Inside Duolingo's Streak: What 600+ Experiments Taught One Retention Team

Shuttleworth, a group product manager leading Duolingo's retention team, describes the streak as a count of how many days in a row you have completed a lesson. Over time it has gained streak freezes that act as insurance, plus challenges, goal setting, rewards and social features. A lot of Duolingo's notifications are tied to it, which makes a pretty simple feature one the team can build on endlessly.

He says the claim isn't just a retention PM talking: apart from the lessons, the streak is Duolingo's biggest feature and, by the numbers, its biggest lever for driving daily active users. Streaks are an engagement hack that he thinks any app could introduce, but only if people already enjoy the product. Right now more than 9 million users have a streak of a year or more, which he likens to a very large city.

Part of the reason is how Duolingo thinks about growth: as much about not losing users as acquiring them. Shuttleworth says bringing in people who then churn is not sustainable, so keeping users coming back gives a much easier base to grow daily actives from.

Duolingo launched with a streak, originally tied to hitting an experience-point goal you set yourself. The problem was that a user could do several lessons in a day, miss a too-ambitious goal and lose their streak. One of the most impactful early experiments moved the streak to "do one lesson a day," which Shuttleworth says drove a big increase in DAUs without users caring less about it.

His principle is to think about your app's unit of use. For Duolingo that is completing a lesson, so a streak built around one lesson stays meaningful and is simple for more users to understand. A later test counted just one exercise toward the streak; DAUs did not move, because the people it captured, who answer a question or two and leave, were the least engaged users imaginable.

Shuttleworth says the team has run over 600 streak experiments in four years, roughly one every other day, ranging from mechanic changes to swapping a single string. About half of them lose, and they still learn from them. He'd be more careful now that 9 million users sit on long streaks, but for a new streak he says to test everything, and he usually shuts down neutral results to avoid adding cognitive load.

A large share of Duolingo's streak experiments target users in their first seven days. Retention curves show that going from a one-day to a two-day streak is a huge jump, the following days are slightly smaller but still big, and by day seven loss aversion kicks in and the curve flattens. A 30-day streak is not dramatically more retained than a seven-day one, so the team focuses on getting users there.

The streak goal feature shows their method: ship the simplest possible version and add to it. It began with copy borrowed from the monetization team's winning message and told users they are seven times more likely to finish the course if they reach a 30-day streak. After that they tested 14 and 50 days, found the lengths appealed to different people, and added an opt-out button, which was almost as big a win as the easier fallback goal.

The most surprising finding concerned friction. When they built a goal picker, they tried preselecting a harder goal for users, expecting better retention, and it lost significantly. Shuttleworth concluded that the act of choosing a goal was where the engagement came from.

Shuttleworth says the experiments are not random: ideas travel between teams, such as the monetization team's winning hook, and the team studies mobile games for inspiration. How strongly you hold a hypothesis should match how costly the experiment is. Copy is cheap, so Duolingo built infrastructure to test it constantly, and one example was changing the goal screen button from "Continue" to "Commit To My Goal," which was a massive win.

Lenny asks how they avoid a monstrous, incoherent product. Shuttleworth says every proposed change goes through a product review that includes founder Luis, the head of product design and product leaders, and the streak team reviews its direction quarterly. Periodically they hit a local maximum and have to throw things out and reset the UI so it is as good as before but simpler, which is a hard experiment to win.

He also acknowledges that Duolingo has an advantage because people want to learn a language, and mobile games streaks feel different. But the streak suits language learning especially well because day-to-day progress is almost invisible, and a growing streak makes that long journey feel tangible. Companies elsewhere need to frame their streaks so users feel good about them.

Shuttleworth says many users don't actually understand how streaks or streak freezes work; his own mother was surprised to find her streak intact after a missed day. His first win at Duolingo was an eight-word explanation on the screen that starts a streak, roughly "start a day to extend your streak, but miss a day and it resets." It was a top-three win for current user retention rate (CURR), the metric that is most effective at driving DAUs, and was on the order of 10,000 DAUs.

Not every winning test shipped. Adding a live XP counter inside the lesson improved results, but Luis asked what the team would do with it next, and they had no good answer. They concluded it was clutter in the most sacred space in the app and shut it down, though Shuttleworth notes that with more XP mechanics today, the view on that could change.

The flame icon also taught them something: UX research in India showed the metaphor of keeping a flame lit didn't resonate there. Duolingo's streak screens now lead with the number, with an odometer animation, and the calendar increasingly looks like a calendar so people understand it is a daily mechanic. Shuttleworth adds that clarity doesn't have to cost delight, such as the Phoenix Duo celebration at milestones.

Shuttleworth shows a knitted "Streak Serenity Prayer" from his co-lead Antonia, which sums up the philosophy: accept the flexibility you need, reach for perfection when you can, celebrate regardless. From a mechanics perspective, flexibility has been the biggest DAU driver. Going from one streak freeze to two or three was a big DAU win, although it hurt CURR and was recovered through returning users, and three freezes were no better than two because they were also training people to take longer breaks.

His rule of thumb is to give more flexibility when a streak is new and less once it is long. Handing out two streak freezes when a user starts a streak was one of the biggest wins, because short streaks are easy to let die and restart. Streak freezes were once bought with gems; Shuttleworth admits there is tension between monetization and retention here and that he would love to remove some of the monetization hooks.

The team also replaced paid streak repair with an "Earn Back" flow in which you complete a few lessons to get your streak back, which felt more deserved and won on retention. Shuttleworth says Antonia is the keeper of the streak's sanctity, because there is a point where users realize the streak means nothing. That would be a one-way door and, for the 9 million users on year-long streaks, an extinction-level event.

Duolingo sends two core streak notifications a day. The practice reminder goes out 23 and a half hours after you practiced the day before, because that beat letting users pick a time, and every more complex logic they tried. The "streak saver" arrives at 10 p.m. if you haven't extended your streak. It could seem spammy, but because users attach positive emotion to the streak, most see it as helpful.

Because flexibility is generous, the team also celebrates perfection. A "Perfect Streak" simply turns your streak gold if you go a few days without a freeze, with no other reward, and it is powerful. Shuttleworth says people care about perfect streaks as much as streaks themselves. The team still holds back such features until after day seven so that new users aren't overloaded with concepts.

On delight, he says animation, haptics and sound all win, partly because they feel good and partly because they make users pause on the streak screen, which he sometimes wants. Haptics were first done by engineers, then designers, and now by an animation studio the company recently acquired. He recalls that trying to hire a dedicated haptics designer was the saddest hiring he has done, given how specific the skill is.

Most Duolingo teams are organized around a metric rather than a feature. Shuttleworth's retention team cares about CURR, and it owns the streak only because the streak has driven CURR better than anything else. Other teams can and do work on the streak, and he pitches retention ideas for leaderboards in the same way. Lenny notes this resembles how Airbnb worked.

Since the streak has several quarters of roadmap, ownership is soft: other teams who want to change it need to fit their ideas into the plan and share context. Running this many experiments, he says, requires process: detailed roadmaps in which each experiment sets up the next, planned months ahead, so engineering and design bandwidth aren't wasted. Much of it lives in Jira, with some Google Docs.

The team also resists building a big V1. They strip a feature down to the core hypothesis and ship it, because it is easy to add bells and whistles that make something win without proving the real idea. That leads to faster shipping, design and approval, plus confidence at every step, and iteration after iteration.

Lenny summarizes the approach as mapping the levers that drive growth, finding the one most connected to it, and mining it for ideas. Shuttleworth says that when he joined, his predecessor told him the team was not even 30% of the way to optimizing the streak, and four years later he says the same. Still, he stresses doing this with intent rather than a bunch of random changes, or you end up at a local maximum.

His main advice for anyone building a streak is to get the core product right first. Loss aversion works on humans, but if people don't want to use your app at the cadence you need, the streak will give you little and distract you. He also repeats a comment from board member Bing Gordon: users care about the streak because Duolingo does, with prominent animated screens after every session and references in messages, so a streak hidden in a corner won't matter much.

Shuttleworth thinks nearly every app could benefit from a streak, although the cadence should fit the product. He cites Peloton's weekly streaks, which suited users who couldn't work out daily, and admits tax software would be a hard case. His closing advice is to stop debating in a conference room and test with users rather than trying to get it perfect the first time.