Advertising attribution: how to measure the value of buying volume?

The author of this article ：Fengyu , tencent IEG Product planning

One 、 Basic concepts

（ One ） Definition

attribution （Attribution） It refers to the use of identifiers to identify the conversion path of users , And judge the effective contact value . The advertising platform is based on attribution results , Pay for it ; Advertisers can be used to optimize the delivery strategy .

picture source ：Google App Online advertising courses

（ Two ） Implement the core logic ： The match between advertising and transformation

Attribution of applied advertising , Rely on getting ads 、 Transform information , And realize the matching of the two ：

1. Advertising information ： The user clicks / Browsed the advertisement , User information that will click on the advertisement ( For example, the user's IMEI perhaps IDFA etc. ） And advertising information （ Like advertising trackingID, Advertising application package name or IOS Of APP ID） Upload and record ;
2. Transform information ： After the advertisement is transformed , Convert this advertisement into user information 、 The attribute information of the advertisement is matched with the user information and advertisement information of the click advertisement to complete the attribution .

picture source ： Zeng Rong's nonsense

As can be seen from the above figure , From advertising to transformation , The core depends on three steps ：

• The first 3 Step ： When the user device clicks on the advertisement , Will the advertisement ID 、 Send user information to attribution platform ;
• The first 7 Step ： Newly installed APP After starting , Will the advertisement ID、 Send user information to attribution platform ;
• The first 8 Step ： Attribution platform for attribution analysis .

Two 、 Record point ： Data required for attribution and how it is used

Users' watching advertisements and conversion usually do not happen in a closed loop APP Inside , To achieve cross user APP The tracking of is actually to connect the first 3 Step （ Advertisement reporting ） And the 7 Step （ Report of transformation ）, And combine the two ID Match and identify the user .

（ One ） Deterministic attribution ：Android

Android Deterministic attribution at the user level , Include equipment ID( Such as Google Of GAID)、Google Play Referrer And consumer identifier ( Such as anonymous login data , Or the email address that has been hashed ) Multiple identifiers including , Accurately identify user equipment , And will click on the advertisement 、 User behaviors such as downloading applications or executing in app events match it .

1. overseas ： Google framework GAID and Google Play Referrer

Overseas Android platforms can be used to track users ID There are mainly ：

1. GAID： namely Google Advertising ID, yes Google Play App store and Android Devices installed by third-party app stores ID;
2. Google Play Referrer： namely Google Play For the same SDK Assigned fixed value , Only for installation of Google Framework Android device , And only track Google Play Application behavior in

2. At home ： Launched by mobile security alliance  OAID

OAID： By the domestic mobile security alliance （MSA） United Huawei , millet ,oppo,vivo And other anonymous device identifiers launched by terminal manufacturers .

In addition to clear users provided by interested parties ID outside , Fingerprint matching can also be used for Android's deterministic attribution . Fingerprint recognition uses device information （ System font settings 、 Hardware properties, etc ） Create continuous and unique ID, Used to identify specific users . This is not in line with Apple's privacy policy .

（ Two ） Deterministic attribution ：iOS

picture source ： after IDFA Time ： Buying volume put 、 advertisements 、 Marketing strategy of attribution analysis - GameRes Hot money

1. iOS 14.4 front ：IDFA

IDFA yes iOS Advertisements ID, stay 14.4 It was in the default on state , Both self attribution platforms and third-party attribution platforms can be used for the deterministic attribution of advertising .

1. iOS 14.5 in the future ：SKAN

The apple in iOS14.5 Upgraded privacy policy （ATT,App Tracking Transparency ）, User IDFA It is off by default , Applications need to track user behavior , You need to send two request pop ups （ Advertisement broadcast and download ）, whole IDFA Low availability .Vungle The test data is 95% The user of opened LAT（Limit Ad Tracking）, Limit advertising tracking .

So apple is IDFA In addition to providing SKAdNetwork（ SK finger StoreKit） , Can be used for non user level deterministic attribution ： Without identifying the user's identity , Measure the application installation and advertising effect .

picture source ：Apple Developer Documentation

Click on the advertisement to download 、 After activation ,SKAN Will start 24 Hour countdown , Count user behavior ; After installation 0-24 After hours, the installation and conversion information will be sent back to the advertising platform （ Such as 24 Hours have conversion events , Then the statistics can be extended 24 Hours , at most 48 Hours ）.

SKAN built-in 64 A bit , It can be used to track different events , picture source ：iOS 14 Growth support programs

SKAN The limitations of include ： Only the early behavior data of users are counted , Cannot measure accurately LTV; Only data at the advertising series level , Excluding advertising materials and other data , Cannot optimize alone ; The data has at least 24 Hour delay, etc .

（ 3、 ... and ） Probabilistic attribution

In the case that deterministic attribution cannot be carried out by means of identifiers, etc , Attribution platform can carry out probability attribution .

The specific implementation mode is ： Attribution platform believes 80% The installation of all takes place within the first hour after the click of the advertisement , Therefore, the user click time will be collected 、 Limited parameters such as installation time and basic equipment information （ Temporary data ）, Speculate the source of the installation within a few hours after clicking .

Branch And other third-party platforms call probability attribution 90% The above users realize probability attribution .

picture source ：iOS14 after , You may ask about attribution

3、 ... and 、 Value point ： Attribution model

From the previous attribution logic part , We assume that the attribution monitoring platform has obtained the required advertising behavior and APP Transform data , The problem to be solved in this step is the attribution model ： How to base on the transformation behavior of users , Judge the value of advertising .

（ One ） Basic model ：

The basic attribution model refers to the set rules , Judge the advertising value . Advertising platforms usually pay and settle according to the rules in the basic model .

1. Final installation （Last Install）/ Channel package

Application installation in domestic Android Market , It mainly depends on channel subcontracting ： That is, different channels / Traffic properties , Put on a different APK package , Advertisers finally activate （Last Install） Which channel package comes from to evaluate the effect of traffic channels .AMS Statistical approx 50% Of game advertisers adopt this scheme , Such as headlines .

Possible problems ：

1. Statistics of advertisers and advertising platforms gap： Due to the influence of Android manufacturer hijacking , The transformation counted by advertisers will be less than that of advertising platforms , As a result, the effect and quality of traffic channels cannot be accurately evaluated ;

picture source ： the B standing UP The master stood out , Reveal the secret of app store blocking mobile game installation | Game view | GameLook.com.cn

1. Can't measure “ assists ” The value of ： For example, users are watching effect advertisements , After content marketing, enter the app store to download , The app store can get all the scores ;
2. Unable to accurately measure the material / The influence of copywriting on transformation ： If the channel package is only subdivided into traffic , Instead of aiming at different copywriting 、 Distinguish between materials , It will be detrimental to the model optimization of advertising platform and the material copywriting optimization of advertisers .

2. Finally, click （Last Click）

Finally, the click model refers to that users have clicked on advertisements on multiple platforms before installation , Finally, it is attributed to the advertising platform that has clicked the most recently since the first installation .

Click attribution is the most common way , Click the attribution window, which is usually 7 God 、28 God （ A special case ：SKAN by 24 Hours ）.

picture source Adjust  Reference documents

Some domestic advertisers will re match based on the last click data from different sources , Do the second attribution of the last click .

Possible problems ：

It is impossible to measure the value of assists ： For example, users watch effect advertisements many times 、 Download after content marketing , The last click can get all the scores .

1. Effective contact attribution （View-Through Ads）

Before talking about effective contact attribution , First of all, we should distinguish the types of advertising behavior data ：

1. Click attribution ： For example, users click on banners 、 video 、 Insert ads, etc . The attribution window is usually 7 God .

2. Exhibition / Exposure / Browsing attribution ： Users see advertisements but do not click on them , The resulting installation can be attributed to the channel that displays the advertisement . thus , The attribution window is short , Only for 24 Hours .

Because clicking has the active behavior of users , Therefore, in the attribution model, click advertising takes priority over display advertising .

picture source Adjust  Reference documents  

At present, foreign Facebook、 Domestic platforms such as Kwai adopt effective contact attribution schemes , The specific logic is ： The advertising platform will let users click + Effective contact （ Like playing 3 Seconds later ） They are sent to advertisers as advertising attribution information .

The effect of effective contact attribution ：

1. For advertisers ： The original natural quantity is classified into the effective contact , Advertisers spend more ;

The picture is from 【 And advertising attribution 】 Analysis on the reason why the attribution of effective contact can be popularized

1. For advertisers ： Improve the conversion rate of Kwai and other buying channels , Will affect brand advertising 、 Search optimization （ASO） The natural quantity attribution of the team , Part of the natural quantity originally belonging to the team will be attributed to the display attribution ;
2. For advertising platforms that use advertising contacts ： Increase the original natural quantity , The effect on the advertising platform becomes better , The overall advertising cost will be reduced ;
3. For advertising platforms that do not use effective contacts ： For example, the buying team put advertisements in Kwai , Yes, click + Play 3s Attribution in two ways （ Grab the original natural quantity ）, Advertising conversion rate increased ; To be specific ：
   1. For flows that have used valid contacts , The flow value is incorporated into the original natural quantity ecpm Will rise , Advertising models can get more （ Click on + Exposure ）; Therefore, for unused effective contacts 、 Just click / Install attributed traffic channels , The effect will get worse .
   2. For example, a Facebook The advertisement has 10w The exposure , CTR 1.2%,CVR 30%, Therefore, the available clicks are converted into 360; In the case of increasing the attribution of effective contacts , For example, exposure attribution has 80 individual , Holistic CVR Up to the 37%. Advertisers keep the same advertising budget as before ,CPA It will go down ;
   3. here , If the advertiser chooses to reduce the total advertising budget , You can also get the same amount of conversion as the original （ Click on + Exposure conversion ）, Therefore, the advertising revenue obtained by the same traffic increases , namely ECPM promote .

Click attribution to upgrade to click + Exposure attribution effect improved , This case refers to ：

【 On advertising attribution 】 Forced the “ assists ” Count as “ Direct score ”, Do you want to be shameless ？

The effect of adding exposure attribution logic to a platform is improved

Other common basic models include ： First click 、 Time decay 、 Location attribution and other models .

（ Two ） Algorithm model ：

A differentiation model based on user data , Algorithm model attribution , Also known as data-driven attribution （Data-Driven Attribution, abbreviation DDA）.

It is different from the basic model based on the set rules ,DDA Use all available path data , Including path length , Exposure sequence and advertising materials , To understand how the existence of a specific marketing touchpoint affects the possibility of user conversion to better allocate credit to any touchpoint .

The commonly used algorithm is Shapley value 、 Markov chain 、Harsanyi Dividend And survival analysis .

1. Shapley Value ：Google Analytics

“Shapley value ” By the Nobel Laureate in Economics Lloyd S. Shapley Put forward , It's a team member （ Marketing contacts ） Distribute team results fairly （ Conversion results ） Methods .

The specific implementation method is ：1） By comparing the conversion probability of similar users who have contacted these contact points with the probability when a contact point does not appear in the path ;2） And the system will compare all the different arrangements of the contact points , Assign different credits to different path locations .

Simplified version understanding ：

adopt “Facebook-Google” Linked CVR yes 2%; adopt “Facebook-TikTok-Google” Linked CVR yes 3%, That can be considered “TikTok” The value that contacts add to conversion is 50%.

meanwhile Shapley The value model will also change the order of links , in the light of TikTok stay Google front 、 The latter cases are modeled separately .

The picture is from ：markov model vs shapley value marketing attribution

Google Analytics Examples of applications ：

Display and transform each channel through the hierarchical diagram of the model 、 Weighted average contribution value at each path position .

Look vertically , The percentage represents at the specified path location , The actual percentage weight of a channel in the current node , for example The contribution of paid search to users' first conversion path is 31%, You can promote users to start transforming by putting such advertisements .

1. Markov chain model

Markov chain assumes that the probability of state transition at a certain time only depends on its previous state . there “ state ” It can be understood as weather ： For example, today's weather only depends on yesterday's weather , It doesn't depend on the weather the day after tomorrow ; In the advertisement , Status is understood as channel , The access channel of the user this time only depends on the last access channel , It doesn't depend on the last visited channel .

The model can remove a certain state （ channel ）, Calculate the contribution value of this state to the final result .

Simplified version understanding ：

1. There is Facebook In the case of channels , adopt “Facebook-TikTok” The probability of link conversion is 30%*20%*80%, adopt “Google” The probability of link conversion is 60%, The total conversion rate is ;
2. remove Facebook after , The overall conversion rate is only Google channel , namely 50%*60%, Judge the value of the channel by the declining conversion rate .

The picture is from ：markov model vs shapley value marketing attribution

1. Uplift： Tencent advertising

Uplift The model predicts the causal effect of some intervention on individual state or behavior , Specifically expressed as the difference between two conditional probabilities .

Specific implementation method ： Divide the target population randomly , Advertising from one group of users , The other group did not intervene （ No advertising ）, Then count the difference between the two groups in conversion rate , This difference can be approximated as the average possible causal effect of people with the same characteristics .

Example of Tencent Advertising Application ：

Compare the number of orders placed by the processing group and the control group 、 Order quantity 、 Conversion rate and other data , Analyze the effect of processing group data improvement , That is to say Uplift Incremental conversion .

Besides ,Adobe It uses Harsanyi Dividend Algorithm ,Facebook Data attribution model has been introduced, but the specific algorithm has not been published , And this year 4 Monthly deactivation .

Four 、 Attribution platform

In the attribution analysis of step 8 , Attribution monitoring platform can play two roles ：

1. Advertising platform ： for example Google Ads Advertising can be launched , Advertisers are also required to report it APP Behavioral data for attribution . This kind of advertisement is used to launch , The platform that also carries advertising attribution analysis is called self attribution platform ;
2. Third party monitoring platform ： for example Adjust,Appsflyer, as well as Kochava And other third-party attribution platforms .

（ One ） Advertising platform / Self attribution platform (Self Attributing Networks, SANs)

When the advertisement is exposed on the traffic , The advertising platform will be exposed 、 Click data is recorded in the background of its advertising attribution . When users download and activate apps from the app store , Advertisers need to send back the activation data of their applications to the advertising background for attribution . install 、 Conversion and other data will be used for advertising billing and subsequent advertising model optimization .

（ notes ：CPC、CPA and CPM It is the mainstream advertising billing mode , therefore , Media channels charge for all clicks that occur during the attribution window period set by themselves , Whether or not this click is the last click . however , Attribution provider sends installation data to channel , Channels can use this data for their own optimization .）

1. Google Ads：

1） Record point ：

be based on Google Click ID Follow up

2） Value point ：

• Attribution model ： Click attribution + Exposure attribution

• Attribution window ：
  • Application advertisements aimed at increasing installation （ACi）：
    • Click on ： Default 30 God , It can be adjusted to 7 God /30 God ;
    • Watch with interest （ Click on or watch 10 Video ads of more than seconds ）： Default 2 God , It can be adjusted to 1 God -3 God
    • Exhibition ：1 God , It's not adjustable ;
    • install ： Default 90 God , It can be adjusted to 1 God - 90 God .
  • Application advertising aimed at improving interactive events （ACe）：
    • Click on ： Default 90 God , It can be adjusted to 1 God -90 God ;
    • Watch with interest ： Default 2 God , It can be adjusted to 1 God -30 God
    • Exhibition ：1 God , It's not adjustable .

The picture is from ： Introduction to the transformation time range of application advertising series

2. Facebook Ads：

1） Record point ：Facebook Account

The user is in A、B Log in the same on the device Facebook Account , Users on Android A I have seen advertisements on the device , stay iOS Of B Install... In the device ,Facebook This installation will also be attributed to .

2） Value point ：

• Attribution model ： Click attribution + Exposure attribution

• Attribution window ：1 Day Click +1 Day exposure ;1 Day Click ;7 Day Click ;7 Day Click +1 Day exposure . There are different attribution windows for different types of advertisements .
  • AEO（App Event Optimization）： Target installation events
  • Conversion： Target in application conversion events

3. Apple Search Ads：

1） Record point ： Do not track users

Advertising orientation ： According to the crowd （ 5000+ user ） clustering , At least 5000+ user

Extended data ：https://searchads.apple.com/privacy

2） Value point ：

• Attribution model ： Finally, click
• Attribution window ：30 God

Extended reading ： How mobile measurement service providers help monitor Apple Search Ads Advanced result

（ Two ） Third party attribution platform (Mobile Measurement Providers, MMPs)

Self attribution platforms usually only have the advertising data of this channel , Ideally , Through the third-party attribution platform, the advertising data of the whole platform can be displayed . The specific process is as follows ：

1. Advertisers report the transformation information to the attribution platform ;
2. Attribution platform sends transformation information to advertising platform ;
3. The advertising platform returns the interactive information between users and advertisements .

（ notes ：MMP It will also provide other capabilities such as traffic anti cheating , For details, please check the introduction of corresponding products ）

1. Record point

A. Deterministic attribution data ：

• IDFA（iOS equipment ） and GPS ADID（ Android devices ） etc.   advertisement ID
• equipment ID , for example IDFV ( in the light of iOS), Android ID ( For Android ) and OAID ( For unusable Goolge Play Service Android devices )
• MMPs self-built ID, Such as Adjust Created on Android reftag id.

B. Probabilistic attribution data ：

• Device type ( Mobile phone, for example ); Equipment name ( for example Samsung Galaxy S7) etc. ;IP Address ( Such as 77.185.208.234)

This section is referenced from Adjust  Reference documents  

1. Value point

• Attribution window ： Click on 7 God , Exhibition 1 God , It can be set by yourself
• Attribution model ： By default, click the model last , It can be set by yourself

3. SAN And MMP The data difference

• Attribution windows are different ：ASA Default 30 God , other SAN（FB/Google） and MMP Adjustable ,
• Attribution data source / Different models ：MMP Record all you have seen 、 Click the advertisement to generate the installed data ,SAN Only single platform data is recorded .

appendix ： It is applicable to the re attribution of reflow users

Re attribution refers to not using for a period of time / Uninstalled apps , However, the reinstallation or opening event of users who return after re marketing and promotion activities , Attribution .

For example, the user is uninstalling the application N God （ The window is set by the advertiser , Usually it is 90 God ） after , Click on the advertisement again , Re Download , It is regarded as re attribution .

picture source Adjust  Reference documents  

other ：

1. Basic concepts can be found MMP Attribution help document （ Adjust  and  AppsFlyer） 
2. Advertising platform 、MMP The specific rules of are updated from time to time , Please refer to the official help documents .

---

# Live broadcast of Tencent technology #

Tencent senior engineers share technical dry goods

Code scanning appointment ,get Broadcast reminder

Looking back ：

Ten years of data analysis , These books are really useful to me  

A word disassembles ： Scams and loopholes in the crazy metauniverse

Web3.0  Industry panorama and representative project research

Pay attention to , See you next time