Name the population before the rate
A registration rate might use all visits, sessions that rendered the entry point, or identified people who became eligible to register. Each can be defensible. They do not describe the same product question.
I prefer a full sentence to a short label: who entered the population, under which conditions, and during which window. The wording can be inelegant. Its job is to make product, engineering, and analysis point to the same records.
The compact metric card
A useful definition fits on a small card. It identifies the unit, eligibility event, success event, observation window, deduplication rule, and exclusions. The examples below are generic and describe no real product.
| Field | Question |
|---|---|
| Unit | Person, device, session, or order? |
| Eligibility | Which records enter the denominator? |
| Success | Which action enters the numerator, and can it repeat? |
| Window | Must eligibility and success occur in the same period? |
| Exclusions | How are tests, errors, and missing identities treated? |
A denominator shift can imitate a behaviour shift
A new entry point, a revised sign-in rule, or an earlier event trigger can admit a different set of records. The rate moves even when the behaviour of the original population does not. Separate numerator and denominator counts, then inspect where the newly eligible records came from.
Some investigations end there. Nothing dramatic happened to demand; the measurement population changed. That is a useful finding, even if it does not make an exciting chart annotation.
Do not hide the seam between versions
When the definition changes materially, the old and new series should not be presented as continuous by default. Record the effective date and affected scope, mark the break on the report, and use any overlap period to understand the difference. If historical data cannot be restated, show the two periods separately.
- Retain both written definitions.
- State whether the numerator, denominator, or both changed.
- Use an overlap window to locate the source of the difference.
- Label the definition version in queries and reports.
Definition first, interpretation second
A stable definition makes a metric comparable; it does not explain the cause of a movement. Product paths, cohorts, releases, and experiments still carry the explanatory work.