Inside Meta’s AI-driven advertising system: How Andromeda and GEM work together
Together, these two systems handle the “What,” “Who,” and “When” of your advertising by splitting the work into retrieval and recommendation.
Andromeda: The Ads Retrieval Engine (The “Hand”)
Andromeda is Meta’s high-speed retrieval system. Its job is to scan tens of millions of active ads in milliseconds to find a shortlist of candidates for a specific user.
- Creative-First Matching: Instead of looking at your “Interests” or “Lookalikes,” Andromeda “reads” the pixels, video frames, and captions of your ad. It identifies themes (e.g., “fitness,” “parenting,” “luxury”) and matches them to a user’s real-time semantic behavior.
- Infrastructure: Powered by custom MTIA chips and NVIDIA hardware, it allows for a 10,000x increase in model complexity compared to 2024.
- The “Bouncer” Effect: If your creative doesn’t provide a clear signal, Andromeda may never retrieve it for the auction, effectively making it “invisible” regardless of your budget.
GEM: The Generative Ads Recommendation Model (The “Brain”)
GEM is the foundational AI model that handles the final ranking and sequencing. It determines which ad from Andromeda’s shortlist is most likely to drive a conversion at that exact moment.
- Sequence Modeling: GEM understands the “time” element of search and discovery. It tracks a user’s multi-step journey (e.g., from curiosity about a product to a final decision) and ensures they see the right ad in the right order.
- Efficiency: Meta reports that GEM is 4x more efficient at driving performance gains than previous ranking models because it evaluates the entire customer journey rather than isolated ad interactions.
What Changed for Advertisers in 2026
The combination of these systems has fundamentally altered the “best practices” for media buying.
| Aspect | Old Meta Era (Pre-2025) | Andromeda & GEM Era (2026) |
| Targeting | Manual (Interests, Lookalikes) | Creative-Led (The ad is the targeting) |
| Account Structure | Granular (Many campaigns/ad sets) | Consolidated (Fewer campaigns, more room to learn) |
| Primary Metric | CTR / ROAS | CPMr (Cost per 1,000 Reach) & Incremental Lift |
| Control Lever | Bid & Audience Tweaks | Creative Diversity & Speed |
2026 Strategy: How to Align with the AI
1. Feed the “Creative Buffet”
Because Andromeda targets based on creative signals, you must provide a wide variety of “hooks.”
- Diversify Archetypes: Don’t just test ten versions of a UGC ad. Test one cinematic founder story, one lo-fi testimonial, one meme-style comparison, and one high-energy motion graphic.
- Semantic Clarity: Ensure your product and the problem it solves are visible in the first 3 seconds. GEM needs to “see” who this is for to categorize it correctly.
2. Simplify for Stability
The “Lattice” and GEM systems learn faster when data is consolidated.
- One Sales Campaign: Use a broad, full-funnel Advantage+ setup.
- No-Touch Windows: Refrain from making edits for at least 7 days or 50 conversions. Frequent tweaks reset the “Sequence Learning” and reset GEM’s understanding of your audience.
3. Treat Budget as a Signal
More budget allows Andromeda to scan larger clusters of users faster. If you are “Limited by Budget,” you are effectively starving the AI of the data it needs to find your winning pattern.