What is Meta Andromeda, and what changed for your ads?
By Denzel Geng · Updated October 13, 2026
Facts on this page verified October 13, 2026 against public sources.
Andromeda is the part of Meta's ad system that decides which ads are even considered for each person. This guide explains it from Meta's own engineering posts and earnings calls, how it fits with GEM and Lattice, and what to watch for in your account now.
Meta Andromeda is the ads retrieval engine Meta introduced in the second half of 2024. For each ad impression, it narrows tens of millions of eligible ads down to a few thousand relevant candidates, which ranking models such as GEM then score to pick the ad that is shown. Meta built it for a fast-growing volume of ad creative, and it groups ads that look or feel alike as variations of the same creative.
See what's winning in your nicheWhat is Meta Andromeda?
Every time someone opens Facebook or Instagram, Meta has to choose one ad from everything eligible to run. It does that in stages. Andromeda is the first big one, retrieval: in Meta's words, it is "tasked with selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates" (Meta Engineering, December 2, 2024). Only ads that make it through retrieval get scored by the ranking models.
Meta says Andromeda runs on NVIDIA Grace Hopper and Meta's own MTIA chips, and that it allowed "a meaningful increase of model capacity (10,000x) for enhanced personalization." That 10,000x refers to the size of the model, not the number of ads. Meta reports that the deployment across Instagram and Facebook "has achieved +6% recall improvement to the retrieval system, delivering +8% ads quality improvement on selected segments."
The reason it exists matters more than the specs. Meta wrote that with generative AI tools for making ads, "the number of ads creatives in Meta's recommendation systems is expected to grow significantly", and that Andromeda is designed to make use of that growing volume. It is the system built for a world where every advertiser can produce far more creative.
When was the Andromeda update?
Meta has never given a single launch date. The engineering post describing it appeared on December 2, 2024. On the Q4 2024 earnings call (January 29, 2025), Meta said: "In the second half of 2024, we introduced … Andromeda", adding that it had driven an 8% increase in the quality of ads people see on the objectives tested (transcript). On the Q2 2025 call, Meta said it was expanding Andromeda's coverage to Facebook Reels (transcript), and on the Q3 2025 call it described combining models across retrieval and early-stage ranking into one, "driving a 14% increase in ads quality on Facebook surfaces" (transcript).
So Andromeda wasn't a one-day switch. It arrived in the second half of 2024 and kept expanding to more surfaces through 2025. Meta hasn't said it reached every advertiser on a particular date, so be wary of any guide that names one, and of blaming a single bad week on "the Andromeda update."
Andromeda vs GEM vs Lattice: what does each one do?
The three names come up together because they work at different steps of the same pipeline. Andromeda decides which ads are considered, the ranking models decide which one wins, and Lattice and GEM are how Meta makes those ranking models better.
| System | What it does | What Meta has reported | Source |
|---|---|---|---|
| Andromeda | Retrieval: narrows tens of millions of eligible ads to a few thousand candidates for each impression | +6% retrieval recall and +8% ads quality on selected segments; +8% ads quality on tested objectives (Q4 2024 call) | Engineering, Dec 2024 |
| GEM (Generative Ads Recommendation Model) | Meta's "largest foundation model for recommendation systems (RecSys) in the industry." It is too large to serve ads directly, so its knowledge is "transferred to hundreds of user-facing vertical models" that rank ads | Meta's blog attributes a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in Q2 2025 to GEM | Engineering, Nov 2025 |
| Lattice | A model architecture that replaces many smaller, separate ranking models with fewer, larger ones that learn across surfaces and goals | About 8% better ads quality on Instagram (2023); ads quality up almost 12% and conversions up to 6% (March 2025); roughly 100 fewer ranking models (Q3 2025 call) | Meta AI, May 2023 · Meta, March 2025 |
All figures are Meta's own, from the linked posts and earnings calls. Meta reports them for its systems overall, not for any individual advertiser.
How does Andromeda treat similar ads?
This is the part that changes your work. In December 2025 Meta explained that its system "groups together ads that share key visual and thematic attributes. When multiple ads look or feel alike, these are seen as variations of the same creative, meaning learnings and delivery optimizations are shared at the creative level" (Meta, Demystifying Creative Diversification). It adds that this applies "when ad images are visually too alike, even if technically different assets."
In practice: ten resizes, recolours or caption swaps of one video are, to Meta, roughly one creative. They don't give the system ten different things to match to ten different people. Meta's ask is ads that are "truly distinct in their visuals, messaging, and formats", made for "different personas or use cases." Meta gives no fixed number of ads or concepts per ad set; its own wording moves away from "a fixed volume of ad creatives per campaign."
You can see both patterns in the Ad Library. During Black Friday 2025, True Classic ran its main sale creative and text as up to 25 separate ads (one of them); by Meta's description, ads that alike are variations of one creative, however many ad IDs they have. Thesis, a nootropics brand, shows the other approach. As of October 13, 2026 it has three live ads with different people, settings and first lines: a podcast-style testimonial, a creator's acted "can't lock in" skit and a creator reading reviews to camera. Those are three concepts.
How to tell when two ads are different enough, and how many concepts to run, is covered in our creative diversity playbook.
What should advertisers watch for under Andromeda?
None of these is new on its own. What changed is how much they cost you when the system is built to group look-alike ads and pick from a much larger pool.
Volume that is really one ad
Thirty ads in an ad set can still be one idea. If the hook, the visual style and the message are the same, Meta treats them as variations of one creative, and you're paying production costs for reach you won't get.
Edits that reset learning
Meta lists "adding a new ad to your ad set" as a significant edit that can send an ad set back into learning, alongside changes to targeting, creative or bid strategy. Dripping in one new ad a day means repeated resets. Launch new creative in batches.
New ads getting no spend
Operators commonly report that when an ad set already has a proven ad, a new one gets little delivery. Meta doesn't document this, and it isn't necessarily a verdict on the new ad. If you need a fair read, test it in a separate, budget-controlled campaign rather than inside the winner's ad set.
Weak conversion signals
The ranking models learn from your conversion events. Meta's own list of reasons for high costs includes this one: "Issues with your data source setup can lead to a lower volume of attributed events and an increased cost per action" (Meta). Check pixel and Conversions API health before you blame the creative.
Enhancements you didn't approve
Advantage+ creative enhancements can change how your ad appears from impression to impression. Before you judge or scale a winner, check which enhancements are on, so you know which version is actually spending. See our Advantage+ creative guide.
Over-reading one bad week
Andromeda rolled out across surfaces over months, and Meta keeps changing the ranking models behind it. A bad week can be the market, a tracking issue or a fatigued ad. Check whether every campaign moved on the same dates before you rebuild anything.
How to adapt your Meta ads to Andromeda
A short version. Each step links to the page that covers it in depth.
1. Audit what you're already running
Group your live ads by the idea behind them, not by file. If most of your ads share one hook and one look, you have fewer creatives than you think.
2. Plan by persona or use case
For each product, list the different people who buy it and why. Build one genuinely different concept per reason to buy: different story, visual and message, not a new crop. Our ad angles guide covers where those reasons come from.
3. Vary the format, too
Give each concept at least two formats, such as a static and a short video, or a UGC talking head and a product demo. Meta asks for ads that are distinct in format as well as message.
4. Test concepts outside your winners
Launch new concepts in batches in a separate test campaign, judge them over several days, and move the winners into your main campaign. The creative testing playbook has the structure.
5. Keep the pipeline full
Distinct concepts wear out too. Keep the next set of angles ready before the current winner fatigues, so you're not starting from a blank page.
Sources
Meta Engineering: Meta Andromeda (December 2, 2024) · GEM (November 10, 2025). Meta AI: Lattice (May 11, 2023). Meta for Business: AI innovation in Meta's ads ranking (March 27, 2025) · Demystifying Creative Diversification (December 16, 2025).
Meta earnings call transcripts: Q4 2024 (January 29, 2025) · Q2 2025 · Q3 2025. Meta Business Help: significant edits · your costs are too high. All accessed October 13, 2026.
Common questions.
Andromeda is Meta's ads retrieval engine. For each impression it selects a few thousand relevant candidates from tens of millions of eligible ads, and ranking models then pick the ad that is shown. Meta introduced it in the second half of 2024 and built it to handle a fast-growing volume of ad creative.
Meta describes it as introduced in the second half of 2024; its engineering post about Andromeda was published on December 2, 2024. Meta then expanded it to more surfaces, including Facebook Reels in Q2 2025. Meta hasn't given a single date when it reached every advertiser.
They work at different stages. Andromeda handles retrieval, narrowing tens of millions of ads to a few thousand candidates. GEM is Meta's large foundation model for ads ranking; it is too big to serve ads directly, so Meta transfers its knowledge into hundreds of smaller models that rank the candidates Andromeda retrieves.
Meta doesn't give a number. Its guidance is about distinctness, not count: ads that look or feel alike are treated as variations of the same creative and share learning. A few genuinely different concepts, each in more than one format, do more than a large batch of near-copies.
Not exactly, and "creative is the new targeting" is an industry phrase rather than a Meta quote. What Meta does say is that the system matches ads to people automatically and groups similar ads together, so the differences between your creatives decide how many different people the system can match them to.
Stop making variations. Start making different ads.
Andromeda rewards ads that are genuinely different. Finding differences that work is the hard part.
The same research as the callout above, aimed at your product, so every new concept starts from evidence instead of another variation.
- Competitors' live Meta ads, with days-running data
- Your niche's organic TikTok and Reels winners
- Distinct angles in your customers' own words
- Built cold from your product URL
Plans from $99/mo, flat — see ViralMojo pricing.