One hypothesis to fix a growth loop
A step-by-step guide for PMs to turn a vague growth idea into a single testable hypothesis that links activation and retention.

Who this is for: product managers who own activation and retention and are tired of vague growth ideas, noisy metrics and experiments that don't show how change works. What you'll have at the end: a single, testable hypothesis, a minimal instrumentation plan and an experiment you can run fast.
Overview: growth loops as causal chains
Think of a growth loop as three steps: trigger → value moment → re-trigger. The trigger is the thing you control (onboarding prompt, invite CTA). The value moment is where a user experiences the product value and is likely to act again. Re-trigger is the next time the loop starts for that user or another. Your goal is a single change that lowers friction at one touchpoint and makes the value moment more likely to cause a re-trigger.
Steps (cards): pick and test one touchpoint
Treat the loop as causal
Map the loop. Write the trigger, the value moment and the re-trigger for the user journey you care about. Keep it to one sentence per node.
Find the weakest link
Use qualitative evidence and simple quantitative checks to pick the touchpoint that most often fails to move users from trigger to value moment and from value to re-trigger.
Write one hypothesis
State mechanism, segment and leading metric. Example form: “If we change X for segment Y, it reduces friction by Z mechanism and increases leading metric M.” Don’t bundle changes.
Instrument one event and one cohort metric
Pick a single event that captures the touchpoint (e.g. invite_sent) and one cohort metric that shows re-trigger (e.g. 7-day re-trigger rate for users who hit value). Instrument before you run the experiment.
Run a constrained experiment
A/B test or cohort split that changes only the selected touchpoint. Keep sample and time tight so you can read mechanism fast.
Interpret by mechanism
Look for changes in the event and cohort metric and in any intermediate actions that prove the mechanism. A lift without mechanism is weak evidence.
Scale or replace
If mechanism holds, run sequential experiments to scale. If null, log what you learned and move to the next weakest link.
Comparison: single-touchpoint vs bundle tests
| Characteristic | Single touchpoint test | Bundle test |
|---|---|---|
| Mechanism clarity | High — you can trace why it changed | Low — multiple changes hide cause |
| Speed | Fast to run and interpret | Slow — needs larger samples and follow-ups |
| Risk of false positives | Lower — less noise | Higher — vanity metric lifts are common |
| Scalability | Sequential scaling possible | Hard to know which part to scale |
A numeric lift is only useful when you can point to the mechanism that caused it. Otherwise it is a guess about what worked.
Checklist before you run anything
- Mapped the loop as trigger → value → re-trigger
- Picked one weakest link that affects both activation and re-trigger
- Wrote a single hypothesis naming mechanism, segment and leading metric
- Instrumented one event and one cohort metric in analytics
- Planned an A/B or cohort split that changes only the touchpoint
- Defined how you will read the mechanism if results are positive or null
Worked example: invite UX to link activation and retention
Discovery: some new users reach the invite screen during onboarding but few send invites. The team thinks invites help both activate the invited user and keep the sender returning. The problem: many ideas on the table and noisy metrics.
1. Map the loop. Trigger: new user reaches invite screen. Value moment: invite successfully sent and invitee accepts (or engages). Re-trigger: sender returns to use product again because an invited colleague interacts or sends another request.
2. Pick the weakest link. Qual and quick quant show low invite_sent rate. The hypothesis: friction on the invite touchpoint (button placement and copy) prevents the sender from initiating re-triggering events.
3. Write one hypothesis using a Hypothesis card. Example headline on the card: “If we move the invite CTA to the primary footer and clarify that an invite takes 30 seconds, new users in the onboarding cohort will send invites more often, increasing 7-day re-trigger among invite senders.” The card states the mechanism (lower cognitive friction), the segment (new users in onboarding), and the leading metric (invite_sent rate).
4. Instrument one event and one cohort metric. Event: invite_sent with attributes: sender_id, cohort_week, test_group. Cohort metric: 7-day re-trigger rate among users who recorded invite_sent (tracked by cohort_week).
5. Run a constrained experiment. A/B test where A is the existing invite UI and B is the single change: move CTA and update the microcopy. No other changes. Run until the cohort that recorded invite_sent reaches a reasonable size for cohort analysis.
6. Interpret results by mechanism. Possible outcomes:
- Invite_sent increases and 7-day re-trigger among invite senders rises. Mechanism supported: reduced friction increases re-triggers. Proceed to sequential experiments (e.g. vary copy, follow-up prompts).
- Invite_sent increases but 7-day re-trigger does not. Mechanism incomplete: lowering friction got more invites, but invites may not produce value for re-trigger. Investigate invite quality and acceptance.
- No change in invite_sent. Null result suggests this touchpoint was not the weakest link. Log the learning and test the next candidate.
7. Next steps. If positive, run experiments that test scaling mechanics like automated reminders or clearer invite benefits. If null, return to the loop map and pick the next weakest link.
Mistakes to avoid
- Testing fuzzy multi-change bundles that hide which change caused a lift.
- Optimising vanity metrics (installs, clicks) rather than the re-trigger you need.
- Skipping cohort or event instrumentation and hoping dashboard aggregates will tell the story.
- Ignoring segmentation; averages hide lift in the right users.
- Stopping at a single numeric lift without validating the mechanism that produced it.
If you want a practical way to capture the hypothesis, use a Hypothesis card to state mechanism, segment and leading metric before you code. It keeps the experiment focused and makes the learning repeatable.


