A/B testing for subscription products, done honestly

I ran the experimentation program on subscription funnels at CaaStle — $2.1M in ARR savings, 20% incremental revenue — and most of it came from testing less, better.

Subscription products are where sloppy A/B testing does the most financial damage, because the metric that matters — retained revenue — arrives months after the metric that's convenient, which is click-through. I spent years running experimentation on subscription funnels at CaaStle across a $30M–$50M ARR portfolio. The program produced $2.1M in ARR savings and 20% incremental revenue growth, and the biggest lesson was counterintuitive: the wins came from running fewer, better-designed tests, not more of them.

The rules that made the number real

  • Measure cohorts through a billing cycle. A variant that lifts trial starts but degrades month-two retention is a loss wearing a win's clothing. Every subscription test we trusted tracked its cohort through at least one full renewal.
  • Pre-register the decision, not just the metric. Before launch: primary metric, minimum effect worth acting on, run time, and what we'll do in each outcome. Post-hoc rationalization is how testing programs become fiction.
  • No peeking promotions. Checking significance daily and stopping at the first p<0.05 is the most common way teams manufacture false wins. Fixed horizons or proper sequential methods — nothing in between.
  • Test where the revenue is. Teams burn quarters testing button colors on the homepage while the cancellation flow — where a well-designed pause option can save a meaningful share of would-be churn — goes untested for years. The highest-ROI tests in subscription are in payment recovery, plan-change, and cancel flows.

What I do for subscription teams

Two engagement shapes. An experimentation audit (two weeks): I review your past twelve months of tests for validity — measurement windows, peeking, sample math — and it's common to find that a third of "wins" don't survive scrutiny; you leave with a corrected playbook and a ranked test backlog. Or fractional ownership of your growth experimentation for a quarter or two, running the program hands-on while training your team to keep the standards after I leave.

Frequently asked questions

How long should a subscription A/B test run?
Through at least one full billing cycle for any test touching conversion or retention — often six to eight weeks total. If your traffic can't support that, run fewer tests with bigger expected effects rather than shortening windows and trusting noise.
What sample size do we actually need?
It falls out of the minimum effect you'd act on — decide that first, then do the power math. The honest answer at many startups is that fine-grained tests are underpowered, and the right move is testing bold changes where a real effect is detectable.
Our test wins don't show up in revenue. Why?
The classic causes: wins measured on leading metrics that don't survive to renewal, peeking-inflated false positives, and interaction effects between simultaneous tests. This exact symptom is what the audit engagement is built to diagnose.
Is A/B testing worth it before we have much traffic?
Below a few thousand conversions a month, mostly no — use painted-door tests, user interviews, and pre/post analysis with holdouts instead. I'll tell you honestly which side of that line you're on.

Related pages

Let's talk about what you're building.

Always happy to chat with founders, builders, and growth operators. 30-minute introductory call. No agenda needed.

Audit your experiment program