Find the retention smile by detecting where the curve flattens

The Retention Smile Is Where the Math Stops Being Pessimistic

A retention curve that trends to zero is a leaky bucket; the team can only outrun it with acquisition spend. A retention curve that flattens at a meaningful level is a business; every cohort adds enduring value. Most teams chase the leak by improving onboarding, raising D7, or tweaking notifications, while the higher-leverage move is to identify the day the curve flattens and find the behavior that produces it.

Amplitude's writing on retention frames the curve as the most diagnostic chart in product analytics. The shape tells you whether the product has habit fit, niche fit, or no fit at all. Amplitude's cohort analysis primer shows how to read the shape across weekly cohorts to separate signal from noise.

Three curve shapes and what they mean

• Smile. The curve drops in the first days, then flattens at a meaningful level. The flatten point is the smile. Users who reach it tend to stay. The product has habit fit for that cohort.
• Slope to zero. The curve trends down across all periods. The product has not earned habit; growth depends on continuous acquisition.
• Bumpy. The curve oscillates because of seasonality, episodic content, or measurement noise. Smooth before drawing conclusions.

The team's job is to push more users into the smile by identifying the behavior that distinguishes smile users from leak users in the first week.

How the Find the retention smile prompt works

Step 1 plots the curve for the last 8 weekly cohorts. Same y-axis, same time intervals, side by side so the shapes are comparable. Eight cohorts is enough to spot a trend without burying the team in data.

Step 2 detects the flatten point per cohort. The rule of thumb is: where retention loses less than 1 percentage point per period. Some cohorts will not flatten (acquisition channel mix is wrong, or the cohort is leak users only). Most healthy products have at least some cohorts that flatten high.

Step 3 defines smile vs leak. The discipline is to define this from the curve shape, not from a single event. Teams that pick "users who used Feature X" up front bias the analysis toward Feature X. Letting the curve define the cohort exposes which behaviors actually predict survival.

Step 4 finds behavioral predictors. The first 7 days is the right window because behaviors there are causally upstream of week 4 retention. The output is 3-5 over-represented events, 1-2 absent-in-leak events, and the time-to-first-X for both groups. Reforge's writing on retention emphasizes that the predictor is usually a small number of behaviors, not a single one, and the time bound is as important as the event.

Step 5 validates with a holdout cohort. The strongest predictor is checked against the most recent two cohorts. If it still separates smile from leak, it is real. If it does not, the predictor was a marker of a particular cohort, not a durable lever.

Step 6 designs the experiment. The hypothesis names the predictor, the treatment is a specific onboarding change or in-product nudge, the sample size detects the smallest meaningful lift, and pre-registered guardrails prevent celebrating a noise win. Amplitude's writing on product onboarding collects examples of treatments that have moved the smile in real products.

Why "improve onboarding" without finding the smile rarely works

Onboarding changes that do not target the smile predictor tend to lift D1 or D3 without lifting the flatten level. The leak slows, the smile does not rise. After a few cycles, the team concludes that retention is "just a hard problem" and switches to acquisition. The audit is what prevents that conclusion.

SVPG's writing on what matters most makes the broader case: outcomes (retention rising) are the only durable measure of progress, and the team should be willing to throw out output (onboarding flow versions) that do not move outcomes.

When to use it

• The retention curve has not moved despite onboarding investment and the team is losing patience.
• A new persona or new market is being added and the team needs the smile shape before committing acquisition spend.
• A leadership review is asking whether the product is leaky or just under-served, and the team needs evidence either way.
• A previous "improve activation" project lifted D1 but not the flatten level, and the team needs a better target.
• A new growth lead is joining and wants a defensible retention diagnostic before installing experiments.

Common pitfalls

• Defining smile by a single event. Let the curve shape define the cohort, then ask what behaviors distinguish.
• Ignoring the time bound. A predictor without a time bound (must happen by day N) is a usage proxy, not an activation lever.
• Stopping at correlation. The validation cohort and the experiment are what prove the predictor is causal.

Sources

• Retention rate - Amplitude
• Cohort analysis primer - Amplitude
• Retention engagement growth: the silent killer - Reforge
• Product onboarding - Amplitude
• The most important thing - Silicon Valley Product Group