Didi Chuxing—How to use data to accurately operate products?

I have been working in operations before, so today I will look at the data with you from an operations perspective. My sharing today is divided into four parts.

01 What is data-driven precision operation?

What is Operations

Let’s first look at what operations are, which is the process of attracting new customers and reviving a product after it goes online. Now it is moving towards commercialization and forming a complete operations chain, which is to make the product work and ultimately achieve commercialization.

In the BAT era, the operation of online products was more important, including the rise of companies such as Didi and 58. Then it gradually extended from online to offline. The scope and object of this operation changed. The product may only be a part of the operation. There may be an offline operation of a business first, and then slowly transformed into an online one. The complexity of the operation objects and data is also slowly expanding.

How data drives precision operations

Regarding the operational effects of data display, friends who may have operational experience know that operations are limited by some indicators, such as conversion rate and repurchase rate, and each link is a display of data. We gradually accumulate these data, display them and compare them in historical records. There will be more such businesses horizontally, and this is the level of data presentation.

Another thing is data programming. We will have different operation plans . By conducting small-traffic tests, we will ultimately decide our strategies and the plans we will use.

The next step is data mining, which may be based on the accumulation of data. Through the accumulation of data, we can do correlation mining and continuously expand the types and strategies of our business. When Didi first started its taxi business, its operations were very similar, with a split-order dispatch model and platform subsidies. But how could it highlight its own advantages under such a similar situation? Ultimately it comes down to efficiency and data operations. By capturing data, we can see all kinds of information in one place, and then deploy different strategies to reflect the advantages of our operations. These new users may attract more orders.

What effects can refined data operations bring? It can be mainly reflected in three aspects: the first is the optimization of the operation strategy, and the next step is what kind of data to use for optimization. In addition, data is needed in the decision-making process, as well as our predictions and user portraits. By comparing user attributes, we find such users in the large market and then use our new customer acquisition tools accurately. Of course, after using refined operations, the added value of our company's gross profit will be around 6%, which is an industry data.

02 How to use data to drive precision operations

The key points of data-driven precision operations

How to use data to drive refined operations? There are several key points in the internal control of data:

The first is the definition of attributes , including the object’s age, gender, occupation, and income level, which can reflect a user’s traffic.

The second is regional characteristics , including many O2O- related companies. The user characteristics of each place may be different. When we were in Chengdu, the carpooling rate was particularly high.

The third is the content of the operation . Different contents have different user feedback and user records.

The fourth is the record of behavior . Users can access some online traces under the influence of other factors, including search behavior and click behavior. We can see the key indicators from these dimensions. When studying a certain dimension, we can fix the other dimensions and then look at the changes in another dimension.

The example I gave may be related to my current job, because I am currently working on Didi’s user market. In fact, Didi has always had a travel route when it comes to travel, which is to provide one-click travel service. Every time you click in, you can reach the desired place and then be directly sent to the destination, including our 300 million passenger users and our driver owners. Our job now is to provide after-market services for the 15 million car owners, mainly services for people and cars. We will look at these people entering the market and the early service model. Of course, this may also interfere with new service models. We will also deepen our sales services to them, and Internet financial services may emerge from here.

Precision Operation Case Sharing

How did we build our model? That is, we have real-time online reach of target groups and car owners, including the scale of users, and offline matching with our service network, which carries the services and is surrounded by new sales between us and users, as well as sales of our car models and insurance. Let’s take a case study of a car company for analysis, including user portraits, real-time data monitoring, our training specialists, and funnel analysis.

Consumers may have online shopping, such as JD.com and Taobao, and there are also some convenience stores offline. Its characteristics are that there are many categories and models. The club is actually a comprehensive service place with limited space and warehousing. This can also be compared online. Because the product space is limited and the online layout is also relatively limited, how to explore or break through the business in this case?

User Profile

First of all, we need to profile the users. What kind of people are the users of car companies? These car owners are generally between 30 and 35 years old, with elderly parents and young children. Their needs may not be consumer products, but practical ones, such as seat backs. Based on these analyses of users, we will narrow the range of merchandise. Since our display center is relatively limited, we may need to reduce our SKUs (stock keeping units). The prerequisite is to conduct massive SKU testing, including online products, and based on its data, reduce the SKUs to less than one thousand.

This user portrait includes several aspects. On the one hand, it is the data on the platform, and on the other hand, it is the business itself. What kind of products are needed and liked, and what the sales situation is like after arriving here, so we did some pilot work. The premise is to display all our product categories and then gradually narrow the scope. We will also have price-related cooperation and may also do offline user diversion. Through these methods, we will gradually clarify the portrait of this group, including some business data and attribute data.

Real-time monitoring

Regarding real-time data monitoring, sometimes after we launch an activity, we find that the click rate is not high, only about 5%. After comparing it with previous data, it returned to 12% the next day. This is real-time data monitoring. Each decision-making cycle must be able to see the previous data, which can then influence the next decision. This is our requirement for data. Because we have both online and offline businesses, we have a particularly strong requirement for real-time performance. Pure online operations may cause data links, while offline operations are more likely to be solved with the use of systematic tools.

Association analysis

Data analysis is often used in e-commerce . Thirty percent of users may buy a body package when they buy a car. At this time, related products will also be derived, and then we will make related recommendations. This is also a complete product system.

Let's talk about the funnel model. In fact, it can basically conform to our psychological process of marketing . It generates interest and desire, and finally takes action, corresponding to online visits, clicking orders, and payment. Each action and each link has a conversion rate. The process of defining the funnel model can be divided into several aspects. The first is the data application process, including our target system, and we cannot find our key factors for optimization. This includes purchase rate, repurchase rate and purchase and sales rate, and these indicators are the main focus.

03 Data-driven precision operation implementation steps

Data-driven precision operation process

If we are faced with a new business and a new product, how do we establish a powerful data-driven and refined operational process? Including the core path of data application, because sometimes everyone does such things every day in operations, and may feel that this is daily work, so this process is very critical, that is, when our entire data comes out, what the user interface is like, and then there is the sales link.

Next, we need to make things quantifiable around our goals, that is, to analyze the target system in a quantifiable way. Based on this business logic, we do quantification. After quantification is completed, we need to collect data and establish a complete data supply chain. We need to know where the data comes from and how it should be collected. Then we need to analyze it in multiple dimensions. Some aspects of the analysis have been mentioned just now. After the analysis, the most important thing is to match it with our operational work. What is the operational action of each data performance? Does such operational action produce good results or bad results? If it is good, we can continue to persist; if it is bad, we can make adjustments.

What may be driving the decision more is the mentality of attaching importance to data. Whether you need to look at the data every day, or insist on looking at professional indicators, and whether your daily work is carried out around the data you look at, this may be something that needs to be done in the early stage, and it is the most critical thing.

04. Data-driven precision operations encounter problems

In fact, we also face several difficulties in the process of data-driven operations: First, there is too much data, not only in terms of quantity, but also in terms of data optimization. We may be faced with a very large data system. How can we collect it in the most structured way possible? It may also not be data of one dimension and is very messy. In addition, the data collection process is relatively slow. Inaccurate data monitoring will also affect the efficiency of the data, including inaccurate data. It is also possible that this situation will actually be adopted.

Solution

We have been discussing whether all the data can be displayed on one screen. It would be great if we had a very powerful system that could display our regular data using some fixed models, so that we don't have to run the distribution of this data every day. Our requirements for data are good scalability and defining user traffic data through certain standardized things. In the process of data extraction, you can open a hole and then use some statements to extract the data you want most. Of course, overall it is reliable, simple, easy to use and beautiful . This is our definition of a good data system, and we are also working in this direction.

Q&A

Could you explain application analysis in detail?

The application analysis based on the original stock data is mainly to make the data chain as long as possible. For example, only the thing he bought has a conversion rate, but it is not extended downward. This is also related to the data system. When the data system of the entire chain can be collected, correlation analysis is a natural thing. Naturally, you can look at the conversion of the purchased things from another perspective, but from the vertical perspective of the user, he bought the thing.

In fact, a major premise of data analysis is that we need a complete data collection process. Now many companies have a confusion, that is, we really need a lot of raw data, but in the process of collecting registration, we require users to leave very detailed information, such as age, income and hobbies, which are very difficult. The problem that troubles operators is how to derive operational strategies when it is difficult to collect raw data?

This may be the source of our circulation data, which mainly includes several aspects. Active research is one aspect, and the background reflection through the user itself is another aspect. Institutions can also provide some data and try to enrich this chain. The richer it is, the higher the price will be. Or we can create such a model through the intention expression designed in the product. Many times, if users think something is fake, they will not buy it, so try to make it valuable, such as letting them buy what they are interested in. During the purchase process, they will show some of their own behavioral characteristics. It is not necessary to send them a questionnaire asking what they like.

When we mentioned user portraits just now, there are many user attributes. How do you ensure their timeliness? For example, my income last year is different from this year. I may have earned 10,000 yuan last year and 20,000 yuan this year.

Because we have a set of real-time data that reflects the user's income level. We just match it with his current income level, rather than using his current data to predict his future. I can see his current income level, including his age structure and so on. This data is not predictive data and is not predicted based on historical data.