Why Sitewide Retention Hides Cohort Decay (and How to Model Around It)
Sitewide retention rates can stay flat for quarters while newer cohorts quietly decay — here's how to spot it in the curves before your LTV forecast lies to you.
Quick answer
A sitewide retention rate is a weighted average across every cohort you've ever acquired. Older, loyal cohorts can prop up the number long enough to hide the fact that cohorts acquired in the last 6-12 months are decaying faster. Always check Month-2 and Month-6 retention by acquisition cohort before you trust any compounding LTV model — including this site's calculator.
Sitewide retention masking cohort decay
When a stable blended retention rate disguises the fact that newer customer cohorts are retaining worse than older ones.
Sitewide retention is the share of customers active in a given period divided by customers active in the prior period, averaged across every cohort on the books. Because older cohorts (acquired during cheaper-CAC, higher-intent eras) often retain at structurally higher rates, their volume can keep the blended number flat for a year or more while Q3 and Q4 cohorts decay 10-20 points faster at the same age. The classic symptom: your dashboard shows 42% annual retention quarter after quarter, but the LTV you forecast from it keeps overshooting actual revenue from recent cohorts. The fix is to stop modelling LTV from a single rate and start modelling from cohort curves.
This is one of the most expensive mistakes a Shopify store can make above €2M revenue. You're forecasting payback windows, contribution margin, and paid-acquisition headroom off a number that's lagging reality by two or three quarters.
Why the average lies
A blended retention rate is a weighted mean. The weights are cohort sizes. If 2021 and 2022 cohorts are large and retain at 55%, and 2024 cohorts retain at 38%, the blended number for 2024 can still print around 45% because the older cohorts have years of compounding volume behind them.
The mechanism is mix-shift, not improvement. Channel mix changes (more Meta prospecting, less brand search), discount depth changes, and product-line expansion all pull new cohorts toward lower-intent buyers. The average doesn't move until the old cohorts age out — at which point you've already overspent on acquisition for 18 months.
The 42% problem
If your annual sitewide retention has been within ±2 points for four quarters straight, you almost certainly have cohort decay hiding underneath. Genuinely stable retention is rare; what's common is decay being masked by a slow-moving denominator.
How to detect it in your data
Pull repeat-purchase rate by acquisition month, indexed to months-since-first-order. You want a triangle table: rows are cohorts, columns are M1, M2, M3, M6, M12. Compare each column down the rows — if M2 retention drops from 22% (Jan 2023 cohort) to 16% (Jan 2024 cohort), you've found decay.
Two signals matter most. First, the M2 column — second-purchase rate is the earliest leading indicator and moves before annual retention does. Second, the shape of the curve itself: if newer cohorts drop faster in the first 90 days but flatten at a similar tail, you have an onboarding problem, not a product-fit problem. Plot the curves on the same axes — see the Cohort LTV Curves page for the standard visualisation.
How to model around it
Stop feeding a single retention rate into a compounding LTV formula. Instead, fit a separate decay curve to each recent cohort (the last 6-8 acquisition months) and forecast forward from those. A simple power-law or exponential fit on M1-M6 actuals is enough — you don't need a survival model for a €5M apparel store.
When you use a retention-lift calculator, plug in the rate from your most recent fully-observed cohort, not the sitewide blend. If your Jul 2024 cohort is at M6 with 34% retention versus the blended 42%, model from 34%. The calculator's compounding will then reflect the cohort you're actually trying to improve, not a flattering historical average.
Rule of thumb
Use the trailing 3-cohort average of M2 retention as your input to any compounding LTV model. It's recent enough to reflect current acquisition mix, and averaged enough to absorb monthly noise from promotions or stockouts.
Experiments worth running
Once you've isolated which cohorts are decaying, target the drop-off point. If M2 is the leak, test post-purchase flows: a replenishment nudge at day 21 for consumables, a fit-feedback email at day 14 for apparel, a cross-sell at day 30 for beauty. Measure lift on the next cohort's M2 column — not on sitewide retention, which won't move for months.
Second test: segment by acquisition channel within the decaying cohorts. Often the decay is concentrated in a single source (Meta broad audiences, a specific influencer drop, a discount-driven Klaviyo flow). Pausing or restructuring that source recovers blended cohort quality faster than any retention email will.
Frequently asked questions
At minimum, six months of post-acquisition data on three consecutive cohorts. Less than that and seasonal noise (Black Friday, summer dips) will dominate the signal. Most Shopify stores already have this in their order history — you just have to slice it by acquisition month.
Churn is a per-customer event — they stopped buying. Cohort decay is a curve property — successive cohorts retain at lower rates than predecessors at the same age. You can have stable churn within each cohort and still have cohort decay if new cohorts start at a lower retention baseline.
Less, but not zero. Even for furniture or mattresses, the M12+ repeat rate funds your referral and cross-sell economics. If newer cohorts refer fewer friends or convert worse on accessory cross-sells, the same masking dynamic applies — it just plays out on a longer timeline.
The calculator takes a single retention rate as input. The correct workflow is to compute your trailing 3-cohort M2 retention separately, then feed that number in — rather than the sitewide blended rate your dashboard shows. The output then reflects the cohort you're trying to improve.
Different channels deliver different buyer intent. Brand search and organic typically retain 1.5-2x better than Meta broad prospecting. As paid scales and brand stays flat, the mix tilts toward lower-intent buyers, and the average retention of new cohorts drifts down — even if each channel's own retention is stable.
It's a close cousin. Simpson's paradox is the formal statistical case where an aggregate trend reverses when you disaggregate. Cohort masking is the operational version: the aggregate looks fine, but the segments tell a worse story. The fix is the same — stop trusting aggregates over heterogeneous populations.
Monthly is fine for stores under €10M revenue. You're looking for trend shifts across 3-6 month windows, not week-to-week movement. A monthly cadence catches mix-shift early enough to act before two more cohorts have been acquired under bad economics.
Then your sitewide rate is overstating decay rather than hiding it — the opposite problem, but still a reason not to trust the blend. The fix is identical: model LTV from recent-cohort curves, not from the average. Either way, the aggregate is misleading you about the customers you're acquiring today.
Yes — Black Friday cohorts retain very differently from June cohorts. Compare year-over-year (Nov 2023 vs Nov 2024) rather than month-over-month when seasonality is strong. For apparel and gifting categories this is mandatory; for replenishment categories like beauty consumables it matters less.
Within ±2 percentage points across consecutive cohorts is normal noise. A consistent 3-5 point decline over three or more cohorts is a real signal. Anything beyond 5 points between adjacent cohorts usually traces to a specific event — a discount campaign, a product launch, or a paid-channel expansion.
Test ideas before you ship them
Run unlimited A/B tests, attach hypotheses to outcomes, and build a searchable archive of what works — and what doesn't.