In this article

The team ships a useful feature. A month later, the dashboard says 30% of customers have adopted it. Someone proposes a bigger announcement; someone else wants to simplify the interface. Neither decision follows from that number until the team knows who could use the feature, who encountered it and what counted as adoption.

Start with a short measurement contract. Name the customer job, the population able to perform it and the behavior that completes it. Then keep first success separate from repeated use. This produces a smaller, more useful discussion than arguing about whether 30% is good.

Count the people who can do the job

Eligibility is a product condition. An account might need a particular plan, a permitted role, an installed integration or enough underlying data before the feature is relevant. A user who cannot reach the feature belongs in a coverage report; counting that person as an unsuccessful adopter obscures a different problem.

Consider an illustrative reporting product with 120 customer accounts. Its new scheduled-export feature is available to 80 accounts with the right plan and a connected data source. Of those, 60 have encountered the feature. Thirty-six create a valid export, and 24 run another valid export during the later observation period. These figures are invented to demonstrate the calculation.

QuestionCalculationResult
How many customers have completed an export?36 / 12030%
How many eligible accounts have completed one?36 / 8045%
How many exposed accounts have completed one?36 / 6060%
How many initial completers have repeated?24 / 3666.7%
How many eligible accounts have repeated?24 / 8030%

Each rate answers a different question. The first describes reach across the customer base. The second combines discovery and use among accounts that can benefit. The third examines the path after exposure. The fourth examines continuation among initial completers. Keep the fifth visible so a strong repeat rate among a small successful subset cannot hide limited overall adoption.

Freeze the eligibility rule before reviewing results. If an integration breaks during observation, retain the account in its original cohort and report the interruption separately. Silently removing difficult cases makes the feature appear more successful precisely when customers are having trouble.

Make exposure an event you can explain

A feature flag being enabled establishes availability. It does not establish that a customer saw the feature. An announcement email being delivered also leaves that question open.

For the reporting example, the team might record exposure when a permitted user opens the exports screen and the scheduling control is visible. Write down that rule, including whether an automated page load counts. If several routes expose the feature, preserve the route as a property so the team can distinguish discovery problems from workflow problems.

Exposure is useful diagnostic information, but it is also behavior. People who seek out the exports screen may already have a stronger need. A high completion rate among exposed accounts does not show that moving everyone onto that screen would reproduce the same result. Use that comparison to find questions for research or a controlled intervention.

Define success at the customer’s boundary

A click on “schedule” can be followed by invalid settings, missing permissions or an empty result. For this feature, a sensible first-success rule is a valid export reaching the selected destination. It should include the account, output identifier and completion time, with retries deduplicated against the same intended export.

This rule is an editorial choice for the example, not a universal event name. A saved design, reconciled payment or resolved support request will need a different completion condition. Ask what the customer can now do with the result. If the answer is still “wait and see,” the event may be too early.

Amplitude’s engagement matrix separates breadth of use from frequency. That distinction is useful here: many accounts trying a feature once and a few accounts relying on it repeatedly are different product patterns. Neither should be compressed into a single adoption score.

Give every rollout cohort enough time

To see how timing changes the reading, consider a separate illustrative rollout snapshot with 60 exposed accounts. Group A received access four weeks ago; group B received it two weeks ago. Each group has 30 accounts, and each has 18 first successful exports within its first 14 days. Group A also has 12 accounts with a successful export in days 15–28.

Group B’s second-window result is not zero. It is unobserved. A calendar-month snapshot that fills that cell with zero would punish the later rollout for time it has not had.

Compare first success within 14 days of exposure across both groups. Compare repetition in days 15–28 only when each group has completed that window. Preserve the launch-date view for operational monitoring, but use elapsed time when asking whether one cohort adopted faster. For the broader distinction between time windows, see how to read a retention curve.

Decide whether the unit is a person or an account

In the reporting product, one analyst may configure an export and another may use its output. If the buying decision concerns team adoption, require a shared account identifier. If the question concerns whether individual analysts can operate the feature independently, keep the person as the unit.

Amplitude’s account-level reporting can aggregate events by groups and count a group funnel completed by different members. That capability does not make the account the correct unit for every feature. Choose it because the customer job is collective, then document which events may be completed by different people.

An account result should still retain enough detail to reveal dependence on one operator. Repeated exports from a single account administrator may be valuable, but they tell a different story from a workflow shared across a department.

Use the pattern to choose the next investigation

In the illustrative results, 20 eligible accounts never encountered the feature. Investigate discoverability and whether those accounts actually need scheduled exports. Among the 60 exposed accounts, 24 did not complete a first export. Examine permission failures, setup abandonment and task relevance. Among the 36 initial completers, 12 did not repeat within the defined later window. Check expected cadence and whether the first output solved an ongoing problem.

Those groups deserve different follow-up. A reminder cannot fix a missing permission. A redesign cannot create recurring demand where the customer needed one export. A quiet feature used successfully every quarter may be doing its job.

End the review with one decision and its supporting population: improve the setup path for eligible accounts that started but failed, for example. Keep the denominator, exposure rule and success event beside the result. The next person reading the chart should be able to reconstruct what changed without guessing what “adoption” meant that week.

Sources

Something we should clarify?

Suggest a correction