Product Operations: How to use data analysis to drive product user growth?

Some time ago, I saw a post asking: Can data analysis drive rapid user growth?

First give the answer - "Yes.", and then tell us how to do it specifically.

Because the data is confidential, I will not use the data of products I have handled to illustrate, but will choose a product that I like to use very much - Kugou, to complete this article. Because it is impossible to obtain accurate product data, I will make analysis based on my own understanding of the product + some third-party data. My purpose is to share common methods of driving growth through data analysis. If there is anything unscientific, please point it out.

Below, enjoy: Common methods of data analysis are [three steps] & [two models]

1. Three steps

The three steps are: determine the core goal, list the composition formula, and confirm the elements.

Common applications in Internet companies are:

Core goal (i.e. North Star indicator) = A*B*C

Take Kugou as an example. Kugou’s slogan is: “You are actually very good at singing.”

The ultimate value of the product is to let everyone sing and let others hear it, so its North Star indicator should be: [Number of original song plays per day]. The number of accompaniment downloads/playbacks, number of friend messages, etc. generated in the process of improving this data are all derived data.

Therefore, the formula is:

Daily original song play times = daily original song number * average single song play times

This is the core data formula. Next, we need to determine the elements that affect the formula. The formula can be broken down into:

Daily original song play times = (daily active users * average number of songs produced per person) * (single song exposure times * exposure - click-through rate)

This formula can be broken down infinitely. According to the scale and functional complexity of the product, the human resources of the operation team can be segmented to the greatest extent possible. We will not continue to list them here, but will analyze them according to the final formula.

It can be found that the elements that affect the core goals are: DAU, the number of songs produced per capita, single exposure, and song exposure-click rate.

It is common sense that in multiplication, the improvement of each element can bring improvement to the overall result. The operation team can make targeted growth plans based on these four elements:

1. DAU

Depending on the definition of the product using different formulas, it can be broken down into many elements.

We simply define the active users of Kugou as registered users who have logged into the app on the same day, which can be broken down into: number of registered users * opening frequency.

The number of registered users can be increased through many methods such as cross-industry cooperation, friend invitation fission, app stores, information flow, etc. These are routine operations and will not be discussed in detail in this article (key point: pay special attention to black market users).

In the long run, the frequency of opening can be improved by setting up effective functional scenarios (for example: clocking in, singing teaching), reasonable channel push (message push reminders, etc.), etc. In the short term, you can use some event marketing and community communication methods to attract users' attention and discussion, thereby temporarily increasing the opening frequency.

2. Number of songs produced per capita

First, we need to do user segmentation: What is the average level of high-quality users (who have produced content)? What is the level of an average user (who has never produced content)?

There must be a part of Kugou users who don’t sing, and only 30% of them may have produced content. So, should we only count the data of users who have produced content?

Of course not.

First of all, such statistics will affect the data formula. We cannot use the average level of high-quality users to expand the statistics of all users. This will cause us to be blindly optimistic about the current situation of users and even make wrong decisions.

So why do we need to layer? Wouldn't it be better to just calculate the total average?

I simply made a funnel for karaoke users. (It’s really that simple…)

According to iResearch data, the monthly active users of WeSing are 160 million. Assuming that active users are defined according to my logic (the following data are all personal assumptions): divided into 4 levels with a ratio of 4:3:2:1, the average output per person is 0.16 songs per day

From a strategic perspective, we need to increase the number of 0.16 to 0.2 or even 0.3.

There are two ways to execute:

1) Allow users from each layer to flow to the next layer.

——That is, let silent users start listening to songs, let users who only listen to songs start singing the first song, and let users who occasionally sing become active. The basic logic is to change the structure of 4321.

2) Increase the average output quantity at each level.

For example: let 0 become 0.1, let 0.3 become 0.5.

Based on the above data analysis and target breakdown, operations can have clearer and more refined strategies.

For example: The goal is to increase the number of users who only listen to music but do not sing, from producing 0 songs per person to producing 0.1 songs per person.

You can plan a corresponding "Your First Song" event by lowering the threshold for participation (for example: do a simple microphone grabbing event, just sing a few lines, no need to sing a full 4-minute song, let the user speak for the first time first) and provide incentives (whether it is emotional incentives - forming a team to grab the microphone, or interest incentives - issuing gold coins).

3. Number of single exposures

Original songs are usually exposed in the following three ways:

1. Social relationship chain - follow/friends, etc.
2. Machine algorithm recommendation - nearby/recommended, guess you like, etc.
3. Fixed exposure position - Discovery - various charts, song request - various charts/categories, advertising position, etc.

Operations can be based on different forms, set different data goals, and plan corresponding operational activities to increase the exposure of songs in different sections.

For example, in the social relationship chain, the key data target is the number of followers/friends. The more friends you have, theoretically, the greater the chance and frequency of your song being exposed.

Of course, this involves the definition of user activity and stratification. If a person’s friends increase from 50 to 100, it does not mean that the number of song exposures will double. But in general, after social relationships are expanded, the original songs released will definitely have a higher chance of exposure in the following/friends sections.

Therefore, the purpose of the operation has changed from [increasing song exposure], which is a goal that we don’t know how to achieve, to [increasing the average number of friends], which is a more specific and executable goal.

The next step is to plan specific activities, whether it is through group microphone grabbing, remote interaction, or emotional interaction matching on the social side of strangers. In the operation mechanism, just pay attention to strengthening the key action of [adding each other as friends]. I will not go into details here.

To put it bluntly, one of the key roles of data analysis here is to find more specific and executable directions on the operational side through data analysis.

4. Exposure-Click Rate

There are two ways to increase exposure-click rate:

1) Find people who are more compatible with and interested in the song to maximize the efficiency of exposure.

For example: recommending a cover of "Qilixiang" by an ordinary user to Jay Chou's fans is definitely more efficient than pushing it to Mayday fans. All the operations staff needs to do is to do user tagging, grouping, and algorithm recommendation, which I will not elaborate on.

2) Increase users’ desire to click and play. Simply put, it enhances the appeal of the song.

This involves optimizing the content, such as: push cover (size, shape, visual, etc. of the header image), title, copywriting (for example: 70% of friends have heard of it/people with scores above 85%, etc.), etc.

High-quality content can effectively improve exposure-click conversion rate, so what operators need to do is a lot of AB testing to see which form of content can best attract users to click, and that's it. Just look at the data.

In summary, the main purpose of the [three steps] of data analysis is to enable the operation team to find the direction of growth and find the entry point of operation through reasonable decomposition.

2. Two Models

The two models are: funnel model and coordinate model

1. Funnel Model

The most typical one is the AARRR model. If you don’t understand it, you can search on Baidu. I won’t go into details.

In addition to the entire process of user acquisition, the funnel model can also be used to analyze a single case. The funnel from top to bottom basically represents the user journey map (in layman's terms: a complete interaction of a user on a product). Mainly used for bounce/churn analysis to identify problems - where are users lost?

Let’s take Kugou as an example. Our goal is to let more people who have never sung before start singing their first song.

Based on this goal, the operation set up an H5 activity - Test your singer's content (the copy may need to be revised...)

The general gameplay may be: choose your favorite song, the system will give you the lyrics, play the original version, long press to record, get the result, and share it to your status.

If I only tell you that the activity page has more than 1 million UVs, but only 500 data are shared to the dynamic in the end, do you know where the problem lies?

At this time, based on the traffic funnel analysis of each H5 page, we can find problems and improve them.

for example:

1) We found that the event page had 1 million UVs, but only 10,000 people completed song selection, which is only 1%. Why?

This is a very unreasonable and far below expected data. Based on this, we have a guess, is the song library not rich enough, or are users too lazy to choose on their own?

Therefore, I will analyze the page embedding data again. For example, 50% of users clicked on the song search function but did not proceed to the next step. This may mean that there is a problem with the richness of the music library or the search matching. For example: Most users click around on the page but never click on the search box? He may be lazy and just looking for system recommendations.

Based on the guesswork, we can do further testing or user research to verify whether the idea is correct and find out where the problem lies.

Of course, content testing should be optimized before going online.

Then let's look at another situation that appears more often in the actual environment:

2) We found that there were 100,000 users who completed the recording and obtained the result page, but only 500 users chose to share the result page to their dynamic feed. Why?

Through the funnel model, we clearly found that the main loss occurred in the "results page-sharing" link. It’s the same as before, first make guesses based on the data, then come up with solutions to verify the ideas.

First of all, 100,000 users produce the result page, so there should not be any major problems in the previous links. Then what is the reason that users “are unwilling to share to the dynamics even if the sharing page is produced”?

Possibility 1: The result page is too ugly and makes people turn it down.

Beauty and ugliness are not absolute. Testing may have been done before the activity goes online, but the page finally selected is not recognized and shared by most users (just like boys and girls, first- and second-tier cities and lower-tier markets have different preferences). What should we do then?

Solution: It is difficult to satisfy everyone. Either do an online test to find out the style that most users like, or make a few different sharing page layouts and let users choose freely.

Possibility 2: No clear sharing guide

The user reads the results page, says OK, and then closes and leaves. There is no clear guidance from the organizer of the event - for example: share to get XXX (benefit inducement), share to the dynamics for friends to see (social currency/shape personal image). Most users are lazy and need clearer guidance. Let them decide freely what to do. In the end, users are likely to decide to do nothing.

There are many more possibilities, which I will not list exhaustively. The main point is that the funnel model can effectively find problems.

2. Coordinate model

The most classic one is the RFM model, which is used to stratify user value based on behavioral data, thereby achieving refined operations. It is used more in the overall operational growth of products and less in individual activities.

Common RFM models are as follows:

Basically, the model is built based on the following three data to divide the user ranges with different values:

• R = Recency
• F = Frequency
• M = Monetary

RFM is common in e-commerce platforms. Let’s take [Kugou] as an example. First, select a few data dimensions. I will choose:

• R = most recent interaction (based on the analysis above, I chose interaction instead of login. The definition of interaction may be playing a song or something like that)
• F = frequency of interaction
• T = duration of single use (for simplicity and understanding, it also means the magnitude of the behavior)

Next is statistical data, modeling analysis, and I don’t have specific product data (the dimensions mentioned are also difficult to obtain through third-party data). Well, that's it for now.

It probably means this:

• Those who have recently interacted on the app (the definition may be within 48 hours), have frequent interactions (maybe logging in more than 10 times a month), and have a long single usage time (single online time of 30 minutes?) are defined as important value customers. We cannot lose them no matter who we lose.
• There has been interaction recently, but overall the interaction frequency is not high and the usage time is very long, which means that he has opened the app recently and the single usage time is very long, but the interaction frequency is not high. This proves that users are very interested in the product and have great potential. They are the focus of development, and more activities should be planned to increase the frequency of their interaction.
• Users who have not interacted recently or have a low interaction frequency, but who have a long single usage time in history, may have loved Kugou in the past but are now on the verge of leaving, and are the targets that need to be won back.

The RFT data model can effectively stratify the value of Kugou users and perform refined operations for users of different value levels. The resources and energy for operations are limited, so of course we have to do more important things.

In addition to user value stratification, there is also stratification based on the life cycle . This is too long to explain, so I will not give an example in this article. I will write about it next time when I have time.

In short, data analysis drives user growth mainly in two aspects:

1. Through the target formula and the split specific data elements, we can come up with a specific and executable operation strategy;
2. Through the analysis of data models, problems can be discovered and refined operations for users can be achieved.

The analysis in this article is relatively basic. Welcome to share your insights on data analysis in the comment section!

Author: Xie Xiaoyang

Source: Xie Xiaoyang