Cohort Retention by Acquisition Source: Which Channels Bring Customers Who Stay

The standard patch is to apply an average customer lifetime to every channel: take blended monthly churn, invert it, multiply by average revenue, call it lifetime value, divide by cost per acquisition. This is worse than no answer, because it looks like an answer.

The arithmetic assumes a constant hazard rate, which retention curves reliably violate. Early-month churn is almost always steeper than late-month churn, because the population that survives month three is systematically different from the population that entered it. A single averaged lifetime overstates the value of high-early-churn channels and understates the value of slow-and-steady ones, which is precisely the comparison you were trying to make.

Worse, applying the blended rate to every channel assumes the thing you set out to test. If channels retained identically there would be no reason to run this analysis at all.

A cohort is a group of customers who share a starting event, tracked forward together. For retention work the starting event should be the first successful payment, not the signup and not the trial start.

Using signup as the anchor mixes trial abandonment into retention, and those are different problems with different owners. Trial-to-paid belongs to the funnel; retention begins when money starts moving. Keeping them separate lets you tell a channel that converts poorly but retains superbly apart from one that does the reverse, and those two channels want opposite interventions.

The second dimension is the source that acquired the customer, taken from the attribution on their first payment. Every subsequent measurement is then a cell in a grid: this source, this starting month, this many months later.

Gross revenue retention excludes expansion, so it is capped at 100 percent and measures only what you lost. Net revenue retention includes expansion and can exceed 100.

For channel comparison, compute both. Net alone can hide a serious churn problem inside one or two large expansions, particularly in a small cohort where a single account upgrading materially can lift the whole line. Gross tells you what the channel loses; net tells you what it is worth. A channel with a weak gross curve and a strong net curve is carrying concentration risk, and that belongs in the decision.

Metrivo's cohort retention exposes customer and revenue modes separately for this reason, so expansion is visible as its own effect rather than folded into a single number.

A drop in a retention curve does not distinguish a customer who left from a customer whose card failed. Both stop paying. Both leave the cohort. On the chart they are the same pixel.

This matters for channel comparison specifically, because payment-failure rates differ by channel through payment-method mix and geography. A channel that appears to retain poorly may be retaining fine and failing at the card. That is a fixable billing problem, not a reason to cut spend, and the two conclusions lead to opposite actions.

Before you draw conclusions from a channel's retention curve, subtract involuntary churn from it. The involuntary churn audit covers how to measure that separately so it can be removed from the retention story rather than silently distorting it.

The tempting source of truth is your provider's subscription status. It is unreliable for retention because status transitions depend on configuration: when a delinquent subscription is marked cancelled is a setting you chose, and it differs across providers, which makes multi-provider retention incomparable.

Derive from payments instead. A customer who paid this month is retained this month. A customer whose expected billing window passed without a payment has churned, whatever their subscription status says. This definition is provider-independent and does not move when someone changes a dunning setting.

Metrivo classifies derived subscription events into first payment, renewal at the same amount, expansion above the previous amount, contraction below it, churn when a subscription stops producing payments within its expected billing window, and reactivation when a previously churned customer starts paying again. Retention is then computed from that classified series rather than from provider status fields.

Cohort month is the calendar month of the first successful payment. Use UTC month boundaries consistently; a customer whose first payment lands at 23:40 local time on the last day of a month can otherwise fall into different cohorts depending on which timezone the query ran in.

Source is taken from the attribution on that first payment and then frozen. It does not update if the customer later arrives through a different channel, because the question is which channel acquired them. Freezing it also prevents a customer from silently migrating between cohorts and making historical curves unreproducible.

Carry the confidence label with the source. A cohort assembled mostly from inferred matches is a weaker piece of evidence than one assembled from confirmed ones, and the distinction is documented under attribution confidence. Where the source is unknown, keep those customers in an explicit unknown cohort. Excluding them quietly biases every other cohort, because unattributed payments are rarely a random sample.

For each cohort and each month offset, the logo retention cell is the count of customers from that cohort with a successful payment in that month, divided by the cohort's original size. The revenue retention cell is the sum of monthly revenue from that cohort's surviving customers in that month, divided by the cohort's original monthly revenue.

Handle reactivation explicitly. A customer who churns in month three and returns in month seven appears in month seven's cell under one policy and not under another. Either is defensible; both must be applied to every cohort. Metrivo classifies reactivation as its own event type, so it can be included or excluded deliberately rather than by accident.

Do not renormalise the denominator as cohorts age. The denominator is always the cohort's original size and original revenue. Dividing by the current surviving population produces a chart that always trends toward 100 percent and means nothing.

Retention tables grow without limit and stop being readable long before they stop growing. A rolling window of 24 months is enough for almost any decision a founder-led SaaS will make, and Metrivo's cohort retention caps at that.

The practical reason to bound it is that very old cohorts describe a product and a market that no longer exist. A cohort acquired three pricing changes and two positioning shifts ago is a historical artefact, not a guide to this quarter's budget.

Even above the floor, a raw channel comparison is confounded, because channels differ in ways that independently drive retention.

Plan mix is the largest. Annual plans mechanically retain better than monthly ones over a twelve-month window, because they cannot churn mid-term. A channel that skews annual will show a better curve regardless of customer quality. Segment by billing period before comparing anything.

Geography drives both payment-failure rates and price sensitivity. Company size drives expansion, which drives revenue retention specifically. Product entry point matters: customers who arrive at a feature page and adopt one workflow retain differently from those who arrive at a pricing page and adopt several. And tenure is a confounder of its own, since a channel you started six months ago has no mature cohorts and will look worse purely from the shape of the table.

The workable discipline is to segment on the one or two confounders that plausibly differ most between the channels you are actually comparing, rather than attempting a full model. If a channel's advantage survives segmentation by plan and geography, it is probably real. If it disappears, you learned something more useful than the original ranking.

The right comparison horizon is not a best practice, it is a cash-flow constraint. A company with eighteen months of runway and no external funding cannot treat a channel that pays back in month fourteen as equivalent to one that pays back in month four, even if the fourteen-month channel eventually delivers more revenue per customer.

State the horizon explicitly and rank within it. If you can finance six months of payback, rank channels by cumulative revenue at month six and note which ones are still improving after it. That way the ranking answers the question you can actually act on, and the longer-horizon information is preserved rather than discarded.

The interesting case is a channel with excellent cost per acquisition and a poor retention curve. The instinct is to cut it. That is often wrong, because the cause determines the fix.

If the channel attracts a segment your product genuinely does not serve, cutting is correct. If it attracts the right segment through a promise the product does not meet, the fix is the landing page and the messaging, and it is cheaper than replacing the channel. If it attracts the right segment who then fail to activate, the fix is onboarding. If it attracts the right segment whose payments fail disproportionately, the fix is billing. Only the first of these is a channel problem at all.

Distinguishing them requires looking at where in the curve the loss happens. Losses concentrated in months one and two point to activation or expectation mismatch. Losses spread evenly across months point to ongoing value or billing problems. That shape is visible on the curve you just built, which is a large part of why building it is worth the effort. The SaaS funnel drop-off analysis guide covers the pre-payment half of the same investigation.

You will not have twelve months of data for a channel you started last quarter, and waiting is not free. The workable interim is to compare young cohorts only against other cohorts at the same age, never against mature ones.

Month-one and month-two retention are available quickly and are genuinely informative about early churn, which is where most channel differences concentrate anyway. A channel losing 30 percent of its customers by month two is telling you something now, and the shape of month twelve will not reverse it.

Record the immaturity in the report. A cell that is empty because the month has not happened yet is not the same as a cell that is empty because everyone churned, and a table that does not distinguish them will eventually be read wrongly by someone who was not in the room when it was built.