After extensive research and development, Xandr is offering a full CPA optimization solution for all line items. We know that advertisers care about users engaging with their brand through actions that lead to a return for an advertiser, like signing up for a service or making a purchase. Previously, traders had to manually optimize to a CPA goal. Xandr now has a machine learning offering that enables traders to achieve optimal CPAs with the click of a few buttons. It will be available to all clients next month, October 2019.
There are generally two different scenarios in which advertisers and traders need CPA optimization. The first is retargeting, where the goal is to reach users that have previously expressed interest in the brand. For example, a retailer advertising to loyalty customers that have made purchases in the past. The second is prospecting, in which the goal is to reach users that aren't a part of the marketer’s core audience but have an unrealized affinity for the brand. For example, a brand selling golf clubs may appeal to golfers that have never engaged with the brand but would if they were aware of the high value of golf equipment sold by the advertiser. The challenge with prospecting line items is discovering inventory and audiences that perform. The challenge with retargeting line items is reaching users and bidding the correct price.
When the Xandr team began working on a solution for CPA optimization, we recognized that these two different problems required two different solutions. We released a solution for retargeting to all clients in March of 2019. (You can learn more about the machine learning algorithms built to solve the retargeting use case here.)
Finding the Right Audiences
The Xandr optimization engine finds useful audiences to an advertiser by leveraging historical conversion data online. Specifically, the following two metrics are used:
- Historical conversion rate of inventory.
- Inventory that looks like inventory generating conversions based on audience overlap.
A line item optimizing to the purchase of sneakers would continue to serve ads on fashion.com because it has a high historical conversion rate. The line item would test showing ads on sports.com because many similar users from fashion.com, which has proven to attract users interested in the brand, go to the site. The line item would not show ads on tech.com because not only have users not converted there but tech.com does not see the same user traffic as other sites that draw interested audiences.
If the actual performance of look-alike inventory (in the above illustration, sports.com) leads to efficient conversions, the line item will continue spending on it. Likewise, the line item will stop spending on inventory that does not meet the goal.
Paying the Correct Price
To pay the correct price for inventory, our system computes two values. One is the predicted value of an impression based on a specific site online and the geographical region in which the impression is being served. The second is the predicted value of an impression based on advertiser frequency and recency. Based on the buyer’s goal and these two probabilities, we produce a bid price that will ensure the line item wins the most valuable impressions at the optimal price.
When setting a CPA goal in Xandr Invest, a trader must explicitly indicate whether she wants to optimize to both post-view and post-click conversions or just post-click conversions. In the case of optimizing to both post-view and post-click conversions, you must label the line item as either prospecting or retargeting so that our system can apply the correct set of algorithms.
Insights and Analytics
As is the case with all our optimization products, Xandr empowers buyers by sharing insights and data to monitor a line item and make decisions. A buyer can view their line item’s performance against its CPA goal as well as view other key performance indicators. Traders can also view the inventory that our system has labeled non-performant and has ceased spending on.
We will be releasing CPA prospecting to all Xandr clients next month, October 2019, as an Open Beta product. Watch our Release Notes page for the specific date. If you have any questions or ideas for optimization – please leave notes in the comments section!