An introduction to Facebook Ads' Machine Learning

We hear a lot about Google’s machine learning and new releases, but it is also important to acknowledge advancements in Facebook’s machine learning. Previously, it was drilled into us the importance of controlling every aspect of a campaign with lots of scientific tests and strict budget controls. But now we have been moving more and more towards utilising Facebook’s machine learning and giving the algorithm space to learn and optimise. We’ve found really positive results from relaxing audience targeting and strict budgets. Read on for the machine learning highlights…

Campaign Budget Optimisation

Previously, it was recommended that ad sets should be split out very granularly to test different audiences against each other with ad set budgets and lots of ads to test. Facebook are now moving away from ad set budgets and it’s all about Campaign Budget Optimisation (CBO). CBO uses a campaign budget to allocate spend to each ad set depending on the performance and reach of each ad set. Testing audiences with CBO means that you can see which audiences have the most potential using Facebook’s machine learning. We’ve seen good results using CBO and Facebook can use its many data points to allocate more or less spend to specific audiences to stabilise CPAs.

Automatic Placements

Running ads on all placements means that the Facebook delivery system has more flexibility to get more and better results. Auto placements are typically the most efficient use of budget to help control costs. Often the placements that wouldn’t be picked manually can bring in some very low CPA results. Facebook aims to get the lowest cost overall so if you look into individual placements and see one with a much higher CPA, don’t panic! Facebook looks at all available opportunities across placements and selects the least expensive ones despite the average cost for that individual placements. When using auto placements, it is also important to use Asset Customisation and tailor creative to the size of the creative on that placement, e.g. always use 9:16 assets for stories, rather than a square with Facebook’s default text underneath.

Learning Phase

Bringing together CBO and Auto Placements, Facebook’s machine learning all ties together with the learning phase. But what is it all about? Each time an ad is shown, Facebook learns about the best people to target, times of day, placements and creatives to show. This learning phase allows Facebook to explore these different combinations, and higher CPAs should be expected in this period. Around 50 optimisation events are needed in a week to exit the learning phase and no more than 20% of the budget should be spent on the learning phase. That means that you should allow Facebook to reach the full extent of its learning before making big decisions. It is important to set a realistic budget based on your current average CPA and how much budget you will need based on that to achieve 50 conversions in a week.

The important thing to note about Facebook’s machine learning is that it needs time. Optimisation is a process and for it to complete, Facebook needs to learn, much like Google. After implementing the best of machine learning and letting Facebook do its thing, you should then review and see what is working and what is not.

Have you had any experiences with Facebook’s machine learning? Did you find performance to improve after implementing?