Predictive Audiences Overview
Predictive Audiences helps you classify an unknown audience into distinct personas, in real-time, using advanced data science techniques.
This article contains product documentation meant to guide you through the setup and usage of this feature. Nothing contained herein is legal advice. Please consult your own legal counsel for legal guidance.
In a marketing context, a persona is an audience segment defined by visitors, users, or potential buyers, who share a specific set of traits, such as demographics, browsing habits, shopping history, etc.
Predictive Audiences models take this concept a step further, by enabling you to use Audience Manager's machine learning capabilities to classify unknown audiences into distinct personas. Audience Manager helps you achieve this by calculating the propensity of your unknown first-party audience for a set of known first-party audiences.
When you create a Predictive Audiences model, the first step is choosing the baseline traits or segments that you want your target audience to be classified by. These traits or segments will define your personas.
During the evaluation phase, the model creates a new Predictive Audiences segment for each trait or segment that you defined as baseline. The next time Audience Manager sees a visitor from your target audience who is not classified for a persona (did not qualify for any of your baseline traits or segments), the Predictive Audiences model will determine which of the predictive segments the visitor should belong to, and add the visitor to that segment.
You can identify the predictive segments created by the model, in the Segments page. Each Predictive Audiences model has its own folder under the Predictive Audiences folder, and you can see each model's segments by clicking the model folder.
To help you better understand how and when you could use Predictive Audiences, here are a few use cases that Audience Manager customers can solve by using this feature.
Use Case #1
As a marketer in an e-commerce company, I want to classify all my web and mobile visitors into various brand affinity categories, so that I can personalize their user experience.
Use Case #2
As a marketer in a media company, I want to classify my unauthenticated web and mobile visitors by favorite genres, so that I can suggest to them personalized content across all channels.
Use Case #3
As an advertiser for an airline company, I want to make sure I classify my audience based on their interest in travel destinations, so that I can advertise to them in real time, within a short retargeting window.
Use Case #4
As an advertiser, I want to classify my first-party audience in real time, so that I can react quickly to trending news.
Use Case #5
As a marketer, I want to predict which customer journey phase my website visitors are in, such as discovery, engagement, purchase or retention, so that I can target them accordingly.
Use Case #6
As a media company, I want to categorize my audience, so that I can sell my advertising space at premium pricing, while offering my visitors relevant ads.
How Predictive Audiences Models Work
When you create a Predictive Audiences model, you go through three steps:
- First, you select a minimum of two traits or two segments that will define your personas.
- Then, you choose a trait or segment that defines the target audience that you want to classify.
- Finally, you choose a name for the model, a data source that will store the predictive segments, and a Profile Merge Rule for the model.
Selection Criteria for Personas
You can choose any of your first-party traits or segments to define your personas. However, for optimal results, here's a set of recommended best practices:
- Choose your persona traits or segments so that each persona has at least a few hundred device IDs .
- We recommend choosing traits or simple segments for your personas, consisting of 1 to 3 traits.
- Choose baseline traits or segments which have minimal overlap.
- Make sure you are capturing granular traits across your digital properties.
Selection Criteria for Target Audience
Similar to persona selection, you should choose your trait or segment that defines your target audience in such way that it has real time users with rich sets of traits, for classification into the right persona.
When selecting the target audience, analyze your your use case and decide which types of IDs you want to classify: device IDs or cross-device IDs. The Profile Merge Rule that you select when creating the model defines the data that will be used to place each user into the predictive segments.
As a best practice, we recommend choosing a Profile Merge Rule that has the same configuration as your target audience Profile Merge Rule, or one that includes the profile type (device profile or authenticated profile) of your target audience.
Predictive Audiences Model Training Phase
Before the algorithm can classify your first-party audience into the right personas, it needs to train itself on your data.
For each persona that you define, the algorithm analyzes its respective audience and evaluates any real time and/or onboarded trait activity for its users in the last 30 days. This step takes place once every 24 hours, to account for changes in your first-party audience.
Predictive Audiences Model Classification Phase
When a visitor who is part of the target audience is seen in real time, the model evaluates whether the visitor is part of the defined personas. For every visitor that does not belong to any of the personas, the model assigns a persona qualification score.
While evaluating first-party audiences and assigning scores, the model uses the default Profile Merge Rule defined in your account. Finally, the visitor gets classified into the persona for which they received the highest score.
Considerations and Limitations
Read through this section carefully before moving on to the implementation phase.
When configuring your Predictive Audiences models, keep in mind the following considerations and limitations:
- You can create up to 10 Predictive Audiences models.
- For each model, you can choose up to 50 base traits / segments.
- Second and third-party data are not currently supported in Predictive Audiences.
- Audience classification is done only for real time first-party audiences. Onboarded first-party audience classification may be supported in a future update.If you add a predictive trait to a regular segment, it becomes a predictive segment. Consequently, all associated profiles are unsegmented.Currently, predictive segments can be activated in real-time destinations only. The Total Segment Population and Addressable Audience of your predictive segments is displayed as 0, and Batch Outbound Data Transfers are not supported for Predictive Audiences. This behavior will change in a future update.
- Predictive Audiences performs audience classification based on your first party traits, from all your first-party data sources.
- Segment evaluation for Predictive Audiences uses the Profile Merge Rule that you choose during model creation. To learn more about Profile Merge Rules see the dedicated documentation .
- Some traits and segments are not supported as baselines or target audiences. Predictive Audiences models will fail to save when choosing one of the following as baselines or target audiences:
- Predictive traits and segments created with predictive traits;
- Adobe Experience Platform traits or segments;
- Algorithmic traits;
- Second and third-party traits.
Data Export Controls
Predictive segments created by Predictive Audiences models inherit the Data Export Controls from the following first-party data sources:
- The first-party data source that you choose when building the model.
- The first-party data sources of your target audience. Specifically, the data export controls of the traits or segments that make up your target audience.
- The Data Export Controls of the Profile Merge Rule that you selected for the model.
The newly created predictive traits and segments will have the same privacy restrictions as the union of the first-party data sources described above.
Traits that have additional restrictions that aren’t part of the Predictive Audiences segment privacy restrictions will be excluded from the training phase, and will not become influential for the model.
Profile Merge Rules
All predictive segments will be assigned the Profile Merge Rule that you selected when creating the model. The Profile Merge Rule that you choose is important for the following reasons:
- It defines which devices and/or authenticated profiles should be taken into account when the model analyzes the influential traits, at the time of classifying a user into a predictive segment.
- It governs which trait types (device level or cross-device level) should be used during the model training step and surfaced as influential traits. Predictive segments are subsets of your target audience.
- If the target audience is a segment, we recommend that you select the same Profile Merge Rule for the model as the one assigned to your target audience, or a Profile Merge Rule that includes the profile type of your target audience.
- If the target audience is a trait, we recommend that you select a Profile Merge Rule that can access the same type of data as the target audience trait (either device profile data or cross-device profile data).
Selecting a Profile Merge Rule that uses both device data and cross-device data maximizes the number of traits that could be used for model training and user classification into the predictive segments.