In this article

Imagine two cohorts of a weekly planning product. Half the first cohort returns in week one; nearly two-thirds of the second does. By week eight, both are at 25%.

Something improved. You cannot yet say that lasting use improved. The newer experience may help people return once without giving them a reason to keep using the product. Or the cohorts may contain different kinds of customers. The line identifies a question; its shape does not settle the explanation.

Reading retention well requires some patience with the denominator. Before interpreting a bend or celebrating a lift, establish who had a chance to return, what returning means, and how much time you have observed.

Read the chart title as if it were a claim

“Week-four retention” leaves most of the measurement unspecified. A usable title would say something closer to: “Accounts that created their first real schedule in August and created another in their fourth subsequent week.”

That sentence identifies the unit, starting behavior, and return behavior. It also reveals what is excluded. Accounts that signed up but never created a schedule are absent. If you are trying to understand the entire new-customer experience, you need a signup-based view as well.

An activated-user curve can improve while fewer signups activate. For example, a difficult setup process might filter out all but the most motivated customers. The remaining group’s retention could look excellent even as the product loses more people before they qualify for the chart.

Choose the return event to match the question. Opening the application measures a return visit. Completing another schedule measures repeated task performance. Neither proves customer satisfaction, but the distinction matters when someone proposes more notifications as a retention initiative.

For products used by teams, check whether the curve follows people or accounts. One administrator returning can keep an account active while its contributors disappear. Conversely, a successful handoff to a colleague can make an individual’s use fall while the account continues getting value.

Two retention methods can give different answers for the same person

Suppose an account uses the product in week one, skips week four, and returns in week six. Under exact-week retention, it is absent from the week-four numerator. Under return-on-or-after retention, its week-six return can qualify it at week four.

Amplitude’s retention FAQ documents this distinction between Return On and Return On or After. It also explains how incomplete observation periods affect the users included in a calculation. Check your tool’s method instead of assuming every retention percentage means the same thing.

Exact intervals are useful when you care whether a recurring job happened on schedule. A wider interval or return-on-or-after view can help with less frequent behavior. But a customer who comes back once after a long gap is different from someone who returns every week. Choose a measure that preserves the distinction you need.

An upward point in exact-week retention is possible: more members of the same cohort may return in week six than week five. It does not necessarily mean the chart is broken. For an aggregate curve, changing cohort eligibility can also affect later points. Investigate the underlying rows before attaching a behavioral story.

Calendar time changes the comparison

A person who signs up shortly before midnight and returns shortly afterward can count as returning the next calendar day. They have not returned after a full day has elapsed.

Amplitude’s time-window documentation distinguishes rolling 24-hour windows from strict calendar dates and explains the role of the project’s time zone. This is a consequential setting for a daily product. Two teams using different settings can produce different numbers from the same events.

For weekly work, think about the job’s actual rhythm. A calendar-week definition may match a Monday planning meeting. An elapsed-week definition may be better for comparing customers who start throughout the week. Record the choice and keep it consistent across the cohorts you compare.

Return-on-or-after charts need another timing check: older cohorts have had longer to return at some later date. Compare with a consistent follow-up horizon when that difference could influence your conclusion. A fresh cohort’s quiet final interval is not equivalent to months of observed absence.

Keep the unobserved cells empty

Here is the opening example with illustrative numbers. Each cell is the share of its cohort completing the defined task in that exact week after entry.

CohortAccounts at entryWeek 1Week 2Week 4Week 8
July20050%40%30%25%
August10065%45%30%25%
September16070%55%Not yet observedNot yet observed

July and August support a narrow conclusion: the early advantage did not persist to weeks four and eight. September has not had enough time to answer that question. Filling its future cells with zero would invent failure; extending its early improvement forward would invent success.

Cohort size belongs beside the percentages. July’s 25% represents 50 accounts, while August’s represents 25. Pooling them yields 75 retained out of 300, or 25%. In other comparisons, averaging percentages without weighting for cohort size can produce the wrong aggregate.

Small cohorts also move sharply when just a few accounts return. Show the counts and uncertainty before calling a modest difference a trend. One unusually large customer can matter even more when the chart measures revenue rather than accounts.

Turn the pattern into a focused investigation

For the July–August pattern, investigate the transition between first use and recurring work. Did the new experience encourage an initial return that still leaves the second task hard? Does the customer need fresh data or a colleague’s participation? Are people completing the job elsewhere after trying it once?

These are hypotheses, not diagnoses extracted from the curve. Inspect task histories and talk with customers who fall on both sides of the transition. Compare acquisition mix and product changes before attributing the difference to onboarding.

A flat tail also deserves scrutiny. It shows that a share of the observed cohort continues meeting the chosen definition over the measured period. It does not establish permanent retention or product-market fit. Check whether the tail is supported by enough accounts and enough elapsed time to be informative.

If the evidence suggests a specific obstacle, design an intervention and evaluate comparable mature cohorts. Preserve the original event definition so that a change in measurement cannot masquerade as a change in behavior. The decision might be to improve recurring setup, test a different segment, or wait for the newest cohort to mature. A retention chart is useful when it makes that next investigation more precise.

Sources

Something we should clarify?

Suggest a correction