Almost every team picks its activation metric the same way: someone asks what "activated" means, and the room settles on the first milestone that is already instrumented. Profile completed. Project created. Invite sent. It is a reasonable guess, it is easy to measure, and it goes on the wall.
Six months later the number is up and retention has not moved. That is not bad luck. It is what happens when you optimise a metric that correlates with success without causing any of it.
What an activation metric is actually for
Activation is a prediction. It claims that a user who does this specific thing, within this specific window, is meaningfully more likely to still be here later. It earns its place by letting you know today whether this week's cohort is going to be fine, rather than waiting eight weeks to find out.
A metric that does that job has four properties:
- Predictive. Users who hit it retain at a materially different rate from users who do not.
- Early. It happens inside a window short enough to act on — usually the first session, first day or first week.
- Influenceable. You can plausibly change it with product work. "Paid twice" predicts retention beautifully and is useless as a target.
- Singular. One event, one window, one definition, so nobody has to ask what it means.
Why the first guess usually fails
The first guess fails in a specific and predictable way: you pick something that engaged users do on their way to succeeding rather than something that makes them succeed.
Profile completion is the classic. Users who stick around eventually fill in their profile, so the correlation in your data is real and strong. Push everyone through a profile wizard and you will move profile completion to ninety per cent and retention not at all, because you changed the symptom.
The three failure patterns to watch for:
- Reverse causation. Retention causes the behaviour, not the other way round. Anything that accumulates over time is suspect.
- Selection. The behaviour marks users who were already committed — connecting a payment method, inviting a colleague — without adding anything itself.
- Too far downstream. The event is genuinely causal but happens in week three, so it tells you nothing about the cohort that signed up on Monday.
Finding the real candidate
The search is mechanical. Take everything a user can do in their first week, and for each one compare retention between users who did it and users who did not.
For each candidate behaviour B in the first 7 days:
did_B = users who signed up 8+ weeks ago and did B
did_not_B = users who signed up 8+ weeks ago and did not do B
lift = retention_week8(did_B) / retention_week8(did_not_B)
reach = count(did_B) / count(all users in cohort)
Rank by lift, then discard anything with reach < 10%
(too rare to move the aggregate) or reach > 90%
(everyone does it; it separates nobody).The shortlist that comes out is usually two or three behaviours, and it is nearly always more specific than the milestone you had. Not "created a project" but "created a project containing at least one item". Not "invited a teammate" but "had a teammate accept". Not "used search" but "used search twice in one session".
Specificity is the signal. Vague milestones capture intent; precise ones capture the moment the product actually did something for the user.
Three tests before you commit
Does it survive segmentation?
Run the comparison separately by plan, platform and acquisition source. A candidate that holds a 3× lift on desktop and a 1.1× lift on mobile is not describing your product; it is describing your desktop users. Real activation behaviours hold their shape across cuts.
Is the dose-response monotone?
If doing the thing once helps and doing it three times helps more, in a smooth line, you are probably looking at something causal. If retention is flat at zero, one and two occurrences and then jumps at three, you have almost certainly found a threshold that selects committed users rather than a behaviour that creates them.
Does moving it move retention?
This is the only test that settles it. Ship a change that raises the candidate for one cohort and watch whether the retention curve for that cohort separates from the previous one. If activation rises fifteen points and week-eight retention does not move at all, the metric was a marker, not a lever. Change the metric; do not keep pushing on it.
This test takes as long as your retention horizon, which is why it is worth starting the moment you have a candidate rather than after a quarter of optimisation.
Write the definition down in one line
An activation metric that needs a paragraph is not going to survive contact with a roadmap discussion. Write it as a single sentence with all three parts visible — behaviour, threshold, window:
A user is activated when they complete two sessions containing a saved report within seven days of signup.
Put that sentence next to the number everywhere it appears. Most arguments about whether activation is going up are actually two people using different windows.
Expect to change it
An activation metric is a hypothesis about your product at its current stage, not a permanent truth. It should change when the product changes shape, when you start serving a different kind of customer, or when the test above comes back negative. What should not change is the discipline: one behaviour, one window, one written definition, and a willingness to retire it when the evidence says it stopped predicting anything.
Reviewing it once or twice a year is enough. Reviewing it every time it goes down is how you end up with a metric that measures nothing except how recently it was redefined.
The takeaway
Do not choose your activation metric in a meeting — find it by comparing early behaviours against long-run retention, prefer the specific candidate over the convenient milestone, and confirm it by moving it and watching whether retention follows. If it does not, you found a symptom, and the honest response is to pick a different metric.