NeoTasteHead of Product2026
Same team, about three times the tests.
Subscription restaurant marketplace, six-figure member base, four countries.
- Instead of
- adding people
- We bet on
- testing faster, with written rules for when to stop.
- Outcome
- about three times the A/B tests, same team
Pace and the rules.
The team already had good foundations: a test journal, an A/A check, a readout template. My job was pace and decision rules. We went from about one test a month to about three, roughly what our traffic allows.
Decision rules
Split by market and platform.
Never move a threshold to fit a result.
Every claim graded as measured, modelled or no claim.
A kill counts as a result.
The step most people didn't need.
Most bookings were redeemed within the hour, usually at the till. Booking was a step most people didn't need, so I made the case to let members redeem directly. The share of bookings that became a visit rose about ten points on both platforms, and cancellations fell by about a quarter.
Fewer people book, and the ones who do come through. Total redemptions haven't moved enough to call yet, so the honest claim is a cleaner funnel, not more visits. It shipped on both platforms.
What we stopped.
- Discount quest. It kept people through the trial and lost them at the first payment. The share who became lasting members didn't change, so it was the same members at a lower price. We didn't ship it.
- Savings popup. Showing members their savings after a redemption lowered conversion from the free period. We removed the popup and kept the savings figure where it does no harm.
The idea I killed myself.
I believed a small group of deal-seeking members was wearing restaurants out. Before anyone built anything, I ran the numbers myself. The group was well under one percent of members, and restaurants that left had no more of them than restaurants that stayed. I paused my own project.
What the engineers did.
Deploys went from over twenty minutes to under five. The iOS download shrank to about a quarter of its size. The slowest onboarding call went from seconds to milliseconds. That was the engineers' work. My part was giving platform work a place on the roadmap next to product work.