User Growth Framework and Practice (1): Building the Basis of User Growth Strategy: KPI Dismantling

This article briefly explains user growth, introduces user growth, and explains how to build a user growth strategy from the KPI dismantling method. Let's take a look~

In recent years, the definition of user growth has gradually blurred, and sometimes it is due to traffic operation. Many businesses integrate the functional departments that need to be used into a team, such as an independent set of operations + products + technologies + algorithms, which is expected to play a role in business growth. It may be that the actual work content is more about daily process management. When I talk about user growth, I mean a methodology that has a significant increment across the business. Others summarize me as a breakthrough strategy. The result of the last business:

I. Preface

It's so simple and straightforward that you'd be surprised that no one is doing such a basic thing? Yes, I was the first to implement most of the strategies I mentioned. Since the companies I've worked for have been very good, I'm sure most of them are in this state as well. Because most people can only vaguely understand the surface of a concept, they often discuss, including cooperative media, and draw concept drawings that are full of imagination, but the logic cannot be self-consistent. It is also possible that your company has a similar middle office product, but unfortunately no one understands the essence, so it does not play a big role, and there is almost no possibility of landing in the business of a large company.

I am based on the experience of the growth of e-commerce users on the top platform. There is not much difference between non-e-commerce businesses, and they can be reused with a little transformation. It's good for medium-sized businesses, and it may not be suitable for small businesses, but some analytical ideas may inspire you. The whole process may seem simple and clear, but if you don't understand the business value and the underlying logic enough, it's likely that the method I'm talking about won't help you much.

I have heard newcomers complain that the company does not have many conditions and cannot carry out their work normally. I have to advise newcomers, and that's basically because you don't play a key role and don't fully understand the business. There must be a way. I did about 300-500 data analyses with in-house product X, and no one else knew where I got it. Even X's product manager didn't know it could be used like this. When you make a result, the other teams will immediately support you.

I have basically no communication with my colleagues in the industry. They all analyze the data, sort out the logic, build the framework, go live and test, and gradually scale up with the help of the algorithm team's capabilities until they are complete. The language may not be uniform, so I hope you can help correct it. It's like when I won the first prize at the Beijing municipal level in high school, a bunch of people came around and asked me which forum I was talking to, and I didn't even know what the forum was at the time.

2. Construct the foundation of user growth strategy: KPI dismantling

The premise is that the company's goals have been clearly defined, whether it is the scale of paying users, the size of orders, or GMV, it is clear. The first step of usage increase is to disassemble and distribute user data in various dimensions from traffic to the target.

Let's start with my definition of engagement (abbreviated E), which is the number and weight of key actions. For example, key actions: register, receive vouchers, complete the beginner's guide. The impact of these on LTV is enormous. It also includes a variety of important actions: follow a KOL, talk to a business, and build friends in the game. The impact on LTV is subtle but can continue to grow.

Each business can calculate the E weight according to its own data, for example, if the registration is 10 points, the attention may be 0.1 points. Then the key process indicators of operation can be summarized as CPE, cost per engagement score. where cost includes the value of its own traffic (organic). It is a misconception that some businesses treat their own traffic as free.

We break it down into two categories: user characteristic data and behavioral data. It not only helps to clarify the current situation and find the first breakthrough, but also helps to specify a complete growth strategy and the starting point for each specific direction, and also provides a foundation for subsequent content optimization.

1. User Demographic Data

Demographic data helps us understand the user's background and persistent attributes, such as geographic location, gender, marriage, family members, devices, telecom operators, etc. Most of them are subject to the latest status. This data is critical for segmenting user segments, customizing creatives, and anticipating user needs.

For example, the distribution of city-> efficiency in a country

From the distribution of the population, it can be seen that the law of data distribution. If you open the map, you can also find that the general coastal cities have their own ports, or the areas that compete with port-based cities, the effect is also good. (Coastal is just one example, and inferences need to be drawn from the past.) ）

For example, a national telecom operator - > efficiency distribution

As a general rule of thumb, such as the city-level data above, large populations tend to work well, but this dimension will tell you that it is not necessarily. It just so happens that the user characteristics of the largest operator do not match the supply side of the service.

If there is a professional data team, data verification is also required. For example, if the registered mobile phone number is hashed into 10 groups, the efficiency of the two groups should be almost the same, and any dimension can be.

2. User Behavior Data

Behavioral data reflects about 15-20 types of user interactions with the product, such as browsing, clicking, and purchasing. The statistical caliber needs to be within the session and recent (can be tentatively set to the last 10 days). This information can be used not only to predict conversion probabilities, but also to track changes in users' interests and preferences in real time to optimize product and content recommendations.

[todo:session add-on behavior to conversion distribution chart]

[todo: Purchase number to conversion distribution chart in the last 10 days]

Counterexample: Simply use a part of the behavior data to define the user data disassembly logic, such as how many payment behaviors are in the history. How many days have you visited the app in the last N days. The number of days in the last N days that have been paid. There is no shopping behavior beyond M days. No access behavior for more than N days. Personally, I understand that this is a sword-seeking, because daily operations directly interfere with some data. The point is that there is no causal relationship between this data and the end goal, and it is not actionable for the user team.

How to find the main line? The variance of the distribution of user characteristic data can be reversed, and the operability of the operation can be considered. Operability For paid channels, budgets can be reallocated on a regular basis. For example, if you create 50 campaigns in the top 50 cities, almost all of the media support targeted advertising at the city level. We can re-evaluate CPT (cost per target) on a weekly or even daily basis and allocate the budget to value depressions.

Operational Operability Examples (Each business is very different, learn to draw inferences from one another.) ）：

• The mainline dimension can be downgraded at the campaign level. [TODO: Own traffic operation method.] 】
• Geographical location, most commonly a city. Based on the city-based content preference strategy, the difference can be as much as 50%. For e-commerce, the availability and efficiency of logistics also need to be considered.
• Equipment, carriers. Targeted service. For example, the content of recharge, recycling, film, and usage tips.
• Gender, marital status. Content preference strategy, the difference can even be as high as 90%.
• The dimension of age also shows a very clear signal. Our business makes certain age groups happier users. In other words, it's not good enough for certain age groups. I judge that there is a shortage of age-specific content and that there is a problem with the recommendation algorithm.
• If there is a large difference in efficiency, see if there is a lack of supply side on the supply side, and it is likely that key goods/content are missing. Let's see if there is a big difference on the traffic side, assuming that the volume is similar, whether the traffic cost is quite different. If A is not as good as B as a whole, and the traffic cost of A is lower, and the conversion process is not much different, it is likely that there is a problem on the supply side. Although we cannot solve the supply problem in the short term, we can control the allocation of resources.

Some data analysis results of service strategy (for everyone's understanding, I have listed many examples, in fact, I basically used two algorithm logics. ）：

• The probability of inertia of the behavior of the geographical location can be up to 90%+. (This is a response to the challenge of the team: how do you know that what you're saying is right, can't the user xxx?) ）
• When the user's APP is separated by N versions from the latest version, the probability of conversion is greatly reduced. I deduce it has something to do with user loyalty.
• The conversion rate of users of a leading operator in a country is significantly lower than that of other operators. It is likely that there is a mismatch between the spending power of such users and our business.
• Category preference, you can clearly see the difference in the distribution of users' conversion rate and retention rate. Only a small fraction of the data is needed to determine the user LTV in advance.
• Category preference is a dimension that also greatly influences content preference. One of the most classic cases, if it is known that users have a demand for wigs, the success rate of pushing hair dye is actually greater than that of pushing beauty products. I believe that smart students can dig out the reasons on their own.
• Trying to say a little more about the devices, maybe there is more similarity between the flagship models, rather than by brand.
• Women can buy fast fashion products for men. But not the other way around. Fast fashion, such as men's clothing and men's bags, at most even 50% of buyers are women. This is consistent with the traffic phenomenon I observed in the malls, where Zara has a 1-storey women's and 2-storey men's clothing in most malls. We will not discuss the 3-layer children's clothing for the time being. On the first floor, all the people shop are women, but on the second floor, you can often see women carefully selecting products.
• Once women have children, the amount of money spent on fashion and beauty is greatly reduced, and the demand is inclined to children and families, kitchen, cleaning, home decoration, and storage.
• Girls' products are basically bought by mothers. The boy's products were bought by the father.
• By identifying the gender of the child, the fertility rate of the country or region can be extrapolated, which is almost the same as the public data.
• Deeper data analysis, i.e., rigid and improved needs, can again improve operational efficiency. Fashion just need for example: clothing, lipstick, foundation, blush, highlighter (three-piece set). For example, high-end cosmetics such as eye shadow, shoes, bags, jewelry, watches. Then unfold. （devils in the details。 Some businesses only consider supply-side classification, not demand-side classification. The products that were originally needed were classified as improved, such as the card issuance to fashion accessories. will make you draw conclusions even the opposite. ）

In the actual landing process, I first used "category preference" as the backbone, and then upgraded to "irreversible life state" user portraits, and gender + marital status. So my final backbone is like this, the business with a larger number of users can continue to be subdivided, so I won't repeat it for the time being:

• Mothers of infants and toddlers aged 0-6 years
• Mothers of children aged 7-18 (demolished here if possible)
• Unconceived women
• Father
• Non-paternal males
• Unrecognizable

About the caliber of KPIs. A basic but important detail, and someone actually made a mistake.

There may be businesses that use a wide attribution caliber, such as 24 hours or even thicker. Business-specific analysis. When it comes to e-commerce alone, session can basically be used as a suitable caliber and can be modified on this basis. However, if the coarse caliber is used, no matter what you do, it will seem to be useless, and even the result will be contrary to business common sense.

When you're running a retail business with your own money and you're very reliant on channels, you're going to adopt a session-like approach. When the channel expects to ask you for a commission with a more coarse granularity, you will refuse. Why is it easy to forget if the money doesn't come out of your pocket? If the CEO sees the wrong goals, the results of the team's efforts can only be random.

This article was originally published by @达太 on Everyone is a Product Manager. Reproduction without the permission of the author is prohibited

The title image is from Unsplash and is licensed under CC0

The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services