By David Linder, Product Manager
One key goal of an ad monetization platform is optimizing yield for developers. To reach this goal, it’s important that every single ad request is filled with the highest-yielding ad available at any given time for any given user.
But optimizing yield is a complicated task — even for mobile ad mediation solutions that offer large amounts of demand. Technical limitations sometimes restrict data on ad performance, making it hard for mediation solutions to predict and serve the highest yielding ads in the queue.
Our mission is to help developers make the most of their ad monetization strategies, so we set out to overcome these challenges. The result is our Predictive Algorithm. In this post, we’ll explain the challenges of yield optimization, the strategy most mobile ad mediation solutions use to operate around these challenges and increase eCPM, and the reason why our Predictive Algorithm is helping developers take optimization to the next level.
The Challenges of Optimizing Yield
Determining which single ad has the highest yield requires access to each ad’s individual performance data. For server-side mediated ad formats like Offer Walls, this works out nicely; the ads are channeled through the ad mediation solution servers, which allows the algorithm to track granular data on an ad level and optimize delivery.
When it comes to client-side mediated ad formats, however, there’s a catch. Collecting data for ad formats like mobile Rewarded Video and Interstitials currently isn’t possible on an ad level, due to lack of interaction between the ad and the mobile ad mediation servers. Because of this, optimization only exists on the network level. In other words, after determining which ad network is, on average, likely to deliver the highest yield for a given request, the mobile ad mediation solution leaves it to the network to decide which ad to deliver.
The Waterfall Model and Its Limitations
The wide-spread approach to handling this lack of ad-level control is the waterfall model. Each time an ad will be shown to a user, the mobile ad mediation solution ensures that available ad networks are requested in a pre-determined order, which is based on each network’s historical average yield. This means that when a user is exposed to several ads, the network with the highest average will be requested over and over again (provided it can deliver the fills).
The problem is this: the later ads in this ad network’s sequence are likely to have a much lower yield than the network average on which the high ranking is based. More importantly, these ads are likely to have a lower yield than the top ads from other networks — even though the other networks might have a lower average ranking.
With the waterfall model, the ads being shown to users aren’t always the highest yielding ads available — sometimes far from it. Needless to say, this represents a substantial revenue loss for developers.
Raising eCPM with Fyber’s Predictive Algorithm
At Fyber (formerly SponsorPay), we have set out to solve this problem using an algorithmic approach that leverages the extensive data at our disposal as a mobile ad mediation platform. While we recognize that full yield transparency on the ad level just isn’t possible today for client-side mediated ad networks, we have seen that it is possible to get closer through mathematical modeling.
The model starts with the assumption that yield declines as the user progresses through the inventory of a given network. Using a continuous feed of traffic data for a dynamic approach, our solution creates unique models for each ad network in combination with each app that uses the network. This is because there are often substantial differences in app usage patterns and network yield patterns.
Finally, these models are applied in real time when the user requests an ad. At that moment, our solution combines the individual user history with the ad network models according to whichever specific app is in question. Then, we make a prediction of the yield of the next available ad in each ad network. Based on these yield predictions, we are able to deliver a list of ad networks ordered by the yield of their next ad in queue (rather than their average yield). Then, the client finds and delivers the first fill in the list.
This ability to locate the top yielding ad available for a particular user, at a particular point in time, has further boosted the performance of our mobile ad mediation solution. It has become an integral part of our optimization of mediated ad networks — and moving past the years-old waterfall model is only the beginning. We will continue to innovate, helping developers discover and execute ever more sophisticated ad monetization strategies.