A program adds 10,000 members in January and 10,000 more in February. By March, one group is far more active than the other. Averages will not show that. Cohort analysis will. It tracks groups of customers who share a start point and compares them at the same age.
What is cohort analysis?
Cohort analysis follows a group over time to see how its behavior changes. Google Analytics defines a cohort as a group of users who share a common characteristic, such as the same acquisition date (Google Analytics Help). The tool then shows how each group behaves over time on an app or site.
A cohort is not the same as a segment. A segment sorts customers by traits at one moment. A cohort fixes the start point and follows the group forward. That makes it the right tool for questions about change, such as whether this year’s joiners stay as long as last year’s. The start can be a campaign, a plan or a first purchase.
What are the types of cohorts?
Teams build cohorts around a start event. The table lists common ones and a loyalty use for each.
| Cohort | Grouped by | Question it answers | Loyalty use |
|---|---|---|---|
| Acquisition | Month or year of joining | Do newer joiners stay as long as older ones? | Enrollment cohorts by month |
| Channel | How they joined | Which source brings lasting members? | In-store, app and partner sign-ups |
| Behavior | A first action | Does the action predict staying? | First redemption or first tier |
| Offer | An offer they received | Did the offer change later behavior? | Welcome bonus vs none |
| Earn vintage | When points were earned | When do points get used or expire? | Points liability and breakage |
How do you read a cohort table?
Rows are cohorts and columns are periods since the start. Read across a row to see one group age. Read down a column to compare groups at the same age. Comparing at the same age is the point, because a cohort that joined last week cannot be compared with one that joined a year ago.
Fader and Hardie show a hypothetical firm with a 2003 cohort of 10,000 customers. 6,334 renewed in 2004, then 4,367, 3,264 and 2,604 in later years (Fader and Hardie, 2010). The yearly retention rates are 63.3%, 68.9%, 74.7% and 79.8%.
Know how the table is counted. Mixpanel’s retention report counts users and not events. A user counts once per bucket, and the clock starts when that user does the starting event (Mixpanel). Amplitude adds that a behavioral cohort groups users who act in a given way, such as watching a help video in a first session. You learn by comparing it with another group (Amplitude).
Why do retention rates rise as a cohort ages?
At the cohort level, retention rates typically increase over time. Fader and Hardie say this is due in large part to a sorting effect in a population of mixed customers. Members who are likely to leave go early, and the survivors are the loyal ones (Fader and Hardie). So a cohort’s retention curve flattens even when no one changes their mind.
Their stylized example shows how. Suppose a third of a cohort keeps 90% of its members each year and two thirds keep 50%. The whole cohort’s yearly retention climbs from 63.3% to 79.8% over four renewals. No member changed. The mix of who remains did (Fader and Hardie).
This matters for forecasts. A single blended retention rate hides the pattern. The authors show that failing to account for it gives a downward-biased estimate of the value left in a customer base. It also understates how much a gain in underlying retention is worth. Our customer retention guide makes the plain version of the point: measure by cohort, because a blended rate hides whether this year’s joiners behave worse than last year’s.
What can make cohorts differ?
Three things can make cohorts differ, and a table cannot tell them apart by itself. Pew Research Center names them life cycle effects, period effects and cohort effects (Pew Research Center). A life cycle effect is about age or tenure. A period effect hits everyone at once, such as a price change. A cohort effect belongs to the group, such as members who joined during a launch.
A loyalty team meets all three. A dip in month 6 can be tenure, since many members drift after the first reward. It can be a period effect, such as a site outage that touched every cohort that month. Or it can be a cohort effect from a sign-up promotion that drew bargain hunters. Read down a column to cancel tenure. Read along a calendar month to spot a period effect. What is left is the cohort.
How do loyalty programs use cohort analysis?
Programs use cohorts in four ways. They track enrollment cohorts to see whether onboarding changes stick. They compare paid and free member cohorts to read the real effect of a fee on frequency, spend and retention, as our paid loyalty guide describes. They hold out a cohort that gets no offer, to measure lift. And they track points by the month they were earned.
Starbucks shows the last one. Green Level Stars expire six months after the calendar month in which they were earned. The oldest accrued Stars are used first. A Star earned on December 15, 2025 expires on July 1, 2026 (Starbucks Rewards Terms). Each month’s Stars form a cohort with its own clock, and breakage and liability can be read by earn month.
How do you build a cohort analysis?
Follow five steps. Pick the start event, such as enrollment. Pick the period, such as a month. Pick the outcome, such as active in the period. Define active, for example a purchase in the last 90 days. Then compare cohorts at the same age. Fader, Hardie and Lee used the same logic on a cohort of new customers of the online music site CDNOW to estimate lifetime value (Fader et al., 2005). Keep the definitions fixed from one report to the next. If "active" changes from 60 days to 90 days between reports, the cohorts stop being comparable.
For illustration, a program enrolls 10,000 members each month. In month 3, 31% of the January cohort are active, 33% of February and 36% of March. Newer cohorts are doing better at the same age. The team checks what changed. If onboarding emails began on February 15, that is the first thing to test with a holdout. Do not assume it caused the gain.
Size matters. With 10,000 members per cohort, the 5-point gap between 31% and 36% is far outside chance. With 500 members per cohort, the same gap would not be. Small cohorts need a longer window or a pooled view.
How GRAVTY handles cohort analysis
Our retention economics guide says GRAVTY logs each earn, burn, tier change and offer response as a timestamped event on a member identity. Lifetime value, churn curves, cohort survival and incrementality start from that ledger. They do not start from samples. So a team can build cohorts on any start event.
Frequently asked questions
What is cohort analysis in simple terms?
Cohort analysis tracks groups of customers who share a start point and compares them over time. A cohort is a group of users who share a common characteristic, such as the same acquisition date (Google Analytics Help). It shows whether newer groups behave better or worse than older ones.
What is the difference between a cohort and a segment?
A segment sorts customers by traits at one moment, and a cohort follows a group from a shared start point. Cohorts answer questions about change over time, such as retention by joining month (Google Analytics Help).
What is an example of cohort analysis?
Tracking customers by the year they joined is the classic example. In Fader and Hardie’s hypothetical, a 2003 cohort of 10,000 customers fell to 6,334, then 4,367, 3,264 and 2,604 over four renewals (Fader and Hardie, 2010). Each yearly retention rate was higher than the last.
Why does retention rise as a cohort ages?
A sorting effect explains most of it. Members who are likely to leave go early, so the survivors are more loyal on average (Fader and Hardie, 2010). The curve flattens without any single member changing their mind.
How do you do cohort analysis for a loyalty program?
Pick a start event and a period. Pick an outcome and define active. Then compare cohorts at the same age. Programs use enrollment cohorts, paid and free cohorts, holdout cohorts and cohorts of points by earn month (Starbucks Rewards Terms). Test any change you think caused a gain with a holdout.
