01 Company claim
The documented mechanism
Duolingo's DAU growth stagnated in 2018, and A/B tests aimed at DAU directly were struggling to show significant impact. The Data Science team built a Markov 'Growth Model'. Every learner is classified into an activity state each day, and transition rates between states (such as NURR, CURR and reactivation rates) are tracked. Using the model, the team simulated a 2% month-over-month improvement in each rate. Raising CURR had the largest effect on DAU. Duolingo then staffed a team to run A/B tests on two questions: could CURR be moved, and did moving it actually move DAU? Only then did CURR become the focus.
02 Kept apart from the practice
Reported outcome
Self-reported by Duolingo: the framework 'helped to grow DAUs by 4x since 2019', and by 2023 DAU growth was 'strong and consistent', with CURR described as the key.
03 Read before copying
Confounders and limits
The 4x DAU figure covers 2019 to early 2023. That window includes many other product, marketing and market changes, and the post doesn't isolate the Growth Model's contribution. Duolingo's mostly organic acquisition and its gamified daily-habit product make retention-state metrics unusually meaningful. Products without daily usage may not behave this way.
04 Over time
Known revisions, reversals or failures
Documented limits in the source: the author says CURR has a ceiling below 100%. With about 90% of DAU in the Current state, aggregate CURR is becoming 'increasingly imprecise' and could become hard to move, repeating the 2018 problem. Duolingo says it is exploring 'bottom-up' unsupervised segmentation as a complement to the top-down model and 'perhaps' an eventual replacement.
05 Editorial interpretation — not a company claim
WeCaseStudy interpretation
WeCaseStudy interpretation: this is a case for narrowing investment to one lever, chosen by modelling. It is equally a case for validating that the lever can be moved, and that moving it moves the outcome, before committing a team. Duolingo's own later caveats show that a narrow focus metric has a shelf life and should get a scheduled revisit.
06 What transfers
Adaptation boundaries
A small company should not assume it has enough historical user data or user volume to train a state-transition model or run conclusive A/B tests. It also shouldn't assume it has cross-device identity resolution; Duolingo's model depends on an 'aliasing' process that keeps returning users from being counted as new. The specific lever (CURR) is Duolingo's. The transferable part is the method of choosing a lever, not the lever.
Claims and their sources
Each material claim has its own type, source, location and review state. A documented policy is not proof that everyone follows it; a vendor describing its own product is vendor-reported.
| Claim | Type | Source and location | Review |
|---|
| CLM-0021 Duolingo reports that in 2018, after several years of strong growth, its DAU metric was stagnating and DAU-focused A/B tests were struggling to show significant impact. | Documented practice | Meaningful metrics: How data sharpened the focus of product teams (Official company publication)
Section 'Getting our DAUs out of a rut': "our DAU metric was stagnating" | Source checked — source checked 2026-10-17 |
| CLM-0022 Duolingo describes its Growth Model as a Markov model that classifies all learners into an activity state each day and monitors transition rates between states. | Documented practice | Meaningful metrics: How data sharpened the focus of product teams (Official company publication)
Section 'Getting our DAUs out of a rut': "It is a Markov Model that breaks down topline metrics" | Source checked — source checked 2026-10-17 |
| CLM-0023 In Duolingo's simulations, increasing Current User Retention Rate (CURR) 2% month-over-month had the largest impact on DAU, and Duolingo then staffed a team to A/B test whether CURR could be moved and whether that moved DAU. | Documented practice | Meaningful metrics: How data sharpened the focus of product teams (Official company publication)
Section 'Finding our new “movable” metrics': "(CURR) 2% month-over-month had the largest impact on DAU" | Source checked — source checked 2026-10-17 |
| CLM-0024 Duolingo states that the Growth Model framework helped grow DAUs by 4x since 2019. | Self-reported outcome | Meaningful metrics: How data sharpened the focus of product teams (Official company publication)
Intro paragraph: "helped to grow DAUs by 4x since 2019" | Source checked — source checked 2026-10-17 |
| CLM-0025 Duolingo says 90% of its DAU fall into the Current User state, making aggregate CURR increasingly imprecise, and that it is exploring bottom-up segmentation as a complement and possible eventual replacement. | Documented practice | Meaningful metrics: How data sharpened the focus of product teams (Official company publication)
Section 'Next up: evolving beyond aggregate metrics': "90% of our DAU fall into this state" | Source checked — source checked 2026-10-17 |
Cite this record
We Case Study. "Duolingo: Growth Model: a Markov model of user activity states, used to find and staff a team on the single 'movable' metric with the most DAU leverage (Current User Retention Rate)". Decision Lab practice record, version 1. https://wecasestudy.com/decision-lab/practices/duolingo-growth-model-curr
Please cite the original company sources listed above for the facts themselves; this record is a structured, scoped summary of them.
All documented practices of Duolingo
