A Datalyze framework · sample data · video companion
How to find what's actually driving your retention.
A six-step diagnostic for figuring out why your D30 number is where it is — and what to do about it. Walked through with sample data so the framework is visible, not abstract.
scroll · 01 of 06
01 / BaselineSet the number you're moving
The starting point
You can't improve what you haven't measured cleanly.
Pull your overall D30 retention curve. Establish the number. Everything that follows is a comparison against this baseline.
50%
D30 retention · all users · last 90 days
retention curve · day 0 → day 30sample data
↳
This is your baseline. Half your signups are still here at day 30. The goal of everything that follows: figure out which slices of that 50% are actually 70%, and which are 25%.
02 / Path 1 — Who they areThe easier win
Demographic cut
Slice by attributes. Look for a clear delta.
Country, channel, plan, device, company size. You're hunting for a segment that already retains better than the average. When you find one, the action is acquisition mix — not product change.
United States
65%+15
United Kingdom
58%+8
Germany
48%−2
India
45%−5
Canada
35%−15
↳
US users retain 30 points better than Canadian users. The lever isn't a product change — it's biasing acquisition spend toward markets where the product already lands. Easiest win in retention work.
Referral
72%+22
Organic search
58%+8
Content marketing
54%+4
Paid search
42%−8
Paid social
31%−19
↳
Referral users retain at 72%. Paid social at 31%. Same product, completely different user. Before touching onboarding, ask whether you're buying the wrong users.
Annual paid
78%+28
Monthly paid
48%−2
Free trial
32%−18
↳
Annual users retain at 78%. Free trial at 32%. Some of this is selection effect — annual buyers are pre-qualified — but a 46-point gap is also a signal that the trial isn't converting the right signal of commitment.
03 / Before you queryPrioritize, don't fish
Hypothesis list
Write your ten hypotheses before you open the data.
A pre-PMF product has 100 behaviors you could test. A post-PMF product has more. You can't run them all. Use product knowledge to narrow before you query, not after.
H01
Users who return 2+ times in the first 7 days retain better at D30 than single-visit users.
H02
Users who hit the core feature within 7 days of signup retain better than those who don't.
H03
Users who invite a teammate in week 1 retain better than solo users.
H04
Users who connect a data source within 48h retain better than users who skip integration.
H05
Users who complete all onboarding steps retain better than users who drop off at step 3.
H06
Users who view their dashboard 3+ times in week 1 retain better than those who view it once.
H07
Users who customize their workspace in session 1 retain better than users on default settings.
H08
Users who act on a notification in week 1 retain better than users who don't engage with notifications.
H09
Users who hit their first "aha" event within 24h retain better than slow activators.
H10
Users in the top-quartile attribute segment who also use the core feature retain better than either signal alone.
↳
You go in with this list. You don't open Mixpanel and hope something jumps out. Ranked hypotheses save you weeks — and produce findings you can actually defend in a product review.
04 / Path 2 — What they doTest each hypothesis
Behavioral cut
Now test each hypothesis as a cohort split.
For each hypothesis, split users into two cohorts (did the behavior vs didn't) and compare D30 retention. The behaviors that produce the biggest deltas are your activation candidates.
retention by behavior · used core feature within 7 dayssample data
Used (within 7d)
72%
D30 retention
Didn't use
38%
D30 retention
Delta
+34pt
Strong signal — ship activation work
↳
A 34-point gap is the kind of signal that justifies onboarding redesign, in-product nudges, and an email sequence aimed at getting more users to the core feature in week 1.
retention by behavior · invited teammate in week 1sample data
Invited teammate (w/in 7d)
81%
D30 retention
Stayed solo
44%
D30 retention
Delta
+37pt
Strongest signal in the set
↳
37-point gap. Collaboration is sticky. Worth wiring an invite prompt into the post-signup flow and a week-1 email — both cheap to ship, both with clear upside.
retention by behavior · 2+ sessions in first 7 dayssample data
2+ sessions (in 7d)
76%
D30 retention
Single session
28%
D30 retention
Delta
+48pt
Habit signal — but lagging, not causal
↳
Biggest delta in the set — but be careful. A second session in week 1 is partly a downstream effect of value, not a cause of it. Treat it as a leading indicator to monitor, not a behavior to force.
05 / Methodology trapThe detail most analysts miss
Time-boxing
Always time-box the behavior. Or the conclusion is wrong.
Same product, same feature, same users. The only thing that changes between these two analyses is the time window. The conclusion you'd draw from each is completely different.
Wrong — no time box
"Do users who ever use Feature X retain better?"
Used
58%
Didn't use
44%
+14pt
Modest delta. "Maybe worth shipping, maybe not."
Right — time-boxed
"Used Feature X within 7 days of signup?"
Used (w/in 7d)
72%
Didn't use
38%
+34pt
Strong delta. "Build the activation loop. Now."
↳
Without the time box, a user who used Feature X on day 1 and one who used it on day 25 count the same. They're not the same user. Skip the time box and you'll either kill a real winner or ship a "fix" that wasn't worth the work.
06 / The loopPush more users into the winning behavior
Action
The work shifts from analysis to moving the population.
Once you know what drives retention, the job is getting more users into that behavior (or segment). PM, design, lifecycle, marketing — all of them are now levers. Then you measure.
A.
Find the winning behavior or segment
The output of steps 01 through 05. A signal big enough to defend in a product review.
B.
Design the push
Onboarding redesign. In-product nudges. Lifecycle emails. Empty states pointing at the activating action. Bias acquisition spend toward the segment that retains.
C.
Run the experiment
A/B test or holdout. Measure the population shift toward the behavior and the lift in D30. If the behavior is causal, retention moves. If it isn't, it doesn't.
D.
Re-baseline and repeat
D30 moves from 50 to 54. New baseline. Next hypothesis from the list. The loop compounds.
↳
Most teams skip steps 01–05 and start at B. They ship "onboarding improvements" without knowing which behavior matters. The diagnosis is the work. The fix is the easy part.