In this hands-on lab, we will learn how to use Audience Manager Look-alike Modeling in order to expand your audiences, while maintaining high target accuracy and maximizing return on marketing spend. We will walk through best practice examples of building effective look-alike models for both publishers and advertisers with TraitWeight (AAM's look-alike algorithm) based on first and third party data.
Adobe Audience Manager (AAM) is Adobe’s Data Management Platform, which allows customers to connect different internal data lakes as well as various data types such as mobile data and device data, to get a complete understanding of the user. AAM currently integrates with all of the Experience Cloud solutions and many other partners in the digital marketing eco-system.
There are four key components of AAM.
Figure 1: AAM key components
AAM, however, is not just a data bank. It also provides you with free out-of-the-box features that allow you to build look-alike (algorithmic) models to expand audiences. In this lab, we will focus on three of these features:
Next, let's get familiar with a few key Audience Manager terms, which will be used throughout this lab:
Figure 2: AAM key terms
It is the ability to take a set of users with a given characteristic and algorithmically find a new set of users that are similar to the original set of users.
Look-alike modeling can be used by both publishers and advertisers who want to maximize revenue while confidently selling and targeting audiences with high accuracy.
Figure 3: Look-alike modeling use cases key
Open the Google Chrome browser and go to the Audience Manager website: https://bank-beta.demdex.com.
For this demo, we are going to use the Beta environment in Audience Manager, which is sandboxed from production. This is the place where we get to experiment with AAM and play with new features and ideas.
Credentials will be provided by lab instructor.
Figure 4: AAM login screen
Once logged in, feel free to click around and get familiar with the menu at top.
Note: AAM credentials will be deactivated after Summit.
Note: The data that you see in your account is simulated. Your trait, segment, and data source ids will be different from the ones in the screenshots, but the names will be the same.
Figure 5: BuildIt company logo
You are the Ad Sales Manager at BuildIt - a popular website and online community for home improvements, interior design and decorating. You have a Garage Rehab Enthusiasts segment audience of 15K users which has sold out and you want to extend this audience with other garage rehab look-alike users from your 1st party data.
Let's go ahead and build your first look-alike model.
Figure 6: Models drop down
Figure 7: Model details
Congratulations! You have just created your first look-alike model. It is that easy! Now you can grab a cup of coffee. Results will be ready within 24hr. Luckily, we have already pre-run the exact same models in each of your accounts, so you don't have to wait.
With steps [3-7], you are selecting your baseline (seed audience) trait or segment.
By selecting Data Sources (step 8), you tell the model which data can be used for modeling.
Trait Exclusion (step 9) provides additional controls in your modeling workflow, allowing you to add the necessary guard rails to the model, based on your domain expertise and regulatory requirements. Use the Exclusions option to select which traits to ignore when creating models from one or more data sources. Here are some use cases you can address with Trait Exclusion:
To view and analyze the model results, open the [PREGENERATED] Garage Rehab Enthusiasts - 1st party model.
Figure 8: Model results
Note: We have generated synthetic data in each of your accounts, so the reach you see may be low. In reality, you will be able to see Accuracy & Reach graphs that go up to 25MM.
Now that we have model results, let's create an algorithmic trait from it. We will generate a new algorithmic trait from the [PREGENERATED] Garage Rehab Enthusiasts - 1st party model.
Figure 9: Create algorithmic trait
Scroll down to the bottom of the Model Results page. Click Create New Trait with Model
Figure 10: Create new algorithmic trait - basic information
Name the trait Garage Rehab Algo Accu >85%
Select Algo-generated Data Source as the Data Source, where the new trait will be saved
Pick the Lab folder as a specific location for storing it
Click on Configuration to expand it
Figure 11: Create new algorithmic trait - select accuracy
Select 85% Accuracy
Click Save
Your algorithhmic trait has been created. It will take several hours for the trait to get populated. Next, you can map it to a destination and activate it.
Learn how to use AAM's new Trait Recommendation feature as a quick way of expanding reach
Oftentimes, publishers looking to sell their inventory turn to Look-alike Modeling to either:
This could be time-consuming and inefficient. Trait Recommendations is a powerful new data science feature in AAM. As traits are added to a segment in the Segment Builder workflow, a table with up to five Trait Recommendations will appear in real-time. The table surfaces recommendations based on a similarity score against the largest traits included in the segment, pulling from any dataset currently available for segmentation. By clicking on a recommended trait, a pop-up will display additional recommendations specific to that trait.
Trait Recommendations help marketers uncover traits that are highly relevant but may not be obvious for segment inclusion. By including these additional traits in audience segments, customers can expand their addressable audiences and improve downstream conversion rates and ROI of segments. As soon as you save the segment rule, the audience is created and can be activated. You lose the more granular result details and activation controls (e.g. Accuracy vs. Reach graph), which are present in Look-alike Modeling, but you gain audience creation and activation speed as well as unlimited use of the feature.
Trait Recommendations can also be used as an exploratory tool in order to see what other traits correlate to your baseline and thus come up with new ways of describing the audience.
Let's review the Trait Recommendations for the Garage Rehab Enthusiasts trait and create a brand new segment with expanded reach.
Figure 12: Segment Builder - create new segment
Click Audience Data -> Segments
Click Add New
Figure 13: Segment Builder - add basic info
Name the new segment Garage Rehab Enthusiasts from Recommendations
Select any folder (e.g. the top level All Segments folder)
Click on Traits to expand it
Figure 14: Segment Builder - add baseline trait
Search for the Garage Rehab Enthusiasts trait
Click Add Trait to include it in the segment
Figure 15: Segment Builder - view recommendations
Click on the + sign to include any of the recommended traits in the segment
Click any of the trait recommendation names to get additional info
Figure 16: Segment Builder - view recommendations pop-up
From the pop-up window, view the second level trait recommendations and add a few more to the segment rule by clicking on the + sign.
Note: If you want to supress recommendations from a given data source, click on the x sign under Data Source.
Note: If your segment rule consists of more than one trait and you want to view recommendations for each individual trait in the segment, you can click on the trait name in the segment rule and a pop-up screen will surface those.
Figure 17: Segment Builder - segment size estimator
Click Calculate Estimates to get an estimate of the population size for your newly-created segment
Click Save
In a nutshell, the main differences between Look-alike Modeling and Trait Recommendations are the following:
Trait recommendations is a quick way to get insights on other traits which are similar to the ones you are using in a segment. Use Trait Recommendations instead of Look-alike Modeling if you are looking for a quick way to:
You are the Digital Marketing Manager at BuildIt and your goal is to broaden the Garage Rehab Enthusiasts audience, then target these new users on other sites. Creating a look-alike model against your 1st party data is great, but the pool of new users that you will be able to reach will hit its limit eventually.
The power of Audience Manager's Look-alike Modeling gets unleashed when you seek to expand your baseline audience against a quality, brand new set of users from 2nd and 3rd party data sources. To make it easy to get that data, Audience Manager provides you with the Audience Marketplace feature. In Audience Marketplace, data sellers list their data feeds and you can choose which you'd like to use by subscribing. Reporting tools let you track feed usage and the overlap between your traits and those in a subscribed data feed.
Now let's see how we can choose 3rd party data and use it for modeling.
Navigate to Audience Marketplace and browse around.
Figure 18: Audience Marketplace
Note: Since this is a beta environment, you can "subscribe" for as many feeds as you wish at no cost.
For this exercise, we will work with two particular data feeds: DataFunnel and DemographicIQ. Your accounts have already been subscribed to them.
Figure 19: Audience Marketplace - pre-subscribed feeds
There are so many feeds! When searching to subscribe to a 3rd party data feed, however, which one should I choose?
The Audience Marketplace stats for each data feed come in handy. Consider data feeds, which have a good number of unique users, and at least some overlap with your data (so your algo model can run successfully). Overlap & unique user counts are calculated for free, but you need to explicitly toggle Segments & Overlap for the feeds of interest (option available once you click on individual data feeds).
Figure 20: Audience Marketplace - overlaps explained
Feeds can be billed at a flat rate or by CPM. If the feed uses CPM pricing, you can enable Modeling at no cost, but once you decide to activate the data, you will need to pay according to your usage.
Figure 21: Audience Marketplace - CPM pricing
If the feed uses Flat Fee pricing, that will be the monthly cost that you need to pay.
Figure 22: Audience Lab - Flat Fee pricing
Once you have subscribed to 2nd or 3rd party data feeds, these data feeds will automatically appear in the list of available data sources for modeling. You can then repeat the steps outlined in Exercise 2.1 to build a model against this data.
Learn how to run active A/B tests on model results via Audience Lab
Now that you have been able to build various models with your own data and with various 3rd party data, let's see how we can measure which models and data providers give you the best outcome. Analyzing the model results may give you some insights, but the best way to measure model performance is by running an active online test.
For this exercise, we will use an Audience Manager native feature called Audience Lab.
Audience Lab allows customers to run A/B testing on their segments against DSPs of their choice in order to measure performance (Click Through Rate - CTR).
For us, it is interesting to know which of the two 3rd party data feeds that we just "purchased" gives us better conversion rate and ROI. For this, we will measure the performance of the audiences created by two look-alike models - using the same Garage Rehab Enthusiasts baseline, but using different 3rd party data.
We will set up two Audience Lab tests in order to measure efficacy of your models. Follow the steps below to set up a test group for your DataFunnel model results. You will then repeat the same steps and set up one more test group for your DemographicIQ model results.
To save time we have:
Figure 23: Audience Lab - drop down
Click Audience Data -> Audience Lab
Figure 24: Audience Lab - create new test group
Click Create New Test Group
Figure 25: Audience Lab - add basic info
Name the Test Group DataFunnel Test
Click Choose Base Segment, then search for the [PREGENERATED] DataFunnel - Accu > 85% segment
When found, click on it
Click Choose Segment
Figure 26: Audience Lab - next
Click Next
Figure 27: Audience Lab - allocate percentages
For Test Segment 1, set the percentage as 10% and for Test Segment 2, set the percentage as 90%
Click Next
Figure 28: Audience Lab - select conversion trait
Click the drop down on Add a Conversion Trait to add [PREGENERATED] My Conversion Trait
Click Next
Figure 29: Audience Lab - add destinations
Under Destinations, select Ad Cloud from the drop down. Drag and drop Test Segment 2 to it. Enter 1 as the mapping value.
Click Next
Figure 30: Audience Lab - finalize test
Click Finalize Group
Figure 31: Audience Lab - comparing CTR results
Aggregate and Trend Reporting on conversion rates and total conversions are also available as you open each test.
Figure 32: Audience Lab - aggregate and trend reports
Learn how to select optimal accuracy when creating an algo trait
Now that you have learned all the key components for building Look-alike models and measuring results, can you try to design a test that measures CTR at different accuracy bands from one of your models?
For example, let's try to measure the performance for three groups of look-alike users:
Hint: Create mutually-exclusive buckets at segment level.
Figure 33: Audience Lab - segments from different accuracy bands
Now that you have learned how to build traits, segments, and models via the Audience Manager UI, you should try to put your new skills in action. Have fun and go ahead - build new workflows to bring better monetization for your company!
As an added bonus, we have included a script that you could use in the future should you decide to automate the creation of traits, segments, and models for your account via the Audience Manager's APIs.
Provisioning script for the summit test site
Thank you for participating! Please rate this lab in the Summit 2019 app!