Not measuring it badly. Not measuring it at all. In October 2020 there was no number for churn, because nobody had built the thing that would produce one.
When we finally had it, it looked survivable. Fourteen per cent of paying customers cancelled every month. Bad, obviously. But a month is small, and fourteen is not a frightening number to look at.
16
Out of a hundred. Everything the company spent on getting customers was refilling a bucket with a hole that size.
51
Voluntary churn 13.9% to 5.4% a month. Involuntary churn halved as well, 7.2% to 3.6%. Same customers. Same product. Same spend on marketing.
The number was not new. The argument was.
Somebody there had almost certainly run that sum before I arrived. What had not happened was anyone putting it in front of the people who set the budget, as a limit on the whole business rather than a support metric, at the moment it could change what got funded.
That gap is not a failure of intelligence. It is what org charts do to knowledge.
So we built the thing that should have existed first. Not a dashboard. An analytics function. Survival models that estimate how long a customer lasts and what shortens it. Cohort studies on people who had already left, to find out who comes back.
I did not run those models. I hired and built the team that did, and I made the case for why a growth budget should pay for them.
Fourteen per cent of people who cancel come back. Eighteen per cent of those who chose to leave. Two thirds of the returners take the lifetime plan.
Then the model told us we were wrong about the product.
Everyone believed activation meant projects. Get a new user to create projects, and they convert. It is written into every onboarding flow in the industry. We had built ours that way too.
Picture the user we thought we wanted. She signs up, puts in her own domain, then a competitor's, then a client's, then two ideas she had in the shower. Eight projects before lunch. By every measure we had on the wall she is the best kind of person who arrived that week.
She never came back.
We fitted a model to what trial users actually did, and it came back with two numbers pointing in opposite directions.
The thing we were pushing people to do was the thing that predicted they would not pay.
It makes sense once you see it. Creating project after project is what shopping looks like. Trying the tool on everything, committing to nothing. Going back and updating one project is what using it looks like. We had built an onboarding that rewarded browsing and called it activation.
Finding that out was the easy part.
A number does not redesign anything. It hands you a design problem, and this one had an uncomfortable shape.
Every convention in onboarding says the same thing. Reduce friction. Let people explore. Show them the breadth of what the tool can do, because breadth is what they are paying for. Every onboarding I have ever reviewed, including the one we had built, is some version of look how much you can do here.
Our data said breadth was the thing killing us.
Which meant the design had to do the thing that feels wrong to a designer: make the new user go narrower. Hold them inside one piece of work. Make starting a second project less inviting than returning to the first. Not by hiding it, but by never making it the obvious next move.
That is a hard sell in a review. It reads as taking features away. It reads as adding friction, which is the one thing everybody in a growth team has been trained to remove.
The only reason it survived the argument is that the argument was not about taste. It was about a coefficient, and the coefficient was significant.
So activation was redefined around depth rather than breadth, and the early experience was rebuilt around returning to one project instead of starting the next one.
Over two years: paying subscribers from roughly 51,000 to 81,000. Accounts from about 150,000 to 310,000. Lifetime plans grew until they were close to half of monthly revenue. First-time trials converted at 44%.
Every company I have joined believed something about its own product that its data denied.
Six companies have hired me. Four made me Chief Product Officer. Every time, the problem described on the phone was not the problem they had, and every time, finding the real one meant doing arithmetic that nobody had thought worth doing.
I could have handed them that coefficient and gone home. It was a good finding. It was worth about one quarter.
What they had two years later was a business that could ask its own questions.
Survival models and win-back cohorts running as ordinary practice, by people who did not need me in the room to do it. I am trying to make myself unnecessary everywhere I go. It is the only version of the job I can defend.
The finding was worth a quarter. The function is still working.