Rebase your first hundred users at month 3
Stop judging early retention too soon by rebasing the curve after early churn clears, then reading whether a real cohort is forming.
You have probably done the unscalable things: hand-written notes, onboarding calls, chasing the same users until the product stopped feeling like a demo. Now the first hundred are gone, and the early numbers look ugly. The temptation is to call it a retention problem before the tourists have left.
Early churn is not the same as product-market fit failure. The people who sign up for a new tool often include the curious, the impatient, and the people who wanted a quick answer. They leave. That is not the whole story.
Rebase the curve at month 3
a16z's review of hundreds of AI companies showed that moving retention and acquisition math from month zero to month three helps founders spot early signs of retention, product-market fit, unit economics, and go-to-market investment. The a16z rebase method compares month 12 retention with month 3 rather than month 0 because most tourist churn has cleared by month 3. Rebase your retention table so the denominator is the users still active at month 3, not the users who ever signed up. Keeping month 0 in the main number mixes real users with early churn and makes every later month look worse than it is.
Define active before you calculate anything. The definition should be the action that makes the user come back, not a login. If the action is too broad, the curve will look healthier than the product is. If it is too narrow, you will mistake a real cohort for a ghost.
Track M12/M3 before you scale
Track the ratio of month 12 retention to month 3 retention. a16z treats that ratio as a way to assess how customers who survive early tourist churn behave over their first full year and as an early signal of long-term retention quality. The output is a single number you can compare across cohorts, not a list of raw counts.
Calculate it for each signup cohort, not just the whole product. A blended number can hide a weak new cohort under a strong old one. If you only have a single cohort, track it monthly and note whether the ratio is stable, rising, or falling.
The ratio removes the early weeks from the retention curve and asks a simpler question: do the people who made a real decision to stay keep using the product? That is the question your first hundred can answer, if you can still talk to them.
Classify the shape, then decide
Classify a curve that flattens after month 3 as a sign that a foundational cohort is forming. Classify a curve that keeps falling as a product-market fit problem, not a channel problem. The shape tells you whether to keep hand-building or fix the product before adding distribution.
Write the classification next to each cohort: flatten-and-hold or perpetual decline. If you cannot write it, your table is too noisy. Add the active definition, the month 3 baseline, and the M12/M3 ratio before you make a decision.
Malika Aubakirova at a16z analyzed real LLM usage on OpenRouter in 2025 and found that products lacking a foundational cohort showed every sign-up cohort churning quickly with no long-term floor. Rodriguez and Immerman at a16z used ChatGPT retention data in September 2025 to document a smiling retention curve as the clearest example. The shape means the early drop is steep, then the line bends and holds.
Top self-serve or month-to-month AI companies already show strong performance on the month 12-over-month 3 basis, which a16z views as a leading indicator of long-term net dollar retention above 100%. SaaS Capital put the 2025 median net revenue retention figure for private SaaS companies at 101%, per Userpilot. Use that as a reference point, not a target for your early cohort.
Do not buy traffic to make the curve look better. Paid acquisition can hide the shape of the curve, and it can make a weak cohort look like a growth problem. You need the unscalable version of the diagnostic: a small set of users you can actually talk to.
Do not add another channel until the M12/M3 ratio is stable. The test is a repeatable manual process: you can onboard a new user without a new playbook, and the next cohort does not need a different promise.
If the curve keeps falling, the fix is usually in the product, the onboarding, or the promise you made. Fix that before you spend money on distribution.
