TL;DR: Meta replaced its hand-tuned ad-ranking system with a transformer in late 2024. The practical shift for marketers: creative volume now beats account structure, broad campaigns now beat narrow ones with stronger creative, and the hook carries more weight in the algorithm than ever before.

How Andromeda works

Meta announced Andromeda in December 2024 as "the most significant architectural change to ads retrieval in our history." The headline is that a hand-tuned layered ranking stack got replaced with a single transformer — the same family of model that powers ChatGPT and Claude, trained on interaction data across Facebook, Instagram, and WhatsApp.

The retrieval-to-ranking pipeline went from this:

graph LR
    A[Millions of<br/>candidate ads] --> B[Layered retrieval<br/>signals + hand-built<br/>ranking models]
    B --> C[~Hundreds of candidates]
    C --> D[Final ranker]
    D --> E[Served ad]

to this:

graph LR
    A[Millions of<br/>candidate ads] --> B[Andromeda transformer<br/>scores tens of millions<br/>in parallel]
    B --> E[Served ad]

Two numbers from Meta's announcement do the heavy lifting:

Metric Change vs. previous system
Ad ranking quality +8%
Model capacity 10,000×

The capacity number is load-bearing. The old system narrowed millions of ads to a few hundred before scoring. Andromeda scores tens of millions in parallel. Practically, more of your ads get a fair look on more impressions, but only if you give the system enough variants to compare.

Why Meta built it

The official Meta framing is "personalization at scale." The unofficial answer: iOS 14 broke the old system.

Before Apple's ATT changes, Meta's algorithm leaned on deterministic conversion signals from your site or app. When those signals went dark for 70–80% of iOS users, the old retrieval system kept ranking ads against a much noisier picture. The transformer can ingest the noisy signals (engagement patterns, dwell time, creative-level engagement) and still match ads to users without the deterministic backbone.

The corollary: the more your creative tells the system about itself (hooks, scene structure, audio, on-screen text), the better Andromeda can place it. A library of 50 near-identical statics gives Andromeda less to work with than 20 visibly different videos.

What this changes for marketers

Andromeda is fully live across Meta's ad inventory as of early 2026. From the operator side, here's what's actually different.

Creative volume is the new account structure

Most performance marketers spent 2018 to 2022 obsessing over campaign structure: how many ad sets, how to split audiences, when to consolidate. Andromeda flattens most of that. Tight audience splits used to give the algorithm a head start; now they starve it of the comparison signal it needs. The accounts winning in 2026 ship more creative variants into broader campaigns and let Andromeda do the matching.

The hook carries more weight

Hook rate has always mattered. Under Andromeda it matters more, because the model uses early engagement signals as one of the strongest features in its representation. A weak hook doesn't just lose viewers. It gives the algorithm a weak signal about who the ad is for.

Audience-level optimizations are mostly noise now

Lookalikes, custom audiences, and exclusion lists still work, but their relative contribution to performance has shrunk. The lift you used to get from a perfectly-tuned lookalike now comes from a better creative library.

Smaller budgets get more from creative diversity than from clever targeting

This was the inversion that surprised most operators. In the pre-Andromeda world, a $500/day budget benefited from narrow targeting because the algorithm needed fewer events to learn. Andromeda's much larger candidate pool means even small budgets get tested against a broader set of users, but only if the creative inputs are diverse.

The Superscale customer marketbirds described this shift directly. "Meta's shift to creative-first raised the volume bar. We needed more variants for smaller-budget clients." Their solution was to frontload a month of ads, get bulk client approval, and let Andromeda iterate. The result was a 540% increase in creative output and a 26% relative CTR lift across their book of clients.

Does Andromeda change how you should structure campaigns?

Yes, but probably less than the panic-Twitter takes suggest. The actual decision tree:

If you're still running tight audience splits (per-country, per-interest): consolidate. Andromeda's representation will do a better job matching your creative to users than your ad-set structure will. Move to broad targeting with multiple ad sets per creative theme rather than per audience.

If you're stuck on ABO (ad set budget optimization): revisit CBO (campaign budget optimization). Andromeda's stronger ranking means the algorithm allocates spend more efficiently across ad sets than manual ABO splits do. The old reason to use ABO (controlling spend on a learning ad set) matters less when the learning phase is shorter.

If you have 5 to 10 creatives per campaign: push to 20 to 40. The 10,000× capacity is meaningful only if you feed it.

If you have 100+ creatives but flat performance: your creatives are probably too similar. Andromeda can't differentiate between 100 near-duplicate statics. You need visibly different hooks, formats, characters, or scene structures.

The agencies and in-house teams seeing the biggest Andromeda-era lifts are the ones who treated the algorithm change as a creative-pipeline change, not a media-buying change.

AI-generated creative under Andromeda

A common concern: if Andromeda rewards creative volume, doesn't that just incentivize spammy AI-generated ads?

The short answer is no, because Andromeda also reads creative-level engagement. Spammy variants get rapidly down-weighted. The system rewards differentiated volume (visibly distinct creative ideas) not duplicate volume.

This is also where AI ad agents like Superscale fit into the Andromeda era. The bottleneck is no longer "how do I produce 40 differentiated ads per month at a sane cost." Tools that auto-generate visibly distinct concepts (different characters, hooks, framings, scene structures) from a single product feed can match Andromeda's appetite without the brand looking like a content farm.

The Taxfix marketing team uses Superscale across four markets and three languages this way: surfacing trends, turning them into testable concepts, and shipping 15+ ads per week into broader Meta campaigns. Their result, +45% CTR and a 20% drop in CPA on Meta UK, is partly an Andromeda story. The algorithm finally had enough material to actually do its job.

How to tell if Andromeda is helping your account

Three things to check on Monday morning.

graph TD
    A[Single ad eating >40%<br/>of campaign spend?] -->|Yes| A1[Library too narrow.<br/>Add 5-10 different variants]
    A -->|No| B[New creatives need<br/>>24h to find audience?]
    B -->|Yes| B1[Algorithm struggling.<br/>Variants aren't<br/>differentiated enough]
    B -->|No| C[CPAs spike when<br/>creative fatigues?]
    C -->|Yes| C1[No fallbacks.<br/>Wider library needed]
    C -->|No| D[Andromeda is working]
  1. Top creative's share of spend, month over month. If one ad is eating more than 40% of the campaign's spend, Andromeda is concentrating because it can't find anything better. The fix is more variants, not better bidding.
  2. Early-life CTR of new creatives. Pre-Andromeda, new creatives needed a 24–48 hour learning phase before they started winning impressions. Andromeda compresses this to 6–12 hours when the creative is differentiated enough. If new ads still need 48 hours to find an audience, the algorithm is struggling to place them.
  3. Creative-refresh cadence vs. CPA stability. If CPAs spike every time a winning creative fatigues, the library is too narrow. With enough diversity, Andromeda has fallback creatives to rotate to.

Andromeda and the agentic shift

The bigger frame: ad platforms moved faster than the people running ads on them.

The old performance-marketing skill set (tight audience splits, manual budget allocation, hand-tuned learning phases) got absorbed into the platforms. What's left for operators is what the platforms can't automate: deciding what to say, to whom, in what format. The creative brief.

That's why "agentic marketing" became the umbrella term for what's replacing classical performance marketing. The agent doesn't replace the marketer. It replaces the manual work between the creative brief and the live ad. Andromeda decides which of those ads gets shown to which user, and which underperformers get killed.

Rebuilding around this shift looks roughly like:

graph LR
    A[Ship more<br/>differentiated creative] --> B[Consolidate<br/>campaign structure]
    B --> C[Let Andromeda match]
    C --> D[Measure CTR/CPA<br/>at account level]
    D --> A

FAQ

When did Meta Andromeda launch?
Meta announced Andromeda in December 2024 and rolled it out across the full ad inventory over the following 12 months. By early 2026 it was the default ad-serving system on Facebook and Instagram.

Does Andromeda apply to TikTok or Google Ads?
No. Andromeda is specific to Meta. TikTok and Google have their own machine-learning ad-serving systems with different architectures.

Do I need to enable Andromeda manually?
No. It runs by default for every advertiser. You can't opt out, and there's no setting to toggle.

Will Andromeda reduce my CPAs automatically?
Not on its own. Andromeda is better at matching creative to users, but it can only work with the creative library you give it. Accounts that haven't increased their creative volume since 2023 will typically see less Andromeda lift than accounts shipping new variants weekly.

Is creative volume the only thing that matters now?
No, but it's the input that's underweighted in most accounts. Targeting, budget allocation, and bid strategy still matter. They just matter less than creative diversity does in 2026.

How many ads should I ship per month under Andromeda?
There's no fixed number, but the operators seeing the biggest lifts (marketbirds, Taxfix, SumUp) are shipping 15 to 40+ new variants per week, not per month.

Sources

  • Meta Engineering, "Meta Andromeda: Personalized Ads at Scale" (December 2024) — engineering.fb.com — source for the "+8% quality improvement" and "10,000× model capacity" figures cited above.