Retention reveals repeated value
Lesson 14 of 24, level 01Index
Signups are up 40 percent this month and the founder is delighted. Week two retention is 9 percent and nobody has looked at it.
Read the shape, not the level
A cohort retention curve shows how a defined group returns over time. A curve that appears to flatten at 25 percent suggests a subset may be finding repeated value, but the interpretation depends on the return event, observation window, and sample size. Continued decline is a reason to investigate repeated value, not a diagnosis by itself.
Read retention alongside activation, the natural usage rhythm, and your business model. More signups can hide weak repeat usage; a short retention chart can also hide uncertainty. Keep the event definition and cohort counts beside the curve.
Retention asks whether a defined group returns to do a meaningful action after a starting event. A curve becomes interpretable only after you specify those events and the product's natural usage rhythm.
Use your chosen course project throughout. The additional examples below are fictional practice cases; transfer the method to your own evidence.
Define the cohort
Group users by the same starting event and period. Compare them at the same age, not simply on the same calendar date. A cohort that started yesterday cannot yet have a week-four result. Keep incomplete cells empty rather than counting them as zero and creating a false decline.
Choose what return means
An app open may be too weak for your product. Pick an action that represents renewed value, and state whether you count return in a particular interval or on-or-after it. Those definitions answer different questions. A monthly invoicing tool should not be judged by an unexplained daily habit benchmark.
Investigate the shape
An early fall can suggest an activation problem; continued loss can suggest weak repeated value. These are hypotheses, not diagnoses from a line. Compare cohorts and interview people who returned and who did not. A flat-looking tail based on a tiny remaining group needs counts and longer observation before a broad claim.
Read a cohort at the same age
- Start: One qualifying event
- Return: A meaningful action
- Age: Equal elapsed time
- Interpret: Counts plus context
The same cohort, two possible stories
Interactive teaching example · All numbers and tickets are fictional.
A flattening tail suggests a group may be finding repeated value. Four weeks cannot establish a lasting plateau. Each week counts return in that interval, always divided by the original 100.
Doing it with AI, and where it breaks
Cluster your churned users by behaviour in their first session and find the one action that separates the ones who stayed from the ones who did not.
Correlation will be presented as cause. The action that predicts retention is often a symptom of already being the kind of user who retains, not the thing that caused it. State that explicitly in your write up.
Your AI workbench
Start with your own notes or clearly labelled practice data. Remove private details before sharing. Replace the placeholders, run the prompt in your chosen AI tool, and keep the output beside its source.
Given these anonymised events [paste], define a starting event, meaningful return event, interval, and cohort. Show the cohort-age table with counts and mark unobserved periods as unavailable. Explain interval versus on-or-after retention. Do not infer causality or invent industry benchmarks.
Before you use the output
- The denominator stays the original cohort
- Unobserved periods are not zero
- The return action reflects the product's usage rhythm
Stuck? Try this next
Ask AI to show one user's inclusion across the table. Manually trace that user before trusting the aggregate. Export the table alongside the chart so the definition remains inspectable.
Keep a brief AI log: input used, useful output, what you checked, and what you rejected. The decision remains yours.
Build it
A cohort retention curve and one testable hypothesis about the drop off.
Google Sheets
What my retention chart can and cannot tell me
Publish a 500–800 word project analysis. Define the starting event, return action, cohort, and interval. Include counts, explain one competing interpretation, and propose a next check. If you used synthetic events, say so beside the chart and do not present them as customer results.
Include: A cohort table and chart with a clear denominator and unobserved periods marked.
- Write in your own voice for someone facing the same product problem.
- Use AI to critique the structure and find unsupported claims. Verify every source, number, and quotation yourself. Add a short note describing how you used AI.
- Remove private customer, company, and participant information. Label practice scenarios and synthetic data clearly.
- Publish on Medium, Substack, or a relevant Reddit community that permits project write-ups and links. Follow its posting rules; share a useful account of the work.
Include this plain attribution with a clickable link:
I developed this project while learning product management at pmcademy.com.
Checkpoint
If you did the build, these take two minutes. If you cannot answer one of them, that is the part to go back to.
Moving on marks this lesson complete. Finish the build first, it is the part that counts.