Skip to content
Primores AI wiki

Meta Andromeda — The Ad-Retrieval Engine That Made Creative the Lever

Meta Andromeda

TL;DR: Andromeda is Meta’s machine-learning ad-retrieval engine — the first stage of its ads recommendation system, which selects a few thousand relevant ads from tens of millions of candidates before ranking and auction. Detailed by Meta in December 2024 and built to support the exponential creative volume that Advantage+ automation generates. Its practical consequence: the engine can discriminate among a vastly larger pool of creatives, which is why agencies now argue creative volume and diversity is the primary controllable lever.

What it is

Retrieval is the first step in Meta’s multi-stage ads recommendation system. Meta’s own description: this stage is “tasked with selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates” — which then pass downstream to ranking and the auction. Andromeda is the engine that does that retrieval.

It is not a targeting product, a campaign type, or a creative tool. It sits underneath all of them, deciding which ads are even eligible to be ranked for a given opportunity. Meta detailed it in a December 2, 2024 Engineering blog post; the system was already in production at that point (the post reports measured production gains), so “detailed/announced December 2024” is the accurate phrasing rather than a precise go-live date.

Why Meta built it

Meta frames Andromeda as a response to its own success with automation: “The continued positive momentum of Meta’s Advantage+ suite further increases the number of eligible ads through automation.” Advantage+ generates far more creative variants and far more eligible ads, so the retrieval stage had to scale to discriminate among them. Andromeda is explicitly built to “support exponential ad creatives growth from Advantage+ creative.”

In other words: automation upstream created a creative-volume explosion, and Andromeda is the retrieval engine sized to find the best ad inside that much larger pool.

The published numbers (with their qualifiers)

From Meta’s Engineering blog (these are Meta’s own figures — keep the qualifiers, or they mislead):

  • +6% recall improvement to the retrieval system, yielding +8% ad quality improvement — “on selected segments”, not a blanket gain.
  • 10,000× increase in model capacity/complexity vs. the retrieval system it replaced (delivered).
  • >100× improvement in feature-extraction latency and throughput.
  • >3× end-to-end inference QPS; 10× inference efficiency via “model elasticity.”
  • Runs on the NVIDIA Grace Hopper Superchip. (Meta’s own MTIA accelerator is named for future integration, not current.)
  • A ~1,000× further model-complexity increase is described as a future projection — not a delivered result.
  • On the advertiser-outcome side, the same blog reports “+7% increase in conversions” from image generation (Meta’s hedge: “we estimate”) and a “+22% increase in ROAS” when advertisers who hadn’t used Advantage+ creative turned its AI features on. Both are Meta self-reported and attribution-based (not incrementality-tested) — see glossary/advantage-plus-creative.

The strategic claim — and who actually makes it

There is a popular framing that Andromeda “commoditizes targeting” and makes creative the new lever. Be precise about attribution:

  • Meta’s own language stresses scaling retrieval to handle the “exponential growth of creatives” and eligible ads. Meta does not say “targeting is commoditized” or “creative is the new targeting.”
  • Agencies make the strategic leap. The inference — if the engine now discriminates among a far larger creative pool, then creative volume/diversity is the primary controllable input — is agency synthesis layered on top of Meta’s technical design, not a Meta claim.

So the honest statement is: Meta built an engine to pick the best ad out of a much bigger pool; practitioners infer that feeding it more genuinely different creatives is the lever that remains in the marketer’s hands. This is the engine-enforced, 2026 instantiation of glossary/creative-is-new-targeting.

⚠️ Separate Meta’s own numbers from vendor numbers. The +6%/+8%, +7%-conversions, and +22%-ROAS figures above are in Meta’s Andromeda blog — Meta self-reported and methodology-opaque, but not invented by third parties. By contrast, figures like “65% higher ROAS testing 20+ ads/month” or a “60% Creative Similarity Score suppression threshold” come from agency blogs (Scaledon, admetrics.io and others), not from Meta — “Creative Similarity Score” is not a Meta-published metric. Treat the agency numbers as directional only, and even Meta’s own figures as attribution-based claims, never guarantees.

Why it matters

For anyone running paid social, Andromeda reframes the work: the platform is increasingly good at finding the right ad, so the constraint moves to supplying a pool of distinct, brand-aligned creatives worth choosing from. That makes a reusable creative-template library — not the single ad — the unit of work. The playbook for that sits in marketing/andromeda-era-creative-strategy.

Sources

By Andrej Ruckij