Andromeda is Meta's deep retrieval system for ad delivery, introduced in 2024, that scores billions of candidate ads against each impression in real time.
Meta's Andromeda is a deep retrieval system built to rank ads at a scale that legacy pipelines couldn't reach. Introduced in 2024, it evaluates billions of candidate ads against each impression decision in real time — a step-change from the roughly 10,000 ads a previous generation ranking system could score.
The mechanism works in two stages. First, Andromeda uses learned embeddings to do a broad sweep across the entire eligible ad inventory, collapsing billions of options into a ranked shortlist. Second, a more compute-intensive scoring pass refines that shortlist before final auction selection. The result is that the system surfaces semantically relevant ads rather than just ads with the highest bid — creative pattern and predicted relevance carry real weight in who wins the impression.
This matters because dynamic creative and ad creative variance now carry more weight than they did before. Andromeda's retrieval stage rewards advertisers who give the system creative surface area to match — multiple formats, angles, and messages — rather than a single locked creative. The Advantage+ suite sits on top of Andromeda: when you turn on Advantage+ Shopping or Advantage+ Audience, you are asking Andromeda's ranking layer to do the heavy lifting on which users to find and which creatives to show them.
In the current 2025–2026 delivery environment, Andromeda is why broad targeting often outperforms narrow interest stacks — the system treats tight constraints as signal penalties rather than precision tools. Practitioners building on the four-layer AI platform model recognize Andromeda as the delivery layer that connects creative signals to audience matching. For a practical account of how this changes campaign structure, the ai ad campaign automation explained post covers the implications in full.
Your job is to feed Andromeda good creative variance; it handles the matching.
Andromeda shifted Meta delivery from constrained audience matching toward broad creative-first selection. In practice, this means practitioners who over-target lose to broader campaigns with stronger creative variance — I've seen narrow 4-audience setups consistently underperform a single broad campaign once the creative set is solid. The system rewards creative diversity, not audience precision.
