The common concern of hundreds of thousands of Internet practitioners!

Author: Liu Liming

Source: Birdbrother Notes (ID: Niaoge8)

Editor: Juvae

With the development of mobile Internet in recent years, “big data” is becoming more and more popular. I will share how to drive business development through data from these four aspects.

01 Data management;

02 Build data operation system from 0 to 1;

03 Data depth analysis;

04 User Management Policies.

01 Data management

1. Misunderstandings of data-based management

① Can more data drive the development of data?

Everybody talks about big data, I want to collect a lot of data, but actually this one is not true.

The first is uneven data quality, such as brushing this behavior;

Secondly, if the data collected is not standardized or standardized, it will be found that the collected data are garbage data and have no substantial role in business analysis.

Third, it is easy to collect data, difficult to use data. Tencent has done so many years, has its own set of data analysis system and methodology. But at the beginning of the establishment of this system, also need a long time of training, practice, iteration, finally to achieve good results. A lot of things like mobile, banking, some traditional businesses, they have a lot of data, but they don’t know what to do with it, and that’s a big problem.

(2) Will the data analysis team find problems and opportunities?

A lot of companies have data analysis teams, and some people think that having a data analysis team can solve their business problems or find more opportunities, but it’s also very difficult to set up.

In many of the companies I work with, the analytics team is separate from the business team. The data analysis team will provide some data analysis results, but only some mechanical report analysis. For example, when the data suddenly increases or decreases, the data analysis team may analyze and draw conclusions from its understanding dimension. However, due to its independence from the business team and lack of understanding of the business, it is difficult to give specific opinions.

For example, in the early stage of our product promotion, I brought in the data analysis team, hoping that they could make effective analysis on business growth. However, I found that if only the data were given to the data analysis team, the results of analysis did not produce substantial growth effect on the business.

Then I made adjustments and set up a project team, and the operation team and the data analysis team merged into a project team, so that the latter knew what the data indicators designed by the former looked like in the actual use process. In the end, this kind of cooperation does work.

③ Is the analysis report the best solution?

If you are an APP maker, you should read a variety of daily papers every day and a variety of weekly papers every week. But seriously think about whether the conclusions given by these reports are in line with the business objectives.

Sometimes the dimension of analysis is too single, and may be simply analyzed from the dimension of time, region, or even version. And if we do a deep enough analysis, those conclusions might not hold.

On the other hand, data quality can interfere with the analysis results, such as the stability and accuracy of data, and some irrelevant data should also be vigilant.

2. Ideas of data-oriented management

Our data management mainly pursues such ideas: data collection – data analysis – action strategy – fast implementation.

① Data collection

Data collection should have a very clear plan, including basic data statistics, user attribute statistics, user source statistics, user behavior statistics, and model data statistics.

Basic data statistics are related to daily business, such as new active, new times and other basic data; The statistics of user attributes include two aspects — social attributes and device attributes. The former refers to gender, age and education background, while the latter refers to models, operating systems and networking modes. User source statistics are some channels, there are versions; User behavior statistics is very important, because it can reflect the path and behavior habits of users using the product, and then describe the user portrait according to the characteristics of users, and even product iteration.

② Data analysis

The data itself is meaningless. You need to analyze the data to get the results you want.

The first is multidimensional crossover analysis, which analyzes data performance from multiple dimensions. For example, “new users” should not only look at the overall growth, but also look at the growth of various regions, versions and even different models.

The second is user group analysis, users need to be divided into groups, and establish the characteristics of different groups, to mature, need to do differentiated user strategies for different users, in order to achieve the optimal effect;

The third product quality is mainly to see the performance of the product, such as the result of the call, whether the product consumption flow, whether electricity consumption, etc..

③ Action strategy

It is more inclined to user cycle management, which mainly includes recruitment, activation, loss recovery and reflux care, which will be discussed in detail below.

④ Fast execution

Finally, fast implementation. If we have data analysis and action strategies, but can’t execute them, the previous work is meaningless. We must finally get to the ground before we can see the results. There should be precise hierarchical management of users, feedback on execution results, and continuous iteration planning based on user feedback.

02 Build the data operation system from 0 to 1

1. Set up ideas

① Indicator Planning

In order to collect the data well, it is necessary to make the index planning for collection, including index definition, dimension setting and update cycle, in which the update cycle involves the allocation of resources, whether it is updated constantly, weekly or monthly.

② Data collection

Data collection is based on what is done after major planning, such as field classification, data burying point, and data reporting. Think about what data to collect and how to report it.

③ Report presentation

After data collection, we can do report presentation. There are a lot of holes here. For example, if we want to make a trend chart for the report, should we use a bar chart or a broken line chart? A list is a detailed similar table of users; Filter controls, if you want to achieve visualization, you need to consider the future in the actual use of which dimensions to filter, such as country, version, channel and so on; Finally, we need to verify the validity and accuracy. After we report the data, if it is a pile of garbage data or inaccurate data, it will not be helpful for the subsequent operation.

④ Data products

Once the basic data collection is OK, we can consider building data products, such as data visualization, iterative optimization, and adding new functions.

2. Construction method

There are two ways of construction, one is self-built, the other is the use of third-party services.

Self-construction has advantages and disadvantages. The first is that it is flexible and convenient to bury where you want to bury it. The second is that it can get through with business data, which is also very important. Because in the use of third-party data statistics, people usually use only basic indicators, in the absence of their own sales data, such paid data is often separated from the business.

Disadvantages one is that the cost of building a data analysis system is very high, and the other is that the self-built system can not get through with operation tools (refers to the advertising platform).

Let me give you an example. When I build a self-built data analysis system, the product has its own user system or label system. If I want to target users who are interested in reading, the final release effect may not be particularly ideal, because there will be a large deviation between the self-built system and other platforms in matching. Tencent itself will also do some advertising, such as toutiao channels, found that we take our own data package to advertising, may not be very good effect. We invested more than 6000 yuan, and the cost of a user soared to 600 yuan, which is a big problem.

The second type is third-party services, many of which are free. Their products are professional and friendly to display. They support operation tools to improve operation efficiency, but the biggest disadvantage is that they are not flexible enough to be connected with their own business data.

3. Iterative optimization and gradual improvement

The data operation system is not built overnight. In the case of limited human resources, we can make selective construction according to the current stage. I divided the whole life cycle of App into four stages, namely start-up stage, growth stage, maturity stage and decline stage.

In the early stages everyone is focused on the overall growth rate of the product, and I think it’s also focused on new activity and user sources. These will help achieve business growth in the initial stage of the product. Product quality is also important. If the App consumes a lot of power, or the Internet connection is too slow, the product may die before it is promoted.

In the growth stage, we should pay attention to the user growth rate and user behavior data. At this time, the product has a certain scale, we need to pay attention to some behaviors of users in the whole product. For example, the number of startup is not enough, the time to stay is not enough long, the use of depth is not enough deep.

In the mature period, the product is basically stable, and the effect of pulling new products in this period should not be particularly good. Therefore, if you look at all large apps, such as Tencent’s iQiyi and Vipshop, internal QQ and wechat, etc., when they reach a certain scale, they will no longer pursue a larger number of users, but activate the existing users, which is very important.

So this stage to dig deeper active users, to prevent loss. We need to do a user churn monitoring, such as churn warning, we need to make a model to monitor, through which user behavior can determine that a user is basically going to lose.

For example, if a credit App chooses to borrow money in installments and then pays back all the money one day, it is likely that he is a lost user. Of course, this is only one dimension, and we need to evaluate it from multiple dimensions. For example, he has not used the App for half a month, which is also a dimension. All models need to judge whether it is really lost from many dimensions.

We’ve done a lot of testing, and once you lose users, it’s hard to get them back, so let’s see where the interest is going and start a new business.

This is the index system of MTA, and the basic indicators include the above.

03 Data in-depth analysis

1. Multidimensional downward analysis

The first is multidimensional downward analysis, which we often use, because when we locate problems, we see changes in the overall data, but in the end to find the problem, we must pick up the pieces and check every dimension.

Some commonly used dimensions, including channel, version, user group, tag, page, region, and then analyze from these perspectives, which users are the most severe decline, which groups have big fluctuations.

2. User portrait insight

This is a set of labels we build, including the first and second level labels, and there is a finer layer below the second level. User profiling insights not only help us get to know our users, but also help us monetize them. For example, knowing what types of users you have can help you analyze which types of ads work best.

3. Funnel transformation analysis

This is also a common method of analysis, because it not only helps you analyze the conversion rate from the first step to the final step, but also the conversion rate at each step. For funnel, a single analysis is meaningless. Without comparison, it is difficult to find the problem. Therefore, a more detailed analysis should be made through trend, comparison and subdivision.

This is the performance of our platform (the data is processed). Our product is for developers, who integrate our SDK, report data, and release it on the App Store. When we look at the data of July, we find that the registration test fluctuates greatly, 40% in May and about 40% in June, but it drops to 21% in July. The data of the last two steps are basically not much different from that of May and June. Therefore, the analysis shows that the change of this ratio may be the same in the numerator but the denominator.

It can be seen that among the growth data of new users in May, June and July, there was a substantial increase in July, but there was no big fluctuation in test users and online users. It was necessary to find out where the problem was, so it was necessary to do driller analysis.

Driller analysis looks first at the growth of each version and then at the growth of each channel. There are many user channels, such as voluntary registration on the official website, as well as the amount brought by wechat and QQ open platforms. Finally, we found that the official website had a large growth volume, which was concentrated in Guangdong. After investigating the activities in Guangdong province in July, we found that there was a round of prize answering activities.

As long as you take part in the questions, you can get the benefits of Q coins and Tencent video members. Many irrelevant users have been attracted to register their accounts and answer the questions, leading to a great increase in the data. Therefore, it is targeted at a problem like “there are many hairy users in guangdong activities in July”.

04 User Management Policies

1. User lifecycle management

User management strategy is to do some operations based on data after data collection and analysis.

It can be divided into 6 stages: potential user stage, novice stage, effective active stage, active decline stage, imminent loss stage, loss stage. Each stage relies on the analysis of specific data, and then makes targeted operation activities.

This is Tencent a gunfight game case, see how they do user management.

1. The Hammocraft

Precise pull new is based on the analysis of historical player sampling, combing 600 pull new fields according to the historical log, to determine which users are most likely to become users, but for startups, there may be no more than 20 fields.

Build new models based on historical players. For example, when launching, target women aged 18 to 20 who like to read, and then do A/B tests. At the beginning, this effect was not good, the experimental group had no obvious improvement compared with the ordinary market. After about 10 rounds of training, this model increased by 30% to 60%, which was quite impressive.

2. Then there is novice care.

Firstly analyze the interest preferences of all players and push some products on this basis. For example, in the novice task, there are different tasks for male and female users and different gifts are sent. This kind of personalized caring reward can improve retention, which is now commonly referred to as precision recommendation, and the same logic is used for thousands of people.

3. Grow actively. This is a shoot-out game, and after analyzing the data, the retention rate of the “with” group is higher than that of the “without” group. This is similar to Facebook retention, where the more friends you add, the higher your retention rate. So adding a social connection to your product is a very reliable way to be active.

4. Anti-loss intervention. First of all, we should monitor the user’s activity. If it drops, it may be the beginning of loss. At this time, we should screen out these users and do some loss intervention, such as sending some care — message push, SMS, etc. If users do not want to open the App, they will not come back until they receive the message, so they can do some user message push.

5. Finally drain back flow. In order to monitor the growth of the whole defection group, some return activities can be designed, but according to experience, once the user has lost, it is difficult to bring back the user, especially after the App has been uninstalled. So it’s better to intervene sooner rather than later.

2. User group management

The first is why do group management. Because clustering can circle out user groups with special attributes and specific behaviors, it also provides a basis for subsequent differentiated operation.

Second, how to create a cluster.

We use statistical indicators, such as age, gender, region, length of use (1-2 hours, 3-4 hours), and status of registration, such as free, trial and paid users. There are also purchase history, visit location are the basis for doing clustering.

Where will it end up? Firstly, data analysis should be conducted, such as the analysis of attributes and behavioral characteristics of different user groups, which can be analyzed in the form of reports or data visualization. There are differentiated operations, can do personalized content, activity push. There is also the precision of new, know which users are high value, understand the characteristics of users, do more effective promotion.

Let’s take another example. This is an e-commerce APP.

The problem is that user promotion is not proportional to user order volume, and the ROI conversion is particularly low.

In view of this case, three subgroups are firstly established. The first subgroup is the mass users, that is, all active users. The second group is transaction users, that is, successful payment users; The third group is high value users, those who pay more than 100.

After stratification was established, population characteristics were compared.

The first is gender. Among the market users, there are more males, and there are more females in the transaction users and high-value users, reflecting that there are more male users in all the users, but the transaction is more female.

Let’s take a look at the crowd preference comparison. Compared with the market, transaction groups and high-value groups are more interested in shopping and finance. In fact, this is an interest label.

Further analysis, the first may be the user drainage channel problem, high value users are female, but the majority of drainage to the large number of users is male;

The second is product positioning, as the products in the App may not be attractive to male users or do not conform to their tastes.

After comparing the two reasons, if the drainage channel is optimized, the effective time is relatively fast; It would be harder to reposition the entire product.

Therefore, choose the first optimization method which is relatively fast, and then look at the data of the whole channel. There are five channels in total. Channel A, D and E all have high transaction rates, and the retention rate of channel D is 27.51%. It can be considered that channel D is A high-quality channel, and channel E also has A good transaction rate. The transaction rate of channel A is very high, but the number of new users is very low, which should be considered more.

In order to verify whether the users of channel A are high-value users, it is necessary to verify whether the users of channel A meet the two strategies mentioned above. It is found that the proportion of females in channel A is as high as 62.36%, and the degree of interest in shopping apps is high, which conforms to the characteristics of high-value users.

After the analysis is done, the strategy can be implemented to verify the effect. First, the proportion of channel investment can be adjusted to increase channel A investment and reduce channel B and C investment. Then, five weeks of testing can be done.

When I looked at the data after the execution, I saw a slight increase in the overall increase, seeing the final conversion rate increase from 7.52% to 9%. This is a complete case study of data analysis, strategy development, and action.

Read More

Email: [email protected]

This article is published by the author authorized by morning Reading class. Please contact the author for reprinting.