Meta Ads Algorithm Explained: Andromeda, GEM & More

How Meta Decides Who Sees Your Ad:

Every time someone opens Instagram or Facebook, Meta picks their ads in under a second. A chain of systems does the work, and each one has a strange name: Lattice, Andromeda, GEM, Adaptive Ranking Model.

Each one fixed a specific weakness in how ads were chosen before. Know the chain and you know what Meta wants from your account.

The hiring picture

Think of one ad slot as one job opening. Every ad is an applicant.

1. Lattice merged many separate hiring teams into one team that shares what it learns.

2. Andromeda reads millions of applications and shortlists a few thousand.

3. The Adaptive Ranking Model interviews the shortlist and picks the winner.

4. GEM is the senior head who trains every interviewer.

5. Sequence learning is how Meta reads a person's full story instead of one snapshot.

Here is each one, oldest first.

Lattice, 2023: one shared brain

Before: Meta ran hundreds of separate models. Each one was built for its own goal and its own surface. What the Reels model learned never reached the Feed model.

Now: Meta folded many of them into fewer and larger models. They share learning across Feed, Stories and Reels and across goals like clicks, video views and conversions. Meta also built it for a world with less granular data after the privacy changes. Meta reported about 8 percent better ads quality in early Instagram results.

What it means for you: A new campaign does not start from zero. Meta carries learning across placements and goals, so split accounts waste that advantage.

Sequence learning, 2024: the story, not the snapshot

Before: Models leaned on static facts that engineers built by hand.

Now: Models read the order and timing of a person's actions. What they viewed, what they clicked, what they bought, and when. Meta says these histories now run to thousands of events.

What it means for you: Buyers move through a journey. Your ad meets them at one step of it.

Andromeda, late 2024: the shortlist

Before: The first stage had limited personalization. It ran on separate stages and many hand-written rules. Advertisers compensated by stacking interests, lookalikes and custom audiences.

Now: A far larger neural network narrows tens of millions of ads to a few thousand for each impression. Meta reports a 10,000x jump in model capacity, 6 percent better recall and 8 percent better ad quality on selected segments.

Meta's creative guidance explains the practical part. Its system groups ads that share key visual and thematic attributes and treats them as variations of one creative.

What it means for you: The shortlist rewards distinct ideas. Build different personas, different motivations and different first frames.

GEM, 2025: the head trainer

Before: Each ranking model learned from its own slice of data.

Now: GEM is Meta's largest ads foundation model. Meta trained it the way large language models are trained, on ads and organic activity across Facebook and Instagram. It passes what it learns down to the models that run live.

Meta's own example: what GEM sees in Instagram video ad engagement improves Facebook Feed predictions. Meta reported 5 percent more conversions on Instagram and 3 percent more on Facebook Feed in Q2 2025.

What it means for you: Signals travel between apps. Run Facebook and Instagram placements together.

Adaptive Ranking Model, 2026: effort where it counts

Before: One-size-fits-all. Every request got the same model effort because ads must be chosen in under a second.

Now: The system routes each request to the model size that fits the person's context and intent. High-intent moments get the heavy model. Meta says it scales to around a trillion parameters, and a Meta VP described it as a distilled version of GEM. It computes a person's signals once per request and shares them across all candidate ads, which makes long histories affordable.

It launched on Instagram in Q4 2025. Meta reports 3 percent more conversions and 5 percent more clicks for targeted users.

What it means for you: Meta spends its smartest model on your most likely buyers. It learns who they are from the events you report, so clean purchase data matters more than ever.

The full chain

Read the table top to bottom and the direction is clear. Meta moved from many small models and hand-set rules to a few large models that learn from data.

How advertising worked before

The old playbook ran on advertiser controls. Stack interests and lookalikes. Split ad sets by audience. Keep about six ads per ad set. Give each ad set roughly 50 optimization events a week to leave the learning phase.

Today many of those targeting inputs work as suggestions. The ad itself tells Meta who to show it to.

What to do with this

Do these five steps:

1. Consolidate ad sets so each one gets enough conversion events to learn.

2. Fill each ad set with distinct concepts. Different person, different motivation, different first frame.

3. Send clean purchase events with real values through Pixel and Conversions API.

4. Run Facebook and Instagram together.

5. Review monthly. Cut weak concepts and refill with new ones.

The result is an account that gives every stage of the chain something to work with.

Where the evidence ends

Every performance number above comes from Meta and covers selected segments or targeted users. No one outside Meta can verify these systems from the inside. The advice in the five steps is my reading of what Meta published.

I run media buying and creative strategy together for e-commerce brands on Meta. If you want this applied to your account, get in touch through the form on this site.

Sources

• Meta Adaptive Ranking Model, Engineering at Meta, Mar 31, 2026: https://engineering.fb.com/2026/03/31/ml-applications/meta-adaptive-ranking-model-bending-the-inference-scaling-curve-to-serve-llm-scale-models-for-ads/

• Meta's Generative Ads Model (GEM), Engineering at Meta, Nov 10, 2025: https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/

• Meta Andromeda, Engineering at Meta, Dec 2, 2024: https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/

• From User Sequences to Scaling Laws, Engineering at Meta, Aug 5, 2026: https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking/

• Meta Lattice, AI at Meta, May 11, 2023: https://ai.meta.com/blog/ai-ads-performance-efficiency-meta-lattice/