On December 2, 2024, Meta’s engineering team published an announcement that quietly reset how ad ranking works for every DTC brand on the platform. They called it Andromeda: a personalized ads retrieval engine that increased their model capacity roughly 10,000x and shipped a 6% recall lift and an 8% ads-quality improvement on selected segments.
Most agencies have spent six months talking around it. Here’s what it actually means for your media buyer next Monday.
What Andromeda is, in plain English
Meta’s ad system has two halves. Retrieval picks a candidate set of ads worth showing to a given user — out of millions of creatives in the auction. Ranking then sorts that candidate set and decides which one wins the slot.
Andromeda is the retrieval half. According to Meta’s engineering blog, it runs on NVIDIA Grace Hopper Superchips and Meta’s own MTIA hardware, with a customized deep neural network designed for sublinear inference cost. The 10,000x capacity jump means Meta can now consider dramatically more candidate ads per impression — and personalize the candidate set far more aggressively per user.
The practical takeaway: retrieval is no longer the bottleneck. Your ad doesn’t fail because Meta can’t find the right user — it fails because Andromeda already considered 12 of your near-duplicate creatives and treated them as the same asset.
Why 30 lookalike ads count as 1
Modern retrieval systems use asset embeddings — vector representations that capture the “shape” of a creative (composition, color, hook structure, talent identity, brand cues). Two creatives that look 95% the same in embedding space get clustered together for retrieval purposes. Meta hasn’t published the exact threshold — but the engineering blog explicitly states the system was rebuilt around personalized retrieval, which by definition requires deduplicating similar candidates so the user doesn’t see four versions of the same ad in one session.
What this means for your account:
- Ten reskinned variants of the same hook, with different colors and music — Andromeda often treats them as one creative for retrieval purposes.
- Thirty cuts of the same talking-head with different captions — same problem.
- A library of 100 ads that all look like the same UGC creator on the same couch — the algorithm sees a small handful of distinct shapes.
Meanwhile, third-party agency analysis and Billo’s post-Andromeda playbook both converge on the same recommendation: shift from volume to distinct volume. Different hooks, different formats, different modalities — not the same hook with a new background.
What “distinct” actually means
Industry analysts disagree on the exact threshold for how different two ads need to be. But on what works in practice, there’s broad agreement. A creative is “fundamentally different” when at least three of the following change vs. the version next to it:
- Hook archetype. Curiosity gap vs. contrarian vs. stat-shock vs. problem-agitation vs. founder-story vs. social-proof.
- Persona angle. First-person founder vs. customer testimonial vs. expert third-party vs. animated character.
- Modality. Live action vs. animation vs. static image vs. carousel vs. UGC selfie-POV.
- Visual composition. Talking-head vs. product close-up vs. lifestyle wide shot vs. screen-recording.
- Script structure. Pattern-interrupt opener vs. identity-callout opener vs. specific-stat opener vs. confession opener.
Same hook with new music? One change. Same composition with new captions? One change. Same UGC creator with new background? Maybe one change. None of those move the embedding far enough to matter.
The volume number nobody can quite agree on
You’ll see “30 ads per cycle” or “20 to 40 variants” thrown around. Where does it come from? Billo and Wonderful both reference 20–40 distinct creatives per cycle as the band where Advantage+ campaigns get the full algorithmic benefit. Meta’s own Advantage+ documentation requires a minimum of 4 creatives but explicitly notes that “more is better” for performance.
Translated: 4 is the floor to even qualify for Advantage+. 20–40 is the band where Andromeda has enough distinct candidates to actually personalize per user. Below 20, you’re feeding the algorithm a thin candidate set. Above 40, you’re past diminishing returns unless your variants are genuinely diverse.
What this looks like in an actual ad account
A 30-ad week that beats a 30-reskin week looks like:
- 6 live-action UGC cuts (3 different creators × 2 hook archetypes each)
- 4 animated/character ads (claymation or 3D character — different format entirely)
- 10 AI-built variant cuts (different scripts, different opens, different visual concepts)
- 6 static image ads (3 different visual compositions × 2 carousels)
- 4 thumbnail variants for video reuse (lift CTR on existing winning videos)
That’s 30 creatives across 4 different modalities with at least 6 distinct hook archetypes. Andromeda reads that as 30 distinct candidates. The same week with 30 reskins of one talking-head? Maybe 4–6 distinct candidates after deduplication.
Where Vibey Volume fits
We built Vibey Volume as the AI-built variant engine for exactly this — 20–40 fundamentally different cuts per cycle, each tagged by hook archetype and persona angle so your media buyer can read what they’re actually testing. Most clients pair it with Vibey Faces (real on-camera UGC for the spots that scale), Vibey Toons (animated/claymation for differentiation), and Vibey Stills (premium static ads + carousels) — the four-modality stack Andromeda was rebuilt to reward.
The brands that win on Meta from here aren’t the ones shipping the most ads. They’re the ones shipping the most distinct ads.
Sources cited above: Meta Engineering — Andromeda announcement (Dec 2024) · Wonderful — 11 Andromeda creative strategies · Billo — Andromeda update playbook