Week 1: instrument, don't optimise
Do not test anything in week one. Instrument the funnel end to end. Confirm attribution. Establish the baseline. Anything you 'optimise' in week one is superstition, because you have no baseline to compare against.
Weeks 2–4: earn category signal
Publish the launch. Prompt reviews from the initial cohort. Respond to every review. Ship the first metadata refresh with the vocabulary you've learned from real reviews. Do not test yet, the category algorithm is still learning who you are.
Weeks 5–8: first optimisation loop
Screenshot test (one variable). CPP for the strongest paid channel. Icon variant if D1 install-through-rate is below your category benchmark. Publish week-eight report with cohort-one performance versus benchmark.
Weeks 9–12: compound or diagnose
If cohort one is above benchmark: double down on the winning motion. If below: diagnose, is it a product problem (retention), a promise problem (screenshots), or a category-fit problem (category selection)? Each has a different fix.
Recovery pattern
Below-benchmark launches can recover with the right diagnosis. Avoid changing everything at once because that destroys attribution; sequence changes so each learning remains legible.