An affiliate that delivers 500 first-time depositors is not automatically worth more than one that delivers 200. The volume gap says nothing about what happens in week two.
The larger group can disappear after the first deposit, consume bonuses, generate chargebacks and leave weak net gaming revenue. The smaller one can stay active for months, make repeat deposits and carry a higher lifetime value.
A period summary hides this. It rolls everyone into one number. Cohort analysis splits the traffic into groups and shows which group stays alive and which goes dark.
What counts as a cohort
A cohort is a group of players who share a trait at entry. Most often it is the date of the first deposit: everyone who funded an account for the first time in the same week or month sits in one group, and that group is tracked over time.
The trait can be something else: traffic source, offer, geo, creative, a single affiliate. For a media buyer the useful cut is date plus source. Mix sources inside one group and the conclusions blur.
Active share. How many players from a cohort return to play after a week, a month, two months. Return to play, not just sit in the database.
Repeat deposits. How many first deposits turn into a second and a third. That gap between FTD and repeat is the signal: a player came to play, or came for the bonus.
NGR per player. Not the total NGR on the offer, but revenue measured against the size of the cohort. It makes partners with different volumes comparable.
Bonus abuse. The share of players who withdraw a bonus and leave. Better counted by cohort than across the whole offer, because abuse usually arrives as one stream.
Chargebacks and refunds. Bad traffic is not visible at once. Payments get reversed later than the first deposit clears. A date-based cohort shows which source throws up a spike in reversals after time.
Setup. No separate BI system is needed. A table will do: a row is a cohort, a column is a period of its life, the cells hold the metrics above. It is built from an affiliate panel export or from your own postbacks, if they are written to storage. Discipline matters more than tooling. The same cohort has to run through the report for months, not be rebuilt each time. Otherwise there is no trend to see.
Payout model. On pure CPA you pay for an event and see nothing after it. Cohort analysis becomes the way to find where you overpay. If a stream with a high count of first deposits shows weak activity and low NGR, the price per FTD in it is inflated, even when the CPA itself looks normal against the market.
On revshare the picture is fairer but late. Revenue arrives over months, and without a cohort split you cannot tell which source produces it. Revshare without cohorts is an average across the offer.
Hybrid deals with CPA plus a share of NGR need both pictures: the event at entry and the behaviour after it.
Where cohort analysis breaks
Attribution lags. A player registers in one period and makes the first deposit later. Assign him to the registration date and the cohort looks dead, when people simply have not reached the cashier yet.
Postbacks go missing. Some events never reach the affiliate panel. A cohort built on partial data gives a false verdict on a source.
Thin data. A cohort with few players is noisy. It says something about a slice of time, not about the stream.
Outside events. A change in bonus policy, new limits, a seasonal dip hit cohorts unevenly. Comparing groups that lived under different conditions is risky.
A summary answers how many. Cohort analysis answers what happened to them. The first is for reporting, the second for buying decisions.
If a source delivers more first deposits but fewer active players by month three, the call is clear: too early to scale. Check the offer terms, the creative, the bonus setup. Do not raise the rate.
The reverse holds too. A stream with fewer FTDs that keeps activity for months and produces repeat deposits is worth protecting. Buy it deeper. Lifetime value is higher even if the entry price looks steep.
