Industries, Travel & mobility
Decisions for booking apps and travel marketplaces.
Personalised paywalls. Smarter activation. Proven lift on the cohorts that matter.
Travel apps live on paywall conversion and D30 retention. They spend $50M–$500M per year on paid acquisition, sit on clean recurring signals, and have growth teams sophisticated enough to want, and verify, causal proof. Metapolicy is the substrate that lets them stop running A/B tests they cannot defend.
What agents handle
Workflow
Paywall variant assignment
Before
Hard-coded 9.99 price for everyone
After
LinUCB picks per user from a frozen-context snapshot
Result
+12% trial-start, +$2.1M ARR modelled
Workflow
Day-3 reactivation push
Before
Send to every dormant user
After
CATE-targeted: only positive-uplift cohort receives the push
Result
−64% sends, +18% reactivation
Workflow
Onboarding step ordering
Before
Static 6-step flow for everyone
After
Bandit-optimised per platform and locale
Result
+9% activation through to first booking
Workflow
Holiday-season discount surfacing
Before
Blanket banner shown to all users
After
Propensity-targeted exclusion of high-LTV users who would convert anyway
Result
+$340K incremental margin per quarter
The challenges we solve
- ATT killed deterministic attribution, your MMP reports are an estimate built on an estimate
- A/B tests under-power on holiday seasonality and traffic mix shifts
- CFOs will not sign off on lift claims without a defensible holdout
- Segment-level CATE is invisible in dashboard tooling, you see the average, not the cohort it helped
- Multi-modal funnels (flights + hotels + activities) make reward attribution noisy
Propensity-logged decisions
Every served variant carries the probability our policy gave it at decision time. That single column is what makes doubly-robust evaluation honest.
Replay your last 90 days, day one
We seed the bandit from your historical logged data via uplift / CATE modelling and reproduce a doubly-robust readout on day one. No three-month ramp.
ESS-guardrailed billing
Effective sample size below n/10 means overlap-limited, flagged, not invoiced. The math itself caps how much we can over-claim.
Worked audit example
What a doubly-robust re-analysis surfaced for Aerial.
Series-C travel-booking app · ~$35M ARR · EU + US · 1.18M trialists across 17 markets · case ID AERIAL-2026Q1-AUDIT-001
What their team reported
Q4 paywall A/B with three price tiers ($9.99 / $12.99 / $14.99). t-test on assignment buckets: p = 0.34. Team called it "no significant effect" and shelved the variant.
What our re-analysis found
Doubly-robust per-cohort: +9.3% lift (CI [+4.1, +14.5]) on top-quartile-popularity routes × peak season, a cell carrying 23% of trial-start volume. The global t-test averaged it away with the off-peak segments.
Recommendation · projected annualised impact
Ship Variant B only to top-quartile-popularity routes during peak season. Suppress otherwise. Same paid-acquisition spend, recovered revenue.
+$2.1M / yr ARRSame shape we'll send back on your last A/B test, in three business days.
Read the full audit PDF →Audit your last paywall test — free.
One CSV, one experiment config. Same-shape readout back in three business days.