Andrew Leng Ly analyst.andrewlengly.com/studies/duolingo-streaks :: duolingo streaks # Do Duolingo’s streaks build language skill? Daily users grew 5.8x while half of B1 still takes 2.2 years of 15-minute days. A data study by Andrew Leng Ly. designed page ......... analyst.andrewlengly.com/studies/duolingo-streaks # Pages / /projects /designer /developer /analyst /resume /card /campus /portfolio /garden /music /pond /sand /keys /timer /build /opening /split /blog /studio # Case studies /case-studies /grid-watch-casestudy /chain-routes-casestudy /nettracer-casestudy /fish-maps-casestudy /campus-it-casestudy /alwaysup-casestudy /third-places-casestudy /sourcing-channels-casestudy /skills-hiring-casestudy /ramp-time-casestudy /internal-mobility-casestudy /tab-split-casestudy # Data studies /studies /duolingo-streaks /chegg-after-chatgpt /tiktok-stem-feed /handshake-and-linkedin /ai-and-student-learning /third-places /oura-and-whoop /protein-boom /campus-rec-and-gpa ── DUOLINGO STREAKS ──────────────────────────────────────────────────────────────────────── analyst work .......... /machine/analyst data study ............ October 2026 question .............. Do Duolingo’s streaks and leagues build language skill, or mostly daily logins? headline finding ...... Daily users grew 5.8x from Q4 2021 to Q2 2026. The median learner in Duolingo’s own study still needed 203 hours to reach half of B1, the intermediate level, which is 2.2 years of 15-minute days. my recommendation ..... Reward mastery checks with the badges and league rank that XP earns today, and report a learning number beside daily users in every shareholder letter. tools ................. CSV tables with a source URL on every row · JavaScript on Node for the derived rates · D3.js for the first dashboard · React and SVG for the charts on this page ── THE QUESTION ──────────────────────────────────────────────────────────────────────────── the client ............ Duolingo’s leadership. In February 2026 the Q4 shareholder letter put teaching better and user growth ahead of near-term profit, and set a goal of 100 million daily users by 2028. the decision they face What should streaks and leagues reward, and what should the company report each quarter, so that a daily-user goal also tracks learning? my starting hunch ..... I came in with the pattern from my study of grades and AI, where students chase the number that gets counted. I expected the streak to work like a grade, with the count climbing and little skill behind it. how the hunch held up I was partly right. The streak asks for one lesson a day, and I found no published Duolingo study that links streak length to a proficiency score. I was wrong about the learning. People who finish units test well on outside exams. Unit 5 finishers matched 4 university semesters in reading and listening, and Unit 7 finishers matched 5. ── THE DATA, 6 TABLES ────────────────────────────────────────────────────────────────────── [01] engagement.csv ................................................................ 39 rows /studies/duolingo-streaks/data/engagement.csv Daily users, monthly users, paid subscribers and daily-user growth by quarter, Q4 2021 to Q2 2026. where it came from Duolingo shareholder letters and SEC filings (8-K, 10-K, 10-Q). why i chose it ... It is the series Duolingo’s 2028 goal is written in. what it cannot tell us What any user learned. The counts are quarter averages from Duolingo’s own analytics, and no third party audits them. [02] metrics.csv ................................................................... 46 rows /studies/duolingo-streaks/data/metrics.csv Every figure that appears once, covering streak counts, efficacy studies, study time, FSI class hours, strategy and market value. where it came from Duolingo earnings calls, blog posts and research reports, peer-reviewed journals, the US State Department’s Foreign Service Institute and market data sites. why i chose it ... Each figure gets an ID, so the other tables and the code can point at the exact row. what it cannot tell us A trend. The streak counts come from separate statements made over three years, and Duolingo worded two of them as “nearly” and “more than”. [03] skills.csv .................................................................... 10 rows /studies/duolingo-streaks/data/skills.csv Pretest and posttest STAMP scores by skill from a 12-week study of Duolingo Spanish (Smith, Jiang and Peters, 2024). where it came from Language Learning & Technology, volume 28. why i chose it ... It tests the same learners before and after 12 weeks of use. what it cannot tell us Cause. The study has no control group, and 48 of the 69 people who started it finished. Duolingo funded it. [04] meta.csv ....................................................................... 6 rows /studies/duolingo-streaks/data/meta.csv Pooled effect sizes with 95% intervals from four gamification meta-analyses, 2020 to 2026. where it came from Educational Psychology Review, Educational Technology Research and Development, Sustainability and SAGE Open. why i chose it ... None of it comes from Duolingo, so it checks the company’s own claims from outside. what it cannot tell us Anything about Duolingo itself. The studies pool classroom and lab tests of points and badges. One row reports Cohen’s d where the rest report Hedges’ g. [05] events.csv .................................................................... 10 rows /studies/duolingo-streaks/data/events.csv Dated events from 2023 to 2026 on streaks, AI-built courses, the share price drop and the turn back to teaching. where it came from Duolingo letters and press releases, plus press coverage. why i chose it ... It puts dates on the strategy changes so they can sit beside the user series. what it cannot tell us Why the share price fell. A share price moves for many reasons, so I make no claim that streaks caused the drop. [06] variables.csv .................................................................. 9 rows /studies/duolingo-streaks/data/variables.csv Nine design variables, each scored 0 to 2 for how far it pushes a learner toward proficiency. where it came from My own coding, with the IDs of the rows each score rests on. why i chose it ... It ties every option to evidence a reader can open. what it cannot tell us Stand as measurement. The scores are judgment, and the evidence IDs are there so a reader can disagree. ── THE PROCESS ───────────────────────────────────────────────────────────────────────────── [01] One row per figure Each figure went into a row of its own with a source name, a URL, the date I retrieved it and a flag for whether Duolingo reported it. That came to 120 rows across six files. [02] Pivot engagement.csv is long, with one row per metric per quarter. I pivoted it to one row per quarter and added daily users as a share of monthly users. [03] Join Streak counts sit in metrics.csv and daily users in engagement.csv. I joined each streak statement to average daily users in the nearest reported quarter, so the Feb 2023 statement sits against Q4 2022. [04] Assumptions “Nearly 8 million” and “more than 10 million” enter as 8 and 10, so the streak shares are approximate. A year is 365 study days. FSI hours count class time only, and FSI students add 17 hours of self-study a week. [05] Check I recomputed year-over-year growth from the daily-user rows and compared it with the rates Duolingo reported. 4 of 6 could be checked, and each landed within 0.2 points. # One quarter before and after the pivot. - Before, three rows of engagement.csv (six of ten columns shown) - After, one row per quarter The quarter draws on two filings. Daily users and subscribers come from the 8-K letter and monthly users from the 10-Q. # The code for the key steps - The pivot from long rows to one row per quarter. derive.mjs ........ /studies/duolingo-streaks/derive.mjs - The join between streak statements and daily users. derive.mjs ........ /studies/duolingo-streaks/derive.mjs # Dead ends google trends ....... I planned a Google Trends series as a supporting signal. I could not pull it, so the study leaves it out and leans on the filed user counts. a gap in the quarters The series has fourth-quarter points for 2021 to 2023 and every quarter from Q3 2024. With no rows for early 2024 I could not recompute growth for the first two quarters of 2025. I kept the reported rates and marked the source. The Q2 2025 rate comes from a Zacks summary of Duolingo’s letters. ── THE 3 FINDINGS THAT CHANGE THE DECISION ───────────────────────────────────────────────── [01] 5.8x · Daily users, 10.1 million to 58.7 million Daily users grew 5.8x in under five years, and daily users as a share of monthly users rose from 23.8% to 41.7%. dataset .......... engagement.csv query ............ derive.mjs, engWide() # Daily users grew 5.8x in under five years Fourth-quarter points through 2023, then every quarter. quarter .......... Users (millions) monthly users .... Q4 2021 42.4, Q4 2022 60.7, Q4 2023 88.4, Q3 2024 113.1, Q4 2024 116.7, Q1 2025 130.2, Q2 2025 128.3, Q3 2025 135.3, Q4 2025 133.1, Q1 2026 137.8, Q2 2026 140.6 daily users ...... Q4 2021 10.1, Q4 2022 16.3, Q4 2023 26.9, Q3 2024 37.2, Q4 2024 40.5, Q1 2025 46.6, Q2 2025 47.7, Q3 2025 50.5, Q4 2025 52.7, Q1 2026 56.5, Q2 2026 58.7 paid subscribers Q4 2021 2.5, Q4 2022 4.2, Q4 2023 6.6, Q3 2024 8.6, Q4 2024 9.5, Q1 2025 10.3, Q2 2025 10.9, Q3 2025 11.5, Q4 2025 12.2, Q1 2026 12.5, Q2 2026 12.7 daily as a share of monthly Q4 2021 23.8%, Q4 2022 26.9%, Q4 2023 30.4%, Q3 2024 32.9%, Q4 2024 34.7%, Q1 2025 35.8%, Q2 2025 37.2%, Q3 2025 37.3%, Q4 2025 39.6%, Q1 2026 41.0%, Q2 2026 41.7% source ........... engagement.csv, derive.mjs, engWide() [02] 2.2 years · 203 median hours at 15 minutes a day At the 15 minutes a day Duolingo aims for, the median learner needs 2.2 years to finish Unit 7, which is half of B1, the intermediate level. dataset .......... metrics.csv query ............ derive.mjs, timeYears() # At 15 minutes a day, half of B1 takes 2.2 years Study hours for each milestone, divided by 15 minutes a day. The Foreign Service Institute (FSI) figures count class hours to professional proficiency, a level above anything Duolingo has tested. Milestone ..................................................... Years of 15-minute days Duolingo Unit 7, half of B1 ....................................................... 2.2 Study hours: 203 · Years at 13 minutes a day: 2.6 FSI Category I (Dutch, Italian) ................................................... 6.0 Study hours: 552 · Years at 13 minutes a day: 7.0 FSI Spanish or French ............................................................. 7.6 Study hours: 690 · Years at 13 minutes a day: 8.7 FSI Category II (German) .......................................................... 9.1 Study hours: 828 · Years at 13 minutes a day: 10.5 FSI Category III (Russian, Hindi) ................................................ 11.1 Study hours: 1,012 · Years at 13 minutes a day: 12.8 FSI Category IV (Japanese, Korean) ............................................... 24.1 Study hours: 2,200 · Years at 13 minutes a day: 27.8 source ........... metrics.csv, derive.mjs, timeYears() [03] 17.7% · Year-long streaks as a share of daily users, Feb 2026 People with a year-long streak have stayed near a fifth of daily users since 2023, and 84% of current users come back the next day. dataset .......... metrics.csv query ............ derive.mjs, streakShare() # Year-long streaks stay near a fifth of daily users Users with a streak of 365 days or more, against average daily users in the nearest reported quarter. Duolingo rounded the counts, so the shares are approximate. Statement date ............................................... Share of daily users (%) Feb 2023 ........................................................................ 18.4% Year-long streaks: 3 million · Daily users, Q4 2022: 16.3 million Aug 2024 ........................................................................ 21.5% Year-long streaks: 8 million · Daily users, Q3 2024: 37.2 million Feb 2026 ........................................................................ 17.7% Year-long streaks: 10 million · Daily users, Q1 2026: 56.5 million source ........... metrics.csv, engagement.csv, derive.mjs, streakShare() ── MY RECOMMENDATION ─────────────────────────────────────────────────────────────────────── Reward mastery, and report a learning number next to daily users. [A] Reward mastery Make the unit mastery check and Duolingo Score gains the things that earn badges and league rank, and run leagues as teams. evidence ......... Findings 2 and 3. Studies of points and badges find a medium-sized average gain in learning (g = 0.49), and shared streaks already lift lesson completion 22%. main risk ........ The pooled effect comes from classroom and lab studies of points and badges. None of them tests Duolingo. [B] Report a learning number beside daily users Add two lines to each shareholder letter, the share of daily users who passed a mastery check that quarter and the median Duolingo Score gain. evidence ......... Finding 1. Daily users lead every letter and the 2028 goal is 100 million of them. Market value fell 64% from June 2025 to October 2026. main risk ........ Duolingo would report the number itself. 78 of the 101 sourced figures in this study already come from the company. - Rewarding mastery changes what a streak is worth and keeps the habit loop behind the 84% next-day return. - With a learning number in every shareholder letter, investors judge Duolingo on learning as well as daily users. ── LIMITS AND NEXT STEPS ─────────────────────────────────────────────────────────────────── # What the data cannot support - 78 of the 101 sourced figures come from Duolingo. Every user metric does. - Nothing here links one learner’s streak to that learner’s test score. The gap between engagement and learning is my reading of two separate sets of numbers. - The efficacy studies tested people who finished, and 90% of the Unit 5 group held a degree. - The data stops at Q2 2026. The Q3 letter was due on November 3, after the study closed. # What I would test with more access - With Duolingo’s own data I would test whether streak length predicts Duolingo Score gains, learner by learner. - I would track the share of daily users who pass a mastery check each quarter, which is the number option B asks Duolingo to publish. - I would add Q3 2026 from the November letter and rerun the build. ── TL;DR ─────────────────────────────────────────────────────────────────────────────────── studied ............... Whether Duolingo’s streaks and leagues build language skill, read from the company’s own numbers and its efficacy studies. found ................. Daily users grew 5.8x, and the median learner still needs 2.2 years of 15-minute days to reach half of B1. # by the numbers 5.8x ................ Daily users, Q4 2021 to Q2 2026 2.2 years ........... To half of B1 at 15 minutes a day 17.7% ............... Daily users on a year-long streak the catch ............. Nothing links one learner’s streak to a test score. The gap is my reading of two separate sets of numbers. next .................. With Duolingo’s own data, test whether streak length predicts Duolingo Score gains, learner by learner. # end of /machine/duolingo-streaks. andrew.leng.ly@gmail.com · leave machine mode