The Retrieval Overhaul Most Advertisers Missed
In December 2024, Meta’s engineering team announced Andromeda, the new ad retrieval engine behind Facebook and Instagram. Meta says the system enables “a meaningful increase of model capacity (10,000x)” for more personalized ad matching, and that its deployment delivered +6% retrieval recall and +8% ads quality on selected segments.
Those numbers aren’t marketing fluff. They come directly from Meta’s own engineering write-up, which describes a deep neural network and a hierarchical ad index built to scale retrieval far beyond what the previous models could handle.
Here’s what that means in practice: retrieval is the gatekeeper. Before your bid gets evaluated, before your budget matters, before your targeting parameters are even considered — Andromeda decides whether your ad is worth showing at all. And it makes that decision based primarily on one thing: your creative.
Retrieval is no longer a filter. It’s a judge. If your creative doesn’t pass Andromeda’s relevance threshold, your campaign never enters the auction — no matter how much you’re willing to pay.
What the Data Actually Shows
This isn’t theory. Here’s what the published numbers say.
Confect analyzed 3,014 eCommerce advertisers across $834 million in spend and 115.7 billion impressions during the 2025 rollout. Their headline finding: ROAS declined 7% on average as Andromeda rolled out. Their explanation points at creative: Andromeda treats near-identical variations as one entity, so accounts running many small tweaks of one concept get fewer real chances to be matched.
| Metric | Source | Finding |
|---|---|---|
| 10,000x model capacity | Meta Engineering, Dec 2024 | Andromeda vs. previous retrieval models |
| +6% recall, +8% ads quality | Meta Engineering, Dec 2024 | Retrieval recall; ads quality on selected segments |
| 7% ROAS decline | Confect, 3,014 advertisers | Average across the 2025 Andromeda rollout |
How Andromeda Actually Reads Your Creative
Understanding retrieval changes how you think about creative production.
Practitioner analyses such as Confect’s describe Andromeda as reading what’s in your ad — the product, the visual style, the mood, even the text overlays. It builds a dense vector representation of each creative and matches it against user-level embeddings built from behavior signals.
The critical detail, as practitioners describe it: Andromeda clusters similar ads into a single entity. Ten variations of the same product shot on a white background? That’s one retrieval ticket, not ten. You’re not diversifying — you’re duplicating.
This is why the old playbook of “make five versions with different headlines on the same image” stopped working. Andromeda sees through surface variation. It wants genuinely different angles, different visual compositions, different product contexts.
Stop thinking about ad variations. Start thinking about ad angles. Each creative should represent a distinct buyer motivation, not a cosmetic difference.
The Practitioner’s Pivot
Here’s the tactical shift, compressed to what actually matters:
Kill the multi-ad-set structure. Five ad sets targeting five different lookalike audiences with five creatives each? Replace it with one to two broad ad sets loaded with 20-25 diverse creatives.
Build for angles, not audiences. Instead of segmenting by demographic or interest, segment by buyer motivation:
- The skeptic — needs proof, data, social validation
- The impulse buyer — needs urgency, scarcity, visual appeal
- The researcher — needs comparison, features, long-form context
- The brand loyalist — needs new product news, exclusivity
Each motivation gets its own creative treatment. Andromeda does the matching. You do the creative thinking.
Volume matters, but diversity matters more. More creatives only help if they represent genuinely different approaches, not 20 color swaps.
What This Means for Performance Teams
The skill set that wins on Meta in 2026 is not media buying. It’s creative strategy.
The person who can brief a shoot, identify an untested angle, write a hook that stops thumbs, and iterate based on Entity ID-level performance data — that person is now more valuable than the person who can build a nested lookalike structure in Ads Manager.
Budget rebalancing follows. Teams that spend most of their effort on audience research and only a sliver on creative production need to flip that balance. The algorithm handles distribution. Your job is to give it ammunition.
The old question was: “Who should see this ad?” The new question is: “What should this ad say — and to which motivation?” Andromeda answers the first question better than you can. Focus on the second.
The audience-first playbook isn’t just suboptimal — it’s actively working against the system that decides who sees your ads.
Creative is the new targeting. Act accordingly.
Want the creative side handled for you? Our Meta Ads service includes 4 static and 2 video ads a month (page in Bahasa Indonesia).