Retention Curve Shapes Decoded: Smiling, Flattening, Decaying
The short version: a retention curve's shape tells you whether your product has a floor. A curve that keeps sliding toward zero means no one sticks. A curve that drops and then flattens at some non-zero level means you've built a habit for a real slice of users. And a curve that drops and then climbs back up (the "smile") means churned users are coming back on their own, which is the rarest and best thing a chart like this can show you.
Most people read retention curves for the wrong number. They fixate on the first data point, the day-1 or week-1 drop, and panic. The drop is mostly noise. The information is in where the curve goes flat, and how high.
What a retention curve actually plots
A retention curve takes one cohort of users, everyone who signed up in the same week, and tracks what fraction of them are still active over time. Week 0 is 100% by definition (they all just signed up). Week 1 might be 40%, week 4 maybe 22%, week 12 maybe 18%. Plot those points, connect them, and the line you get is your retention curve.
The x-axis is time since signup. The y-axis is percent of the original cohort still active. "Active" is whatever you decide it is, and that choice matters more than people admit. If "active" means "opened the app," your curve looks generous. If it means "completed a core action," it looks honest. Pick the definition that maps to real value, ideally the same metric you'd anchor a north-star metric tree on, and then never quietly change it.
One more thing before the shapes. A single cohort is a bumpy line. What you want is the shape across many cohorts, or a smoothed average, because week-to-week wobble from a small cohort will fake a trend that isn't there.
The three shapes, decoded
There are three shapes worth naming. Churnkey's 2024 retention guide lays out the same taxonomy, and it's become the standard vocabulary, so I'll use it.
| Shape | What the line does | What it means | Verdict |
|---|---|---|---|
| Decaying | Drops and keeps dropping toward zero | No stable core; every cohort eventually leaves | Product-market fit not found |
| Flattening | Drops, then levels off above zero | A habitual core exists at the plateau height | Healthy; the goal for most products |
| Smiling | Drops, bottoms out, then climbs | Churned users are resurrecting on their own | Rare, usually network effects |
The decaying curve is the one that ends careers, and it's the one founders most often lie to themselves about. If your curve is still visibly sliding at week 20, there is no plateau, there's just a slower approach to zero. A slower slide is not a floor.
The flattening curve is what you're actually aiming for. Somewhere in the first four to eight weeks the curve stops falling and settles onto a plateau. That plateau is your core: the people who found something they keep coming back for. Sequoia Capital's retention primer makes the point bluntly, that a curve flattening at a number greater than zero is the sign the product has struck a balance between churn and resurrection. The higher the plateau, the more of each cohort you keep forever.
The smiling curve is the unicorn. The line drops, bottoms out, and then turns upward, because more previously-churned users come back each period than new ones leave. That usually takes a network effect: the product gets more useful as more people use it, so someone who bounced in month one has a reason to return in month six. Sequoia notes these ascending curves show up in exceptional products during hyper-growth. If you have one, protect it. If you don't, don't fake one by resetting cohort definitions until the math smiles for you.
A worked example with 12 users
Benchmarks in the abstract are slippery, so here's a tiny concrete dataset. Twelve users sign up in the same week. We track "completed a core action" over three weeks.
- Week 0: all 12 are active (100%).
- Week 1: 5 are still active (42%).
- Week 2: 4 are still active (33%).
- Week 3: 4 are still active (33%).
Read the shape, not the drop. Yes, we lost more than half the cohort in week one. That's expected; some users were never going to stick regardless of what we built. The interesting event is weeks 2 and 3 holding steady at 4 users. The curve flattened. Those four are our core, and the plateau height for this cohort is 33%.
Now imagine a second product with the same week-1 number but a different tail:
- Week 0: 12 (100%). Week 1: 5 (42%). Week 2: 3 (25%). Week 3: 2 (17%).
Same start, same week-1 retention, completely different product. The first one found a floor at four users. The second is still bleeding and has no floor in sight. If you'd stopped reading at week 1, you'd have called these two products identical. They are not.
The 2x2 I actually use: plateau height vs. decay slope
Here's the framing I wish someone had drawn for me in year one. Every retention curve has two properties that matter independently, and it helps to think of them on two axes:
- Plateau height — how high the curve flattens out. This is what fraction of every cohort becomes permanent. High plateau means a big durable base.
- Decay slope — how fast the curve is still falling as it approaches the plateau, or whether it's approaching one at all. A steep, still-falling slope late in the curve means no plateau exists yet.
Cross them and you get four quadrants:
| Shallow / flattened slope | Steep / still-falling slope | |
|---|---|---|
| High plateau | The dream: big base, stable. Pour fuel on acquisition. | Mirage: looks great early, still leaking. Fix the leak before scaling. |
| Low plateau | Real but small: a niche that sticks. Decide if the niche is big enough. | Danger: small and shrinking. This is the pre-death pattern. |
The quadrant most teams misread is top-right, high plateau but still-falling slope. Early cohorts look fantastic because the curve hasn't had time to reveal it's still sliding. Acquisition covers the leak. Then growth slows, the leak surfaces, and everyone's surprised. There's nothing to be surprised about; the slope was never flat, you just hadn't waited long enough to see it.
Why flattening beats day-1 retention
This is the opinion I'll plant a flag on, and I know some growth folks will push back: for durable growth, a flattening curve at a modest height beats a taller curve that's still decaying. Full stop.
Day-1 retention is a hook metric. It tells you whether onboarding lands, and it's easy to move with a better first-run experience, a push notification, an email. All useful. But a great day-1 number sitting on top of a decaying curve is a bucket with a hole, filled faster. You can win the first day and still lose every cohort.
The flattening matters because growth compounds off the floor, not the peak. If your curve flattens at 20%, then one-fifth of everyone you ever acquire becomes permanent, and permanent users stack. Acquire 1,000 a month against a 20% floor and your base climbs by roughly 200 every month, forever, before you've improved anything. Against a curve that decays to near-zero, that same 1,000 a month buys you a treadmill.
The benchmark data backs the "don't over-index on the early number" instinct. Mixpanel's 2024 Benchmarks Report, drawn from more than 7,700 companies and 11.7 trillion events, found average week-one retention across industries fell from about 50% to 28% during 2023. Week-one numbers moved a lot, industry-wide, for reasons that had little to do with any single product's health. The plateau is the steadier signal.
And plateaus vary wildly by category, which is why a raw percentage means nothing without context. Lenny Rachitsky and Casey Winters's benchmark study of leading growth practitioners pegged good six-month retention at 25% for consumer social and 70% for enterprise SaaS, with "great" at 45% and 90% respectively. A 30% plateau is mediocre for a workplace tool and excellent for a consumer app. Read your curve against your category, never against a universal number.
Where this goes wrong
A few failure modes I've watched play out more than once.
Reading a single noisy cohort as a trend. Twelve users bouncing around week to week will draw any shape you want to see. Smooth across cohorts or wait for a bigger sample before you name the curve.
Changing the "active" definition mid-analysis. Loosen the bar and every curve lifts. Someone will be tempted to redefine active from "core action" to "any open" right before a board meeting. Don't. You're not measuring the same thing anymore, and you'll have poisoned every historical comparison.
Calling a slow decay a plateau. The most expensive mistake. A curve at week 12 that's declining by two points every four weeks is not flat, it's just slow. Extend the time window. If the late-stage slope isn't near zero, you don't have a floor yet.
Averaging away a bimodal product. Sometimes one blended curve hides two populations: a segment that sticks hard and a segment that never activates. The average looks like a gentle decay; the truth is a high plateau for one group and a cliff for the other. Segment before you conclude. The fix lives in the segments, not the average.
A short FAQ
How many weeks of data do I need to judge the shape? Enough for the curve to stop moving. For many products that's the first eight to twelve weeks, but a slow-burn B2B tool might not reveal its plateau for two quarters. If the late slope is still visibly falling, you're not done watching.
Is day-1 retention useless, then? No. It's a good early-warning signal for onboarding, and it's fast to move. Just don't mistake it for the health of the whole curve. It's the first data point, not the verdict.
What if my curve genuinely never flattens? Then you haven't found product-market fit for that cohort yet, however good the top of the funnel looks. A decaying curve is the clearest "not yet" a chart can give you, and the honest move is to treat acquisition spend against it as expensive research, not growth.
The whole discipline here is one habit: stop reading the drop, start reading the floor. Where the curve goes flat, and how high, is the number that decides whether you have a business or a bucket with a hole in it.