A must-read for 20 million self-media professionals: The new logic of content distribution is revealed

This is an article written for the 20 million self-media people. We also hope that it will be helpful to platform operators and traditional media people who are undergoing transformation. The full text is over 14,000 words, and it takes at least 30 minutes to read it. It is recommended that you read in a quiet environment - of course, if you have a quiet mind, the environment is not a problem.

If you don't want to go to great lengths, you can also read the following article excerpts:

1. Today, the power of content production has shifted from professionals to self-media people. This is just like the early days of reform and opening up, when a number of individual restaurants emerged alongside the "state-owned restaurants".

2. The power of content distribution has also shifted - from professionals to social media to algorithms ( artificial intelligence ). First, we have the user portrait , and then we match the content. This is the basis of the algorithm. Compared with traditional media and portals, it is like the transition from a planned economy to a market economy.

3. However, there are problems with relying solely on any of the three distribution models: editorial, algorithmic, and social . Therefore, large platforms are moving towards a “three-in-one” distribution model. WeChat and Yidian Zixun are typical representatives of exploration.

4. It is precisely because of the distribution logic of "one thousand faces for one thousand people" that a large number of self-media content in the long-tail field was born. It is precisely because there are more than 20 million self-media accounts that the platform is prompted to conduct more in-depth research on "one thousand faces for one thousand people". The transformation of content production power and the transformation of content distribution power promote and depend on each other.

5. When self-media people post on a platform, they just want to gain two things - profit and fame. Profit refers to income represented by money; fame refers to the comprehensive data of indicators such as content display, clicks, sharing, collection, and user stay time. What the platform needs to do is to first define high-quality content and then allocate profits and fame to it.

6. The core of algorithm distribution is threefold: user portrait, article portrait, and algorithm model. User portrait: Who are you and what do you like? Article portrait: What kind of article is this? What subject matter? What field? Is it good or bad? Algorithm model: How to achieve the best match between user portrait and article portrait. The core difficulty of the algorithm is to understand human nature.

7. If you don’t study “ cold start ” persistently, your users will soon leave. Cold start is the content you see for the first time when you download an APP, just like the first time you fall in love.

8. Editing is guided by values, while algorithms are guided by data. Values ​​ensure fairness and data ensure efficiency. What we need to do is to find a balance between fairness and efficiency. Yidian Zixun follows the "socialist path with Chinese characteristics".

9. The editors of the future will be compound editors who need to understand algorithms, self-media, and short videos - because these are standard features of large platforms. Editors who only want to be blue-collar workers will eventually be eliminated.

10. "Studying things and exploring their principles" is a very painful process. But once you figure it out, you can grasp the laws that govern how things work. If you can achieve "unity of knowledge and action", you will be invincible.

It should be noted that this article is selected from an internal sharing of mine. Because some of the content involves important data, the relevant details have been deleted. But behind every point of this article with a well-organized framework and structure, there are countless explorations, setbacks and failures. This is where the fun of "unity of knowledge and action" lies.

1. Today, the power of content production has been distributed to everyone

As shown in the figure above, since the birth of mankind a million years ago, communication has consisted of the above four links: content acquisition (production), content review, content distribution, and user interaction. To this day, these four links have not changed at all, but in the past 2-3 years, the connotation of each link has undergone fundamental changes, especially in content production and content distribution. For example, in the portal era, we emphasized massiveness and speed; today, we emphasize richness and precision.

First, let’s look at the changes at the content acquisition level (or content production level).

In 2016, I proposed a concept - the transformation of content production power. In the past, who had the power to produce content? It is those of us here, and our partners - including Xinhua News Agency, CCTV, China Newsweek and many others. In other words, media professionals have the power to produce content. These professionals are performing their duties and actually represent the media organizations behind them. But now, every natural person has the right to produce content, and this is self-media.

According to incomplete statistics, by mid-2017, there were more than 20 million registered WeChat public accounts, of which more than 2 million were active accounts. On the Yidian Zixun platform, when I took over in August 2015, there were less than 5,000 self-media outlets. Now, there are nearly 500,000 Yidian accounts (including our friendly Phoenix account), with nearly 200,000 articles published every day.

Let's look at the following data: In the past week, on the Yidian Zixun App platform, the display volume, clicks , shares, comments, collections, and reading time of self-media have exceeded 90% (the rest is the content accessed by RSS crawling and other methods); two years ago, these data were less than 30%.

Therefore, in such an environment where "everyone is a content producer", we must face today's situation seriously, coordinate operations and refine operations.

2. How to operate self-media: redefine high-quality content

How to operate self-media? First of all, we must have such a mind - all high-quality content in the world should be used by me, and all outstanding self-media people in the world should be our authors. Break the mindset that everything must be original, because readers (users) don’t care whether your content is original or not. They only care whether the content is good-looking, important, and of interest to them.

The core of self-media operation lies in whether it can provide "profit" and "fame" for high-quality content. To this end, we need to think about the following:

1. What is premium content?
2. What kind of benefits can be provided to high-quality content?
3. What kind of name should be given to high-quality content?

I have been in the media industry for 20 years, and the definition of high-quality content is something I think about every day. If we look at it from the perspective of articles (or videos, pictures, live broadcasts , etc.), it’s nothing more than eight words: topic selection, interviews, writing, and packaging (mainly including titles and pictures). Regarding good topics, good interviews, good writing, and good packaging, these have been explained many times. If you have any questions, you can read the two books "Beyond the Portal" and "The Tao of Self-Media", and I will not repeat them here.

But let me insert an example here, which is the definition of high-quality short videos by Tencent colleagues.

It's really a classic - the fact that they can come up with such a summary proves that they are thinking seriously. Short videos are very popular right now, and I hope everyone can give more ideas in this regard.

The second dimension for defining high-quality content is the self-media account. It is equivalent to the "column" in the print media era and is a collection of articles. Whether a public account is of high quality can be defined from the following six dimensions.

The first is originality

"All articles in the world are plagiarized, and the cost of plagiarism is reduced by a pair of scissors and a knife." With computers and the Internet, the cost of plagiarism has become much lower. Because plagiarism among self-media accounts is very serious, high-quality accounts must be original.

But we must be clear about how to define and identify whether a public account is original. Searching for information online to supplement the background of the article, or analyzing known content and extracting your own opinions - this is not considered plagiarism. Therefore, it is necessary to be able to define the overlap ratio between titles, pictures, and articles in particular. Precisely because the volume of self-media is so large, this is not something that can be done by humans. Instead, we need to sort out and compare them through a copyright system. The technology should be able to identify who is the first to release a product on the entire network, how many duplicates there are, and clean up the "number-making party". Now, the "number-making party" has formed an industry - they are very good at capturing hot spots, and quickly searching for popular articles, piecing them together, and then making a party title to cheat clicks. Once the content produced by the "number-making party" enters our platform, it will divide up a large amount of traffic and subsidies, resulting in the prevention of truly original and high-quality content from entering, and will also trigger a large number of copyright lawsuits and complaints.

The second is verticality

To date, a lot of content on Yidian Zixun is accessed through RSS crawling. Especially some large content sources, such as Xinhua News Agency. We will also rate them according to the quality of the manuscript sources, from level one to level six, from bad to good. So the question is, Xinhua News Agency’s content includes both classic and general ones, how do you define its level? But this definition is important because it is directly related to the weight of content distribution.

Therefore, we need to break down the content of Xinhua News Agency, for example, into Xinhua current affairs, Xinhua finance, Xinhua military, Xinhua entertainment, Xinhua culture, etc. If Xinhua Current Affairs is the best, we would give it a rating of six; if Xinhua XX is the worst, we would only give it a rating of two (I am just making an analogy, Xinhua Entertainment is still very good content). Xinhua's current affairs can be further broken down, such as personnel changes, officials' fall from power, major meetings, policy interpretations, etc. For each area, we can define its level.

When you break it down to the last detail, you'll find all sorts of "numbers". The finer the split, the easier it is to distribute - in addition to clearly distinguishing between good and bad, there is another very important point: it can be distributed more accurately. We know that the distribution logic of Yidian Zixun is different for different people. If our account is defined very accurately, the recommended articles will also be very accurate. We can send relevant content to people who are interested in it and increase the click-through rate of the article - this is an efficiency issue, which I will explain in detail later.

The third is activity

The number and proportion of posts published by an account per week and month, especially high-level accounts, is its activity level. But for more refined operations, we need to compare the number of posts, the quality of posts, and the time of posting with those on the platforms of our "competitors". We must be clear: if a current affairs article is published on the opponent's platform first, we can tolerate a certain delay; if a financial article is published, we can tolerate a certain delay... For some content, if it is published one or two hours late, it will not be of much significance even if it is published. This idea and method of refined operation must be implemented throughout every aspect of the work.

The fourth is attention

This is the first time I formally proposed this concept. The so-called attention level refers to the number of user subscriptions to a self-media account, as well as how close the relationship between the account and the user is after the user subscribes.

Why are people willing to open their own media accounts on WeChat? In addition to the large number of people using WeChat, the subscription relationship of WeChat public accounts is very clear - after fans subscribe, whenever a post is made, the user will be given a very clear prompt. On other platforms, this feature needs to be enhanced.

We first need to create the subscription functionality. This is relatively simple. Just add a button on the product article page. You can also directly recommend the official account on the homepage. There are many methods, but the relationship after subscription is very complicated: for example, do you want to PUSH? NetEase has this function - when you subscribe to an account, you will be pushed the content sent by this account. But if you subscribe to 100 accounts, which one should you push? While pushing, do you want me to recommend it to you on the homepage? Recommend to the first screen or the second screen? How often is it recommended? What is the relationship between this recommendation and your subsequent clicks, shares, and other behaviors? In short, this involves algorithms, is very complicated, and cannot be accomplished overnight. But if the subscription function is not established, self-media people will never find the relationship between themselves and users (or readers), and they will be unwilling to post content here.

The fifth is popularity

This is easy to understand, it is the identity of the self-media author, or the status in the industry. We divide self-media people into three categories: celebrities, experts, and ordinary people. A celebrity refers to someone who is well-known in society, such as Yao Chen; an expert refers to someone who is well-known in a certain field, such as Wu Chenguang , and his influence is limited to the media circle. For these two types of people who come to Yidian Zixun to start their own media, we weight the distribution of content. That is to say, for articles of the same quality, we give priority to those written by celebrities. We are also working hard to develop some ordinary accounts on the platform into expert or celebrity accounts. This is the responsibility of the platform.

Finally, the quality of the content

An account is made up of articles and other content, and the quality of the content determines the quality of the account. The definition of content quality has been mentioned before, so I will not repeat it here. But it should be emphasized that behind each dimension is a system - including products, algorithms, and insights into data details. For editors, these are all new things, and the road to transformation also begins from now on.

3. How to operate self-media: What is “profit”?

Let’s take a look at the following table:

Please take a look at the table compiled by a self-media person. Including Yidian News’ Yidian account, Toutiao ’s Toutiao account , NetEase’s NetEase account, UC’s Dayu account , Penguin’s self-media platform , and WeChat account. Through this table, you can see what benefits the platform brings to self-media people.

When I think of “profit”, my first understanding of it is money. Self-media people need money. Because everyone here has a salary. For example, for me, the self-media account "A Little Morning Light" is just a hobby for me. However, many self-media people do not have salaries, and they can rely on platform income to provide basic living security. Therefore, how much money you can bring to them is the basic point for whether they choose to publish articles on your platform.

Various platforms have started competing in this regard. You give me 10 million, I give you 100 million, the next one claims to have 1 billion or 2 billion, and some even put up a sign of 10 billion. But is more money really better? I don't think so. If all the money is distributed to the "number-making party", it will be counterproductive. Therefore, we must first define high-quality content, original content, and exclusive content. Secondly, we must have a very scientific and reasonable distribution mechanism.

How should the money be allocated? There are many channels and fields on this platform, and their contribution to the platform’s traffic should be a very important factor in the distribution of money. Various data can be reflected. Currently, entertainment is the largest contributor to traffic on a certain platform, so we allocate the largest proportion of money to entertainment. We will continue to break down entertainment into celebrities, variety shows, movies , and TV, and we will break it down based on traffic. This ratio is not fixed. It should be analyzed according to the actual situation on a daily, weekly and monthly basis, and then the allocation model should be adjusted. We can also artificially increase the investment in a certain field - if we want to expand and strengthen a certain field, or we have a judgment that there is a serious shortage of high-quality content in this field and users have huge potential needs.

The indicators such as originality and verticality of the self-media mentioned above are also important factors in measuring the allocation of money. For example, a certain account has very good article data, but all of them are reprinted or even plagiarized. We can't give money to such an account. Because it breaks the rules of the platform, its traffic brings "negative energy". Tolerance will lead to the departure of original and high-quality accounts. This is the same as managing a business. Bad money cannot be allowed to drive out good money.

"Benefit" includes resources in addition to money. When the money given to self-media people is insufficient, other resources can be provided as compensation. For example, we can insert a link in the article and click on this article to go to it? Use this link to enter the e-commerce platform to place an order, and then let the user and the platform share the profits. We can also provide advertising space and let self-media people operate and profit from it themselves - of course, the prerequisite is that we must have a supporting advertising review team. We cannot bring in false advertising and turn users into Wei Zexi.

At present, Yidian is seeking resource docking and cooperation on a large scale, which is consistent with the overall strategy of our company. In March this year, we connected Visual China's image resources, allowing self-media people to choose their favorite pictures from 20 million pictures for illustration. They are high-definition, free and without having to worry about copyright. In July this year, we connected GIF images, allowing self-media people to choose from 1 million animated pictures. One point is that, apart from WeChat, it is the only platform that provides GIF image services for self-media.

Of course, the various resources we open up are proportional to the contribution of self-media people on this platform. For example, who is eligible to get the money, who is eligible to use the pictures for free, etc. Also because of the temptation of profit, self-media talents will continue to write and write good articles. To put it in one sentence, we need self-media to have more "assets" on this platform so that they can generate stickiness. You can think about WeChat, what can’t you live without? That’s because you have a lot of assets in WeChat - including friends, wallets, information flows , etc.

4. How algorithms control “name”

In addition to profits, the platform also needs to provide "fame" for self-media people.

Fame is influence. If we talk about it from the perspective of a single article, it includes: number of clicks, number of shares, number of comments, number of favorites, length of user stay, etc.; for a public account, it is the number of subscriptions and the stickiness between the public account and readers after subscription.

So, who has the power to distribute content? In the era of traditional media, only editors had this power. For example, the number of clicks depends on the location of the article, the exposure duration, as well as the title and pictures. This is the power of the editor. The number of shares depends on the quality of the article, the number of comments depends on whether it can stir up social emotions, the number of collections depends on whether the article is valuable, and the length of time users stay depends on the quality of the article. All of these are operated by people.

But today, with the rise of social media, everyone in social media, except professionals, has the power to speak and forward; the application of artificial intelligence gives algorithms the power to distribute content as well.

Since Yidian Zixun is an app that relies on artificial intelligence for distribution, we will skip the social aspect and move on to how algorithms distribute content.

Here, I would like to tell a little story to help everyone understand a few concepts: artificial intelligence, machine learning, and algorithm-driven content distribution.

This story tells how I watched Japanese AV action movies under the nose of my boss.

My workstation is near my boss's office and he has to pass by it every time he goes out. But I really want to watch a Japanese action movie during work hours. So, the first thing I had to do was install a camera facing the door of his office. As soon as he leaves the office, the camera will take a picture of him and transfer it to my computer. I put a screen saver with the word "Yi Dian Zixun Algorithm Distribution Logic" on my computer. Whenever the camera focuses on him, the screen will automatically jump from the action movie to the page of "Research Algorithm Distribution". My boss will think that I am working hard.

Here comes the core question - why does my screen change as long as the camera captures him? Because I developed a facial recognition system. I used every possible means (including searching on Baidu , bribing his secretary, etc.) to collect 100,000 photos of my boss, and then input them into my computer. The computer learned from these photos (learning is divided into supervised learning, unsupervised learning, semi-supervised learning, deep learning, etc., in short, it is learning desperately), so it can judge the similarity between the photos taken by the camera and the photos I input. Once the similarity reaches a certain value, it is determined that the boss is coming. So I gave instructions to the screen and quickly switched from AV to the page for researching algorithm distribution.

This is called artificial intelligence ( AI ). Artificial intelligence is the biggest concept. The process of inputting photos into the computer and learning to identify whether this person is the boss is machine learning. Artificial intelligence has many applications, such as "face recognition" mentioned in this short story, as well as driverless cars and health diagnosis. The famous AlphaGo is the application of artificial intelligence in the field of Go. The application of Yidian Zixun is to distribute information through algorithms to achieve personalized and personalized services for each individual.

The greatest artificial intelligence company in the world is Google . They have an organization called "Google Brain", which mainly studies the application of artificial intelligence; China's Baidu is also moving towards becoming an artificial intelligence company. They have accumulated a large amount of data through searching, which is just like the process of collecting photos mentioned in the previous story - without big data , artificial intelligence cannot be realized.

Now let’s get back to the application of artificial intelligence in information dissemination. There are three elements to achieve precise distribution: one is user portrait, two is article portrait, and three is algorithm model.

• User portrait: Who is he? What do you like?
• Article portrait: What is this about? Is it pictures, text, video or photo album? Is it current affairs, society or finance? Is this content good or bad?
• Algorithm model: How are user portraits and article portraits matched?

First look at the user portrait

As shown in the figure above, living environment is the foundation in user portraits. The living environment includes two levels, the first is the geographical location. Yidian Zixun can already locate business districts and neighborhoods, while Baidu Maps and Didi Taxi can already locate where you are standing in real time. Food, clothing, housing, transportation, water, electricity, gas, and even products and services provided by the government are all related to your living environment and geographical location. Therefore, the more accurate the geographical location, the better.

In terms of time scenarios, the things pushed in the morning should be hard and information-oriented content, because after a night, everyone needs to see what happened last night. Moreover, office workers are busy in the morning and do not have time to read long articles. At night, it is the opposite. Users have time and want to relax, so it is necessary to push softer and more entertaining content, including some beautiful articles - only in the quiet night can they appreciate the flavor of them.

The second level is the mobile phone environment

It also contains two meanings. One is the mobile phone model. The reading habits of users of different mobile phone models are different, such as Redmi users and Apple users. Even for phones of the same brand - OPPO A57 and OPPO R11, users have different reading habits, which requires us to gain insight through data. In addition, we need to have a deep understanding of which software is activated in our mobile phones and how often it is used without affecting user privacy.

The third level is user information

If you are a registered user, your gender, age, identity, education, etc. can all serve as the basis for us to judge your interests. In addition, in order to describe a user portrait more accurately, Yidian Zixun designed an opening screen questionnaire asking users what they like. More than 50% of users made a choice, so we were able to understand their basic interests. Of course, there are also people who are unwilling to fill out or even disclose their geographic location. Therefore, this poses a difficult problem for us - what information should be provided to users during cold start?

Cold start refers to the process when a user opens an App for the first time after downloading it. Just like the first meeting in love, its importance is beyond doubt. In a cold start, the most important thing is the first screen displayed after the user opens the app, which may have 5 pieces of information. Tests show that if a user clicks on one of these five ads, his or her next-day retention rate will be twice as high as that of those who do not click. Therefore, it is not excessive to spend a lot of effort on creating these five contents. To this end, we conducted a large number of comparative tests - including trying to push information to related mobile phone models, including pushing information to related geographic locations, with some successes and some failures. In the cold start, because the user portrait is not clear enough, the basic principle should be to push important events and high-probability areas (such as football, not golf), rather than pushing niche content to gamble, otherwise you will easily fail.

After the user finishes reading the cold start first screen content, five situations may occur:

1. There is a click drop-down;
2. No click drop-down;
3. No click to slide up;
4. There is a click to slide up;
5. Just leave.

We should launch different strategies for these different behaviors. This requires very detailed data observation to summarize the patterns. Since then, user behavior has become diversified - such as clicks, sharing, collections, comments, user stay time, etc., which are the basis for the algorithm to determine which content to push.

This is a very complex calculation process. For example, a user clicked on "Wu Chenguang" five times and searched for "Zou Ming" (Editor-in-Chief of Phoenix.com) once. If only one article is recommended to him, should "Wu Chenguang" or "Zou Ming" be recommended first? In terms of quantity, there must be more "Wu Chenguang", but searching is a deep behavior. He took the initiative to search, which shows that he is very interested in "Zou Ming". In this case, who should be recommended?

For example, he clicked "Wu Chenguang" five times and "Zou Ming" twice. But there is no more high-quality content about Wu Chenguang, and there is only a good article about "Zou Ming". So should we promote a very watery article about "Wu Chenguang" or a good article about "Zou Ming"?

Also, he clicked on the picture and text of "Wu Chenguang" five times. Should he promote a video, audio or Q&A? Or if a user likes "Wu Chenguang" very much, and today a scandal about "Zou Ming" suddenly breaks out, should we promote the hot topic of "Zou Ming" or continue to promote "Wu Chenguang" based on the user's interests?

Therefore, the complexity of user portraits is not a simple matter of group division. It is a profound insight into human nature. The meaning behind an action, such as clicking, is completely different. Some people click because of real interest, while others click because of clickbait.

Look at the article portrait again. For editors, article portraits are relatively simple because this is our specialty. In the era of traditional media, we profile articles - such as news, correspondence, features, and also the investigative reporting and explanatory reporting that I have been trying.

Of course, the complexity of article portraits today is far greater than in the era of traditional media. Because the article here refers to the content in general, rather than a simple textual concept. It includes:

• Genre - is it text, pictures, video, audio, live broadcast, Q&A, or short content like Weibo?
• Author : The author actually represents the self-media account behind it, including its level, popularity, number of subscriptions, etc.
• Tags ——This is the understanding of the field described in the article, such as sports—football— Chinese Super League —Guoan. We give the power of labeling to self-media authors, who can label the content they produce. At the same time, our review editors and channel editors will correct the labels; the algorithm will also make further adjustments to the labels. Finally, make a comprehensive evaluation and give the article the most accurate positioning. If the algorithm is very smart and has done deep learning, it can also analyze more about the characteristics of the article, that is, understand the central idea and main points of the paragraphs like a human. Of course, this is difficult. In particular, it is more difficult to recognize videos and pictures.
• Content quality - it has two dimensions, the first is editorial judgment, and the second is algorithmic judgment. As mentioned before, the editor's judgment is nothing more than four dimensions: topic selection, interviewing, writing and packaging; while the algorithm's judgment is nothing more than based on data, clicks, sharing, collections, user stay time and so on.

We have user portraits and article portraits. We will now start matching. The basis for matching is the algorithm model. We can break down the model in detail and divide it into many modules - such as spam filtering module, hotspot module, local module, etc. The algorithm selects and sorts articles in each module; articles in different modules also compete and are sorted. Finally, the best ones are selected and distributed to users.

We take the hotspot module as an example and make a specific analysis.

Hot spot is a word that is easy to understand. It refers to news that has just happened and has attracted a lot of attention. It is different from other non-timely information. First, we need to tell the algorithm what the hot spots are. There are many ways here - for example, monitor Baidu. Baidu hot search terms may be hot topics. In Yidian Zixun's news channel, the editors also pin hot topics to the top. We can also manually define some articles as hot topics: for example, those with the prefix "Xinhua News Agency Express" are considered hot topics. This is the first step. Let the algorithm understand the hot spots, this is machine learning.

The second step is to “recall” the hot content. The recall here is different from the recall of problem cars . It means gathering related hot articles in a certain pool. We call this the "hotspot pool". We will divide the hot spot pool into several levels. Three years ago, I described the rules for major emergencies on Sohu.com in "Beyond the Portal", from level three to level three. After Lai Yidian Zixun became the editor-in-chief, further revisions were made. Now, we input the standard case into the computer and let the algorithm judge the size of the hotspot to decide to what group to push it to.

Caption: In the book Beyond the Portal, the editors have detailed grading standards for handling major emergencies. Now we want to let machines learn to make more accurate distribution.

Next, let’s take a look at the entire process of article distribution. Note this diagram:

Articles are divided into two types - RSS source articles and self-media articles. But no matter what kind of article it is, it needs to be filtered by the anti-spam module to remove harmful political information, vulgar content, false advertising, "clickbait" articles, etc. Therefore, the anti-spam module should have these functions. If you ask how these functions are implemented, the answer is still the same as mentioned above - input cases or keywords , let the machine learn, find features and draw inferences.

After spam filtering, algorithms or editors label and categorize the articles, and the articles are gathered in various pools. As I said just now, this process is called recall. According to the article's label classification and different attributes, they are assigned to various pools - including the hot module and local module we just mentioned, as well as good articles selected by editors and entered into the boutique pool module. If some articles have local attributes and are high-quality, they will be placed in both pools at the same time. There are also some articles that are tagged with sports or finance and are placed in the sports or finance pool.

Finally, when it comes to the distribution process, it is mainly targeted at a certain user, and among so much content, some content will be released first and some later. That's sorting. The content that users see when they scroll down the home page is sorted. The same is true for other channel streams. The specific sorting rules depend on the algorithm, that is, the relationship between user portrait, article portrait and algorithm model. We have discussed this just now.

There are probably several points on how to measure the effectiveness of distribution:

• The first is accuracy - the things recommended are what users like;
• The second is speed - good content should be pushed as soon as possible;
• The third is quality – for articles in the same field, the ones with high quality must be promoted first. Otherwise, our high-quality content will have no traffic, and will not gain profits and influence. As a result, excellent self-media people will stay away from this platform, thus forming a vicious circle.
• The last is breadth – preventing users from being trapped in information silos. You need to have accurate interest detection to predict what you like and what you don’t like. This is what we are talking about, ensuring users have more “right to know”.

5. Trend: editorial, algorithmic, and social distribution in one

After finishing the micro operation, let’s go back and look at the macro. We just said that there are already three people who have the power to distribute content, and the latest trend is that the three distribution modes of editing, algorithm, and social are converging.

Please look at the following three examples:

A. WeChat’s “Take a Look” feature

We know that WeChat is the most typical social tool , and its " Friends Circle " function is actually an information flow with strong social attributes. You can see at a glance what your friends are doing, watching, and paying attention to through Moments. On June 15, 2016, a small feature of WeChat - "Hot Articles in Moments" was quietly launched. According to a WeChat insider, in the initial period, only about 10% of users activated this feature; even today, nine months later, the feature is not open to all users. Judging from the location of this product, “How deep is the courtyard?” is it hidden behind the third-level page? .

But the developers gave this product an atmospheric slogan: “Discover a wider world”. Through this product, they try to present the best content from more than 24 million public accounts on WeChat to 650 million users in the form of "WeChat Headlines".

The source of “hot articles in Moments” is the hyperlink content (mainly in text and image format) forwarded to Moments by all WeChat users. The WeChat team explained: “This feature analyzes the behavior habits of WeChat users in depth and, based on the user’s choice, will prioritize (a) articles from public accounts that the user follows and (b) original articles that their friends have read. It will also (c) select high-quality articles from the entire WeChat platform.”

The weight of the public account in the hot article is high quality, which is specifically reflected as follows:

(1) Basic weight index

Certified accounts (such as the official WeChat account of the Beijing News) have a higher weight than personal accounts, original accounts have a higher weight than non-original accounts, and accounts followed by users and friends of users have higher weights, but they have no direct relationship with the number of fans of the public account.

From this logic, it can be seen that the WeChat team attaches the utmost importance to original content. Among its 24 million public accounts, about 1/10 are active accounts, and the number of certified original accounts has exceeded 500,000. WeChat uses various methods to support original creations - for example, the reward function is only open to original authors, and this time it is the priority recommendation right in the information flow. The so-called "certification number" is also one of the original logos.

(2) Fan activity

It includes the click rate of the article (open rate), fan messages and message interaction frequency, like rate of the message content, open rate of the original text, etc. (note that it is all "rate", not "quantity").

(3) Public account search frequency

If the official account often appears in the user's search results and is clicked and followed, the recommendation weight of the official account can also be increased.

(4) Keyword matching

The matching degree of the public account name and the introduction keywords; the matching degree of the public account name and the description keywords and the content keywords, the higher the matching degree, the greater the weight.

And in the evolution of this product, we also found:

1) Many articles have labels in the lower left corner. Mainly include:

• Already followed: It means that you have subscribed to the official account that generated this article.
• Friends are reading: It means that your friends are reading this article (but what proportion of click-through rate will you get this recommendation, and you have not obtained specific data yet).
• XX people are reading: For example, people from Beijing or Hebei are reading this article.
• XX elites and XX enthusiasts are reading: It means that a certain circle (such as IT people) are reading this article. XX people should come from the user's registration information.

2) In terms of the arrangement logic of the article, you can refer to the following figure:

3) The algorithm logic of personalized recommendations is increasingly integrated into recommendations. For example, if you click on articles about "Zodiac" twice in a row, "Hot Articles on Moments" will continuously recommend articles about zodiac on the first screen that is third and fourth screen that is fourth.

For WeChat, even without social logic, it is unmatched in just obtaining information from a certain user - when you register, it obtains all personal basic information; when you join the group, it verifies your identity information; when you swipe the content in your circle of friends, it gains your interest in reading; when you consume, it obtains your financial information, which is the most core information of a user, and your expenses represent a hobby to a large extent.

This may be the "most dangerous" part of the product. It marks a new recommendation logic. The distribution of content with social and circle attributes is gradually being combined with the distribution with algorithms and personalized recommendations attributes.

Two months ago, the "hot articles in the circle of friends" had been replaced by "Look at it". The entrance has become shallower and covers all users, and the function of "WeChat Headlines" has begun to take shape.

B. The distribution concept of "edit + algorithm" of Yidian News

The distribution concept of Yidian News is similar to that of Toutiao, but it does not mean Toutiao. The person in charge of Toutiao once said that Toutiao has no editor-in-chief, no editor, and is not responsible for providing values ​​(in fact, Toutiao also has editor-in-chief, and there are many editors). But a little information cannot be like this - we emphasize the integration of editing and algorithms, and we must solve problems together - especially the issues of fairness and efficiency.

Judging from all data, the efficiency of algorithm distribution must be higher than that of manual labor, especially click-through rate. Because the algorithm can be distributed according to your interests and gives you what you want to see. But it is also because of this that it will lead to the flood of content of the "Title Party" and the flood of saliva articles, and will also lead to users falling into an island of interest. This damages "fairness". The fairness mentioned here refers to important content that has certain values ​​but users may not pay so much attention to. For example, Chinese scientist Tu Youyou won the Nobel Prize. Compared with those celebrity scandals, if you rely solely on algorithms, few people will definitely read this article. But can we only rely on algorithms to promote such an important content? For example, if Jia Yueting was appointed as the chairman of LeTV , if the algorithm makes a judgment, it would be a financial news or technological news, and it would not be pushed to the public. But we can't do this because we are the media. We have a responsibility to promote the public's right to know and social progress.

Especially in hot events, algorithms still have natural weaknesses. Because algorithms determine hot spots, it is necessary to increase user click-through rates and improve hot words on platforms such as Baidu. There is a process, maybe 15 minutes or more. But the editor can judge the hot spots within 5 seconds. Therefore, in major emergencies, Yidian News already has a very big advantage. Our PUSH on major events is the first in the commercial network media almost every time. Behind this is manual push. But as mentioned above, the establishment of a hot spot pool is also to allow the algorithm to learn the editor's judgment as soon as possible and surpass the manual as soon as possible.

6. About Products and Transformation

Finally, let’s talk about the product.

Water is equivalent to content. We need to put it in a cup, and the cup is the product. The cup is well designed so that we can feel comfortable when drinking water. At least not leaking.

In my understanding, there should be three systems for the product:

• The first is 2C - for users;
• The second is 2B - for self-media authors;
• The third is 2E's - the operational tools provided to editors.

Taking self-media products as an example, 2C is mainly the basic experience of users, and this also involves the core functions of some products - such as our search subscription function. If a self-media account and a channel are duplicated, the sorting of search results (SUG) is problematic.

On the 2B side, when I first started working as a self-media, I proposed 12 words, "Entertainment is easy, publishing articles quickly, having a great impact, and making more money." At present, the core content of the B-side is the issue of the points system - it can more conveniently provide fame and profit to high-quality self-media people, so that more assets can be left on our self-media platform, such as free pictures, data statistics, one-click import, video functions, etc., all have a relationship with the points system.

The core of the 2C side is that it can monitor and query more quickly.

A good product is not about repeatedly preaching to users, but it must be that the user is willing to use it. All users are lazy, stupid and white (not that everyone is bad, but I am the same). The more convenient and practical the product functions are, the more they like it. Otherwise, point-and-shoot cameras would not have been invented.

Regarding algorithm problems, self-media problems, and product problems, they are all things that editors in this era need to deeply understand. Some editors complained to me that the channel is "fewer people and more lives" and that it is very complicated to work as focused as when working in print media. I said this is a good thing. From a micro perspective, it is an improvement to your personal level - a soldier who does not want to be a general is not a good soldier, and an editor who does not want to be an editor in chief is not a good editor; from a macro perspective, this is a requirement of the new media era - we are not brick movers, we are architectural designers.

Some classmates said that I had never done it before, or that when the company recruited me, I was asked to do another position, but now I can't adapt. I said I had never done the self-media before, but I must do it now because this is the requirement of the times.

If you really can't do it, there is only one way: leave. The same is true for me, the company will not set up positions based on its person. Especially for a startup company like this.

Some people also say that I don’t understand algorithms and are not interested in this. If you don't understand the algorithm, you will be eliminated because it represents advanced productivity. We cannot return to the old mode of manual editing just because there is a problem with the algorithm. It's like a carriage does not run faster than a car, but a car will definitely bring more traffic accidents. We cannot say that we will not build cars today because of a car, or like Cixi, let a few horses pull the locomotive and run forward. This is a regression of the times.

I would like to emphasize again that the advanced nature of the algorithm lies in improving the efficiency of distribution. Because it is based on data. Data is a reflection of user behavioral habits. If we are not sensitive to data, it means we do not understand the user. How can users improve distribution efficiency? For real transformation, on the one hand, we must adhere to the editor's philosophy, values ​​and sentiments, and on the other hand, data is very important for the results of operations. Never ignore every percentage point of change.

Everyone should also fully understand our position in the company. In the past, we emphasized that "content is king", but this is a one-sided concept. As shown in the picture above, the content is just one link. The content is placed on top of the product, and this is the relationship between water and cup. The lowest "one" is the channel . Without channels like Apple, OPPO, Xiaomi, and Huawei, no one would know any information. The "1" in the middle is a technology or algorithm, just like a deliveryman, which accurately pushes out the content. Users are above this "king". Only by having users can they dominate the world.

In 1992, the year I entered university, China underwent essential changes, from a planned economy to a market economy, which was called "a socialist market economy with Chinese characteristics." What is a planned economy? The state stipulates how many tons of steel, how many tons of cloth, how many tons of rice are produced, and then allocates it to everyone. How much do you have and how much he has, you need to buy cloth tickets and grain tickets. It’s different now. We produce according to market demand. The market needs to produce how many tons of steel and how many tons of rice, and then use price leverage to mediate supply and demand. This transformation is the same as the transformation of our media today - content production and content distribution have undergone essential changes, so we must transform. With the thinking of planned economy, there is a dead end. You see, in the past, in the era of planned economy, restaurants were all state-owned - this is equivalent to our institutional media; now, state-owned brands still exist, but most restaurants are already "self-employed", which is equivalent to our self-media. How to manage and use it is the new challenge given to us by the times.

Finally, we will return to the topic - today we are talking about the issue of "Tao". What is Tao? It is the law of the operation of everything. Those who master the Tao can be invincible. So how to master the Tao? The seniors have already told us that it is to "exploration of things and principles".

It comes from Wang Yangming, the founder of "Xinxue". The eight core words of "Mind Learning" are "Studying things and exploring the principles, and integrating knowledge and action." The so-called "grid" is analysis; for example, grid bamboo is used to analyze its structure. After placing it, you will understand its growth pattern.

The process of "exploring things" is very boring. What I talked about today is a summary of what I have been “informing” for two years. Many things were hard to figure out at the beginning, just like Wang Yangming was suffering before he thought clearly about the "nature is human desire."

But when people are most in pain, they are often when qualitative changes occur. My memory is on duty in the office on the night of the 30th of the Spring Festival this year. Firecrackers rumbled outside, and I felt extremely distressed inside. I remembered the Spring Festival two years ago. It was like a night of brilliant fireworks that I completed the preface to the book "Beyond the Portal". Later, this book became a classic work in the media industry. During the Spring Festival in 2017, I also wanted to make something, so I didn’t know where the inspiration came from, so I hand-drawn this picture. When I finished drawing the picture, my eyes suddenly became clear. Everything in this speech is on this picture. I laughed to the sky because I reached another realm, and a new unknown was revealed to me.