Predictive audiences are extremely powerful predictors of a consumer action – online purchase, offline purchase, coupon download, and so forth. They are built using a “seed” of consumers who have taken the required action, and our model examines over 600 attributes within that seed. We then look for these attributes in other Oath users to find additional consumers who are likely to take the same action in the near future.
Predictive audiences offer the following advantages:
- Performance - predictive audiences outperform generic audience modeling audiences by 4-8x. These audiences are integrated directly into our bidder’s algorithm.
- Flexibility - advertisers can choose their own confidence level to best-fit for performance and scale, and select from a variety of data to build their seed audience.
How Predictive Audiences are Built¶
DSP’s predictive audiences are built on a new machine learning platform. We use the breadth and depth of data from Oath’s web properties, mobile apps, ad serving, demographics, and geographic signals into our advanced machine learning model. Elements such as recency and frequency are also incorporated.
Our machine learning engine finds the top correlated data points which predict conversion. We then score each user against these “predicted” activities which lead to conversion. Users who score positively are assigned to one of 8 tiers, with each tier corresponding to a percentage probability that the user will convert.
During a performance campaign using predictive audiences, the DSP will act as follows:
- A bid request comes in from user 123.
- DSP checks whether user 123 is part of one of the predictive audience tiers activated for a line item within that campaign.
- If user 123 has been put in one of the 8 tiers as “likely to convert”, the bidder is instructed to place this user at higher priority for auction.
- DSP submits a high bid for that impression.
Even for a user within a predictive audience audience tier, the bid is still limited to the maximum bid and frequency cap settings for that line item.
When Audiences become Available¶
When a predictive audience becomes available depends on a number of factors.
- The audience will not start scoring until 100 hits have been reached. This is a requirement for the audience to start building.
- The maturity of the pixel or segment (see below).
Pixel-based audiences are based on pixels that measure either conversion or re-targeting events.
- If the predictive audience is based on a new pixel (created less than 7 days ago from the date on which the audience is created), it can take up to 7 days for the predictive audience machine learning system to build out the eight tiers and make them available for targeting.
- If the predictive audience is based on a mature pixel (created 7 days or more before the predictive audience), then it can take up to 48 hours for the predictive audience machine learning system to build out the eight tiers and make them available for targeting.
For segment-based audiences (based on purchase receipt, 3rd party data, email, or device ID), have the following constraints in effect:
- If the predictive audience is based on a new segment (created less than 7 days ago from the date on which the audience is created), it can take up to 7 days for the predictive audience machine learning system to build out the eight tiers and make them available for targeting.
- If the predictive audience is based on a mature segment (created 7 days or more before the predictive audience), then it can take up to 72 hours for the predictive audience machine learning system to build out the eight tiers and make them available for targeting.
How Users are Assigned to Tiers¶
Each user within the Oath universe is scored individually against hundreds of criteria analyzed when building predictive audiences. (Users are not grouped and scored as a group.)
After scoring, users are grouped into 8 tiers based on their probability to convert (take the action desired by the advertiser). The division into 8 tiers to simplifies the processing required of the DSP bidder.
- Tier 1 has the highest probability to convert, but the lowest scale.
- As you move through tiers 2-8 the scale increases but the probability drops.
Predictive audiences are re-scored daily.
Lookalike vs Predictive Audiences¶
The following table should help clarify any confusion about the differences between lookalike and predictive audiences:
|Objective||Find users who are similar to the seed population.||Estimate likelihood of a single user to convert in the future.|
|Time||Time of conversion is not used.||Time of conversion is used.|
|Output||A single set of users most similar to the seed audience.||Multiple tiers of users based on their probability of conversion.|
|Integrated with AdLearn||No||Yes|
|Best use case||
|Advertiser Type||Brand advertisers||Performance advertisers|
Build a Predictive Audience¶
A seed audience is the audience on which a predictive audience is based. Create a custom seed audience or third-party seed audience as follows:
- Navigate to Advertisers and select the advertiser for whom you want to create the audience.
- Navigate to Audiences > Audience Builder > Predictive Audience.
- Select the seed data you want to use to build your seed audience. You can select from one or more of the following seed audience types: pixels (see Pixels Overview), Device ID Audiences, Email Address Audiences, Purchase Receipt Audiences, and third-party segments.
The minimum size for a seed audience is 100. There is no maximum.
- Select the device type you want to target: desktop or mobile (and for mobile, app or web).
- Select the country.
- Enter a name for the predictive audience.
- Select Build Audience to save the predictive audience.
- After one day, you should be able to see the volume and performance of the 8 tiers for the predictive audience you created. Once populated, you can roll over an individual tier to see both the number of users it contains, as well as the likelihood that an individual in that tier will convert (A Performance value of 0.5, for example, indicates that we estimate an individual in the tier is 50% likely to convert.)
- After two days, the audience should be available for Audience.