01 · Measurement
77% of orders. 48% of the budget.
iOS and Android, instrumented end to end. AppsFlyer as the single source for installs, orders and revenue. In-app events specified once, shipped through GitHub, and sent server side to every ad platform. On iOS, SKAdNetwork conversion values and App Tracking Transparency consent. One event spec, so a purchase counts once, wherever it is reported.
With every channel measured the same way, the difference between them became clear.
TikTok turns installs into orders at nearly twice Meta’s rate. That decides where the next dollar goes.
02 · Structure and budget
Budget follows efficiency.
Boosted posts replaced by conversion campaigns, one platform at a time, each live the week its tracking was. One campaign per operating system, each bought on its own economics.
One operating system acquired customers at roughly half the cost of the other, four weeks running. Budget moved toward it in capped steps, with a clean read after each.
Capped moves
Clean reads
Learning intact
03 · Creative
Image for efficiency. Video for volume.
A standing A/B program, not one-off launches. Each format won on a different measure, so the roster runs both.
Image variants run against the winning video.
04 · Funnel
18–19% install to order. Five straight weeks.
Conversion at every step, all channels, week 4. A steady end-to-end rate makes week-over-week growth readable.
77.1%
Install to product view
41.5%
Product view to cart
05 · AI agents
Autonomous reporting. Every week.
Platform exports go in. Agents built on Claude read every file in code, cross-check the sources against each other, and write the report leadership reads. Every figure keeps its source, so each week compares cleanly with the last.
- 01ParseEvery export, read in code
- 02ReconcileSources checked against each other
- 03ReportLeadership’s weekly numbers