How to formulate a new user recommendation strategy? It's right to meet on a blind date

A new user, literally speaking, is a brand new user, which is like a blank sheet of paper for the product. Everything you do around new users is called cold start.

In actual work, I deeply feel that the cold start of new users is exactly the same as a blind date between men and women. Both are trying their best to leave a good impression on the other party (user). Today, I will have a blind date with our users, and the goal is to keep him (her) by our side.

Three steps of blind date:

1. Get to know the other party (user)—know yourself and know your enemy
2. How to meet the needs of the other party (user) - cater to their preferences
3. How to meet the needs of the other party (user) - use the right method

1. Get to know the other party (user) - know yourself and know your enemy

When we learn that we are going to have a blind date with someone, we need to try every possible means to find out information about the other person in the early stages. We may gain nothing, or we may gain something, such as appearance, education, work, family situation, etc.

The same is true for operating new users. In the early stages, you need to use all means to get to know the users. Practice has shown that it is not easy to get to know them.

1. Completely unknown, unable to obtain user information

Users have zero knowledge of the product, and the system recommends no information to refer to. At this time, you need to be fully prepared, test carefully, and avoid touching minefields. Or use a universal methodology to deliver content that everyone might like.

2. Coarse-grained understanding, obtaining partial user information

(1) Using technical means to obtain

1. Use GPS and IP to obtain user location information and know the user's geographic location;
2. Get the list of apps installed on the user's phone and guess their interests.

Advantages: high credibility and high availability.

Disadvantages: Manual sorting and integration are required, and the operating cost is high, such as the definition of population and interest mapping standards for the APP list.

(2) Use product cards to actively collect

1) Set up interest-based opening screen (age + gender + interest category) cards to collect information actively selected by users.

Advantages: Users actively choose, high availability

Disadvantages: Low proportion of users actively choosing, low user coverage; credibility of user selection needs to be evaluated;

(3) Obtaining information by guessing from channel sources

User source channel: Based on the coarse-grained portraits of users of different pre-installed models, artificial modeling of new users is carried out. For example, Xiaomi's users tend to be lower-class and middle-aged and elderly, OPPO's users tend to be young women, and Huawei's users tend to be mid- to high-end.

User growth: Based on different delivery materials, manually define fuzzy portraits to attract new users. For example, if you put out "square dance" materials, the new users you attract may be more likely to be middle-aged and elderly women; if you put out "crosstalk" materials, the new users you attract may be more likely to be middle-aged and elderly men.

Advantages: Clear understanding of the user group and strong practicality.

Disadvantage: Once user cognition is biased, irreversible consequences may occur.

3. The system obtains user behavior information

When users enter the product and receive corresponding behavioral feedback, the system machine can quickly capture and utilize this information, such as click behavior, subscription behavior, search behavior, comment behavior, forwarding behavior, collection behavior, etc.

Advantages: Real feedback on user behavior, directly reflects user psychology, and is highly usable.

Disadvantages: It is a double-edged sword and may backfire if used improperly.

2. How to meet the needs of the other party (user) - cater to their preferences

After we understand the other party (user), we should think about how to please her and what gifts can make her happy. So for new users, it is necessary to position the product based on its own tone and provide things that can make users feel happy. These things can be collectively referred to as content.

1. Different product forms will have different content supply and display forms

(1) Information flow information products

Single picture, three pictures and text + large picture, and single picture video (short video + small video) are the mainstream content forms and also the mainstream content consumed by users. Different users may have different preferences for the two content forms. For example, low-end users have a higher demand for short video consumption.

Representative products: Toutiao, Tencent News, Yidian Zixun, etc.

(2) Community products

Short content is the mainstream content form, which includes pictures, texts and videos, supplemented by comments to meet users' demand for content consumption.

Representative products: Zuiyou, Pipixia, Xiaohongshu, etc.

2. Integrate content in different product forms

(1) Integrate and package high-quality content using product forms such as themes and topics to maximize the value of a user’s single click.

Representative products: Zuiyou, Jike, etc.

(2) Use rankings of different dimensions to aggregate high-quality content , thereby increasing exposure of high-quality content and retaining new users.

Representative products: Bilibili, etc.

3. How to meet the needs of the other party (user) - use the right method

When we have collected gifts (content) that meet his (user's) interests, then how do we send them to the other party (user)? This is a question worth exploring.

In daily work practice, the "content pool" method is widely used. It is an invisible medium that links users and content. It is like a bridge that allows the right content to meet the right users. The actual strategy application of the content pool varies for new user cold start in different scenarios.

1. For completely unknown new users

Point to think about: If there is no information available on the system machine, then the focus should be on the content itself.

What kind of content do people like to watch?

In a broad sense, it is what is usually called top content, universal content, such as social hot spots, entertainment gossip, funny humor, leisure and entertainment, history and culture, positive emotions, etc.

From a fine-grained perspective, it may be more inclined towards popular and high-quality products, such as distinguishing based on dimensions such as title, cover image, content tone, richness, completeness, and readability. We should abandon clickbait, vulgar and low-quality content involving sex, and embrace high-quality content with good titles, rich content, and strong readability.

Thinking more deeply, content consumption is also a process of satisfying the basic needs of human nature. Only by seeing the essence through the phenomenon and grasping the basic needs of human nature can we truly meet the needs.

Let’s look at two examples:

From this, we can repeat the deduction and divide the basic needs of human nature into seven categories:

1. Profit-seeking mentality
2. Gender Psychology
3. Lazy
4. Vanity
5. The need for empathy
6. Social Currency
7. sense of security

We can explore high-quality content based on the underlying needs of human nature and combine it with the quality standards of the content itself to meet users' content consumption needs. (I won’t write about this in detail. I want to write about it separately later. Welcome everyone to discuss it together.)

Practical application case 1: Never-boring content experiment

We once tried to start a cold start project by simply focusing on general high-quality content to find content that users will never get tired of. Use content pools for aggregation and use algorithm machines to recommend content to users according to certain strategic rules.

We use a guaranteed pool approach, which takes effect for the first 10 swipes of new users. Each swipe will produce 2-3 pieces of content, with pictures and videos recommended at a ratio of 2:1 (because there is a fixed ratio of pictures and videos in the feed). From the data results, process indicators such as average clicks per person are good, and users will click on this part of the content. However, result indicators such as next-day retention are poor. There may be the following reasons:

1) Experimental strategy design: other interfering factors were not eliminated and the experimental test was conducted in a pure environment

There are about 10 recommended contents in the feed stream, but only 2-3 of them are covered by the experiment, and the remaining 6-7 are occupied by advertisements and recalled contents from other channels. The result is that users click on the recommended contents in the experimental content pool, but the recalled recommended contents from other channels are of low quality, which arouses user disgust and results in a poor overall user experience.

2) Content selection: Data and content quality are viewed in isolation, and there is insufficient research on the deep-seated needs of content consumption.

The overall idea of ​​content mining is exactly the same as above, which is to mine high-quality content from the perspective of human experience. From the practical process, it is easy for everyone to fall into a vicious circle of limited thinking when looking for content.

It’s right for everyone to look for content ideas, but in practice, they simply rely on editing experience, pay too much attention to data, pay too much attention to article quality, and view the two in isolation, without thinking more deeply about what content can better meet underlying human needs.

Content with good data such as click-through rate is a posteriori data distributed by machines, but it only captures some weaknesses of human nature, such as gender psychology (recommending vulgar and pornographic content). We need to break out of the limitations of machine recommendations and look at human needs more comprehensively. For example, the human need for empathy is to experience the feelings of others in life. There are many contents that meet such needs, such as character inspiration, counterattack content, positive energy content, etc. In the era of content explosion on the Internet, how to grasp human needs and provide them with high-quality content is the key to success.

Practical application case 2: Interest exploration

Interest probing, as the name suggests, is to tentatively recommend different categories of content to users. This is a relatively common cold-start content recommendation strategy.

For the interest testing experiment, we use the method of guaranteed pool + interest testing pool. The content pool can be divided into four states: birth, activation, normal, and death.

Birth: Initial state of the content pool

Activation: When different categories of content are recommended for different users, the content pool enters the activation state and continuously outputs content.

Normal: When users respond to the exploratory content with behavioral feedback (clicks, comments, shares, etc.), the content pool begins to enter a normal state and further expands the exploratory categories.

Death: As the trial deepens and the machine recommendation provides timely feedback, user feedback becomes more positive, entering a positive cycle. The content pool then begins to enter a state of death and gradually enters the machine personalized recommendation sequence.

There is one thing we need to note here. During the transition from activation to normal, we adopted an aggressive strategy. When users clicked on the trial category content, the algorithm machine promptly gave click feedback, and the weight of the feedback content was high, causing the user's interest to gradually become single, and no trial content of other categories could be swiped, which also caused the experimental results to fall short of expectations.

We studied Toutiao's probing mechanism at the time and summarized it as follows:

The heuristic mechanism can be broken down into:

There are currently exploratory categories such as A, B, C, D and other head categories, and 1, 2, 3, 4 and other vertical categories. The content coverage of top categories and vertical categories ranges from coarse to fine granularity to ensure the richness of the content.

User’s first browse: Top category content accounts for 50%, vertical category content accounts for 50% (the specific proportion can be flexibly adjusted according to the actual effect), and the top content of each category is displayed first.

As the refresh progresses, there will be two types of user behaviors:

1) If the user has behavioral feedback, click feedback should be given in a timely manner, and you should also continue to test content in categories that have not been recommended, as well as content in categories that have been recommended but not clicked by the user.

1. Test the content of categories that are not recommended, and ensure that each swipe accounts for about 30%-40% (the specific proportion can be flexibly adjusted according to the actual effect), and the granularity of the test recommended content can gradually change from coarse to fine, to test the user's interest limit.
2. When recommending content in categories that have not been clicked, if the user swipes 2-3 times in a row without any behavioral feedback, the recommendation weight should be lowered or even stopped.

2) If the user does not provide behavioral feedback, the user may leave after 2-3 visits (based on research on new users, users will only give 2-3 visits on average), so it is extremely important to test the category mix and select content in each category in the first 2-3 visits.

2. Have a coarse-grained understanding of users

Manually process the acquired user information and convert it into usable information. For example, convert the obtained APP list and user growth delivery materials into crowd information, and then study the crowd content preferences. This can be sorted and collected through market research reports and user portrait information already available on the platform, and then target content aggregation, strategic experiments, and effect regression are performed. The general experimental process is as follows:​

Practical application case: Using crosstalk material to attract new users, the general experimental process above can be broken down as follows:

Information processing and transformation: People who like to watch crosstalk and comedy are more likely to be middle-aged and elderly men

Research on people’s content preferences:

1) Research method - use the existing portraits of the platform + the population research reports in the market to sort out and summarize.

1. The portrait side can be cut into from dimensions such as size category, keyword, content source, entity, etc.
2. The research report can be summarized from specific information points

2) Research results: The content preferences of middle-aged and elderly men are more inclined towards military history, health and wellness, gardening, rural life, social positive energy, community hot spots, love and marriage, etc.

Target content aggregation:

1) The content pool recalls content through high-quality content sources and channels in various categories

2) The content pool is divided into crowd content pool A, which aggregates content that people who like crosstalk may like to watch; and crosstalk-related content pool B, which aggregates high-quality crosstalk video content.

The demands for building two content pools:

1) The crosstalk-related content pool is used to meet users' content consumption needs for crosstalk videos. By releasing crosstalk materials, new users are attracted, which shows that there is a strong demand for crosstalk.

2) After satisfying users’ strong demand for crosstalk, use the crowd content pool to further meet their demand for other content.

3) In the actual strategy, the two content pools recommend A+B combination to collaboratively meet users' content consumption needs.

Online strategy experiment: Set recommended positions for the two content pools A+B, and increase the number of effective positions. For example, the recommended positions for the crosstalk-related content pool B are 2, 4, 6, and 8, and the recommended positions for the crowd content pool A are 5, 7, 9, and 10. The first 10 swipes to attract new users for crosstalk materials will take effect (the specific number and position of swipes can be flexibly adjusted according to the actual effect)

Data effect regression:

1) Overall data performance: focus on the actual number of users covered by the strategy, average clicks per person, average stay per person, penetration rate, and retention on the next day, 2 days, 3 days, etc.

2) Content pool performance data: focus on the actual exposure, clicks, and dwell time of the content in the content pool.

Go through the entire process, flexibly adjust the content pool content and optimize the online recommendation strategy based on the data results to improve new user retention.

Finally, I would like to end this article with a quote from our leader: new users are a mystery, and it is too difficult to understand them. The only way is to keep exploring and summarizing, and seek methodological truth in practice.

The same is true when we go on blind dates with people. It is very difficult to truly understand a person. It takes a long time of getting along with each other, and constant exploration and summary, before we can gain their true love.

May everyone find true love! I hope that all new users of our products can be retained firmly!

Author: Ke Ran

Source: WeChat public account "Ke Xiaowang"