Meta Andromeda is the AI system that decides which ads even get considered for you to see. It is the first, “retrieval” stage of Meta’s ad delivery: it shortlists candidates from every eligible ad before ranking models pick the winner. On 28 October 2026, Meta detailed a rebuilt Andromeda that it says scores over a billion ads per second per GPU server. For advertisers, Meta’s direction points to creative variety as a bigger lever than narrow manual targeting.
Key takeaways
- What it is: Andromeda is Meta’s machine learning system for personalized ads retrieval, the first stage of its ads recommendation pipeline (Engineering at Meta, 28 October 2026).
- What changed: Meta says the 2026 version processes over one billion ads per second on a single GPU server while scoring every eligible ad, about 100x the throughput of the 2024 system.
- Meta’s results claim: “Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook.” Meta gives no baseline or method for this figure.
- What to do: feed the system genuinely different creative, give each ad set enough conversions to learn from, and use Ads Manager’s new creative Breakdown to see which images, videos and text variations work.
What is Meta Andromeda?
Every time someone opens Facebook or Instagram, Meta has to choose a handful of ads from an enormous pool of eligible ones, in a fraction of a second. It does this in stages. Retrieval comes first: it narrows the huge pool down to a short list of strong candidates. Ranking comes next: larger, more expensive models score that short list in finer detail and decide what actually appears.
Meta Andromeda is the retrieval stage. Meta first described it on 2 December 2024 as a “proprietary machine learning (ML) system design for retrieval in ad recommendation.” At that launch, Meta said retrieval narrows tens of millions of ad candidates to a few thousand, and that Andromeda was co-designed with NVIDIA Grace Hopper hardware and Meta’s own MTIA chips. Meta reported a +6% recall improvement and a +8% ads quality improvement on selected segments at the time.
Why it matters to you: if your ad never makes the retrieval short list, the ranking stage never sees it, no matter how good your bid or offer is. Andromeda is the gatekeeper.
What changed in the October 2026 Meta Andromeda update?
On 28 October 2026, Meta’s engineering team published a detailed post on what it calls “the 2026 Meta Andromeda.” The headline claims, all Meta’s own figures:
- Scale: over one billion ads per second on a single GPU server, scoring every eligible ad in the candidate pool. Meta says that is a 100x lift over the 2024 system, which handled tens of millions of ads per second per GPU card.
- Smarter model, less compute: a new “mixture-of-experts” design with several specialized user models, which Meta says uses about 300x fewer inference-time FLOPs than the 2024 retrieval model.
- One pipeline on one server: index matching, scoring and top-K selection now run together on a single GPU server, which Meta says cuts network traffic by about 100x.
- NVIDIA collaboration: Meta says it worked with NVIDIA from early 2026 on optimized GPU kernels, with 2x to 5x speedups per component (Meta’s benchmark figures).
- Conversions: “Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook.” As PPC Land points out, Meta gives no baseline, period definition or method for that number.
Meta also explains why it needed the rebuild: generative AI creative tools have helped advertisers create hundreds of millions of new ads, so the pool of candidates has grown to an “unprecedented scale.” Meta says the pool continues to grow, driven by Advantage+ automation and generative AI creative tools. In other words, Andromeda was rebuilt to handle far more ad variations, not fewer.
| Meta Andromeda | 2024 launch | 2026 version (Meta’s figures) |
|---|---|---|
| Role | Retrieval: narrows tens of millions of candidates to a few thousand | Retrieval that scores every eligible ad “with near ranking-level model performance” |
| Throughput | Tens of millions of ads per second per GPU (figure Meta gave in 2026) | Over 1 billion ads per second per GPU server (about 100x) |
| Model design | Custom deep neural network, hierarchical index | Mixture-of-experts with specialized user towers; ~300x fewer inference FLOPs |
| Reported impact | +6% recall, +8% ads quality on selected segments | +2.4% conversions on Facebook “since 2025” |
How does Andromeda work with GEM and GenRecs?
Andromeda is one part of a bigger system. On 28 October 2026, Social Media Today reported Meta’s update on three AI systems in its ads stack. Meta said Andromeda can now consider a wider breadth of potential product matches for each business, and that GEM (its Generative Ads Recommendation Model) can process even more data. Social Media Today adds that GEM assesses people’s ad engagement history and more signals, and that GenRecs can now better match ads to people’s preferences, “even with brand-new ads with little engagement history.”
| System | What Meta says it does | Latest number Meta has shared |
|---|---|---|
| Andromeda | Retrieval: shortlists candidate ads for each person | Q3 2026 improvements: +1.4% conversions on Facebook |
| GEM | Processes even more data to judge relevance | No separate figure in this update |
| GenRecs | Matches ads to preferences, including new ads with little history | No separate figure in this update |
| Generative retrieval (Facebook Feed) | Launched in Q3 2026 | Nearly 2% more conversions across targeted segments |
| Ads ranking overall | Combined ranking improvements | Q2 2026: 8.3% more ad clicks and 15.7% more conversions on Facebook |
All of these are Meta’s own figures, reported without independent verification. They describe averages across Meta’s whole ad system, not what any single account will see.
Does Meta Andromeda make ads more expensive?
Andromeda doesn’t set prices; the auction does. But it is worth knowing the context. In its third-quarter 2026 results, published on 21 October 2026, Meta reported that ad impressions grew 10% year over year and the average price per ad rose 14%. Revenue was $64.60 billion, up 26%.
In our view, that combination means you can’t count on cheap reach. The practical lever you control is relevance: ads that match the right people tend to convert better for the money you spend, and better retrieval only helps if you give the system good, varied options to match.
What should advertisers do differently because of Andromeda?
Meta’s own messaging points in one direction. Its December 2024 Andromeda post says “increased ad diversity can improve people’s experience with ads and drive better advertiser outcomes.” Meta Ads specialist Jon Loomer sums up the official recommendation as: creative diversification will help advertisers get the best results. Here is how we would turn that into action:
- Diversify concepts, not just colours. Five versions of the same image with different backgrounds give the system little to work with. Test different angles: a demo video, a customer story, a price-and-benefit static, a founder talking to camera, a carousel of use cases.
- Mix formats. Include vertical video for Reels and Stories alongside images, so retrieval has options for different placements and people.
- Give each ad set enough data. In a guide reported on 25 October 2026, Meta recommended “roughly fifty conversions a week” per ad set and said ad sets below that ran about 28% more expensive per purchase in its analysis. In our view, consolidating small ad sets is better than splitting budget thinly. Our guide to the Meta Ads learning phase explains why.
- Stop over-editing. Summarising the same Meta guide, Social Media Today notes that many advertisers react to performance swings too early, which gives the system less time to learn. Decide your review dates in advance.
- Read the new creative Breakdown. Social Media Today reported on 27 October 2026 that Ads Manager’s Breakdown can show how each image or video performed when delivered as single media, and how each text variation performed. Ads opted into multi-post displays such as “Show as carousel” or “Show as collection” aren’t included.
- Guide, don’t micromanage, targeting. Meta’s Marketing API now lets advertisers set an age and gender audience for each creative element (mainly via third-party tools). See our explainer on choosing an audience for each ad creative.
- Test properly. Use A/B tests or Conversion Lift studies when you need a clear answer, as Meta itself suggests. Our A/B testing guide covers the basics.
Is Meta Andromeda a reason to stop using detailed targeting?
Not automatically. Andromeda is built to find the right people from the signals it has, and Meta’s direction is clearly toward broader delivery guided by creative. In our view, the safest approach for most small and mid-sized advertisers in Lebanon and the Gulf is to test broad or Advantage+ audiences against your current setup with a clean A/B test, rather than switching everything overnight. Budget structure matters too; see CBO vs ABO in Meta Ads.
What should you do now?
- Audit your active ad sets: how many truly different creative concepts does each have?
- Plan at least three new concepts (not variations) for your next campaign, in at least two formats.
- Check each ad set’s weekly conversions against Meta’s “roughly fifty” guideline and merge ad sets that fall short.
- Open the creative Breakdown in Ads Manager and note your top and bottom performing images, videos and text.
- Set fixed review dates and avoid daily edits during learning.
- Review the Advantage+ creative AI terms before you lean harder on Meta’s generative tools, and see what Meta’s AI business assistant can now do in Ads Manager.
FAQ
What is Meta Andromeda in simple words?
It is the AI system Meta uses for the first step of ad delivery, called retrieval. It quickly shortlists which of the eligible ads could be shown to a person, before ranking models choose the final ads. If your ad isn’t shortlisted, it can’t win the auction for that person.
When did Meta launch Andromeda?
Meta first described Andromeda on 2 December 2024. On 28 October 2026, it published details of a rebuilt version it calls “the 2026 Meta Andromeda.”
What is new in the October 2026 Andromeda update?
Meta says the new version scores over one billion ads per second on a single GPU server while scoring every eligible ad, about 100x the 2024 system’s throughput. It uses a mixture-of-experts model that Meta says needs about 300x fewer inference FLOPs than the 2024 model.
Does Andromeda improve ad results?
Meta says Andromeda has delivered a 2.4% increase in conversions on Facebook since 2025, and that Andromeda improvements in Q3 2026 increased Facebook conversions by 1.4%. These are Meta’s own figures without a published baseline, and results in individual accounts will vary.
What is the difference between Andromeda and GEM?
Andromeda handles retrieval, shortlisting candidate ads. GEM, Meta’s Generative Ads Recommendation Model, judges relevance using people’s ad engagement history and other signals, according to Social Media Today. Both work alongside GenRecs, which helps match new ads with little engagement history.
How many creatives should I run for Andromeda?
Meta doesn’t publish a magic number. Meta has said increased ad diversity can drive better advertiser outcomes, so focus on several genuinely different concepts and formats rather than many near-identical copies. Meta also recommends each ad set gets roughly fifty conversions a week.
How can I see which creative works best now?
Ads Manager’s Breakdown option, reported on 27 October 2026, shows how individual images and videos performed when delivered as single media and how each text variation performed. Ads opted into multi-post displays such as “Show as carousel” or “Show as collection” aren’t included.
Are Meta ads getting more expensive?
Meta’s Q3 2026 results show the average price per ad rose 14% year over year while ad impressions grew 10%. Andromeda doesn’t set prices, but better-matched, varied creative is the main lever you control to get more results for the same spend.
Sources
- Engineering at Meta: Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design (28 October 2026)
- Engineering at Meta: Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine (2 December 2024)
- Social Media Today: Meta says its expanded processing capacity is driving ad results (28 October 2026)
- PPC Land: Meta says Andromeda lifted Facebook conversions 2.4% since 2025 (28 October 2026)
- Social Media Today: Meta adds individual performance data for ad creative (27 October 2026)
- Social Media Today: Meta says advertisers should let campaigns gather data (25 October 2026)
- Social Media Today: Meta offers creative targeting for ads (21 October 2026)
- Meta via PR Newswire: Meta Reports Third Quarter 2026 Results (21 October 2026)
- Jon Loomer (jonloomer.com): Meta Andromeda creative diversification (accessed 29 October 2026)
Information as published by Meta and the sources above as of 29 October 2026. All performance figures are Meta’s own. Meta rolls out Ads Manager features gradually, so what you see in your account may differ.
Learn to run Meta Ads that work with Meta’s AI
Gradezilla’s Digital Marketing & AI Course covers Meta Ads, tracking, creative testing and AI tools in live online sessions with Bashar Amin, including preparation for the Meta Certified Digital Marketing Associate exam. Classes run over 8 weeks on Tuesdays and Thursdays, 6–9 PM, and every session is recorded. Check the course page for the next start date.
Written by the Gradezilla team · Reviewed by Bashar Amin