Platform update
Meta's Andromeda update, explained for D2C brands
Meta's AI retrieval system changed what its ads auction rewards. Here is what Andromeda actually is, what it makes obsolete, and the five moves that matter for D2C accounts in 2026.
Andromeda is not a feature you opt into — it is the machinery underneath every Meta ad you run. It is Meta's AI-powered ad retrieval system: the stage that decides, out of millions of eligible ads, which few thousand are even considered for a given person before the final ranking picks a winner. Meta began rolling it out in late 2024, and industry coverage places the full global rollout through 2025 — so by 2026, every account is living with its consequences.
What actually changed
Before systems like Andromeda, advertisers steered delivery with inputs: stacked interest audiences, lookalike percentages, bid tweaks. The retrieval model inverts that. It reads the creative itself — the imagery, the copy, the product — and the conversion signal coming back from your pixel and server events, and does the matching on its own.
The practical consequences reported consistently across the industry:
- Creative is the targeting now. The system infers who an ad is for from what the ad is. A handle photographed on a beautiful door finds renovators; the interest stack you layered on top mostly just constrains it.
- Diversity beats duplication. Coverage of the update describes Meta scoring creative similarity — near-identical variants collapse into one ad in practice. Ten genuinely different concepts give the retrieval system ten chances; ten crops of the same concept give it one.
- Fatigue is punished faster. Tired creative does not just convert worse, it gets retrieved less. A steady supply of fresh concepts is no longer a nice-to-have; it is how the account keeps its place in the auction.
- Signal quality compounds. The system learns from your conversion events. A clean purchase signal — pixel plus server-side events, deduplicated, with values — teaches it the right lesson. A polluted conversion pool teaches it the wrong one, at scale.
What this means for a D2C account
Our read, from running D2C accounts through this shift:
- Go broad, structure by product line. Keep campaign structure legible by product economics — one line per ad set, budgets as decisions — and leave the audience broad. The structure is for you to steer budget; the matching is the machine's job now.
- Build a creative pipeline, not a creative project. A quarterly "campaign shoot" cannot feed a system that rewards fresh, diverse concepts. Curate the library you already own, cut everything in 1:1 and 9:16, brief creators for a steady stream of review-style assets, and treat creative production as a monthly rhythm.
- Clean the conversion pool. Remove low-value events (store visits, chat opens) from the optimisation signal, send values with purchases, and run server-side events alongside the pixel. The retrieval system is only as smart as what you feed it.
- Judge it against the store, not the dashboard. None of this changes who grades the homework. Platform-reported revenue remains a claim; the number that decides budgets is store revenue over total spend. An AI-optimised account can over-report exactly as confidently as the old one did.
- Stop paying for audience archaeology. Time spent micro-managing interest stacks and exclusion lists is better spent on the two inputs the system actually listens to: creative and signal.
A note on the numbers you will see quoted
Most Andromeda coverage comes from vendors and agencies, not from Meta directly — including widely-repeated figures like a 22% ROAS lift from AI-generated creative variations. Treat those as directional. The structural shift itself — retrieval-based delivery that reads creative and conversion signal — is consistent across every serious account of the update, and it matches what we see in accounts we run.