App Delivery Growth: Introduction to Attribution Models and Frameworks

The advertising industry has always had a famous saying from John Wanamaker, the father of the department store industry: "I know that half of the advertising dollars are wasted, but I don't know which half." Why do we say “I don’t know which half is wasted”? The reason is simply that it is impossible to measure the conversion effect brought by advertising traffic. "Traffic" and "conversion" have always been the two key words surrounding the growth of App delivery. During the delivery process, advertisers often don't understand a series of questions such as "Where did the money go? Which channels are more effective? Which ones have low ROI? How to allocate the advertising budget more reasonably in the next stage?"

Every advertiser hopes to obtain the maximum advertising revenue within a limited budget and spend the budget wisely. In fact, these issues are closely related to the advertising attribution model and logic.

Therefore, understanding attribution models and logic is one of the essential skills for every advertiser/product/operator.

background

Attribution is a very complicated matter, because there may be thousands of influencing factors behind the occurrence of a phenomenon. What we need to do is to find the most critical factors through attribution, and then gradually improve on these key factors.

The key points of attribution: make bold guesses and verify carefully. A bold guess is to list all possibilities and key paths. Careful verification means verifying every possibility of the entire link based on the results, and finding the one with the greatest possibility or the most convincing one.

Let's assume a scenario: suppose I was watching a TV series on Tencent Video on OTT and saw a pre-roll ad for Product A. Then I searched for information about Product A on Baidu on my PC, clicked on its paid ad and reached the details page of Product A to learn about the product.

The next morning, I felt that product A was particularly suitable for me, so I searched for it on my mobile browser. By clicking on its paid advertisement, I was redirected to the app store, which allowed me to download and install the shopping app.

After downloading and installing the app, I forgot to open the app to make a purchase. While I was watching TV series on my phone at night, I saw an advertisement for this product on the iQiyi App. I clicked on the advertisement and woke up the subsequent App for activation and registration.

Then I searched for the product in the shopping app and saw the search bidding ad for the product in the app. I clicked on the ad, went to the product details page and placed an order. I found matching product B in the order details of product A, and ended up purchasing them together.

Okay, now let’s sort out my shopping journey: OTT advertising -> paid search advertising on PC -> paid search advertising on mobile -> app market -> download shopping app -> mobile OTV advertising -> wake up shopping app -> activate shopping app -> paid search advertising in shopping app -> order product A -> purchase product B at the same time.

The entire shopping journey is very long, so the question is, in this case, to which advertising campaign should the last two orders be attributed?

At the same time, to which channel should the download activation (new users) be attributed? Everything that follows is about attribution.

Attribution is about "cause" and "effect". Let's first analyze the relevant causes and effects in the above shopping journey case.

Cause: Usually refers to the touch points where users learn about the product when the advertisement is placed. These touch points include exposure, clicks, or any other ways that can reach the user. Mainly answer the following questions:

• when: What time?
• who: Who?
• where: On what channel and on what device?
• what: What did you do?

Result: usually refers to the event that advertisers hope to bring about conversion effects through advertising, such as downloading and activating the shopping app, purchasing product A in the advertisement, and purchasing product B at the same time.

Now that we have both the “cause” and the “effect”, we will formally begin to introduce the relevant methodology of attribution.

User ID

During this shopping journey, users visited three terminals: OTT, PC and mobile. Although it is the same real user (biometric identifier) ​​on these three ends, in programmatic advertising, we need to identify the same user through different features.

OTT

OTT (Over The Top) advertising refers to video services based on the open Internet, and the terminals can be televisions, set-top boxes, etc. Currently, the operating system of OTT devices is mainly Android, and Apple TV is not ruled out at that time. On the Android platform, the device identifier of OTT is mainly MAC, and on TVOS (Apple), the device identifier is mainly IDFA.

PC

For PC advertising, since it cannot obtain device-related information, the user’s identity is mainly IP+UA. There are also ways to identify users through cookies, but this requires cookie mapping.

Android

For mobile advertising, the commonly used device identifiers for Android are IMEI, OAID, MAC, and AndroidId, and the attribution priority is IMEI>OAID>MAC>AndroidID. Since the Android system no longer supports the acquisition of IMEI after the Android Q version, the current Android device number is mainly based on OAID. One thing to note here is that the OAID formats of different manufacturers are different, so you need to be careful when using them.

iOS

For mobile advertising, the commonly used device identifiers on the iOS side are mainly IDFA, IDFV, and CAID, and the attribution priority is IDFA>CAID>IDFV. After IOS14, in order to further protect user privacy, the acquisition of IDFA requires explicit authorization by the user, resulting in a lower IDFA acquisition rate. Major mainstream media outlets have also released response plans, which are basically CAID. However, apps that integrate the CAID SDK will be rejected when submitted to Apple for review.

Attribution Logic

Now that user identification has been explained, let’s talk about attribution logic. Attribution logic plays a decisive role in attributing the "effect" to a certain "cause". The commonly used attribution logics in the industry mainly include the following:

First Interaction Model

The first interaction model emphasizes the impact of the first interaction on the user. Regardless of any subsequent behaviors, the conversion results are attributed to the first interaction channel that established the connection.

However, this model may have problems such as the user behavior cycle being too long, resulting in inaccurate predictions. Therefore, this model is suitable for new brands to build awareness in the early stages of brand building or product promotion.

In the above shopping journey, through the first interaction model logic, the conversion of activating the order will be attributed to "mobile paid search advertising". In fact, the user’s real first interaction occurs on the OTT side, but it is difficult to connect users on the OTT side and the mobile side. If it can be connected, then this conversion can be attributed to “OTT advertising.”

Last Model

The last interaction model emphasizes the impact of the last interaction on the user. Regardless of any previous behaviors, the conversion results are attributed to the last interaction channel that established the connection.

However, a significant problem with this model is mis-swiping and theft. By constantly simulating interactive behaviors, traffic from other channels can be unified into your own conversions. In the above shopping journey, through the final interaction model logic, the conversion of activated orders will be attributed to "mobile OTV advertising".

Therefore, this model is suitable for businesses with few conversion paths and short cycles, or in other words, the purpose of advertising is to guide users to purchase, and often the final step is the most critical.

Average Model

The average distribution model emphasizes that all interactive behaviors of users throughout the entire journey need to be recorded and analyzed, so the final conversion results should be attributed evenly to each link. It is suitable for companies where advertisers want to stay connected with users throughout the journey and maintain brand awareness. In this case, each channel plays an equal role in the customer’s consideration process. But the problem with this attribution model is that it cannot distinguish the effectiveness of different channels.

Time Decay Model

The time decay model recognizes that the user's final conversion is affected by every behavior throughout the entire journey, but emphasizes that the closer the behavior is to the moment of conversion, the greater the impact. This attribution model is an upgraded version of the average distribution model. This model is suitable for situations where the user decision cycle is short and the sales cycle is short. For example, if you are doing a short-term promotion and only run advertising for two days, then the ads for these two days should be given a higher weight.

Customized Model

The custom model emphasizes that the final conversion effect can be customized and attributed to each channel in the conversion chain. When using it, the proportion needs to be set for different channels. The problem with this conversion model is how to set the ratio for different channels. It requires strong data analysis capabilities to support the entire model and must stand from a fair and just perspective to prevent the initiation of gray industries.

summary

Although each attribution model has its own advantages and disadvantages, when using it, as long as the selected model is in line with the business scenario and is fair to all channels, it will be fine. After all, any channel that has participated in the conversion process can have a certain impact on users.

Attribution Window Period

In addition to the attribution model, the attribution window period will also affect which ad delivery results in the final conversion. The so-called attribution window cycle refers to how long the "cause" time should be traced back when a user converts. This "time" is the attribution window. It is not difficult to understand that any conversion result needs to be traced back to a certain time period. Different time period settings will result in different "causes" that can be involved.

So how long should this window period be set to? Generally speaking, the correct period of the attribution window should be determined according to the goals of the promotion campaign. Different promotion campaigns have different attribution periods. Even for the same promotional campaign, the attribution window period will be different when promoted on different platforms, such as PC and App. You can formulate different window cycles based on the average conversion process of each company.

Attribution Method

Now that we have explained users, models, and window cycles, let’s talk about how to associate these “cause” events with conversion results. Common attribution includes the following:

Unique device number attribution

For In-App delivery, device number attribution is mainly used, provided that the device number is unique and can be associated in different scenarios. When a user interacts with the device, such as viewing or clicking on an ad, the advertising platform obtains the device ID of the device and sends the device ID back to the advertiser through a monitoring link. When a user completes a conversion on the advertiser's side, the advertiser's data analysis platform can match the user's advertising behavior on the delivery channel based on the device number, thereby measuring and attributing the delivery effects of different channels.

IP+UA Attribution

IP+UA attribution is fuzzy attribution because it does not have a unique device identifier. It is used to supplement attribution when the device number cannot be obtained. Its attribution principle is similar to device number attribution, which means that when a user interacts with an advertisement, the user's IP and UA (User-Agent, including the user's operating system, mobile phone model, browser information, etc.) are collected and matched with the user's IP and UA at the time of conversion to achieve conversion attribution.

Channel Package Attribution

The main application scenario of channel package attribution is on the Android side, where the pre-defined "channel number" is written into the APK installation package.

When the advertisement is delivered, the download link is the App package link with a "channel number". After the user downloads and activates the App, the channel number can be read from the installation package for attribution.

This attribution method is simple and is not limited to the acquisition of device numbers. However, this method has the problem of installation package coverage: Android phones generally have a system-level app store. Installation packages with specified "channel numbers" can be easily intercepted by the app store, forcing users to download them from the app store. As a result, the final conversion effect is attributed to the installation package in the app market.

In order to prevent this situation from happening, the issue of anti-coverage will be introduced here, which will be introduced in subsequent articles. Let’s put an eye on it first.

Scheme Attribution

Scheme attribution refers to clicking an ad in a media app and waking up the advertiser's target app through a scheme link. At this time, the scheme link carries relevant performance parameters.

When a user performs a conversion, the performance parameters on the scheme are matched to achieve conversion attribution. The usage scenario of this attribution method is to attract activity, because it relies on the user having installed the advertiser's promoted app.

For in-App delivery, PC and H5 can also derive landing page attribution, that is, splicing relevant performance parameters into the landing page link of the advertisement. In this way, when the user converts, the advertiser can also obtain the performance parameters on the landing page for attribution.

Clipboard attribution

Clipboard attribution means that when a user interacts with an ad in an H5 environment, the unique identifier of the interaction behavior is written to the clipboard. When the user subsequently converts, attribution is performed by reading the clipboard content.

The problem with this attribution method is user privacy. After all, the Ministry of Industry and Information Technology has been strictly investigating operations involving user privacy.

summary

As different attribution methods, the core point is mainly how to associate "cause" and "effect" through a unique identifier.

This unique identifier can be device-related information, channel package information, or a unique identifier for a particular ad delivery. As long as the unique identifier can be quickly collected when user ad interaction and conversion occur, the collected unique identifier can be stored locally or in the underlying data warehouse.

Attribution framework design

After introducing the overall methodology, I will talk to you about the overall design of the attribution framework and see which modules need to be designed for a complete attribution.

The above picture basically shows a complete process from advertising interaction to conversion attribution, with points 1-6 marked. Let’s break down the modules corresponding to these 6 points.

Monitoring module

The monitoring system corresponds to 1 in the above figure. This module mainly collects user interaction behaviors through links. The parameters collected by the links depend on the macro replacement parameters supported by different advertising platforms. The general parameters are: device number, timestamp, IP+UA, and some parameters involved in advertising delivery.

After the monitoring link receives the data, it is stored in the underlying data warehouse. When a conversion behavior occurs subsequently, the corresponding interactive behavior data is matched in the data warehouse.

Scheme Links

The Schem link corresponds to 2 in the above figure. The main function of this module is to wake up the App to the specified page and record the wake-up performance parameters.

When the target App is awakened, the awakening performance parameters are stored locally; when the user converts, the local parameters are read and sent to the downstream conversion attribution service, based on which Scheme attribution is performed.

One thing to note here is that when storing performance parameters locally, parameter verification is required and illegal parameters need to be discarded; at the same time, it is necessary to combine attribution logic, such as the first interaction model, which only needs to be stored once.

Android packaging module

The Android packaging module corresponds to 3 in the above figure. The function of this module is to package and manage the Android channel package and output the APK file of the channel package.

JS-SDK module

JS-SDK corresponds to 5 and 6 in the above figure. The main function of this module is to record the content through JS-SDK when opening the H5 page, and write the relevant content to the clipboard and cookie.

The recorded content includes but is not limited to IP, UA, URL link, unique advertising ID, etc. When the user opens the App, whether it is activated or awakened, the corresponding clipboard content can be read. Display different contents according to the corresponding content.

Activity Configuration Module

Configure the page that is presented to the user when the user opens the App. Scene restoration or interactive customization can be achieved by associating monitoring links and clipboard contents.

This module is relatively important. The first page a user enters into the App is directly related to the user's conversion. In this section, you can configure some relevant strategies to show users the content that they are most interested in.

For example, display the advertising products that users see, so what you see is what you get; display special zone content for new users to increase the possibility of first orders, etc.

Attribution Module

The attribution module is the most important module in the overall framework. It is mainly responsible for the management of attribution strategies and the display of attribution results.

Different attribution strategies can be configured for different scenarios; the attribution results can be displayed through data reports.

Summarize

This article introduces the attribution system from the dimensions of user identification, attribution logic, attribution method, and time window, and also gives a rough attribution framework design.

Author: Baozi

Source: Daily Notes on Commercialized Products