How The Optimization Algorithm Works

The usual journey of any media buyer is a cycle of running a campaign, making mistakes, learning from said mistakes, adjusting targeting options and repeating. This cycle can equate to thousands of dollars spent just to “probe” the traffic and collect whitelists.


The ultimate weakness of this cycle is how much time we waste on the testing process. Each action is supported by hours of checking reports and analyzing which traffic segments bring the best results.

This step by step optimization revolves around analyzing traffic segments which are responsible for campaign performance. An example of traffic segments for a campaign is for example “sites” as displayed below:


In the above diagram, we see a simplified form of the optimization process where the advertiser will pause Site X and Site Z for not delivering optimal results.

In the below diagram, however, we can see extended granularity by adding “Browser” type to each “site”. Instead of pausing all traffic from Site X, the advertiser can pause only the underperforming “Browsers” from Site X and keep “Browser A” which is profitable.


As per example above, we can see that in fact, even though Site X didn’t look profitable at first, it was profitable under a specific browser (combination Site X – Browser A). So, is Site Z in fact not profitable? To check that let’s add once another dimension – device. As you can see in the below diagram, with the additional “Device” dimension, certain traffic from Site Z can be profitable.


In conclusion, pausing Site X and Site Z would have been harmful for the campaign’s performance because there are in fact profitable segments when we go deeper.

What does that mean for you? Our Auto-Optimization feature will identify these granular segments responsible for delivering results.


That means our machine-learning model will decide to bid more on specific traffic segments for you (like Site Y – Browser E – Device A), and will drop bids once it’s certain that specific combinations don’t deliver good results (for example - Site X, Browser B – Device B). As per this example you see only 3 levels of granularity, but our Auto-Optimization feature works on almost 20 dimensions!

  • Bundle/Site

  • Widget

  • Ad Space ID

  • ID

  • Ad Exchange ID

  • ISP Connection type

  • Region

  • City

  • Device Model

  • OS version

  • Browser version

  • Browser language

  • Hour

  • Country

  • Creative ID

Also, our algorithm is able to differentiate between traffic responsible for “visits”, “clicks”, or “conversions” so you can set the Auto Optimization feature to favor certain types of user action. If you still want to manually optimize campaigns, our Auto-Optimization feature can work in the background as you pause/adjust your bids.

It’s always learning, and the more data it will eat, the smarter it gets

Unfortunately, we are not fortune tellers (not yet). That means to make auto-optimization work, your campaign will need to gather minimum events needed to identify profitable segments of traffic. The first version of the optimization algorithm for your campaign is created after 100 successful events  (visits for CPV and iCTR, clicks for CPC and conversions for CPA optimization). Once your campaign reaches the minimum events, your campaign’s Auto-Optimization algorithm will be initialized with your specified goal. Your campaign will take its first optimization steps, but with each added impression – our optimization algorithm will get smarter and make better and more granular campaign adjustments.

Below, you can see a timeline of this process:


Day  2 – 100 events reached, first version of the algorithm appears.

Day 2-8 – campaign collects more and more events with different parameters, effectively increasing algorithm accuracy.

Day 8 onwards – Auto-Optimization is on full-steam, buying 80% of its inventory from the best-performing dimensions according to the algorithm.

Why will only 80% of the traffic be bought accordingly to the optimization algorithm? This is strictly because we know that traffic streams can surprise any buyer, disappearing placements, spikes of traffic can affect your ROI. Therefore the campaign will always buy 20% of its traffic from not verified traffic streams to possibly enrich it’s dataset and learn even more. It’s always learning, and the more data it will eat, the smarter it gets.


So, in the end, what we did was take the classic optimization process of any old media buyer and set it on hyperdrive by adding a few dozen layers of granularity and dimensions to the traffic we optimize.

If you’ve got DSP campaigns running now without Auto-Optimization, stop what you’re doing, log in to your account and turn it on for your campaigns ASAP! Then, take a seat and let our algorithms do your optimization work for you.