Subverting the traditional growth model: What did Duolingo do right as its DAU soared?

How to maintain sustained growth or break through the growth dilemma is a question that almost every company needs to think about. In this article, the author shares the growth case of Duolingo, a language learning platform, which achieved DAU growth by adopting a more detailed user segmentation and data model. Let's take a look at the analysis of this article.

Growth is an eternal theme for any business.

Whether it is a startup or a mature enterprise, after a period of rapid development, they will hit a stage ceiling. Whether they can continue to innovate and maintain growth is the key to whether a company can become great or mediocre.

Duolingo is a magical existence. As a language learning platform, its DAU has been soaring in a less popular track. In Q4 2022, Duolingo's MAU reached 60.7 million, compared with only 42.4 million in the same period last year, a year-on-year increase of 43%, and DAU increased by 62% year-on-year to 16.3 million.

From 2018 to 2022, Duolingo’s DAU increased by 450% | Image source: Lenny’s Newsletter

In 2022, when it is difficult for major leading platforms to expand their user base and increase revenue, how can Duolingo achieve rapid growth? The answer is to subvert the traditional AARRR model and adopt a more detailed user segmentation and data model to find growth levers.

Today, we will take Duolingo as an example to explore how we can learn from its successful experience, apply it flexibly and break through the growth dilemma.

Without further ado, let's go straight to the main text. Enjoy:

1. Advantages and Disadvantages of the Traditional AARRR Model

When it comes to user operations and growth, the AARRR growth model is a well-known framework in the industry, also known as the Pirate Model, which is similar to what pirates do, achieving predatory growth. It is a five-level funnel model. It includes the following steps:

(Image source: drawn by myself)

1. Acquisition: How you acquire new users, attract their attention, and make them aware of your product.
2. Activation: Convert new users into active users, generate interest in your product or service, and maintain a good first impression.
3. Retention: Ensure that acquired users continue to use and return to improve user retention rate.
4. Revenue: Convert users into paying customers and achieve business goals through purchases.
5. Self-propagation (Refer): Let users recognize the product and spread it spontaneously, bringing in more new users.

The beauty of the AARRR growth model is that it provides a simple and methodical way to evaluate user operations and growth strategies.

• Easy to understand: Dividing the user life cycle into 5 stages makes it easier for us to find corresponding strategies in each stage.
• Easy to implement: It is a recognized standard, so most companies can quickly understand and implement it.
• Easy to track: Since the goals of each stage are very clear, its progress can be tracked through specific indicators.

However, the AARRR model also has its disadvantages:

• Oversimplification: Only five stages are involved, whereas in real life, user situations are often very complex and may require more specific operational strategies.
• Lack of focus: Not distinguishing which users are more important, so too many resources may be consumed to meet non-essential user needs.
• Lack of guidance: There is a lack of specific guidance, and other models or data are often needed to implement operational strategies.

Because this type of model is too simple , each step depends on the success of the previous step, which may make the model unsuitable for certain products or services in some cases.

Therefore, when Duolingo encountered a growth bottleneck in 2018, it decided to explore a new growth model from the perspective of life cycle management , leverage it, and shift to refined operations .

2. Deconstruction of Duolingo’s life cycle model

1. Why do we need lifecycle management?

As the growth of mobile Internet users approaches saturation, the cost of acquiring new users increases, making the retention of old users particularly important. The basic concept of the user life cycle broadly includes five stages:

1. Introduction stage: the user acquisition stage, converting potential user traffic in the market into own users.
2. Growth stage: registered, logged in and activated, have begun to experience the product's related services or functions, and experienced the Aha moment.
3. Mature stage: In-depth use of product functions or services, contributing more active time, advertising revenue or payment, etc.
4. Dormant period: mature users who have not generated valuable behavior for a period of time.
5. Lost period: users who have not logged in or visited for a period of time.

If user growth is viewed as a system, the purpose of this system is to continuously increase the user scale and user value; active users are the stock, and new users and lost users are the traffic.

User lifecycle management can be seen as a feedback method of this system , that is, strengthening the system's reinforcement loop while suppressing the regulatory loop. This helps us see the value of user lifecycle management from a global perspective.

The goal of operations is to increase the number of users and conversion rate , but focusing only on these superficial indicators is not enough to achieve the best results. Many novice operators ignore the necessary thinking and only focus on superficial growth or conversion methods for the direction of operation strategy.

For example, every day we think about how to use event prizes to attract new users and how to use discounts to promote payment, but we don’t think deeply about some more important issues : how to measure whether the monthly growth target is reasonable? How to maximize the GMV of existing users? What is the stratification of our existing users? Which users can help us achieve our goals, etc.

If we only stay at the surface of the operation, we will not be able to effectively predict the results, resulting in the failure of the activity. At the same time, it is not helpful for personal growth, because we can only get the data after the activity, but we cannot know: how many old users were activated in the project, how many new users participated in the product's paid conversion, etc. It is impossible to trace and guide the subsequent operation strategy.

Therefore, we need to conduct in-depth thinking and data analysis to manage the user life cycle in order to develop more effective operation strategies .

2. How to build a data observation model

After many practices and adjustments, Duolingo built a data observation model based on "user activity" : users with different levels of activity are stratified, key users are clearly identified, and they are guided and intervened to keep them active in the product.

Every user who has used the product will be classified into a specific user group on a specific calendar day. This also means that users at different levels are always mutually exclusive. Different arrows represent the conversion ratio of users at each level (including CURR, NURR, RURR and SURR, but the time dimension is day instead of week). In this completely closed-loop model system, we can see that new users are the only breakthrough.

(Image source: Lenny's Newsletter & self-drawn)

The sum of the yellow, green, and blue user groups constitutes the product’s DAU (daily active users), which are:

• New Users : Users who log in to the App for the first time;
• Active Users : Users who have logged in today and have logged in at least once in the past 6 days;
• Reactivated Users : Users who logged in today, have not logged in in the past 6 days, but have logged in at least once in the past 7-29 days;
• Resurrected Users : Users who logged in today but have not logged in for at least 30 days.

The bottom three user groups are users who are not logged in today but have different levels of participation data in the past.

• Possible lost weekly active users : users who did not log in today but logged in at least once in the past 6 days [Possible lost weekly active users + DAU (daily active users) = WAU (weekly active users)]
• Monthly active users that may churn : users who have not logged in in the past 7 days but have logged in within 30 days [Monthly active users that may churn + WAU (weekly active users) = MAU (monthly active users)]
• Lost users : users who have not logged in for the past 31 days or more [MAU + Lost users = total number of users]

If you think this picture looks complicated, I will add a timeline to deform it and it should be easier to understand.

(Image source: drawn by myself)

From the above formula, we can see that common DAU, WAU and MAU can be calculated by adding up users at different levels, which means that Duolingo can perform data modeling for these users. This is a key feature of this model system.

Furthermore, by adjusting the different retention/churn rates represented by the arrows, we can simulate the combined effect of different retention rates over time. In other words, these rates are levers that product teams can use to drive user growth.

• Active User Retention (CURR) : The percentage of active users who logged in this week among those who logged in in the past two weeks.
• New User Retention (NURR) : The percentage of new users who registered in the past week and logged in this week;
• Retention of recalled users (RURR) : The percentage of users who logged in this week out of the users who were recalled in the past week (active within 30 days).
• Retention of returning users (SURR) : The percentage of users who have not logged in for a long time (at least 30 days) but returned in the past week and logged in this week.

3. How is it better than the traditional model?

Compared with the traditional AARRR model, Duolingo's lifecycle model is more flexible and has the following advantages:

• Distinguish different types of users: The layered life cycle model can help us better distinguish different types of users, allowing us to better understand their needs and behaviors, and also adopt different strategies for each group.
• Capture more detailed information: Since users are divided into different groups according to their active time, it is possible to dynamically understand information such as user activity, usage frequency, and loyalty.
• Create more growth opportunities: The layered lifecycle model also allows us to better understand which users have potential for growth, so that we can formulate corresponding strategies accordingly.

3. How to flexibly apply and break through the growth dilemma

Duolingo's life cycle model provides us with a new way of thinking. While using Duolingo's successful experience, we also need to flexibly apply its methods and develop growth strategies that suit our own.

To sum up in one sentence: So lucky, you can apply it directly.

For products such as communities, entertainment, and information , who can resist such a MECE data model? Imagine that next time the boss asks why the data went up or down early in the morning, we don’t have to fumble around in various data tables, and can immediately give an accurate answer.

Not only can you find data analysts to build data dashboards and complete daily data monitoring , but it also brings more advanced gameplay: data prediction and mining core levers.

For example, after building the data model, Duolingo's growth staff kept daily records to observe the daily changes in different user groups and retention rates over the past few years.

With this data, they can simulate future data and perform analysis to predict which levers will have the greatest impact on user growth. The figure below is the result of their first simulation data estimation, which shows the impact of different retention/churn rate data changing at the same rate on MAU and DAU.

It is clear from the results that CURR (active user retention) has a huge impact on DAU , which is 5 times higher than the second most influential data. They later realized the truth. From the current user level, active users with different levels of participation will always be classified as "active users" in the end.

(Image source: Lenny's Newsletter)

Based on this analysis, Duolingo confirmed that CURR was an indicator that must be overcome in order to achieve a breakthrough in user growth. A series of strategies were launched: gamification mechanism to increase the overall user usage time, strengthening the message push function to increase user participation, and optimizing the continuous check-in mechanism to positively motivate users to be continuously active.

(Image source: Duolingo product experience screenshot)

After four years of hard work, Duolingo's CURR increased by 21%, which means that the daily churn rate of their core users decreased by more than 40%.

How to carry out refined operations of user segmentation? These general strategies can also be used as reference:

Image source: I drew it myself

① "Activate" potential users

• Activate new users, optimize the onboarding process for new users, and enhance the first-time product experience;
• Increase the silent new user guidance channels: add multiple triggering methods - SMS, push, email, service number;
• Set up appropriate incentives: red envelopes for new users, discounts, etc.

② " Retention and monetization" healthy and active users

• Start by optimizing product functions to improve user retention;
• Further stratify users, compare and find the differences in retention rates of different users, and improve retention in a targeted manner (channel segmentation, age segmentation, mid-week/weekend segmentation, user search terms, user click-through rate);
• Improve user engagement by increasing user usage frequency and intensity, thereby improving retention (guided by incentive system);
• Flexible pricing: algorithmic discounts, regular events;
• Optimize the core payment path: test and optimize to increase the conversion rate of each step and ultimately maximize "monetization".

③ "Discover and intervene" users at risk of churn

• Direct negative behavior occurs: for example, a member who is about to expire clicks on the renewal page but does not renew; for example, a large amount of data and documents are exported; for example, a user gives an "unsatisfied" rating when reviewing;
• Changes in behavior patterns: For example, the user logged in several times a day before, then once every three days, and then not logged in for a week;
• Data model prediction: Based on a large amount of data, a "churn prediction model" can be built to generate a "churn possibility" score to help pay attention to users at risk of churn in advance;
• Storage of user data: data expiration reminder if not logged in for a long time;
• Increase conversion costs: For example, the points and level system greatly "increases conversion costs";
• Lock in users in advance: For example, plan an activity to purchase multi-year memberships to "lock in users in advance";
• Loss warning mechanism: Give corresponding incentives in time to activate the "loss warning mechanism".

④ "Recall" lost users

• Consider whether to conduct user recall: Is it worth doing? Why do they come back? Can they be retained? If the weight is added to make the product significantly improved, it will be very helpful to bring back certain types of lost users, and users will have a reason to come back.
• Select recall targets: There are 4 types of churn targets: churn due to non-login, churn due to non-activation, churn due to new users, and churn due to long-term users. Usually, the second and third types of users have a greater chance of being recalled because they are reachable and have not yet deeply experienced the value of the product.
• Recall A/B testing: From the recall object, recall timing and frequency, recall channel to copy design, sending time, and jump path, all can be tested and debugged to achieve better results.
• Measuring the recall effect: Direct results can be judged from the push opening rate, click-through rate, and 24-hour visit rate; while long-term results can be judged from whether there is a core behavior (Aha moment) and the retention rate of recalled users. Long-term results are the ultimate goal of recall.

4. Write at the end

The traffic dividend is disappearing, and refined operations are the way to go.

Duolingo’s successful experience tells us that the traditional AARRR model is not the only operating model. We need to flexibly apply different operating strategies and data models based on our own product characteristics and business needs.

Only by deeply understanding user needs and behaviors can we develop growth strategies that meet user expectations and achieve sustained growth in the fierce market competition.

The above is the content of "Subverting the traditional growth model and innovating against the trend with life cycle management" . The reference comes from user growth work practice & research. If you have different opinions, please leave a message in the comment area below for discussion.

Next time I would like to talk to you about practical strategies for recalling old users of various platform products!