8%
Fixed28.3%
Adaptive26.9%

Adaptive conversion is lower across the reported repeated-seed range. Adaptive uses less follow-up cost per 1,000 trialists across the reported range.

Simulation on synthetic data. The uplift reflects the simulator's assumptions. It shows how the pilot is measured, not a result.

Adaptive versus fixed, with every run retained

24 seeds per cohort · costs and buyers normalized to 1,000 trialists · empirical 2.5th–97.5th percentile component ranges

Cohort / policyFollow-up cost / 1,000Incremental buyers / 1,000Cost / incremental buyer
Main · fixed24 / 24 runs retained3,265,495₫3,151,199₫–3,363,879₫43.131 to 57.575,806₫54,834₫–108,445₫ · component envelope
Main · adaptive24 / 24 runs retained812,294₫734,623₫–881,017₫18-2.2 to 38.545,198₫Unbounded across the component range; see cost and buyers separately.
Underpowered · fixed24 / 24 runs retained3,237,651₫3,069,622₫–3,445,069₫45.82.9 to 72.170,691₫42,552₫–1,182,161₫ · component envelope
Underpowered · adaptive24 / 24 runs retained670,947₫608,760₫–742,918₫18.3-19.3 to 41.836,619₫Unbounded across the component range; see cost and buyers separately.

Main cohort · adaptive minus fixed

-2,453,200₫ follow-up cost / 1,000Paired range: -2,609,789₫ to -2,280,058₫

-25.1 additional buyers / 1,000Paired range: -38.5 to -10.1

Adaptive conversion is lower across the reported repeated-seed range. Adaptive uses less follow-up cost per 1,000 trialists across the reported range.

Underpowered cohort · adaptive minus fixed

-2,566,704₫ follow-up cost / 1,000Paired range: -2,798,746₫ to -2,351,606₫

-27.5 additional buyers / 1,000Paired range: -63 to 1.2

The repeated-seed range does not establish a conversion advantage for either policy. Adaptive uses less follow-up cost per 1,000 trialists across the reported range.

Cost per incremental buyer = sum of cost across all runs / sum of incremental buyers across all runs. Zero and negative-lift runs remain in both sums. A ratio range is unbounded when its incremental-buyer component range reaches zero; the component envelope is not a calibrated 95% confidence interval. These are simulated repeatability ranges, not real-world results.

A complete story, in three minutes