AI Innovations in Meta's Ad Ranking Driving Advertiser Performance

At Meta, we harness the power of AI to drive better experiences for people, and better results for your business – not only with our advertising tools, but also on the behind the scenes in our ads infrastructure and ranking systems to make your ads more relevant, effective, and personalized. We’ve adopted innovative approaches to machine learning modeling in ads, by developing and utilizing cutting-edge technologies that combine the best of human insight with the speed and scale of artificial intelligence.

Here are four AI innovations in ads ranking that are driving advertiser performance:

1. Meta GEM: The Super Brain

Generative Ads Recommendation Model, GEM, is a powerful new machine learning model trained on thousands of GPUs to optimize results for ad products, delivering an improvement in ad performance. GEM enables the ads system to rapidly process large amounts of data to deliver highly relevant and personalized ads, and has enabled a paradigm shift for recommender systems. We recently rolled out GEM more broadly in our ads system following the success in improved ad conversion results in Meta Reels earlier this year.

Said differently: Imagine having a super brain that can read an entire library of books in seconds, understand the relationships between all the characters, remember every single detail, and connect the details into an understanding of the sequence of events a person goes through across all types of activities. That's what GEM does for Meta's ad system: catalogs, analyzes, and connects trillions of pieces of information, making it incredibly intelligent and effective. With GEM, Meta's recommendation system learns from an enormous amount of data, recognizes subtle patterns, and provides the most relevant ads to the right person at the right time with low latency.

Results: During the initial launch on Meta Reels this year, GEM has increased ad conversions by up to 5%¹ .

2. Meta Lattice: The Giant Library

Meta Lattice is our ad ranking architecture that allows us to generalize learnings across campaign objectives and surfaces in place of numerous, smaller ads models that have historically been optimized for individual objectives and surfaces. This is not only leading to increased efficiency as we operate fewer models, but also improving ad performance. These significantly larger models are able to learn more because they can represent steps that people take in their purchase journey across surfaces and campaigns. For a deeper understanding, read more here.

Said differently: Think of Meta Lattice as one giant library. In the past, we had many small libraries, each dedicated to a specific subject like history or art. Each library had its own set of books and librarians who were experts in that one subject. Now, with Meta Lattice, it's like we've combined all those small libraries into one library, connecting information from all subjects. It doesn't just have more books; the expert librarians can also learn from all the books it has and apply that knowledge to different subjects. This is much more efficient because we only maintain one library, and it can help people find better information faster, no matter what they're looking for.

In the world of ads, this means we can use one powerful system to improve how ads are shown to people, making them more relevant and effective, rather than relying on many smaller systems that each focus on just one type of ad product or placement destination.

Results: Meta Lattice has increased ad quality by almost 12% and increased ad conversions by up to 6%.

3. Meta Andromeda: The Personal Concierge

Meta Andromeda is an innovative end-to-end hardware, software, machine learning co-designed system introduced in 2024, with Meta Training and Inference Accelerator (MTIA) and NVIDIA Grace Hopper Superchip. This more efficient system enabled a 10,000x increase in the complexity of models used for ads retrieval, the first step in the ranking process where we narrow down a pool of tens of millions of ads to the few thousand we consider showing someone. The increase in model complexity enables running far more sophisticated prediction models to better personalize ads. And, as businesses upload more and more creatives to support a diversification strategy, Meta Andromeda works behind the scenes to power more complex models that allows Meta to pick the right creative to deliver more personalized ads that are relevant and interesting. For a deeper understanding, read more here.

Said differently: Imagine having a personal concierge who knows your tastes so well that they don’t just understand that you covet shoes, but that you like to wear red flip flops at the beach. Meta Andromeda learns your preferences, so Meta can show you ads that are more relevant and interesting.

Results: Meta Andromeda has led to an 8% increase in ads quality, increasing the impact of Advantage+ automation and creative GenAI tools for advertisers.

4. Sequence Learning: The Memory Game

Sequence Learning is an AI modeling technique that enables our ads systems to consider the sequence of actions a person takes before and after seeing an ad. This allows a deeper understanding of the purchase journeys people take, allowing us to deliver more personalized and relevant ads, at the right time. Sequence model learning techniques do this by more deeply understanding patterns around a conversion to better infer sequence of ads for optimizing future conversions for purchase. For a deeper understanding, read more here.

Said differently: For example, previously with traditional aggregated data models, if a user converted on one ski resort ad, they may continue to see other ski resort ads. With recent changes in our ads learning model, after purchasing a ski resort room a person would now see ads for ski equipment, lift tickets or ski luggage, providing more relevant ads personalized to the purchase journey.

Results: Since we adopted Sequence Model Learning last year, we’ve already seen a 3% increase in conversions based on testing within selected segments.