The Hidden AI Layer Driving Advantage+ Performance

.

The Hidden AI Layer Driving Advantage+ Performance

Every time someone opens Facebook or Instagram, Meta has just a few milliseconds to decide which advertisement should appear. That decision is no longer powered by a single algorithm. Instead, it passes through three AI systems Andromeda, Lattice, and GEM, that work together to retrieve, rank, and continuously improve ad delivery.

Together, these models have replaced hundreds of independent machine learning systems with a unified AI architecture that learns across Facebook, Instagram, Reels, Stories, and Threads. The result is faster retrieval, smarter ranking, and more personalized advertising at a scale that would have been impossible with Meta's previous infrastructure.

Why Meta Rebuilt Its Advertising Stack

Meta's earlier advertising system relied on hundreds of separate models, each designed for a specific platform surface or campaign objective. A model trained for Feed couldn't share what it learned with Reels, and improvements made for Instagram rarely benefited Facebook. As the number of advertisers and active ads grew into the tens of millions, this fragmented approach became increasingly difficult to scale.

The system also depended heavily on manually engineered features and rule-based logic, making it slower to adapt to changing user behaviour. Maintaining separate infrastructure for every objective increased computational cost, delayed updates, and limited how quickly Meta could improve ad performance.

Rather than simply building a larger model, Meta redesigned the entire advertising pipeline around three specialised AI systems, each responsible for a different stage of the decision-making process.

1. Andromeda: Finding the Right Ads

Before Meta can choose the best ad, it first has to narrow millions of active ads to a shortlist of the most relevant candidates. That's Andromeda's job.

Whenever someone opens Facebook or Instagram, Andromeda retrieves a few thousand likely ads using a deep neural network that learns behavioural patterns directly from data instead of relying on fixed audience segments or manually defined rules.

Powered by Meta's MTIA chips and NVIDIA's Grace Hopper platform, it compares millions of user-and-ad combinations simultaneously while generating latent signals in real time. Rather than using a fixed set of user traits, it builds a fresh understanding of each person's behaviour and intent for every ad request.

To make retrieval even faster, Andromeda organises ads into a hierarchical tree, grouping similar ads together and eliminating entire branches of irrelevant candidates before narrowing down to the best matches. It's like finding a book by first locating the right section of a library instead of searching every shelf.

The result is a far more capable retrieval system. Meta says Andromeda supports models that are 10,000× more complex, improves ad recall by 6%, and increases ad quality by 8%.

2. Lattice: Deciding the Winning Ad

Finding relevant ads is only half the challenge. Thousands of strong candidates still remain, and only one can occupy the available advertising slot.

This is where Lattice takes over.

Lattice is Meta's unified ranking model, replacing hundreds of surface-specific ranking systems with a single architecture that works across Feed, Reels, Stories, and other Meta experiences. Because every platform now learns from the same model, insights gained from one surface can immediately improve ranking decisions on another.

For every shortlisted advertisement, Lattice evaluates dozens of signals, including predicted engagement, conversion probability, advertiser bid, placement suitability, and the overall value of showing that ad to a specific person at a specific moment. Rather than simply maximising clicks, it balances advertiser performance with long-term user experience.

Behind the scenes, Lattice operates at enormous scale, using trillions of parameters trained on hundreds of billions of interactions across Meta's platforms. This shared learning allows it to recognise patterns that isolated models would never have discovered.

The result is a ranking system that delivers ads that are not only more relevant but also more likely to drive meaningful business outcomes. Meta reports improvements of around 12% in ad quality and up to 6% higher conversions after introducing Lattice.

3. GEM: The Intelligence Behind the Models

If Andromeda retrieves ads and Lattice ranks them, GEM makes both systems smarter.

Unlike the other two, GEM doesn't participate directly in real-time ad delivery. Instead, it acts as Meta's large-scale foundation model for recommendations, learning behavioural patterns across the entire platform—including Feed, Reels, Stories, and organic interactions—before transferring that knowledge to Andromeda and Lattice through a process known as knowledge distillation.

Because GEM learns from billions of user interactions rather than advertisements alone, it can identify emerging interests and behavioural trends much earlier than traditional advertising models. Those insights are continuously distilled into the smaller, faster models responsible for real-time retrieval and ranking.

In effect, GEM performs the heavy learning offline so Andromeda and Lattice can make faster and more accurate decisions when an ad request arrives.

During its initial rollout on Reels, Meta reported up to a 5% increase in ad conversions, demonstrating the value of applying foundation-model learning to advertising.

4. How the Three Systems Work Together

Although each model has a distinct role, they operate as a single pipeline.

Imagine a sportswear brand promoting a new running shoe across Reels, Feed, and Stories.

When someone who has recently engaged with marathon training content opens Instagram, Andromeda immediately narrows millions of advertisements to a few thousand likely matches, including two of the brand's video creatives.

Lattice then compares those shortlisted ads using predicted conversion, advertiser bid, placement suitability, and overall user value before selecting the strongest creative for that specific Reels impression.

Behind both decisions is GEM, which has already learned from millions of similar user journeys. Rather than making predictions from scratch, it continuously improves the retrieval and ranking models with richer behavioural understanding.

Once the user watches, skips, clicks, or converts, that interaction becomes another learning signal that eventually feeds back into GEM, allowing future retrieval and ranking decisions to become even more accurate.

Fig4.1 Andromeda, Lattice, and GEM: How It All Connects

What This Means for Advertisers

Meta's AI now handles much of the work involved in retrieving, ranking, and optimising ads. As a result, advertisers can spend less time on manual campaign adjustments and more time on the quality of their inputs.

To get the most from these systems, advertisers should focus on:

• Creating varied, high-quality creatives.
• Maintaining accurate conversion tracking.
• Giving campaigns enough time and budget to learn.
• Using automated features where they make sense.

While AI can make faster optimisation decisions, it still relies on clear business goals, strong creative, and reliable data. Those are areas where human judgement continues to play an important role.

As Meta's advertising systems continue to evolve, success is likely to come from combining AI-driven automation with thoughtful planning and high-quality campaign inputs.