Reading cohort retention curves: cliff, shoulder, plateau
Every subscription retention curve has three regions: a cliff in the first one or two renewals, a shoulder where the slope eases through roughly month 6, and a plateau where survivors churn slowly and mostly involuntarily. The broadest public anchor for subscription commerce comes from Recharge, whose 2023 report averaged 45% subscriber retention at month 6 and 33% at month 12. A healthy curve flattens. A sick one keeps sliding at the same rate forever.
A retention curve is the most information-dense chart a subscription store can draw, and most merchants never draw it. They stare at a single blended churn percentage instead, which averages the three regions together and hides all of them. This post is about the shape: what each region means, what healthy and sick look like against the public data that exists, and which lever moves which segment of the curve.
Draw it before you read it
If you have not built the curve yet, it is one spreadsheet exercise: group subscribers by signup month, count survivors at each month of age, divide by the cohort's starting size, and plot the percentages. The full column-by-column recipe, plus the LTV math that sits on top of it, is in LTV math you can do in a spreadsheet - we will not repeat it here. One drawing rule matters for reading: plot each cohort as its own line, or at minimum split old and new cohorts. A single averaged line mixes January's mature behavior with June's incomplete data and produces a shape that belongs to no actual customers.
The cliff: months 0 to 2
The cliff is the steep drop across the first renewal or two, and it is where most of your lifetime churn happens. This is not a WooCommerce quirk. Baremetrics' LTV guide flags the same pattern in SaaS cohorts: a cliff right after the first month, sharp enough that they recommend discounting naive LTV projections for it. In physical subscriptions the mechanics are concrete: the first box either matched expectations or did not, the default cadence either fit consumption or did not, and the customer either formed the habit before renewal number two or the charge arrived as a surprise.
In the illustrative curve we used in the LTV post - calibrated to Recharge's cross-vertical averages - a 100-person cohort drops to 76 after the first renewal and 66 after the second. Half of everyone who will leave in the first year has left within 60 days. That concentration is the single most useful fact about the curve, because it tells you where intervention capacity should go.
The cliff is also where cancel-flow offers do their work, since cliff churn is dominated by voluntary, reasoned cancellation. The scale of what is recoverable there is documented: Churnkey's State of Retention 2025, built on 3 million cancellation sessions, reports that when offers were presented, discounts were accepted 53.9% of the time and pauses 19.2%. We have covered which offer belongs with which cancel reason in pause vs discount.
The shoulder: months 2 to 6
The shoulder is the region where the slope should be visibly easing, month over month. Customers who cleared the cliff have some habit formed, but the subscription is still being re-evaluated: product piling up, seasonal use, a card statement review. Shoulder churn is quieter than cliff churn and more often logistical - wrong cadence, wrong quantity - than dissatisfaction.
Read the shoulder as a derivative, not a level. The question is not "how many are left at month 4" but "is each month's loss smaller than the last". If month-over-month losses are shrinking - 10 points, then 8, then 5, then 4 - the curve is bending toward a floor. If losses hold constant at 6 or 7 points every single month, you do not have a shoulder. You have a straight line, and straight lines are the signature of a sick curve, because they extrapolate to zero.
The plateau: month 6 onward
The plateau is where the curve goes nearly flat and every remaining subscriber is habituated. Churn here should be a slow leak, and its composition changes: a large share of it is involuntary - failed payments, expired cards - rather than decisions. The public data on that split is consistent. Recurly's churn benchmark research (July 2026 network data) puts median annual churn at 3.60% overall, of which 1.25 points are involuntary - roughly a third of all churn. For the ecommerce segment specifically the split is 2.87% voluntary against 1.38% involuntary.
The practical consequence: plateau churn responds to payment tooling, not to offers. Churnkey's 2025 report found insufficient funds alone caused 42.3% of declines, and that a full recovery stack - retries, dunning email and SMS, failed-payment walls - recovered about 70% of the involuntary churn it detected. A merchant who points a discount offer at the plateau is treating a billing problem with a pricing concession.
Healthy vs sick, against public data
The honest caveat first: public cohort-curve data for subscription commerce is thin. Recharge's cross-vertical averages - 45% retained at month 6, 33% at month 12, from 2022 merchant data - are the widest anchor we know of, and category spread around them is large. Replenishment products hold dramatically better floors than discovery boxes, so check the WooCommerce churn benchmarks post for category-level context before judging your own numbers. The early ChurnStop install cohort is nowhere near large enough for us to publish curves from, so shape signals below are what the cited sources and the math support, not a proprietary dataset.
| Signal | Healthy curve | Sick curve |
|---|---|---|
| Month 1 drop | Steep but bounded; the worst month by far | Mild month 1, then losses that never shrink |
| Slope by month 6 | Visibly flattening; each loss smaller than the last | Constant 5+ points lost every month |
| Month 6 level | At or above the ~45% cross-vertical average | Well below it with no category excuse |
| Month 12 level | At or above ~33%, curve near flat | Sliding through 20% and still falling |
| Plateau churn mix | Mostly involuntary; small, stable leak | Still voluntary-heavy after month 6 |
| Cohort over cohort | Newer cohorts sit above older ones | Newer cohorts sit below older ones |
Two of these deserve emphasis. The flattening test beats any absolute level: a store at 38% month-6 retention with a hard flat plateau is healthier than one at 48% that is still losing 5 points a month, because only the first curve has a floor. And the cohort-over-cohort test is your only real experiment readout: if the cohorts that signed up after you fixed onboarding or added a save flow sit above the ones before, the fix worked. If every new cohort lands lower, your acquisition mix is deteriorating faster than your retention work is compounding.
Same store, different cohorts
Curves differ across cohorts for reasons that have nothing to do with your retention work, and you need to name those before crediting or blaming a change.
Acquisition mix is the big one. A cohort acquired through a 40%-off promotion contains a different kind of customer than one acquired at full price from organic search. The promo cohort cliffs harder because some fraction subscribed for the deal, not the product, and leaves the moment list price appears on a statement. If you ran a big discount push in May, the May curve sitting below April proves nothing about your product or your flow.
Seasonality is the other. Cohorts acquired around gift-heavy periods behave differently: gifted subscriptions and impulse holiday signups carry weaker intent than a self-selected February buyer. Compare December against last December, not against October.
The clean way to handle both is annotation, not adjustment. Keep a row under the curve noting what acquisition looked like for each cohort - promo depth, channel mix, anything unusual. When a cohort deviates, check the annotation before the retention theories. This is also the strongest argument for judging your curve on the flattening test rather than absolute levels: acquisition noise moves the whole line up or down, but it rarely changes whether the line finds a floor.
Which lever moves which region
Match the intervention to the region where your curve deviates, because each one only reaches certain months.
| Curve problem | Likely cause | Lever that reaches it |
|---|---|---|
| Deep cliff, months 0-2 | Expectation mismatch, cadence default too fast | First-order experience, cadence choice at checkout, save flow at first renewal |
| No shoulder bend, months 2-6 | Product piling up, wrong quantity | Skip and swap options, pause offers, cadence editing |
| Leaky plateau, month 6+ | Failed payments, expired cards | Retries, dunning, card updater - billing tooling, not offers |
| Whole curve shifted down | Acquisition quality, discount-hunting cohorts | Fix the traffic mix before the flow |
A save flow is a cliff-and-shoulder tool. It intercepts voluntary cancellation decisions, which is where those concentrate. If you want to estimate what bending the cliff is worth in revenue terms, the save rate impact calculator runs that math against your own subscriber counts.
Four traps that fake a curve
Each of these produces a plausible-looking chart that misreads reality.
- Right-censoring. A cohort that is 4 months old has no month-6 data point. If your spreadsheet treats missing as zero, or averages incomplete cohorts into the curve, recent months look catastrophically bad. Plot cohorts only as far as they have lived.
- Pending cancellation counted as active. WooCommerce Subscriptions holds cancelled-mid-term subscribers in a pending state until the prepaid term ends. For curve purposes they have churned the day they clicked; counting them as active shifts your cliff artificially late and understates it.
- Mixed billing intervals. An annual subscriber cannot churn in month 3. Blending annual and monthly plans into one curve flattens the cliff by construction. Draw one curve per billing interval, always.
- Small cohorts. A 25-person cohort moves 4 points per person. Do not read shape from it; merge two or three adjacent months into one cohort until each group is at least 50 to 100 subscribers, and treat anything smaller as directional.
What to plot this week
- One line per monthly cohort, monthly billing only, pending-cancels counted as churned, cohorts under 50 merged.
- Mark three numbers on it: month-1 retention, month-6 retention, and the last three months' slope.
- Ask the three questions in order: Is the cliff bounded? Does the slope shrink through the shoulder? Is the plateau leak involuntary?
- Fix the region that fails, with the lever from the table above - and re-read the cohort-over-cohort ordering in three months to see whether it worked.
The blended churn number on your dashboard is the average of three different stories. The curve tells you which story is yours.