A row of old brass lanterns, each lit with warm amber flame, receding into dense fog on a dark stone path at night Growth
AI-generated, Working Theory
Growth · ◉ Evergreen

Activation is a guess you instrument, then a lever you prove.

by · ·5 min·Working Theory

Your activation metric isn't found in the data — it's drawn, instrumented, and proven. How to find the candidate, log the road to it, and test whether it's a lever or a thermometer.

Everybody wants to know their “aha moment.” Fewer people notice that the aha moment isn’t a thing you discover lying in the data, like a fossil. It’s a line you draw — and the quality of your growth work is mostly the quality of where you drew it and how honestly you checked it.

Start with the uncomfortable truth: you cannot improve a moment you haven’t defined as an event. “Users get value early” is a feeling. It doesn’t have a timestamp, it doesn’t fire in your analytics, and you can’t tell whether it went up last week. So the first job is to turn a feeling into something a machine can count. That means committing to three specifics that feel arbitrary and aren’t: an action (what did they do?), a count (how many times?), and a window (by when?). “Created two documents in the first seven days.” “Invited one teammate on day one.” Ugly, precise, loggable. The ugliness is the point — vagueness is what lets a team argue about activation for a year without moving it.

Where does the guess come from? Not from a whiteboard. You find it by contrasting the people who stayed with the people who left, looking only at their earliest sessions — before you know which group they’ll become. What did the retained cohort do in week one that the churned cohort didn’t? You’ll usually get a few candidates. Pick the one that is early (so you can still influence it), common enough among keepers to matter, and rare among leavers so it actually separates them. That contrast is your hypothesis. It is not yet a truth.

signup step 1 step 2 activated? (the line you draw) retained nudge vs holdout — does pushing to it CAUSE retention?
Instrument the whole road, not just the finish line: each step you can see is a step you can fix — and the dashed line is a claim you still have to prove causal. Original diagram · Working Theory

Here’s the part most teams skip, and it’s the difference between instrumenting activation well and instrumenting it uselessly: don’t just log the finish line — log the road to it. If your candidate is “created two documents,” then instrument opened the editor, created the first, created the second as separate events. Now the endpoint stops being a single pass/fail number and becomes a funnel you can see through. When activation is soft, the funnel tells you where — people open the editor and never type; people make one doc and stop. A bare activation rate tells you that you have a problem. The instrumented road tells you which step is the problem, which is the only version of the information you can act on.

Then measure it the boring way: cut by weekly cohort, plot the share that crossed your line, and plot each cohort’s retention beside it. If the line is real, the cohorts that activate more should retain better, consistently, over time. If they don’t, your guess was wrong — go back and pick a different candidate. This loop, run a few times, is how a vague “aha” turns into a metric you’d actually stake a roadmap on.

And now the trap that a companion idea on this site already named: correlation is not the finish line. The fact that activated users retain better does not mean pushing more users across the line will make them retain — you might just be measuring a thermometer, reading the temperature of intent that was already there. The people who create two documents in week one may have been your keepers no matter what; the second document is a symptom of fit, not a cause of it. So build the causal test in from the start, don’t bolt it on later: take a cohort, actively nudge them toward the activation action — onboarding prompt, reminder, a little friction removed — hold out a comparable cohort, and watch whether the pushed group retains better weeks later. If it does, you’ve found a lever, and you should pull it hard. If retention doesn’t move, you’ve confirmed you had a thermometer, and the honest response is to stop optimizing the number and go find the real driver.

The whole discipline, compressed: guess the moment from the keeper-vs-leaver contrast, define it as an ugly precise event, instrument the road to it and not just the arrival, prove the correlation across cohorts, and only then test whether it’s a lever you can pull or a temperature you’re merely reading. Most teams do step two and declare victory. The value is in the last step, and the last step is the one that can tell you your favorite number was never a lever at all.

Sources

  • Sean Ellis and the "aha moment" / North Star metric framing
  • Reforge and Amplitude activation frameworks
  • leading vs. lagging indicators
  • correlation vs. causation and holdout/experiment design

Liked this? Get the next one in Working Theory.

Going weekly in August (it's in beta now). One genuinely interesting read on building, the brain, and the science most people missed.

Subscribe →
Got a reaction, a counter-example, or something I missed? Reply by email — I read everything.
◉ join in

Where have you hit this — in a product you use, or one you're building?

Threads open here soon. For now, the conversation lives two clicks away — discuss on GitHub, or just reply by email. I read and answer everything.