Andromeda, GEM and Lattice: How Meta Now Decides Which Ad to Show
Meta's ad delivery now runs on three AI systems: Andromeda retrieves candidate ads, GEM predicts engagement and Lattice ranks and prices them. Here's how they fit together and what it means for your campaigns.
Key takeaways
- Andromeda is Meta's ad retrieval engine. It picks the shortlist of ads that compete for each impression, and its global rollout was completed in October 2025.
- GEM (Generative Ads Recommendation Model) predicts how likely each person is to engage with each candidate ad. Meta published a technical paper on it in November 2025.
- Lattice ranks the scored candidates and sets the auction price.
- The practical upshot: creative variety now does much of the targeting. Fewer, broader campaigns with many genuinely different creatives tend to work best.
If your Meta ads have behaved differently since 2025, with broad campaigns outperforming tightly targeted ones and creative fatigue setting in faster, the reason is a rebuilt delivery system. Meta now decides which ad to show using three AI systems working in sequence: Andromeda, GEM and Lattice.
You don’t need to understand the machine learning. You do need to understand what each stage rewards.
The three stages
| Stage | System | What it does |
|---|---|---|
| 1. Retrieval | Andromeda | Picks the shortlist of eligible ads that could compete for an impression |
| 2. Prediction | GEM | Predicts how likely this person is to engage with or convert from each candidate ad |
| 3. Ranking and auction | Lattice | Ranks the GEM-scored candidates and sets the price in the auction |
Andromeda: who gets considered at all
Andromeda is the retrieval engine. Every time someone opens Facebook or Instagram, Meta has to narrow millions of possible ads down to a manageable shortlist in milliseconds. Meta announced Andromeda in December 2024 as the engine behind Advantage+ automation, and its global rollout completed in October 2025.
Andromeda matches ads to people largely from signals in the ad itself (the visuals, text, format and topic) combined with what Meta knows about the person. That is why your creative now does so much of the targeting.
GEM: how likely is this person to act?
GEM, the Generative Ads Recommendation Model, scores each shortlisted ad. Meta began rolling it out in Q2 2025 and published a technical paper on November 10, 2025, claiming it is about four times more efficient than its predecessor (that is Meta’s own figure).
Lattice: ranking and price
Lattice takes the scored candidates, ranks them and determines the auction outcome. Better predicted performance can mean winning impressions at a lower cost.
What this means for your campaigns
1. Creative is the new targeting
If Andromeda retrieves ads based on what they show and say, each distinct creative effectively reaches a different pocket of people. One ad featuring a busy mother and another featuring a gym-goer will be retrieved for different audiences, even inside the same broad ad set.
2. Variety beats volume
Ten versions of the same video with different captions look like one concept. Meta is likely to treat near-identical variations as one idea. Aim for genuinely different concepts:
- Different hooks (problem, result, testimonial, how-it-works)
- Different formats (UGC video, static, carousel, founder-led)
- Different people and settings
- Different angles on the offer (price, quality, speed, trust)
Our Meta Ads creative tips for the Andromeda era go deeper.
3. Broader, simpler structures
Splitting budget across many narrowly targeted ad sets starves each one of data. Most accounts now do better with fewer campaigns, broad or Advantage+ audiences and more creatives per ad set. See how to structure Meta ad campaigns in 2026.
4. Feed it clean conversion data
Every stage optimises toward the events you send. Weak or duplicated tracking means the system learns the wrong thing. Set up the Conversions API alongside the Pixel and check your Event Match Quality.
5. Expect faster creative fatigue, and plan for it
When the system finds the people most likely to respond to a creative quickly, it can also exhaust them quickly. Budget for a steady flow of new concepts each month.
Related changes in 2025–2026
- October 8, 2025: Meta blocked the creation of new Advantage+ Shopping and App campaigns through Marketing API v24.0, moving to unified Advantage+ campaigns.
- January 2026: Meta removed more detailed-targeting interest categories.
See the full Meta Ads update history for context going back to iOS 14.5.
The bottom line
Andromeda, GEM and Lattice put Meta’s AI in charge of who sees your ads. Your job has shifted to giving it the right raw materials: many genuinely different creatives, clean conversion data, and a simple structure that lets it learn. Check your account against our Meta Ads audit.