Why we need cohorts
If you compare all January users with all August users today, the January group had a lot more time to accumulate activity. Such a report will almost inevitably make the old source "better." Cohort analysis corrects this error: groups are compared at the same age.
Simple cohort
Take the week of the first FTD as the start date. Then look at the results of each group after 7, 30 and 60 days.
| cohort | FTD | D7 active | D30 active | NGR D30 |
|---|---|---|---|---|
| August 3-9 | 80 | 38 | 24 | $5 400 |
| August 10–16 | 92 | 46 | — | — |
For the second line, the D30 has not yet arrived, so it is not yet possible to compare it with the first line according to the NGR D30.
Add a source.
Next, break cohorts down by SubID. This shows both how the August cohort behaves and which campaign brought a more sustainable audience.
Add GEO only after the source
Too many incisions on a small volume creates noise. If a cell has five users, one person changes the percentage by 20 points. Start with large segments and split only where there is volume.
What's useful for RevShare
Cohorts help to see if the quality has deteriorated after scaling. For example, the July campaign could give 100 FTDs and a strong D30, and the August campaign after the budget increase - 180 FTDs, but a weak D30. Without cohorts, volume growth would look like a success.
What matters for CPA
Even if the payout is fixed, a cohort quality analysis helps to understand why the approval rate changes and which sources the affiliate program may consider more sustainable.
Minimum report
- week of FTD;
- source;
- GEO;
- number of FTDs;
- confirmation;
- D7/D30 activity, if available;
- NGR on a fixed horizon.
The relationship of cohorts with long value is discussed in detail in the material Retention and LTV.
Choose one “age” for comparison.
For a weekly report, it is convenient to record, for example, D30. Then each cohort is evaluated only after 30 days. New cohorts are left with an empty value, rather than getting an artificially low result.
Notice changes in terms
If the rate, promo or landing has changed in the middle of the month, add a label to the corresponding cohorts. The difference can then be linked to the context rather than attributing it to the date itself.
How to detect degradation at scale
Build a FTD cohort next to it and its D30 result on one FTD. If volume grows and quality per unit declines systemically, the source can extend beyond its strong audience.
Don't draw conclusions from one cohort
One period may be atypical. For a full budget reallocation decision, it’s best to see a recurring pattern across multiple cohorts, especially if it’s a long RevShare.
Add the cohort size next to the percentage
A D30 of 40% in a cohort of ten users and 32% in a cohort of a thousand is a very different level of confidence. The size of the group should always be visible next to the retention.
Rate comparison
If the terms of RevShare change, do not mix the cash result of the cohorts without adjustment. To assess the quality of the audience, it is rateter to look at the NGR on the FTD or other metric before applying the rate, if available.
Cohort comment
Add to each week one short note: a major promotional period, a change of creativity, a new placement. A few months later, these comments explain the schedule better than the team's memory.
