The biggest change to Meta ads in the past two years didn’t show up in Ads Manager. It happened in the retrieval layer — the system that decides which ads even get considered before ranking begins. Meta’s Andromeda algorithm introduced a 10,000x increase in model complexity for ad retrieval (Engineering at Meta, 2024). That’s not a typo. The old system looked at a narrow slice of candidate ads. Andromeda blew the doors open.
Most advertisers are still running campaigns like it’s 2023 — splitting by placement, micro-targeting audiences, throttling creative volume. Meanwhile, Meta rebuilt the entire engine underneath them. In this post, I’ll break down exactly how Andromeda works, how it connects to Lattice and GEM, and what you should actually change in your account structure right now.
TL;DR: Meta’s Andromeda algorithm evaluates ad candidates with 10,000x more model complexity than the old retrieval system, improving ad quality by 8% (Engineering at Meta, 2024). Campaign consolidation, creative diversity, and Advantage+ adoption aren’t optional anymore — they’re how you give Andromeda more surface area to find winners.
What Is Meta’s Andromeda Algorithm?
Meta’s Andromeda algorithm is the retrieval engine that decides which ads are eligible for ranking — and it processes candidates with 10,000x more model complexity than its predecessor, delivering an 8% improvement in ad quality on selected segments (Engineering at Meta, 2024). Think of it as the bouncer that decides who gets into the club.
Here’s the thing most people miss: Meta’s ad delivery has two stages. First, retrieval — the system pulls a shortlist of candidate ads from millions of possibilities. Then ranking — it scores those candidates and picks the winner. Before Andromeda, the retrieval layer was relatively simple. It used basic signals to narrow the field to roughly a thousand candidates.
Andromeda changed that math completely. It applies deep learning models at the retrieval stage, evaluating far more candidates with far more sophistication. The system also achieved a 100x improvement in feature extraction latency and throughput, plus 3x higher queries per second and 10x better inference efficiency (Engineering at Meta, 2024). It’s not just smarter — it’s faster.
Why does this matter for you? Because retrieval is the gate. If your ad doesn’t make it past Andromeda, it never reaches the ranking stage. It never gets shown. And with a 6% recall improvement (Engineering at Meta, 2024), Andromeda is pulling in more relevant ads that the old system would have missed entirely.
So what does Andromeda actually reward? Ads that are genuinely relevant to users — not ads that game narrow targeting. The model evaluates creative quality, user-ad fit, and engagement probability at a depth that wasn’t possible before. If you’re still running the same three ad variations in a hyper-targeted ad set, you’re leaving retrieval surface area on the table.
Citation capsule: Meta’s Andromeda retrieval engine increased ad candidate evaluation complexity by 10,000x while achieving 100x faster feature extraction and 10x inference efficiency, resulting in 8% higher ad quality on selected segments (Engineering at Meta, 2024).
How Do Andromeda, Lattice, and GEM Work Together?
Andromeda doesn’t operate alone. Meta’s ad AI is a three-layer stack, and the numbers tell the story: Lattice improved ad quality by 12% and conversions by 6% (Meta for Business, 2025), while GEM — the Generative Ads Model — drove +5% conversions on Instagram and +3% on Facebook Feed (Engineering at Meta, 2025).
Here’s how the stack works. Andromeda sits at the top — it’s the retrieval layer. It picks which ads even get considered. Lattice is the ranking layer. Once Andromeda hands over candidates, Lattice scores them using trillions of parameters trained on hundreds of billions of examples (Meta for Business, 2025). GEM is the foundation model underneath both — it’s the brain that accelerates how the entire system learns.
What makes this stack powerful is how they feed each other. GEM achieved a 23x increase in effective training FLOPs and proved 2x more effective than standard knowledge distillation, running on 16x more GPUs (Engineering at Meta, 2025). That training power flows up through Lattice and Andromeda. GEM is also 4x more efficient at driving ad performance gains than previous approaches (Engineering at Meta, 2025).
Lattice does something especially interesting for campaign architecture. It optimizes jointly across Feed, Story, and Reels placements, delivering roughly 8% higher ad quality on Instagram through that cross-surface optimization (Meta AI Blog, 2024). This is why splitting campaigns by placement actively works against you now. You’re fragmenting the signal that Lattice needs.
Have you restructured your accounts to account for this yet? If you’re still running separate Reels campaigns and Feed campaigns, you’re fighting the system instead of feeding it. And that brings us to what this AI-powered delivery system is actually doing with your spend — and why the DTC strategy playbook looks completely different in 2026.
Citation capsule: Meta’s AI ad stack operates as three interconnected layers — Andromeda for retrieval (10,000x complexity), Lattice for ranking (trillions of parameters, +12% ad quality), and GEM as the foundation model (4x efficiency gains, 23x training FLOPs) — with each layer improving the others’ performance (Engineering at Meta, 2025).
What Does Meta’s Q4 2025 Data Tell Us About AI-Driven Ads?
The financial results make the AI impact undeniable. Meta reported $59.9 billion in Q4 FY2025 revenue — up 24% year-over-year — with ad impressions growing 18% and average price per ad rising 6%, while Lattice drove a 12% YoY improvement in ads quality (Futurum Group, 2026).
But here’s the number that should reshape how you think about Meta’s direction. AI-driven redistribution of ads delivered nearly 4x the revenue impact of simply increasing ad load in H2 FY2025 (Futurum Group, 2026). Meta isn’t growing revenue by showing more ads. They’re growing it by showing better ads to better-matched users. That’s Andromeda and Lattice at work.
The generative AI adoption numbers are staggering too. Over 1 million advertisers used Meta’s generative AI tools, producing 15 million+ ads in a single month, with image generation alone driving a 7% conversion lift (Engineering at Meta, 2024). The system is getting more data to learn from every day.
I’m seeing this play out in my own accounts. Campaigns that I consolidated in Q1 2026 — merging what used to be 6-8 ad sets into 1-2 broad Advantage+ campaigns — are outperforming the fragmented structure by 15-25% on CPA. The AI stack rewards simplicity because it gives Andromeda and Lattice the full picture. When I split things up, I was essentially blindfolding the system and asking it to optimize with one hand tied behind its back.
What’s driving your ROAS gains right now? If you haven’t checked whether your structure is aligned with how this AI stack actually works, the answer might be “less than it should be.” For a deeper look at how AI is reshaping Meta ads analytics, I’ve written a full breakdown.
Citation capsule: Meta’s Q4 FY2025 results showed AI redistribution of ads delivered nearly 4x the revenue impact of ad load increases, contributing to $59.9 billion in quarterly revenue (+24% YoY) as Lattice drove 12% year-over-year improvement in ads quality (Futurum Group, 2026).
Why Is Creative Diversity Now a Retrieval Signal?
The data on this is striking. Common Thread Collective analyzed 53,000 ads during the Cyber Five shopping period and found that only 1,100 ads — roughly 2% — drove 56% of total revenue. UGC-style ads were 2x more likely to become high-spending winners (Common Thread Collective, 2025).
Here’s what’s happening under the hood. Andromeda evaluates candidates based on predicted relevance and engagement. When you give it a diverse creative set — different formats, different hooks, different visual styles — you’re giving it more shots at matching the right ad to the right user at the right moment. A single polished brand video gives Andromeda one option. Twenty variations across UGC, static, and short-form video give it twenty.
Static ads saw a +158% increase in volume year-over-year in CTC’s dataset (Common Thread Collective, 2025). But that volume increase came with diminishing returns per static ad. The takeaway isn’t “run more statics.” It’s that format variety is what wins because Andromeda can match different formats to different placements, audiences, and moments.
Here’s how I think about it: creative diversity isn’t just a performance tactic anymore — it’s a retrieval strategy. Each distinct creative format gives Andromeda a different retrieval pathway. A founder-talking-to-camera UGC ad and a product-on-white static image appeal to fundamentally different user states. Andromeda sees these as separate candidates. More formats equals more retrieval surface area. You’re not just testing creative — you’re expanding your footprint in the retrieval layer.
If you want to understand how to build a creative testing system that feeds Andromeda properly, or how to identify the 5% of ads that actually win, I’ve covered both in detail. The game has changed. Creative volume without format diversity is just noise.
Citation capsule: In an analysis of 53,000 ads during Cyber Five, only 1,100 (2%) drove 56% of revenue, with UGC-format ads proving 2x more likely to become high-spending winners — evidence that Andromeda’s retrieval layer rewards creative format diversity over volume alone (Common Thread Collective, 2025).
How Should You Restructure Campaigns for the Meta Andromeda Algorithm?
The clearest signal comes from Advantage+ performance data: Advantage+ Shopping campaigns delivered an average ROAS of 4.52x compared to 3.70x for manual campaigns — a 22% improvement (Engineering at Meta, 2024). That gap will only widen as Andromeda, Lattice, and GEM keep improving.
Here are five specific changes I’d make — and that I’ve been making in client accounts since Q4 2025:
1. Consolidate campaigns aggressively
Fewer campaigns means bigger candidate pools for Andromeda to evaluate. When you fragment into 10 campaigns with 5 ad sets each, you’re giving the retrieval system 50 tiny pools instead of one large one. I’ve been merging accounts down to 2-4 campaigns total. The system performs better with more data concentrated in fewer places.
2. Maximize creative format diversity
Don’t just make 20 variations of the same concept. Run UGC, static images, short-form video, carousels, and catalog creative within the same campaign. Remember — CTC’s data showed UGC ads were 2x more likely to become winners (Common Thread Collective, 2025). But the real gain comes from variety, not any single format. For more on AI creative analysis systems, I’ve written a full guide.
3. Use Advantage+ Shopping over manual
That 22% ROAS improvement isn’t theoretical. Advantage+ gives Andromeda and Lattice full control over audience selection, placement, and budget allocation. If you’re still clinging to manual campaigns for “control,” you’re optimizing for your comfort, not your results. Here’s my full Advantage+ Shopping guide.
4. Stop splitting by placement
Lattice optimizes jointly across Feed, Story, and Reels — delivering ~8% higher ad quality on Instagram through cross-surface optimization (Meta AI Blog, 2024). When you force placement splits, you break that joint optimization. Let the system decide where each ad performs best. Read more about Advantage+ audience targeting and why it matters.
5. Feed the system with conversion data
GEM gets 4x more efficient at driving performance with more signal (Engineering at Meta, 2025). That means your Conversions API implementation, your event setup, and your attribution windows all directly affect how well this AI stack serves you. Don’t starve the model. My budget allocation guide covers how to maximize signal quality.
I restructured one DTC client’s account from 12 campaigns down to 3 in January 2026. Their CPA dropped 18% in the first two weeks and stayed down. The creative diversity recommendation was harder to implement — they’d been running mostly polished brand video. We added iPhone-shot UGC and simple static images. Within a month, those “lower quality” creatives were outspending the brand videos 3:1. The Andromeda algorithm didn’t care about production value. It cared about relevance.
Citation capsule: Advantage+ Shopping campaigns delivered 4.52x average ROAS versus 3.70x for manual campaigns — a 22% improvement — while Lattice’s cross-surface optimization added ~8% ad quality on Instagram, making campaign consolidation and automated placements the clearest path to performance gains under Meta’s Andromeda algorithm (Engineering at Meta, 2024).
What Comes Next for Meta’s AI Ad Stack?
Meta is doubling down. GEM’s GPU allocation grew 16x to support a 23x increase in effective training FLOPs (Engineering at Meta, 2025), and the pace isn’t slowing. With 1 million+ advertisers already using generative AI tools, the flywheel is spinning faster every quarter.
The trajectory is clear. Meta is building toward a system where the AI handles targeting, placement, bidding, and creative generation end-to-end. The role of the media buyer shifts from campaign configuration to creative strategy and measurement frameworks.
Does that mean media buyers become irrelevant? Not even close. It means the job changes. The people who understand how Andromeda, Lattice, and GEM work — and feed those systems the right inputs — will outperform the ones clicking buttons in Ads Manager. The death of manual targeting is just the beginning. The entire workflow is being rebuilt, and understanding the architecture gives you the edge.
Related: Meta Advantage+ Shopping vs Manual Campaigns: When AI Targeting Beats Human Setup.
Related: Psychology + AI Creative Generation: Using Behavioral Science at Scale in Meta Ads.
Frequently Asked Questions
What is Meta’s Andromeda algorithm?
Andromeda is Meta’s AI-powered ad retrieval engine that decides which ads qualify for the ranking stage. It processes candidates with 10,000x more model complexity than the previous system, achieving 8% higher ad quality and 100x faster feature extraction (Engineering at Meta, 2024). It works alongside Lattice (ranking) and GEM (foundation model) in Meta’s three-layer AI stack.
How does Andromeda affect my ad costs?
Andromeda doesn’t directly set prices, but it determines which ads compete in the auction. Better retrieval means more relevant ads reach ranking, which tends to lower CPMs for high-quality creative. Meta’s Q4 FY2025 data showed AI redistribution delivered 4x more revenue impact than ad load increases (Futurum Group, 2026), suggesting the system rewards relevance over spend.
Should I switch all campaigns to Advantage+?
For most ecommerce brands, yes. Advantage+ Shopping delivered 4.52x ROAS versus 3.70x for manual campaigns — a 22% improvement (Engineering at Meta, 2024). The exception is if you have strict geo or audience exclusion requirements that Advantage+ can’t accommodate. Start by migrating your highest-spend campaigns first.
Does creative diversity actually matter for performance?
The data says yes. CTC’s 53,000-ad study found UGC ads were 2x more likely to become high-spending winners, while only 2% of ads drove 56% of revenue (Common Thread Collective, 2025). Diverse formats give Andromeda more retrieval pathways. Volume without variety is just noise.
When will Meta fully automate ad creation?
It’s already happening partially — 1 million+ advertisers used Meta’s generative AI tools to create 15 million+ ads in a single month, with AI-generated images driving 7% conversion lifts (Engineering at Meta, 2024). Full end-to-end automation is likely by late 2026 or 2027. The smart move is to learn the system now, not wait.
Key Takeaways
The Meta Andromeda algorithm isn’t a future state — it’s already running. Every ad you publish today goes through this three-layer AI stack. Here’s what matters:
- Andromeda evaluates candidates with 10,000x more model complexity — your ads need to earn retrieval, not just ranking.
- Consolidate campaigns — give the system bigger candidate pools and more concentrated data.
- Creative format diversity beats creative volume — UGC, static, video, and carousel all serve different retrieval pathways.
- Advantage+ Shopping outperforms manual by 22% on ROAS — stop holding onto manual control that doesn’t help.
- Stop splitting by placement — Lattice optimizes across surfaces jointly, and fragmenting that signal costs you money.
The advertisers who understand Andromeda, Lattice, and GEM — and structure their accounts to feed those systems properly — will consistently outperform those who don’t. This isn’t about chasing a trend. It’s about aligning your campaigns with the infrastructure Meta has already built. The restructure isn’t optional anymore. If you’re ready to start, begin with your creative testing system — that’s where the biggest gains are.