ANALYTICS & REPORTING

Cohort retention reading

Use when you have a cohort table and need to know what it actually says about retention.

cohort-retention-reading.md

Download .md

You are reading a cohort retention table for {{PRODUCT_OR_SITE}}.

Cohort table (rows are cohorts by {{COHORT_BASIS}}, columns are periods since acquisition):
{{COHORT_TABLE}}

Cohort sizes: {{COHORT_SIZES}}
What counts as retained: {{RETENTION_DEFINITION}}

Answer in this order:
1. Does the curve flatten, and at roughly what period and what level? If it does not flatten in the data available, say that rather than extrapolating.
2. Are later cohorts better, worse, or the same as earlier ones? Quote the specific cells you are comparing.
3. Which cohorts are too small to draw a conclusion from? Name them and give the cohort size.
4. Which apparent trends are explained by cohort maturity (recent cohorts having fewer observed periods) rather than by behavior?
5. Three questions this table raises that it cannot answer.

Rules:
- Never describe a difference as significant unless you show the two numbers and the cohort sizes behind it.
- Treat any cohort under {{MIN_COHORT_SIZE}} as directional only and label it as such.
- Do not project a lifetime value figure. If asked to, explain what assumptions that would require.
- Plain language. No statistics jargon without a short definition in the same sentence.

Fill in before running

Replace each placeholder with your own detail. The more specific you are, the less the model invents.

{{PRODUCT_OR_SITE}}
{{COHORT_BASIS}}
{{COHORT_TABLE}}
{{COHORT_SIZES}}
{{RETENTION_DEFINITION}}
{{MIN_COHORT_SIZE}}

Getting a better result

Include the raw cohort sizes; percentage-only tables are how small cohorts become false trends.
Define retention explicitly - "returned" and "did a core action" give opposite-looking curves.
Ask it which cohort you should re-check in 30 days once it has matured.

Questions about this prompt

When do I use this rather than reading the cohort table myself?

When the table is being used to argue that retention is improving. Most such arguments compare a mature cohort against an immature one. This separates that maturity effect from genuine behaviour change, which is the reading people get wrong, and it is why the prompt asks for cohort sizes rather than percentages alone.

What do I need in front of me?

The cohort table, the raw cohort sizes, the cohort basis, an explicit definition of retained, and a minimum cohort size you will draw conclusions from. Define retained precisely: came back to the site and completed a core action produce curves that look nothing alike and support opposite decisions.

What comes back?

Whether and where the curve flattens, whether later cohorts beat earlier ones with the specific cells quoted, which cohorts are too small to read, and which apparent trends are cohort maturity rather than behaviour. Question four is the useful one. It usually removes the finding somebody was about to put in a deck.

What is the mistake that costs me here?

Asking it to extend the curve into a lifetime value figure. It declines and explains the assumptions that would require, which is correct: a projected LTV from a curve that has not flattened is an assumption wearing a number. Pasting percentages without the sizes underneath is the other way small cohorts become trends.