As the product manager of search, recommendation and other strategies of the traffic platform, he will design many strategies to allocate traffic, or increase the click rate and conversion rate.
These strategies often have to go through multiple rounds of iterations before they are relatively mature, so there will be some loopholes in the iteration process.
For those who gain traffic on these platforms, that is, do growth, there are opportunities to capture these loopholes. The discovery of these opportunities also often requires some reverse thinking.
Today on the fifteenth day, I will share some very interesting opportunities that I have learned in the past.
1. Click model and click right escalation
We have talked about the iteration and evolution of the search system in the past. For the search system, with the launch of the product, users will accumulate a large number of click samples for the system.
These click samples help the system to gradually transition from some baseline models to click models , that is, Learning to Rank (LTR) methods.
In other words, by learning from samples, the system can build a model that increases the weight of high-click-through-rate content in search results.
Because generally speaking, the higher the click-through rate, the more likely it is that the user actually wants the search result.
But on the other hand, as an SEO party, you can also use this logic to improve your website’s search ranking and get more traffic.
For example, constantly search for target keywords, and then click to visit your own website in the search results, leaving a sample of clicks to the system.
This process can be done manually by yourself, or it can be done in batches by machines, and it can also be distributed to crowdsourcing.
This is click right escalation , that is, through this click behavior, the weight of the target content is increased. Some friends who did SEO in the past gained a lot of traffic through this method when search engines were immature.
From a more general point of view, in fact, headline party, abstract induction, etc. are also a kind of click-to-raise. At the same time, as one of the most common cheating methods, click to raise privileges can also be applied to various product modules such as advertisements, search prompts, and hot lists.
Therefore, if you just do a simple click model, it is easy to be exploited by such loopholes and escalate the rights of inferior content. For platforms, evaluating content quality and identifying fraudulent traffic is a long way to go.
2. Social recommendation and precise delivery
The social recommendation of WeChat video accounts has been a topic of discussion since the day it was launched.
But in fact, Douyin has made many attempts in social recommendation for a long time. If you use Douyin a lot, you may occasionally see such recommended tags:
It is very reasonable to think about it. The so-called people gather together and things are divided into groups. In addition to the videos posted by friends, other short videos such as likes, comments or favorites by friends are likely to be liked by you.
However, as a short video application without a social network itself, Douyin needs to start from various clues to explore the relationship country email list between users if it wants to tap the original social relationship of users.
For example, the relationship with the WIFI network, the relationship obtained after a third-party authorized login such as Weibo, the user's mobile phone address book obtained by authorization, etc.
As long as you have a slight intersection, you can use it as a weight to increase the probability of your friends' favorite videos and increase the accuracy of your recommendation.
For example, if user A's phone number is stored in user A's address book, usually two people know each other, or the probability that two people know each other is much higher than random two people in the database. .
Therefore, recommending a video that B has liked to A, or recommending a video that A has liked to B is a reasonable social recommendation logic.
But as I said before, this is just the usual case. If A stores B's mobile phone number, does B really know A? Or even, does A really know B?
The problem lies here.
Probably last year or the year before, someone did something like this:
First of all, through some gray or black means, a group of relatively accurate mobile phone numbers of users, such as all mothers, or all college students, are obtained;
After getting these mobile phone numbers, through the import address book function, this "friendship" is authorized to Douyin in batches;
What do you start doing next? Simulate a user starts to swipe the video, and other videos are randomly skipped. However, the video uploaded to the account published or maintained by oneself is full of interaction, like, follow, comment, full play, and full operation.
As a result, after these "friends" are online, many people will find that these carefully prepared and manipulated videos "accurately" appear in their video streams due to social recommendations.
It’s time to drain traffic later, and it’s time to convert. It is equivalent to using a free method, increasing a lot of exposure, and achieving a precise delivery effect similar to paid advertising .
These people also quietly earned seven figures back and forth.
Of course, for the platform, in the iterative process of the recommendation strategy, the system can be further improved to prevent this "traffic stealing" behavior.
For example, treat friend relationships in the address book as one-way instead of two-way, or use second-degree relationships to further filter.
But I have to say that this kind of customer acquisition idea can be regarded as very strange.