The Meta Andromeda update of 28 October 2026: why creative variation wins
On 28 October 2026 Meta explained how it has rebuilt Andromeda, the engine that decides which ads even get to compete for a spot in your feed. No new button in Ads Manager, but new rules of the game. Here is what actually changed, what Meta itself says and what we see in the accounts we run.
TL;DR – the essentials in 30 seconds
- What: Meta has rebuilt Andromeda, the first step in choosing an ad, as a "Mixture-of-Experts" model: several specialised experts on different kinds of users instead of one general model.
- The numbers (Meta's own): about 300 times less computation than the 2024 version, over one billion ads per second per server and +2.4% conversions on Facebook since 2025.
- What it means for you: ads that say different things reach different people. The same image in five colours does not.
- Our take: if you can't produce creative variation, you're losing ground now. We see rising CPM and falling CTR in accounts that cling to a rigid visual identity, and we believe smaller, faster companies can take market share from bigger ones.
What happened on 28 October 2026?
Meta's engineers published a technical deep dive into the new Andromeda on their engineering blog. Andromeda launched in December 2024 and is the part of Meta's system that does the rough sorting: out of all the ads that could be shown to a person, it picks a shortlist. Only then do the heavier ranking models take over and decide which ad finally wins.
Why rebuild it? The number of ads is exploding. Meta itself writes that Advantage+ and its generative AI tools have helped advertisers create hundreds of millions of new ads. The more ads there are, the more expensive it gets to evaluate them all, and the more it matters that the rough sorting is accurate.
- All adsthat may be shown to you (your age, your country, the advertisers' settings).
- 1. Andromedasorts at lightning speed and picks a shortlist. This is where the update happens.
- 2. GEMMeta's big ranking model scores the shortlist carefully: how likely is a click and a purchase?
- 3. Auctionbid, likelihood and quality are weighed together. One ad wins the spot.
How the new Andromeda works – without the engineering jargon
The old Andromeda basically had one model describing you as a user and one describing each ad, and then worked out how well they matched. The problem: a single description of "the user" easily becomes a compromise.
The new version borrows a trick from large language models like ChatGPT, called Mixture-of-Experts. Instead of one general user model there is now a group of specialised experts which, according to Meta, read different signals: which ads you have interacted with and in what order, which groups of users you resemble, the meaning of content you engage with (interpreted by language models) and how you behave across Facebook, Instagram and Reels.
Between the ads and the experts sits a router. For each ad it decides which experts are the most relevant to ask. The result, according to Meta, is a first sort that is almost as accurate as the heavy ranking – at a fraction of the cost.
- Expert ATime-pressed parents of young kids
- Expert BBargain hunters
- Expert CQuality nerds who read reviews
- Expert DReels scrollers who buy on impulse
- Expert EGift buyers
This is our interpretation, not something Meta spells out: if the router picks experts based on what an ad means, then five ads with the same message in different colours look like one and the same ad. They end up with the same experts, reach the same people and compete with each other. Five ads with different messages, on the other hand, open five different doors.
So what do five different concepts look like in practice? Here is the same made-up product – the BRYGG coffee maker – in the five concepts you saw in the router. Click through and compare what actually sets them apart.
Five ads – five different reasons to buy, formats, senders and moods. Not the same ad in five colours.
The numbers: 2024 vs 2026
All figures below come from Meta's own engineering blog. No external party has reviewed them.
| What | 2024 | 28 Oct 2026 | What it means |
|---|---|---|---|
| Capacity | Tens of millions of ads/sec per GPU card | Over 1 billion ads/sec per GPU server | More ads can be evaluated for every impression |
| Compute cost | Baseline | About 300 times less computation | Room for smarter models without slowing down |
| The scoring itself | 2.4 billion ads/sec | Over 10 billion ads/sec on one GH200 (5.3x, optimised together with NVIDIA) | Faster scoring of every ad |
| Rule filtering | – | Over 100 billion ads/sec | Advertiser settings are checked faster |
| Network traffic | Candidates are sent between servers | Everything happens on one server, about 100 times less traffic | Cheaper and more stable for Meta |
| Precision | 16-bit computation | 8-bit, up to 2x faster with comparable accuracy | More done with the same hardware |
An important nuance: most of this is efficiency for Meta. Their servers getting 100 times faster does not mean your ROAS gets 100 times better. The figure closest to your day-to-day is +2.4% conversions on Facebook since 2025 – an average across the whole platform, not a promise for your account. Meta also doesn't say exactly when the different parts went into use or whether all advertisers got them at the same time.
Andromeda, GEM and GenRecs: three parts of the same engine
Andromeda didn't come alone. In October 2026 Meta published three technical posts that belong together:
GEM – the brain of the ranking
Generative Ads Model is Meta's big central model that learns from Facebook, Instagram and Reels at the same time. Meta reports 18.9% better conversion and 12% higher CTR in its own measurements. GEM is too heavy to review every ad – which is why Andromeda is needed first.
GenRecs – new ads no longer start from zero
GenRecs Tokenization replaces anonymous ID numbers with a stable "vocabulary" describing what an ad or product actually is. Together with GEM and Meta's Adaptive Ranking Model: +0.35% conversions across all surfaces.
Andromeda – the first sort
Experts, a router and new hardware-level code mean more ads can be evaluated more intelligently for every impression. Meta is already moving on to NVIDIA's next platform, Grace Blackwell.
- Last season’s running shoe12,400 clicks · 380 sales
- Men’s trail shoe8,100 clicks · 210 sales
- Running jacket5,600 clicks · 140 sales
- Frying pan9,900 clicks · 300 sales
For e-commerce this may be the most concrete news. A new product in the catalogue used to be a blank slate, and the algorithm happily kept pushing old bestsellers with a long history. When the product is instead described with a shared vocabulary, it can borrow knowledge from similar products straight away. That makes good product data – images, titles, descriptions, categories – worth more than ever.
Our take: if you can't create variants, you're out
This is not an update you can wait out. Meta has built an engine that looks for the right ad for the right person, and that engine needs fuel in the form of different creative ideas. Those who can't deliver that will lose to competitors who find it easier.
We see it clearly. Accounts with real creative variation – different angles, different formats, different senders – get better results. Accounts that hold on to a rigid visual identity and don't dare to experiment with their Meta ads are losing, and it mainly shows in two places: CPM goes up and CTR goes down.
And here is an opportunity many miss: this is a bigger shift where smaller companies can take market share from larger ones. A small team that can shoot an honest phone video today and test three new angles tomorrow suddenly has an edge over a big organisation where every ad has to pass three approvals. The algorithm doesn't care how big your brand book is.
Observations from the Meta accounts our Meta team (Sandra and Therese) run every day – not a controlled study. Results vary between industries and accounts.
What to do now
There is nothing to switch on in Ads Manager – the change happens inside Meta's own systems. What you control is what you feed the system. This is how we prioritise in the accounts we run in paid social:
Different ideas, not different colours
Test different reasons to buy (save time vs save money), different use cases (everyday vs special occasions), different proof (product demo vs customer video) and different offers (single product vs bundle). Let every ad say something of its own.
Make your product data rich
Good images, clear titles, complete descriptions and the right categories. The better the system understands what you sell, the faster new products can find buyers.
Fewer, broader campaigns
Lots of small, narrow ad sets give the system less to choose from. Consolidate the budget and let Advantage+ do its job. Hard limits like country and age apply just as before.
Measure what pays the salaries
Send purchases via the Conversions API, track CPM, CTR and frequency per concept – and evaluate against customer acquisition cost and contribution margin, not just ROAS in Ads Manager.
| Area | What many did before | What we do now |
|---|---|---|
| Creative | Lots of near-identical variants (colour, font, button text) | Fewer but clearly different concepts, angles and formats |
| Catalogue | Focus on a few bestsellers | The full range with rich product data |
| Structure | One ad set per interest, age and placement | Consolidated Advantage+ campaigns with broad targeting |
| Follow-up | CPC, CTR and short-term ROAS | CAC, customer quality and contribution margin – plus CPM and CTR per concept as an early warning signal |
The Andromeda checklist
0/8 DONEWhat the update doesn't mean
- Not a new setting. There is nothing to switch on. Campaigns that work well shouldn't be rebuilt in a panic.
- Not a guarantee. +2.4% is an average for all of Facebook. Your account may do better or worse.
- Not free reach. As the system gets better at finding relevant ads, more good ads meet in the auction. Competition gets tougher for those who stand still.
- Not the end of your brand. A clear identity is still valuable. But it should make room for variation – not stop it.
Frequently asked questions about Meta Andromeda
What is Meta Andromeda?
Andromeda is the part of Meta's ad system that makes the first selection: out of all the ads that could be shown to a person, it picks a shortlist that is then ranked more carefully. Andromeda launched in December 2024 and was rebuilt according to Meta's write-up of 28 October 2026.
What changed in Andromeda on 28 October 2026?
Meta has moved to a Mixture-of-Experts model with several specialised user experts and a router that chooses which experts evaluate each ad. At the same time the code has been optimised for NVIDIA's GH200 hardware so the whole selection runs on one server. According to Meta this means about 300 times less computation, over one billion ads per second per server and +2.4% conversions on Facebook since 2025.
Do I need to change anything in Ads Manager?
No. There are no new settings – the change happens inside Meta's own systems. What affects your results is which ads, which product data and which conversion signals you give the system.
Why does creative variation matter more with Andromeda?
When the system matches ads to different kinds of users based on what the ad is actually about, different messages reach different people. Ads that only differ in colour or font effectively look like the same ad and compete with each other. That is our interpretation of how the system works, and it matches what we see in the accounts we run.
Is it enough to change the background colour or headline?
Rarely. Small changes can still affect how an ad performs, but they don't open up new audiences. Go for a few clearly different ideas instead: different reasons to buy, different use cases, different proof and different offers.
Does Andromeda favour smaller companies?
We think so. Speed and creative range weigh more than budget size when the system looks for the right ad for the right person. A small team testing new ideas every week can take market share from bigger players with slow approval processes.
Why is my CPM rising and my CTR falling?
There can be many reasons – season, competition, tired ads. A pattern we often see is accounts with little creative variation: the same message is shown over and over to the same people, interest drops and it gets more expensive to get through. Start by checking how many genuinely different concepts you have running.
How does Andromeda relate to GEM and GenRecs?
Andromeda does the first sort, GEM (Generative Ads Model) ranks the shortlist and GenRecs Tokenization gives ads and products a shared vocabulary so new ads can benefit from knowledge about similar ones. Meta published technical write-ups of all three in October 2026.
Sources
- Meta Engineering, "Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design", 28 October 2026.
- Meta Engineering, GenRecs Tokenization – transforming ads recommendation with a stable model vocabulary, 19 October 2026.
- Meta Engineering, "Meta's Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI", 13 October 2026.
- Meta Engineering, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine", 2 December 2024.
Figures about Meta's systems are taken directly from the sources above. Advice and opinions in this article are Zkond's own.