The Machine Learning Systems Behind Meta

Learn about machine learning and how it powers Meta's ad platform.

When Meta first began offering ads, its platform was no different than any other. You could buy banner ads that went all over the web, or you could buy banner ads that went into the right rail on Facebook. Within a few years, Meta better differentiated itself based on the company’s unique strength: world class engineering and cutting edge use of algorithms, artificial intelligence (AI), and machine learning (ML).

By 2018, Meta and Google dominated the digital media landscape, with about 40% of all media budgets going to Meta and 40% going to Google. Why was Meta so popular? What was Meta doing differently that made it a favorite place to try to turn media dollars into business growth?

Meta’s Machine Learning system was powerful, efficient, and best-in-class:

  • The ad auction

  • Automatic creative optimization

  • Automatic placements

  • Campaign budget optimization (CBO)

  • Look-a-like audiences (LaL)

  • Dynamic creative

  • Dynamic pixel-based audiences

With each one of these tools, you weren’t placing bets and gambling, but instead you were putting ad dollars into a learning engine that was being fed massive amounts of data and finding the balance of great performance and great efficiency.

Meta kept getting more budgets and kept producing great outcomes. Things grew to the point where Meta was receiving $100B+ in annual revenues and was a core component on every brand’s media plan, big or small.

Then Apple released iOS14.

In April of 2021, when iOS14 was released, the data being shared back to Apps stopped being automatically set to the most generous settings. Instead, iOS14 empowered users to choose which apps could capture their data and allowed users to opt-out of data sharing all together.

92% of iOS users decided to opt out of data sharing within Facebook and Instagram. This massively reduced the amount of data being fed into Meta’s ML models. And when you put an ML system on a data diet, it stops working with much efficacy.

The rollout and impact of these changes to data sharing happened in waves over a 12 month period. Because Meta is an identity driven platform, all of the data it receives gets matched back to individual users. At some point, Meta’s platform policy team made an extremely conservative decision: they decided that when data gets matched back to a user who had opted out of iOS14 data sharing, that data should get dropped (deleted!).

It didn’t matter what device the data came from, it mattered who it came from. This led to a massive drop in data that could be used. By Meta’s own estimates, 60% of all data coming from external websites via the website pixel was being dropped.

CAPI offers a build-your-own solution:

Meta had launched tools that enabled developers to build their own pixel and be in greater control of their data sharing with Meta. The Conversions API (CAPI) became the best way to work around Meta’s strict platform policies. Really it was a way of empowering brands to opt-in to share their own 1st party data. Why was that compliant? Because the data was from their customers, it was captured responsibly, and brands had permission to share it back to Meta based on their website terms and services.

CAPI connections grew in popularity between April 2021 and April 2022, providing countless brands with a real solution to the growing issue of dropped data and overall signal loss.

Where things stand today.

Today there’s two groups of brands: those who are using advanced CAPI connections to restore the data flow feeding Meta’s ML systems, and those who are not using advanced CAPI connections and have poor, erratic performance with their Meta ads.

Any data that falls within Meta’s terms of service is subject to strict platform policies: data that matches back to an opted-out user gets dropped. This is true for data captured by Meta’s website pixel and for CAPI solutions that Meta owns/maintains like the built-in Shopify CAPI connection.

Most people think of Meta’s Events Manager as the best line of sight into the black box of data being shared and used by Meta. However, Events Manager shows stats, scores, and rates of all the data that gets shared to Meta. It doesn’t tell which data is getting dropped, and which data is making its way through. So you get a deduplication score for all your data, even though 60% gets dropped; you get an Event Match Quality score for all your data, even though 60% gets dropped.

Who is impacted by Meta’s signal loss problem?

The worst part about Meta’s signal loss issue is that it disproportionately hurts smaller brands. Massive global brands (thing Coca-Cola, Nike, etc) are sending hundreds of millions of signals per month - if 60% of the data gets dropped, there’s still a massive scale of usable data that feeds the Meta ML system. If you’re a small brand and you send a few thousand purchase signal per month to Meta, you can't afford to have any of that data get dropped.

The clearest signs that your ad account is being impacted by signal loss is if you have any of these symptoms within your ad campaigns:

  • Erratic performance with huge swings day over day

  • Massive over-reporting and under-reporting of your purchases

  • Slow exit from the learning phase of campaigns

  • Overall poor performance

How to fix Meta’s signal loss problem.

To fix the signal loss problem at the core, brands need to create their own CAPI connection or work with a partner that creates it on their behalf. The provenance of the data (its source origin) is critically important, as the data must be certified 1st party data that follows all the appropriate rules and guidelines.

Connecting an Advanced CAPI connection has an immediate, positive impact. It restores the flow of data between your website and Meta’s ad platform. Brands using an Advanced CAPI connection should expect to see:

  • Stabilized ad performance

  • More accurate ad attribution

  • Rapid exit from the learning phase

  • Ad performance improvements of 20% or more

Popsixle was created to help solve Meta’s signal problem. We strive to create both the most advanced CAPI connection, as well as make it as easy as possible to use. Find out more if you qualify for a free Popsixle 14 day trial and get back to growing with Meta.


Similar posts

Keep up to date on the latest in adtech