How Duolingo Turned a Free Language App Into a $7.7B Business

This is Duolingo. Around 34 million people use the language learning app every single day, some for well over a decade, making the company around $500 million annually and making it one of the top education apps in Apple’s App Store for the past decade. And yet. Most people who use Duolingo do not pay. All of its lessons are completely free to access.
Only around 9% of its users pay and less than 8% of its revenue comes from advertising. So, how does a free educational app manage to also have a valuation of $7.7 billion? And is the company’s strategy sustainable? The answer lies in AI, A/B testing, and this little green owl. This is the economics of Duolingo.

If you’ve ever used Duolingo, you’ll have noticed
it’s less like a
traditional language lesson
and a lot more like a game.
Users have lives.
Get five answers wrong in a day
and you’ll run out of lives for that day. Lessons have progress bars and are laid out in a journey. Players gain experience points and streaks encourage users to play every day. The company says that’s part of what helped attract some three million users within the first few years. When we launched, we had a thing that was mobile, free, and fun, and because of that, very quickly it overtook all methods to learn languages. But before long, Duolingo ran into an issue with its mission to try and deliver free language lessons. One of our board members who had just invested in Duolingo, she said, “Look, you can’t continue doing this. You gotta make revenue.” We had not given ourselves permission to make any money from our app. It’s hard to build up new capabilities in an organisation. We were in a bit of a pickle because we wanted Duolingo
to be a free way to learn languages. So, in 2017, the company introduced advertising,
which now accounts for under 8% of its revenue. It also launched a subscription tier where users could pay $10 a month to remove ads and get unlimited lives. The company went public in the summer of 2021 and saw a boom during the pandemic as users practiced languages
while travel was limited. Today, analysts say Duolingo is around 12 times bigger
than its main rival Babbel by downloads with some eight million users who are now paying for the app. And while that’s only a small fraction of its overall user base.
The majority of the money is made from this, you know, under 10% of the active users. They give us about 80% of our revenue. But while Duolingo’s growth
through gamification and subscriptions may seem pretty simple, there’s a lot more going on behind the curtain. Every move the company makes is meticulously crafted down to the pixel, tested on unknown users, and refined using AI,
rigorous A/B testing, and constant tweaks to its app.

Whenever people are using the app,
we measure how long they use it for,
do they come back the next day, et cetera.
And we try a lot of things. We try, when do we send you a notification, what we say in the notification? Do we make the lessons a little harder? Do we make them a little easier, a little longer, a little shorter?
Take Duolingo’s push notifications. If you forget to do your lessons and risk ending your streak, you’ll get a little notification
from Duo, the company’s mascot,
reminding you to return to the app.
The messages on the push notifications
started out simple enough, but quickly evolved to become the stuff of memes,
including one particularly effective guilt trip. We decided that the last notification we would send you was telling you, “Hey, these notifications don’t seem to be working. We’re gonna stop sending them for now.” Sending that got people to come back a lot because they felt guilty.
It’s like your mother telling you like, “Okay, well, I’ll stop, like, guilt-tripping you.”
Duolingo says that this led to a 3% rise in retention. But crucially, the messages that you receive aren’t the same as everybody else’s. AI is used to start figuring out
basically, like, which ones of these are more effective for particular users or at particular times of day or whatever and letting the more effective ones rise to the top and then get sent to more people.
To figure out which messages work best,
Duolingo built something called a bandit algorithm, named after a reference to a one-armed bandit, a term for a slot machine in a casino. Say there are, you know, several different slot machines.
In this case, think of them as several different notifications we might send people. And we’ve tried each one some number of times and seen what it paid out each time. So, we have limited information
about all these different options. And so the algorithm has to figure out, do I just go back to the one that has so far given me the most money and just go back to that one every time? Or how do I balance that with exploring different options to see maybe they’re actually better, but I just got a little unlucky earlier? And each time we send out a notification, we track, did it work? Basically, did they come back to Duolingo, do their daily practice?
For example, the algorithm found out
that the notification time for the language you’re learning worked better on Chinese
learners than English learners. It’s mostly chosen by AI. We must have about a
hundred different things that we can send you. They’re written by people
who are pretty funny and some of them are also written by AI. And that
could be part of the secret to Duolingo’s success. The company gets so much data back from those attempting its free lessons
that it has a huge database of information
that it can feed to its AI algorithms. They do something like 13 billion exercises every week. We are throwing all that data to AI and using that to personalize
the learning experience for everyone
to make sure people are learning effectively and staying engaged.
At any given moment, the company is running hundreds of A/B tests like these
where one user is given a different experience to another to see which one wins out. Some of the tests are AI-powered, others aren’t. And for most tests, around 5% of its user base act as unknowing guinea pigs. I don’t know how many A/B tests we’ve run at this point.
It’s probably more than 10,000.
We’ve A/B tested our way into
getting more people to pay us,
into getting more people to use Duolingo,
more people to recommend it
to their friends, et cetera.
So, without having run those,
we would not be a successful company.
I’m certain of that.

So, if Duolingo’s algorithms
and A/B testing mean
that the company knows exactly
what makes its users tick,
why are subscriptions only at 8%,
and could they be higher?
After all, while the company
may have 12 times more
downloads than Babbel,
analysts say its revenue is
only two to three times that
of its rival.
We could tomorrow massively increase
the fraction of subscribers
by doing certain things.
For example, right now, you get one ad
at the end of a lesson.
We could play five ads
at the end of a lesson.
That would get a lot
more people to subscribe.
We could maybe even force
people to subscribe.
And it is true, we would
have an amazing next quarter.
However, that would
also stifle our growth.
But analysts say
that while Duolingo’s
leading market position
may have been achieved
by giving away much of
its lessons for free,
that growth strategy may start to slow,
especially as some users leave the app
after mastering a language
or giving up on it.
And that may lead to Duolingo
trying to increase the number
of paying customers.
So, the company is exploring some new ways to get users to part with their cash.
Its new AI-powered Duolingo Max tier
costs around $30 a month
and allows players to role-play
with AI-powered bots, something the company had previously tried to offer with humans.