If you run Meta ads, you've probably heard the word "Andromeda" thrown around in the last year. Usually in a sentence like "Andromeda changed everything" with no explanation of what Andromeda actually is.
So let's fix that. This is a plain-English walkthrough of what Meta actually built, drawn from Meta's own engineering publications, and what it means for the way you make ads.
First, the One-Sentence Version
Andromeda is the retrieval engine at the front of Meta's ads pipeline. For every single impression, it scans tens of millions of candidate ads and narrows them to a few thousand finalists, and it does that by reading the actual content of each ad: the pixels, the video frames, the text overlays, the captions.
Not the audience settings. The ad itself.
That one design decision is why everything you knew about Meta targeting stopped working the way it used to.
How Ads Used to Get Delivered
For a decade, the formula was simple. You defined an audience (interests, lookalikes, custom lists), you made an ad or two for that audience, and the algorithm optimized delivery within the box you drew. Your job was to find the right people. Creative mattered, but audience selection was the primary lever.
Meta flipped the core question. The system no longer asks "who should see this ad?" It asks "which ad should this person see?" Those sound similar. They are not. In the first model, your targeting decides who's in the room. In the second, everyone is in the room, and your creative is competing against tens of millions of other ads for each individual impression.
The Three-Stage Pipeline
Meta's ads stack now works as a pipeline of three systems, each handling one phase of the journey from millions of candidate ads down to the single impression you see in your feed.
Stage 1: Andromeda (Retrieval)
Announced in December 2024 and fully rolled out globally by October 2025, Andromeda is the front door. Every time an impression is about to be served, it filters tens of millions of eligible ads down to roughly a thousand candidates using a hierarchical index built on semantic similarity and user intent.
The key detail: Andromeda doesn't match on audience labels like "Interest: Golf" or "Lookalike: Purchasers." It matches on what your ad is. A video of a mom solving a mess in her kitchen produces a fundamentally different retrieval signature than a studio product shot with a discount code, even if both are selling the same thing to the same nominal audience.
Meta built Andromeda explicitly to handle the flood of creative volume coming from generative AI tools. The system is designed to index massive numbers of distinct creatives. Which is exactly why it rewards advertisers who supply them.
Stage 2: GEM (The Foundation Model)
In November 2025, Meta detailed GEM, the Generative Ads Recommendation Model. Think of GEM as the brain that learns offline. It trains across thousands of GPUs on sequences of thousands of behavioral events per user, spanning both organic content and ad interactions across Facebook and Instagram.
Here's the part advertisers should tattoo somewhere visible: creative representation is an explicit input to the model. The content of your ad (what's in the image, the hook, the format) helps GEM figure out who to show it to, not just predict whether they'll convert. GEM learns which creative patterns work for which kinds of people based on observed behavior, then distributes that intelligence downstream to Meta's entire fleet of ads models.
In other words: your creative is no longer just the thing being delivered. It's the targeting data.
Stage 3: The Adaptive Ranking Model (Real-Time Execution)
The newest piece, launched in late 2025 and detailed by Meta in March 2026, is the executor. It takes GEM's learned intelligence and makes the final per-impression call in under 100 milliseconds, billions of times a day. Rather than running one model on every request, it dynamically routes each impression to the most effective model for that person's context and intent, with multi-card GPU serving enabling trillion-parameter scale.
The division of labor, in one line: GEM does the deep learning at training time, the Adaptive Ranking Model executes that learning at impression speed, and Andromeda curates the candidate pool they both work from.
Taken together, Meta describes this as roughly a 10,000x increase in model complexity for ad matching versus the old system. All three stages reward the same thing: genuine creative variety. Andromeda indexes it, GEM learns from it, and the ranking layer routes impressions to it.
Why "More Ads" Isn't the Takeaway
The lazy reading of all this is "make more ads." That's not quite it, and getting this wrong is expensive.
The system identifies each truly distinct creative and consolidates near-duplicates. Twenty videos of the same person in the same kitchen making the same point, with different headlines, don't register as twenty ads. They register as one, and delivery gets limited accordingly. You paid for twenty edits and bought one.
A practical rule of thumb we use: for two ads to count as genuinely different, they should differ on at least two of these four dimensions.
- Format: static image, video, carousel
- Persona: who is on screen and who the ad speaks to (the mom, the professional, the skeptic)
- Environment: kitchen, office, outdoors, gym
- Benefit: saves time, saves money, better quality, status, health
To be clear on sourcing, since we try to be careful about this: the three-stage pipeline above is documented by Meta's engineering team. The two-of-four rule is an industry heuristic built on top of how the system behaves, not something Meta publishes. It's the same relationship we covered in Give Meta What It Wants: Meta provides the mechanics, and strategy is the layer you build on them.
What the System Rewards (and Penalizes)
Once you understand the pipeline, the best practices stop being arbitrary rules and start being obvious consequences.
Rewarded:
- Creative diversity. Meaningfully different concepts give retrieval more distinct entries to match against distinct user moments.
- Consolidation. Fewer campaigns and ad sets pool your conversion data, which gives GEM a cleaner learning signal. Fragmented accounts split their own signal.
- Signal quality. The models learn from your conversion events. Clean pixel plus CAPI tracking with high Event Match Quality is the algorithm's eyesight.
- Patience. The learning happens over days, not hours. Constant mid-week edits reset it.
Penalized:
- Near-duplicate creative. Consolidated and delivery-limited, as above.
- Micro-targeting. Interest stacks and narrow audiences constrain the exact matching process the pipeline was built to do better than you.
- Daily meddling. Budget and bid changes every morning keep the system in a permanent learning state.
What This Means for Your Account, Practically
Three changes follow directly from the architecture.
1. Your account structure should be boring. For most DTC accounts, this means one broad prospecting campaign and one retargeting campaign. Location and age constraints only. No interest layers, no lookalike stacks. The pipeline performs the matching; your targeting settings can only restrict its options.
2. Your creative process is now your media buying. When performance dips, the old reflex was to adjust targeting. That lever is gone. The two levers that remain are creative and tracking. A systematic creative process, one that deliberately varies persona, format, environment, benefit, and funnel stage, is what "optimization" looks like now. This is exactly why we build every account around a creative matrix rather than ad-hoc batches of ads.
3. Your intervention cadence should slow down. Daily: observe, check for anomalies, resist touching things. Weekly: retire genuinely failed ads, launch new concepts. Monthly: audit what won, and brief the next round of creative from those learnings. Ads also fatigue faster under this system, typically burning out in two to four weeks, so the refresh pipeline never really stops.
The Bottom Line
Andromeda is not a mysterious algorithm update to be gamed. It's a published, documented architecture with a clear logic: Meta built a system that reads creative content semantically, learns which content works for which people, and routes every impression to the best match in real time.
That system has one honest requirement of advertisers. Give it genuinely diverse, strategically sound creative to work with, because it can only match messages that exist in your creative set. If every ad you run says the same thing to the same imagined customer, the smartest ad delivery pipeline ever built has exactly one thing to deliver.
Creative is the targeting now. Build accordingly.