Optimizing Meta Ads: Understanding Machine Learning and Data Signals

Very few Meta ad buyers actually know how the machine learning and look-a-like models work within Facebook. Here's a simple explanation that might change the way you buy ads.

Imagine you are doing a 1000 piece puzzle. You might start by sorting the pieces by color. Or you might separate the edge pieces from the center pieces. These are sorting algorithms, albeit very simple ones.

Meta's machine learning system is running similar sorting algorithms, but it does so in an infinitely more complex way.

▶️ Instead of sorting 1000 puzzle pieces, Meta sorts 3.5 Billion total users

▶️ Instead of sorting by color, it can sort on unlimited parameters across 20 years of collected data

Meta's system wants to figure out which of its 3.5B users are most similar to the small sample of past converters it has trained its learning model against for you specific ad.

To make this process work optimally, you want to:

✅ Send high signal data back to Meta via an advanced CAPI server-side data connection (like Popsixle)

✅ Send 50+ conversions per ad set per week so Meta has enough data volume to build an accurate learning model

✅ Use broad audiences so Meta has large pools of candidates to choose from

✅ Use diverse looking creatives and placement-optimized creatives so that your ads look great no matter where Meta serves its ad impressions

To get these powerful systems to turn your ad dollars into revenue, ad buyers need to stop assuming machines think like marketers.

Instead, we need to understand how the machines think, particularly when it comes to sending conversion event data that these systems learn from and optimize against.

Data is one of several key components to success, and its the one that most people neglect to prioritize.

Drop a comment or DM if you need help upgrading your data connection or fixing a dataset/pixel problem on your site.