How does Xiaohongshu carry out refined growth operations?

In an environment where it is increasingly difficult to acquire customers, how to retain users is an issue that every company must consider. When faced with user churn, how can companies find the causes and obtain the correct solutions to the problem? As a lifestyle sharing platform, Xiaohongshu has grown from 20 users to over 100 million users in 4 and a half years. How did they retain users?

1. Xiaohongshu’s growth path

Xiaohongshu is a pan-category lifestyle sharing platform. As of December 2018, the number of users has exceeded 180 million, which is a relatively fast growth. Looking back to the end of 2014 and the beginning of 2015 when I first joined Xiaohongshu, it only had about 20 people. Now it took us 1,644 days to reach over 100 million users.

The topic I want to share today is: a data analysis conducted within our company on the poor retention of young users.

2. Question: Why is the retention rate of young users relatively poor?

Many guests have mentioned that the cost of customer acquisition is constantly increasing. In the AARRR model, when customer acquisition becomes more and more expensive, how can we ensure the final profit? How to find a balance between profit R and increasingly expensive A?

For example, in the past 1,000 yuan could attract 100 users, with a retention rate of 10%, and 10 people stayed; now 1,000 yuan can only attract 50 people. If you still want to keep 10 people, what should you do? We could only increase our retention to 20%, so that in the end 10 people would stay. As traffic becomes more and more expensive, we must pay more attention to retention issues.

When our analysis team was studying the retention rates of different user groups, they found that the retention rate of users from channels such as information flow was very low. They have one characteristic, that is, they are young , and most of them leave after reading an article or clicking on a note, so the retention rate is very poor.

3. Ask Hypotheses and Data Questions

We made an assumption at the time: we thought that younger users might still be in junior high or high school, and many schools do not allow students to bring their phones to class, so they can only play with their phones on weekends, so the retention rate of younger users might be poor. This assumption sounds reasonable and logical, but is it really the case? Let’s see what the data shows.

We proposed three dimensions and questions for analysis:

First, are there differences in the performance of different young users?

Prior to this, our internal age classification was that those under 18 were considered young, but we felt that this classification was a bit too general because those under 18 included three school ages: elementary school students, junior high school students, and high school students.

There are actually quite big differences between different stages of students, so the age dimension itself needs to be further subdivided.

Second, what content do they want to see when they come to Xiaohongshu? Can they see content they like?

 Every user has expectations for a new product or platform. When they download the APP, activate it and register, they always hope to find something valuable on this platform. If it is not found, the probability of user loss is very high.

Third, does their feed contain the content they want to see?

The homepage of the Xiaohongshu product is a double-column note feed generated by our recommendation system. When new users register, they will select some of their own interests, and then we will recommend notes on related topics to users based on the interests selected by the users.

The accuracy of recommendations directly affects the user experience. To give the simplest example, when I was choosing my interests, I chose fitness, but you recommended a trip to me. That would be far from my expectations. Users would feel that this platform does not have the information I want to see, and would naturally leave.

4. Compare real data with assumptions

I think it must be one of these three problems that leads to poor retention. Next, let’s take a look at the specific analysis of these three dimensions.

a. Are there differences in the performance of different younger users?

The first dimension is whether there are differences in the performance of different younger users. We made a table called "Retention distribution of users of different age groups". We reclassified users under the age of 18 into three categories according to their school age: elementary school students aged 12 and below, junior high school students aged 13-15, and high school students aged 16-18. At the same time, we looked at the next-day retention and weekend retention of these three categories of users.

We were surprised to find two phenomena:

The real difference in next-day retention is among primary school students aged 12 and below and junior high school students aged 13-15. The retention rate of high school students is not much different from that of our overall user base.

Users with poor next-day retention also have poor retention on the weekend, and there will be no rebound on the weekend. Therefore, we came to two conclusions:

Our previous assumption that the poor retention rate of younger users is because they cannot use their phones on weekdays due to school is not true. Even on weekends when they can use their phones, they do not come back.

Young users cannot be divided simply by age, but must be divided according to their school age. Therefore, in all subsequent data analyses, when we need to break down by age, we will take this into consideration.

We also made a second table, which is the "Distribution of new channels for different age groups". The reason for doing this analysis is very simple. We spent a lot of money to buy many young users. If we can't keep them, it would be a waste of money. The marketing department can make some optimizations for channels with a particularly large number of young users, such as increasing age targeting.

From the data, we found that Baidu SEM and Guangdiantong’s information flow are channels for attracting a large number of young users, accounting for more than 60%. Therefore, stricter age targeting should be set in the delivery of Baidu SEM and information flow. For the time being, don’t push advertising to elementary and middle school students because they won’t be able to stay even if they come.

b. What do they want to see on Xiaohongshu?

The second question is that these young users come here after seeing the advertisement. What do they want to see?

There are two ways to solve this problem:

User interviews: You can interview 100 or 1,000 users to understand their reasons for visiting Xiaohongshu. However, the sample size is limited, and the results may not represent the thoughts of all users. Moreover, it is very troublesome to implement and requires a lot of manpower costs.

A better way is to look at the user's searches. Searching is an active and powerful behavior. What he searches for means what he wants to see. So we created a search profile for users of different ages: under 15, 16-18, 19-23, and 29-33.

Once this table comes out, we can basically know what people at each stage care about.

We found that users under the age of 15 mainly search for simple drawings, anime, avatars, and many celebrities; at the age of 16, they start paying attention to dressing, skin care, and weight loss; and at the age of 19, they start paying attention to makeup.

This data is actually very consistent with the user usage scenario. Elementary and middle school students are not allowed to wear makeup in school, and most of them wear school uniforms, so they search for things like anime and wallpapers.

We even found that many young users come to Xiaohongshu just to download wallpapers and beautiful pictures as avatars. As a middle-aged person like me, I may never know about this demand if I don’t look at the data.

When junior high school students grow up to 16 years old and enter high school, they will slowly start to pay attention to their dressing and wear some light makeup. When you are older and get married, you will care about wedding, recipes, decoration and so on.

We now know what our younger users want to see when they come to Xiaohongshu, but do they see the content they want here? What indicators are used to measure this information?

Students who have done searches know that the search click-through rate is the most direct way to measure whether users click on search results. Of course, it represents the degree of satisfaction with the search results to the greatest extent.

Therefore, we pulled two more tables: "High-frequency search terms with no clicks exceeding 40%" and "High-frequency search terms with no clicks below 20%".

From these two tables we find:

The click-through rate of searches for celebrity names such as Dilraba Dilmurat is not high, and the proportion of young users who do not click on them is relatively high.

The high-click words after the search are mainly concentrated on weight loss, skin care, and manicure.

We were quite surprised by this result. What kind of celebrity content do young users want to see when they come to Xiaohongshu? Do you want to read gossip? Or looking for a fan club?

These are indeed contents that we do not have. What Xiaohongshu provides to users is some more real information about the celebrities themselves outside of their work, such as what cosmetics they use outside of work, what snacks they like to eat, etc., but it seems that these things may not be what young people want to see today.

c. Is their feed what they want to see?

The previous two questions let us understand what young users are interested in, what they search for, and what content Xiaohongshu cannot satisfy them.

The third question is, when they passively receive information on the feed, is this content what they want to see?

I pulled three more tables.

The first one is the “User Interest Characteristics” distribution table. Those who have used Xiaohongshu know that when a newly registered user first enters the APP, the system will ask you to select some tags of interest as your startup data. At the beginning, the content we push to you is based on the interest tags you selected.

So we want to see, what are the choices of users aged 13-15? What do 30-year-old users choose? It is obvious from this table that users of different age groups have several differences in tag selection:

The top four tags account for only 20% of users aged 13-15, while those aged 30 and above account for 30%;

The tail is not long: the total amount of the 4-5 tags ranked behind is only more than 2%, but the total amount of the 4-5 tags that older users need to follow is less than 1.5%;

Young users have relatively diverse interest choices, which is more in line with Xiaohongshu’s idea of ​​marking my life.

We know what tags our young users have chosen, so how do we measure whether the content recommended to them is what they care about?

We measure it along two dimensions: content richness and distribution match.

Content richness refers to whether users can see enough notes in this category when they select a tag of interest. We found that "Fashion and Outfit", which was chosen by the most people, also had the most note exposure, which is reasonable; but in categories such as music and games, many people also chose them, but the exposure was very small, which shows that users cannot see enough content they like in the music category.

Regarding the distribution matching degree, I selected some categories with more content on our platform and displayed them in the form of a heat map. Next, I will analyze it from the two dimensions of exposure distribution and preference distribution.

1. The so-called exposure release, for example, in the category of clothing, the platform distributes exposure content, and the exposure among users aged 13-15, 19-23, and over 34 years old is similar.

2. Preference distribution refers to the degree to which users like the content distributed and exposed by the platform. The standard for measuring the degree of user liking is whether the user likes, comments, or collects the content.

This number is very obvious. Although they are both clothing categories, there is not much difference in distribution and exposure, but the degree of user preferences varies greatly. If the color contrast of each block in the heat map is very obvious, it means that there is a problem with the distribution mechanism.

This data gives us some inspiration:

1. The exposure of the top notes is almost the same, but the preference levels are very different;

1. When distributing, even for the same category, different distribution strategies are required for different age dimensions;

This actually proves that we need to carry out refined operations when allocating traffic.

5. Experimental Conclusion

Okay, let’s summarize and look at the three questions raised earlier.

Q: The first question is, are there differences in the performance of different young users?

A: The users with low retention rate are middle school students and primary school students under the age of 15, and most of these users are purchased through SEM and information flow. The marketing department needs to target age information more accurately on the delivery side.

Q: What kind of content do they want to see when they come to Xiaohongshu? Can they see content they like?

A: A large number of young people want to come to Xiaohongshu to watch anime, avatars, celebrities or study-related content.

Judging from the search performance, our celebrity content does not meet their needs very well. The research team needs to conduct user research on this issue to find out what kind of celebrity content they want to see.

The marketing department (especially SEM) can still try more topics such as weight loss, acne removal, skin care, and liquid foundation. Because the content of these topics has a relatively high click-through rate when searching on Xiaohongshu. Advertising should be consistent with product attributes.

Q: Are their feeds streaming the content they want to see?

A: In the categories that many young people prefer, there is too little content exposure and their needs are not well satisfied. In the future, the operation team needs to focus on supplementing the content of these categories. Of course, categories with insufficient content can also be temporarily removed to prevent new users from having too high expectations after registration and thus losing them.

On the distribution side, global popular content needs to be differentiated for different age groups, and the algorithm team needs to adjust the current distribution strategy.

This is a very simple case. We perform data analysis from various dimensions several times a week. In fact, we can also analyze from the user's perspective, for example, is the user using iOS or Android? If it is Android, is it OPPO, VIVO, Huawei, or Xiaomi? If it is these, can we tell whether it is a high-end or low-end machine?

We can switch to different angles to do data analysis. These data analyses can be used to guide the company's next actions, whether it is adjustment or implementation.

6. Xiaohongshu’s Growth Experiment Concept

Below I will briefly introduce the data analysis dimensions we commonly use on Xiaohongshu.

We will use two tools in Xiaohongshu. The first one is called data platform. Data drives growth. This sentence has been said many times. So how can we truly drive growth? That is to see the problem and find direction from the data.

Xiaohongshu treats all product managers equally. When they join the company, they will be given a book teaching you how to write SQL and a platform for running data. Everyone can do it themselves and have enough food and clothing.

The second is the experimental platform. This is Xiaohongshu’s internal experimental platform. Experimenting is a form of awareness.

Let me tell you something that happened a long time ago, about the avatar frame in a mobile phone APP that guided users to upload avatars.

The engineer who worked on this feature posted his own WeChat avatar on it. Later, when everyone was analyzing the low proportion of users who uploaded avatars, they joked that the engineer's avatar was too ugly. Then, in order to prove this, everyone did an experiment.

We uploaded the avatars of all engineers, PMs, and data analysts who participated in the development of this feature and tested it to see whose avatar would increase the proportion of uploaded avatars when used as user guide.

This story is also a small thing. But in fact, there is such a culture within Xiaohongshu that no one knows whether a problem is right or wrong before getting the data, so let's just do an experiment.

Source: 199it