APP channel promotion: How to conduct channel evaluation efficiently?

How can channel acquisition practitioners eliminate false traffic from their own channel data, identify high-quality traffic, and maximize the ROI efficiency of channel delivery through continuous channel composition optimization? In this article, the author will analyze two real cases based on Umeng+ Mobile Statistics (U-App AI version) to explain: How to efficiently conduct channel evaluation so that the fake volume gray market of Apps has nowhere to hide?

In the second half of the Internet, the performance of App in attracting new customers is as follows: on the one hand, the explicit costs continue to rise, and on the other hand, abnormal channels and false traffic increase the hidden costs.

To address this issue, Himalaya built a real-time channel delivery monitoring system with the help of Umeng+ Mobile Statistics (U-App AI version): identifying abnormal channels and false traffic, and automatically monitoring alarms to intuitively display details, providing timely decision-making support for management and operations teams, while forming a closed loop of iterative optimization of delivery strategies. In practice, it has saved hundreds of millions of yuan in channel delivery costs, greatly promoting the healthy growth of channel customer acquisition.

In the context of an economic downturn, such a "lean customer acquisition system" is very important for Internet companies.

At present, the competition among various applications and media in the market for channel promotion has reached a fierce state. From the traffic dividend to the stock era, the channel promotion of App products is becoming more and more difficult, which can be manifested in the following three aspects:

1. App stores have built high traffic barriers: the commercialization of app stores of major mobile phone manufacturers has become increasingly serious, and they have begun to pursue profit maximization. The cost of app promotion has risen sharply. The era when the average activation cost was 2-3 yuan is gone forever. Currently, the activation cost of a financial application is even as high as dozens of yuan.
2. Channel media is a mixed bag: mobile media promotion, information flow advertising, and native advertising have exploded. Faced with an extremely competitive and mixed media environment, accurate evaluation of conversion effects and identification of abnormal traffic have become pain points for advertisers.
3. The data capabilities required of channel acquisition practitioners are getting higher and higher: Data science plays a vital role in channel promotion. Making the most of data to drive business growth is a new test for growth managers, product managers, and operation managers.

But as the saying goes, "channels are the foundation of existence, and products are the foundation of life", Internet companies cannot avoid problems, but need to face them head-on.

The author believes that among the common channel acquisition methods, whether free, paid or in exchange, there are certainly certain differences in the natural attributes of different channels (that is, the user activation of the App will naturally vary due to the channel attributes).

But leaving aside uncontrollable factors, the key to solving the above three problems lies in how channel acquisition practitioners can eliminate fake traffic in these channels, identify high-quality traffic, and maximize the ROI efficiency of channel delivery through continuous channel composition optimization.

The author has built a channel evaluation system for this purpose, as shown in the figure below:

The App channel evaluation system is divided into: channel anti-fraud evaluation system and channel quality evaluation system.

• Channel anti-fraud system: takes the real channel volume as the target and is used to identify fake channel volume.
• Channel quality evaluation system: It can be divided into channel effectiveness evaluation system, channel retention evaluation system and channel ROI evaluation system, which correspond to the three goals of new additions, activity and revenue respectively, and aim to quantify channel efficiency, quantify channel retention quality and quantify input-output efficiency.

1. Real case analysis

Case 1: The channel as a whole adds fake volume according to proportion and time period

In December 2018, when I was reviewing the quality of an account in the DSP channel, I found that this account had been sticking to the KPI in all established indicators for three consecutive months, neither higher nor lower.

My long working experience has made me realize that there may be problems here. Then we went deeper into the channel package dimension (one delivery plan corresponds to one channel package) to view the data details of each indicator.

Through analysis, we found that every once in a while, there will be several channel packages with normal short-term retention rates such as 1 day/3 days, but extremely low medium- and long-term retention rates such as 7 days/14 days/30 days (as shown in the figure below). At the same time, these channel packages with low medium- and long-term retention rates are not fixed, but change in rotation at irregular intervals. At this point, it can be basically confirmed that this channel has introduced fake quantities while controlling the overall quality.

This type of advertiser can be said to be "cunning": they realize that it is easy to be discovered if the entire amount is fake, so they choose to retreat in order to advance. While meeting the KPI as a whole, they add a certain proportion of fake volume to the total amount, mixing good and bad together and increasing the difficulty of detecting the fake volume.

For example: 5 fake quantities are mixed in for every 100 real quantities. In addition, they are elusive in the time dimension. They may adulterate channel package A one month and switch to channel package B the next month. If they are discovered, they will use the excuse of channel instability.

Attribution analysis can help companies identify which channel the additional users ultimately come from, but when the channel budget is sufficient and the relevant indicators meet the KPI, channel operators often tend to overlook the false traffic that exists in different channels.

As for these abnormal channels that are a mixture of true and false, if we do not trace the quality of the traffic in each channel, we will be condoning the long-term impact of false traffic on channel delivery, affecting the healthy development of channel customer acquisition, and ultimately causing immeasurable losses to the company. In the above example, it was because the author tracked the data in the channel package dimension that he was able to finally identify the fake volume.

Case 2: Using some low-end models to increase traffic

In a special analysis of the distribution of new users in various channels, the author found that the distribution of new users in some channels was very abnormal:

Theoretically, different channels cover different user groups, and the distribution of users' phone models will also be different. For example, for users of the Huawei App Store channel, the vast majority of their phones are Huawei phones. Excluding the app stores that come with these mobile phone manufacturers, under normal circumstances, the model distribution of new Android users should be diversified, and the four major mobile phone brands of Huami, OV should account for a relatively large proportion.

However, the author found that among the model distribution of new users in certain channels, the ones that topped the list were low-end models from one or two less well-known mobile phone brands, and accounted for as much as 10%-20% (as shown in the figure below).

The author further analyzed the user behavior data of these channels, such as the number of launches, usage time, retention rate, and paid conversions of new users, and found that the number of launches was basically 1, the usage time was less than 10 seconds, the retention rate after 7 days was 0, and the payment rate was 0. At this point, it can be basically determined that the model was used to increase the traffic.

Since channel service providers have agents at all levels during the advertising process and traffic acquisition is opaque, there are more and more gray areas for cheating. The seemingly normal advertising retention actually hides the crisis of fake traffic.

The above two cases are both common cheating methods in CPA and CPD payment forms - using various means to increase downloads, activation and retention. Although the cost of cheating is relatively high, it is also difficult to block. In addition, there are also cheating methods such as brushing exposure under CPM and CPC payment methods and brushing orders under click CPS payment methods.

In short, each type of advertisement has a corresponding payment form, and each payment form has its own appeal for fake traffic benefits.

Although identifying fake channel traffic has always been a major pain point for growth managers and channel promoters, it is impossible for channel promoters and data analysts to have enough time and energy to check the detailed data of each segmented channel one by one and identify fake channel traffic. Even if they can be checked one by one, it will take a lot of time. By the time the problem is found, the related losses may have been caused and cannot be recovered.

So, is it possible to establish a mechanism and build a sound system that can accurately identify fake channel traffic, automatically alarm, and promptly remind relevant personnel?

2. Solution

1. Data monitoring of the entire channel delivery process

The author believes that no matter which dimension is used to identify abnormal channels and false traffic, it is first necessary to associate, integrate and cluster the advertising exposure and click data, user attribute data and user behavior data through certain data collection methods, so as to obtain complete, comprehensive and accurate basic data, and thereby realize the tracking and analysis of user data.

With complete and comprehensive data, on the one hand, we can look for traces of fake traffic from different breakthrough points based on the attributes of the channel; on the other hand, by analyzing and retaining data on a complete link of fake traffic, we can establish strong evidence to support fake traffic.

As shown in the figure above, the data of the entire advertising process can be divided into three categories:

1. Advertising data : advertising exposure, advertising clicks;
2. User basic attribute data : activation period distribution, regional distribution, device terminal distribution, network and operator distribution, and App version distribution of newly added users;
3. User behavior data : retention of new users, number of app launches, app usage time, and business conversion (such as playback time and payment behavior).

2. Channel Anti-fraud Evaluation System

The data itself does not explain the problem, but the interpretation of the data can explain the problem.

In other words, data is only unprocessed raw material that represents objective things. It must be processed and formed into logic before it can be called information. By summarizing and systematizing a large amount of information, it forms knowledge and guides practice.

After achieving data monitoring of the entire channel delivery process, the next step is to use this data to identify abnormal channels and fake traffic - that is, to establish a channel anti-fraud evaluation system, which includes:

1. Obtain data from various dimensions on various platforms;
2. Setting of monitoring indicators and thresholds;
3. Automatic alarm reminder and channel details display;
4. Complete abnormal channel identification and false traffic monitoring.

Third-party monitoring platforms are also improving their anti-cheating capabilities. For example, Umeng+ will conduct anti-cheating at four levels:

1. Restructure the communication protocol, add confidential algorithms, and raise the threshold for cracking the protocol.
2. Add ID change tracking strategy to lock out fake IDs.
3. Upgrade algorithms and control strategies on the server side to more effectively identify brushing behavior.
4. It is to increase the appkey protection strategy, launch technologies such as device fingerprint and simulator recognition, especially to crack down on the volume-brushing behavior on Android devices, and vigorously crack down on the wool party.

3. Implementation steps

(Figure: Abnormal channel monitoring process)

In the channel anti-fraud implementation steps, after a new channel is opened, it must first go through the channel management system and successfully enter the channel package and account-channel matching relationship before it can enter the channel anti-fraud evaluation system.

After obtaining relevant data, set monitoring indicators for abnormal channels, and train thresholds based on historical sample data. Automatically alarm for suspected abnormal channels that are less than the threshold to achieve "instant information arrival". At the same time, display channel details in the report to guide channel optimization.

1. Channel management system

The channel management system is a necessary preliminary preparation in the channel anti-fraud process, which mainly includes channel package management and account-channel matching management.

1) Channel package management: Before launching the channel, channel operators need to add UTM tags to each channel, which is called a channel package, and use channel attribution to identify the channel source of new users. At the same time, these channel packages are classified according to various dimensions to prepare for the subsequent statistics of channel data in various dimensions;

2) Account-channel matching management: Channel operators need to match the accounts/plans in the channel delivery data background with the channels in the App conversion data, in order to achieve continuity between the channel delivery background data and the App user conversion data.

2. Get relevant data

Umeng+ Mobile Statistics (U-App AI Edition) provides diverse data on user channel dimensions, such as:

• The basic user attribute data such as the activation period distribution, regional distribution, device terminal distribution, network and operator distribution, App version distribution of new users, the user behavior data such as the retention status of new users, the number of App launches, and the length of App usage provide a data basis for the channel anti-fraud system.
• The open API interface provides great convenience for obtaining relevant data;
• The quasi-real-time nature of the data makes it possible for subsequent automatic alarms and to achieve refined and efficient risk control.

3. Set monitoring indicators and thresholds for abnormal channels

According to the attributes of each channel type, set abnormal channel monitoring indicators for each channel. At the same time, historical relevant indicator data are used as samples to train and set thresholds. When the indicators exceed the thresholds, they are identified as suspected abnormal channels.

This article only takes the retention indicator as an example:

(1) Data preparation

The author selected all channel packages with historical activation volume greater than 100, and prepared the new volume of these channel packages and the 1/3/7/14/30-day retention rates.

(2) Determine indicators

• Retention rate indicators: 1/3/7/14/30-day retention rate;
• Week-over-week retention rate fluctuations: The decrease in 1/3/7/14/30-day retention rate compared to the same day last week for the same channel.

(3) Setting thresholds

The deciles of the weekly retention index and the weekly year-on-year retention rate fluctuation index in the historical sample data are calculated as shown in the following table:

According to the above standard within 90%, and combined with the actual business situation, the threshold of the monitoring indicator is adjusted as follows:

That is, among the above retention monitoring indicators, if the actual value of a certain channel (activation volume greater than 100) is less than the threshold, it is defined as a suspected abnormal channel.

4. Automatic alarm robot

The company uses DingTalk as an internal communication and collaboration tool. The DingTalk group robot is an advanced extension function of the DingTalk group. It can aggregate information from third-party services into group chats and realize automated information synchronization.

Based on the threshold of abnormal channel monitoring indicators, the information of suspected abnormal channels is synchronized to the DingTalk group, and an alarm is automatically triggered to remind channel operators in a timely manner.

5. Reports visualize details and guide optimization

All historical suspected abnormal channels can be displayed more intuitively and in detail through the Tableau visualization tool, including various relevant data indicators, to guide channel operators to adjust channel delivery in a timely manner.

At the same time, it serves as strong evidence for negotiations with media or advertising agencies, and corresponding countermeasures such as requesting additional volume or refusing payment can be taken.

IV. Conclusion

Faced with the endless stream of abnormal channels and fake traffic, channel acquisition practitioners don’t need to panic. Through the above series of examples and solutions, we found that:

1. Abnormal channels and fake traffic are not without traces. As long as they are digitized and online, they can be traced.
2. We have found that in-depth data analysis with the help of powerful tools is the most direct and reliable way to identify abnormal channels and fake traffic, and this method will play an increasingly important role in the increasingly complex media environment.

The author summarizes his own experience in the industry into two points, hoping to provide some reference and reference for the majority of APP advertisers who are eager to improve their data-driven capabilities:

(1) Use advanced tools to provide strong data support.

Based on my proficient application of data analysis tools and monitoring execution experience, I believe that App advertisers should actively cooperate with platforms that can realize multi-dimensional data analysis. On the one hand, it can save a lot of time for basic data collection and processing; on the other hand, it is also easier to realize the iterative extension of the indicator system.

For example: In my work, I have used the diversity, openness, and quasi-real-time nature of the user channel dimension data of the U-App AI version to help my company achieve one-stop multi-channel and multi-dimensional analysis.

(2) Based on the understanding of the industry and actual work, establish and improve a timely channel evaluation system.

A complete and sound channel evaluation system is the magic weapon for the healthy development of enterprise channel customer acquisition. The channel anti-fraud evaluation system, identification of true channel volume, and channel quality evaluation system correspond to the three goals of new additions, activity, and revenue respectively.

At different stages of development, companies will have different goals and focuses - they may focus on the channel effectiveness evaluation system to pursue user scale, or focus on the channel retention evaluation system to emphasize user activity, or focus on the channel ROI evaluation system to increase user benefits. Channel acquisition practitioners need to handle this flexibly according to the company's development stage.

Author: Lou Yongjin

Author: Lou Yongjin