May 4, 2026
Duolingo and the Wrong Metric
Moving a metric 2% a month sounds irrelevant. At Duolingo, that number was worth five times more growth impact than any other variable. The discovery came from stopping to push DAU and asking a different question.
Moving a metric 2% a month sounds irrelevant. At Duolingo, that number was worth five times more growth impact than any other variable the team was tracking.
The discovery came from stopping to push DAU directly and asking a different question: what moves it indirectly?
DAU was too large, with too many variables and too few identifiable levers. So they built a Markov model that segmented the entire user base into seven engagement states and tracked the transition rates between them.
The metric the model identified as the main lever was CURR, the retention of users who were already active. They created a dedicated team focused solely on moving that number. Over four years, it rose 21%, churn among their best users fell 40%, and DAU grew 4.5x.
Before the wins, two mistakes that teach more.
They tried copying the move counter from Gardenscapes, which consumed months of work and produced neutral results. The mechanic worked in the game because each move was strategic. In Duolingo, you either know the answer or you do not, so the counter was just noise. Then came a referral program modeled after Uber, which excluded exactly the users who refer most: those already on the premium plan.
Former CPO Jorge Mazal documented the lesson: adapt when you adopt. Before copying a feature, understand why it works in that context.
What came next worked because the team understood the mechanisms before copying the forms. Leaderboards exploit social comparison: competing against someone with similar engagement retains more than competing against friends who already stopped using the product. The streak capitalizes on loss aversion, with the psychological weight of losing 100 days of history growing every day it continues. Notifications were optimized with deliberate volume restraint and adjustments to timing, copy, and localization that apply variable reinforcement without destroying the channel.
By 2023, 90% of active users were already in the same segment. The metric that had saved growth had become a new monolith, too large to move, and the cycle started again.
Every metric that works carries an expiration date. What forces the right decision today may be covering the wrong problem tomorrow.
Sources: Jorge Mazal, Lenny's Newsletter · Erin Gustafson, Duolingo Blog