Reading Retention Curves Without Fooling Yourself on LTV

Most LTV numbers are confidence laundered into a forecast. A model fits a decay curve to a few months of orders, extrapolates the tail to infinity, multiplies by average order value, and hands you a figure big enough to justify almost any acquisition cost. The number feels rigorous because it came out of math. It is still mostly a guess about a future you have not observed yet — and it is denominated in revenue, which you do not get to keep.

The fix is not a better model. It is a discipline: read the survival curve for what it actually shows, stop counting money you do not bank, and refuse to spend on the part of the tail you have never seen.

For the adjacent growth decisions, compare Is Your Loyalty Program Incremental, or Taking Credit? and then use How to Use Data-Based Marketing Tools to Scale E-Commerce Sales to pressure-test the operating plan.

Predicted LTV is the vanity figure

Predicted lifetime value has three compounding problems, and they all push the number in the same direction — up.

• It extrapolates beyond your data. If you have eight months of history, everything past month eight is the model’s opinion, not your customers’ behavior. Curve-fitting can be optimistic about tails because a smooth function does not know your category has a natural ceiling on repeat purchase.
• It is gross, not net. Revenue LTV ignores cost of goods, shipping, payment fees, returns, and the discounts you used to drive each repeat order. The money a customer “is worth” and the money that reaches your bank are different by a wide margin.
• It hides cohort mix. A blended LTV pools a strong early cohort with a weak recent one and reports a flattering average that describes no real customer.

A predicted figure is fine for a board slide. It is dangerous the moment it sets a target CPA, because you will pay real acquisition cost today against modeled contribution you may never collect.

Read the survival curve, not the spreadsheet cell

A cohort LTV retention curve is a survival curve: take everyone acquired in the same window, then plot what fraction is still ordering in month 1, month 2, month 3, and on. The single LTV number is a summary of that curve. The curve is the truth; the number is the rumor.

Two things to separate immediately, because they get conflated and the conflation flatters you:

Logo retention versus revenue retention

Logo retention asks: what share of the cohort placed any order this period? Revenue retention asks: what share of the cohort’s original spend did this period generate? These diverge constantly. A cohort can lose buyers every month while the survivors spend more, so revenue holds even as the customer base thins. That looks healthy on a revenue chart and is quietly fragile — you are increasingly dependent on a shrinking core. Plot both. When logo retention falls faster than revenue retention, your “LTV” is being propped up by a handful of heavy repeaters, not by a durable base.

Where the curve actually flattens

Every healthy repeat-purchase business has an inflection where the steep early drop-off gives way to a near-flat plateau — the loyal residual that keeps buying. That plateau is the only part of LTV you can defend, because it is the part you have measured rather than projected. The honest read is: find the period where the period-over-period retention delta stops shrinking and goes roughly flat, and treat the cumulative contribution up to that point as your evidence-based LTV. Everything to the right of the flattening is forecast, and forecast is a planning input, not a banked asset.

If your curve has not flattened yet — common for younger cohorts or long purchase cycles — then you do not have an LTV. You have a slope and a hope. Say so out loud rather than letting a model fill the gap with optimism.

Discount revenue to contribution

This is the step that separates operators from dashboards. Walk each period of the curve down to what you actually keep:

1. Start with revenue per surviving customer in that period.
2. Subtract cost of goods sold.
3. Subtract variable fulfillment — pick, pack, shipping, the slice of returns and refunds that hits that cohort.
4. Subtract payment processing and the repeat-purchase incentives you spent to trigger those orders.

What remains is contribution. Sum contribution across the periods up to where the curve flattens, and you have an honest cohort LTV retention curve: a survival curve denominated in money you keep, truncated to the horizon you have observed.

A quick worked example, with clean ratios so the logic travels. Say modeled revenue LTV lands at a 4.0 multiple of first-order value. Apply a 45% contribution margin and you are at 1.8 in contribution terms. Now truncate the projected tail to the period you have actually observed — keep, say, 70% of that — and the defensible figure is closer to 1.25. Same cohort, same data, but the number you should underwrite acquisition against is roughly a third of the headline. The 4.0 was never wrong arithmetic; it was the wrong question.

The discipline matters most where margins are thin. A category running 35% contribution and heavy discounting to drive repeats can have a beautiful revenue curve and a contribution curve that flattens barely above the first order — meaning there is almost no banked future value to spend against, no matter what the model’s tail says.

How to build the honest version, step by step

1. Fix the cohort window. Group by acquisition period — weekly for fast cycles, monthly otherwise. Never blend cohorts of different ages into one curve.
2. Plot both retention lines. Logo and revenue retention on the same axis. Watch the gap.
3. Convert each period to contribution, not revenue, using the subtraction above.
4. Find the flattening point. Mark where the period-over-period retention delta stops shrinking. That is your observed horizon.
5. Sum contribution to the flatten point. That cumulative number is your defensible LTV.
6. Label the tail as forecast. If you need to include projected periods, show them as a separate, clearly-marked band — never folded silently into the headline figure.
7. Re-read monthly. Newer cohorts shift the picture. A figure that was honest two quarters ago can be stale now.

This is exactly the kind of read that is easy to get wrong by hand and easy to fool yourself on under deadline. A read-only operator layer like Bach can hold the cohort split, separate logo from revenue retention, and force the contribution discount before any number reaches a CPA target — surfacing the honest figure for you to approve rather than letting a flattering blended average set the spend.

What changes when you do this

Your allowable acquisition cost drops, and it should. You will pay against banked contribution to the flattening point, not against an extrapolated revenue tail. Channels that looked profitable on modeled LTV frequently stop looking profitable on contribution LTV — that is the read working, not failing. And you stop the common self-inflicted wound in performance marketing: scaling spend into a payback that only exists in the model’s imagination.

The takeaway: an honest cohort LTV retention curve is contribution-denominated and truncated to where the curve flattens. Predicted LTV tells you what a cohort might extrapolate to. Contribution-to-flatten tells you what it has actually given you. Underwrite acquisition against the second number, and treat the first as a hypothesis you are still testing — never as money in the bank.