Andromeda vs GEM vs Lattice: How the Meta Ads Algorithm Works

Read enough about Meta ads and you will meet three names: Andromeda, GEM and Lattice. They are often used interchangeably, as if each were "the new algorithm". They are not. Each does a different job, and knowing which is which helps you understand why some changes in your account move results and others do not.

This article explains the stack in plain language. For a practical plan, read our complete guide to Meta Andromeda.

The journey of one ad impression

Imagine someone opens Instagram and scrolls to a spot where an ad can go. In a fraction of a second, Meta works through these steps:

1. Eligibility and index matching. Which ads can be shown to this person, based on each advertiser's targeting, budget, schedule and policy status?
2. Retrieval (Andromeda). From the eligible ads, which few thousand are most likely to be relevant?
3. Ranking (Lattice and related models). For each shortlisted ad, how likely is this person to click, convert or engage, and how good is the ad?
4. Auction. Combine the bid, the predicted action rate and ad quality into a total value. The highest value wins.
5. Learning. What happens next, whether the person ignores, clicks or buys, flows back as training data. GEM and the serving models learn from it.

Andromeda: the retrieval engine

Andromeda is the shortlisting stage. Meta introduced it in December 2024 and rebuilt it in October 2026. Its job is recall: making sure the ads most likely to work for a person are in the shortlist, out of tens of millions of candidates.

• What it looks at: signals about the person and rich representations of each ad, including what the creative is about.
• What it outputs: a shortlist of a few thousand ads for ranking.
• What Meta has reported: in 2024, +6% recall and +8% ads quality on selected segments; in October 2026, over a billion ads scored per second per GPU server and a 2.4% increase in Facebook conversions since 2025.
• What it responds to in your account: the variety of your creative (different ads reach different people), how broad your eligibility is, and the quality of your conversion data.

Read the full history in the Meta Andromeda update explained.

Lattice: the ranking architecture

Ranking takes the shortlist and makes precise predictions. Historically Meta ran many separate ranking models, each specialised for a goal, a surface or an ad type. Lattice, which Meta introduced in 2023, is an architecture that consolidates many of those models into larger shared ones that learn across objectives and surfaces.

• What it looks at: detailed features about the person, the ad and the context, plus what similar people did with similar ads.
• What it outputs: predictions such as the chance of a click or conversion, and quality estimates, for each shortlisted ad.
• Why consolidation matters: a model that learns across many goals and surfaces can use patterns from one to improve another, which helps campaigns with less data.
• What it responds to in your account: your optimisation event, the clarity of your conversion signal and the post-click experience that makes conversions happen.

GEM: the foundation model that teaches the others

GEM, the Generative Ads Recommendation Model, is different again. Meta described it on 10 November 2025 as the largest foundation model for recommendation systems in the industry, built in the style of large language models and trained on thousands of GPUs.

GEM is too large to serve every ad request directly. Instead, it passes what it learns to hundreds of smaller models that do the serving, through techniques such as knowledge distillation. Meta reported that this transfer is about twice as effective as standard distillation, and that GEM lifted ad conversions by 5% on Instagram and 3% on Facebook Feed in Q2 2025.

• What it looks at: huge amounts of history across people, ads and surfaces, including sequences of behaviour over time.
• What it outputs: knowledge that improves the serving models, rather than ad decisions itself.
• What it responds to in your account: indirectly, everything. Your results become part of what GEM learns from.

Side by side

• Andromeda: retrieval. Question it answers: which ads are worth considering for this person? Introduced December 2024, rebuilt October 2026.
• Lattice: ranking architecture. Question it answers: how likely is this person to take the action this advertiser wants? Introduced 2023.
• GEM: foundation model. Question it answers: what general patterns in people and ads can make every other model better? Detailed November 2025.
• Advantage+: not a model at all, but the family of automation settings in Ads Manager (audience, placements, budget, creative) that give these models more room to work.

What this means for your campaigns

Creative feeds retrieval

Because Andromeda reads what your ads are about, varied creative gives it more ways to match you to different people. That is the reason behind Meta's advice on creative diversification.

Conversion data feeds ranking and GEM

Ranking predicts the action you optimise for. If you optimise for a shallow event, you get people who do shallow things. Send events closer to revenue, such as qualified leads or purchases, through the Conversions API. For B2B teams, see Meta ads for B2B SaaS.

Structure decides how much data each decision gets

Every ad set is a separate learning problem. Many small ad sets mean thin data everywhere. Consolidate where you can; see Meta ads account structure.

The post-click experience still counts

Ranking learns which ads lead to conversions. A slow or confusing landing page lowers those conversions, which lowers predicted value in future auctions. Our SaaS website design guide and conversion rate optimization guide cover the fixes.

Frequently asked questions