---
title: "Meta Andromeda update 2026: how it works"
description: "Meta rebuilt Andromeda, the system that shortlists which ads you might see. What changed in October 2026, how it works, and what to do with your Meta ads."
url: https://dumbcmo.com/blog/meta-andromeda-2026-how-it-works
category: "Paying to be seen"
tags: ["news", "meta ads", "facebook ads", "instagram ads", "andromeda", "advantage+", "ai"]
published: 2026-11-03
updated: 2026-11-03
reading_time: 16 min
author: "The DumbCMO"
---

# Meta Andromeda update 2026: how it works

> **TL;DR:** Andromeda is Meta’s first filter: it picks a few thousand ads worth judging from millions. The 2026 version scores every eligible ad, and Meta credits it with 2.4% more Facebook conversions since 2025. No new button. Your job: genuinely different ads, broad targeting, clean signals.

[Notebook](https://dumbcmo.com) › [Blog](https://dumbcmo.com/blog) › Paying to be seen › Meta Andromeda update 2026: how it works

Every time someone in Pune or Pittsburgh opens Facebook or Instagram, Meta has a split second to decide which ad to show them. There are millions of ads it could pick. Yours is one of them.

On 28 October 2026, Meta published a long engineering post about how it rebuilt the part of that system that makes the first cut. It’s called Andromeda. The post is full of words like “mixture-of-experts” and “fused GPU kernel”, and LinkedIn has already turned it into fifty carousels titled “Andromeda changes EVERYTHING”.

It doesn’t change everything. But it does explain, quite clearly, why some things that used to work on Meta ads don’t work as well anymore. Let’s go through it slowly.

## In normal-people words

**Andromeda** is the first filter in Meta’s ad system. When someone opens the app, it narrows millions of possible ads down to a shortlist of a few thousand. Only ads on that shortlist get judged properly and can win the slot.

**What changed in 2026:** Meta rebuilt it so it can properly score *every* eligible ad, not just the obvious ones, at over one billion ads per second on a single server. Meta says Andromeda has delivered **2.4% more conversions on Facebook since 2025**.

**What you have to do:** nothing to switch on. There is no new setting. But it rewards advertisers who give it **genuinely different ads**, broad audiences and good conversion data, and quietly ignores ten versions of the same ad.

> **The DumbCMO:** No new button. No new feature to buy. Just a better shortlist, and a stronger reason to stop making ten copies of one ad.

## What Andromeda actually is

Meta picks ads in stages, a bit like hiring.

First, someone skims a huge pile of CVs and keeps a shortlist. Then the shortlist gets proper interviews. Then someone gets the job.

In Meta’s system:

- **Stage 1 is retrieval (The first, fast step that pulls a shortlist of possible ads out of millions).** This is Andromeda. Meta says it narrows “a pool of tens of millions of ads down to the shortlist we consider showing someone”.
- **Stage 2 is ranking (The slower, more careful step that predicts how likely each shortlisted ad is to get a click or sale, and orders them).** Bigger, more expensive models (Meta’s main one is called GEM) score the shortlist in detail and decide what wins.

The key point: **if your ad doesn’t make Andromeda’s shortlist, the clever ranking models never even see it.** It doesn’t matter how good your ad is.

## The chai stall version

Imagine a wedding caterer with 10,000 dishes in the recipe book. The guest walks up to the counter. There’s no time to cook all 10,000 and let them taste.

So an assistant glances at the guest, pulls out 30 dishes that might suit, and only those 30 go to the head chef, who picks the best three.

Andromeda is the assistant. The head chef is ranking. **If the assistant never pulls your dish, the head chef never tastes it.**

Meta’s own analogy, from its business blog: Andromeda is a personal concierge. The old concierge checked the popular shelves they knew best. The 2026 concierge “now walk[s] the entire store” and reads every label, in the same amount of time.

## What’s new in 2026

Meta first announced Andromeda in December 2024. That version already used AI models 10,000 times more complex than what came before, and Meta reported a 6% improvement in recall (finding the right ads) and 8% better ad quality on selected segments.

The October 2026 version is a rebuild. Here’s what Meta says changed, in plain words.

**Andromeda 2024 vs 2026, per Meta’s engineering posts**

| What | 2024 version | 2026 version |
| --- | --- | --- |
| How much it looks at | Narrowed the search with a layered index (Meta: checked “the most popular shelves”) | Scores every eligible ad |
| Speed | Tens of millions of ads per second per GPU card | Over one billion ads per second per server, about 100x the 2024 system |
| Model design | One general model for everything | A team of specialist models (mixture-of-experts), near ranking-level quality |
| Cost to run | Baseline | About 300x fewer computer operations per prediction than the 2024 model |
| Where the work happens | Targeting checks and scoring on separate machines, talking over the network | Targeting checks, scoring and picking the top ads all on one GPU server |
| Result Meta reports | +6% recall, +8% ads quality on selected segments | +2.4% Facebook conversions since 2025 |

**Why Meta rebuilt it.** Meta says its AI creative tools helped advertisers make “hundreds of millions of new ads”. More ads means a bigger pile to search, in the same split second. The old system handled tens of millions of ads per second on one GPU card. Meta wanted billions.

**The newer numbers.** In a business-facing post at the same time, Meta added that in July to September 2026 it made Andromeda 10 times more complex “without increasing computing costs or hardware needs”. That lifted conversions by 1.37% on Facebook and 0.24% on Instagram Reels. On its earnings call, Meta rounded the Facebook figure to 1.4%.

**Built with NVIDIA.** From early 2026, Meta and NVIDIA worked together and, in tests, sped up each part of the pipeline 2x to 5x. Those test gains aren’t all in live use yet. Work is now moving to NVIDIA’s newer Grace Blackwell chips. Interesting if you sell GPUs. Not something you need to act on.

> **The DumbCMO:** “Fused kernel co-design” means they made the computer do the work in one trip instead of five. That’s it. You can go back to your ads.

### A note on the big numbers

You’ll see “1 billion”, “10 billion” and “100 billion ads per second” quoted side by side. They measure different things:

- **Over 1 billion per second:** the whole system in real use, end to end.
- **Over 10 billion per second:** just the scoring part, tested on one NVIDIA GH200 chip.
- **Over 100 billion per second:** just the step that checks targeting rules.

None of these is a promise about your results. They’re speed records, not sales figures.

## How it works, step by step

Here is the journey of one ad request, based on Meta’s 2026 engineering post. Simplified, but not made up.

**What happens in the split second after someone opens Instagram**

1. **Someone opens the app.** Meta gets a request: this person, this place in the app, right now.
2. **Build a picture of the person.** Several “expert” models each look at the person differently: what ads they’ve engaged with lately, what similar groups of people like, what topics they care about, and what they do across Meta’s apps.
3. **Check who’s allowed.** Every ad’s targeting rules (location, age, exclusions and so on) are checked. Ads that aren’t eligible for this person drop out here.
4. **Score every eligible ad.** Each remaining ad gets a quick score, with live context like which app and placement this is. A small gate decides how much to trust each expert for this person and ad.
5. **Keep the top few thousand.** The highest-scoring ads become the shortlist. All of this happens on one GPU server, in milliseconds.
6. **Hand over to ranking.** Heavier models, like GEM, rank the shortlist properly. Then the auction picks the winner.

## Jargon Jail

**Locked up: “Mixture-of-experts”**

An AI setup where, instead of one giant model doing everything, several specialist models each give an opinion and a small “gate” decides whose opinion counts most for this case. **Think of it as asking five shopkeepers instead of one, and listening hardest to the one who knows this customer.**

## What it means for your ads

Important: Meta’s engineering post is about computers, not campaign tips. It doesn’t tell you how to run ads. Below, we keep two things apart:

- **What Meta says**, with a source.
- **Our read**, which is inference from how the system works and from what experienced advertisers report. Treat that as reasoned opinion, not Meta’s words.

### 1. Different ads beat more of the same ad

**What Meta says:** its system “groups together ads that share key visual and thematic attributes”. When ads look or feel alike, they’re treated as variations of the same creative, and learnings are shared “at the creative level – not just the individual ad level”. Meta separates *creative iteration* (same visual, different button text) from *creative diversification* (genuinely different ads for different kinds of buyers). In April 2025, Meta wrote that the focus has “shifted from niche targeting to creative diversification as the best lever”.

**Our read (inference):** ten ads with the same photo and different headlines are, to Meta, roughly one ad. They don’t give Andromeda ten chances to find ten kinds of buyer. Five ads that are actually different (a customer testimonial video, a product close-up, a price-led offer, a “how it works” demo, a founder talking) give it far more to work with.

**Same ad in a new shirt vs a genuinely different ad**

| Looks like variety but isn’t | Actually different |
| --- | --- |
| Same photo, five headlines | A video of a customer using it |
| Same video, different background music | A close-up product photo with the price |
| Same ad, red button vs green button | A short demo of how it works |
| Same model, three crops | An ad aimed at gift buyers, not self-buyers |

### 2. Broad targeting gives the system room

**What Meta says:** targeting rules are checked first, in what the engineering post calls index matching, which is “enforcing advertisers’ targeting rules”. Only then are eligible ads scored.

**Our read (inference):** so your targeting still matters. It decides which people your ad is even allowed to compete for. Stack narrow interests and you shrink the pool before Andromeda’s scoring ever starts. With broad targeting, the creative does more of the targeting: the ad itself signals who it’s for. Keep the hard limits that are real for your business, like the cities you deliver to.

### 3. Advantage+ is where Meta wants you

**What Meta says:** Meta ties eligible ad growth to “Advantage+ automation and generative AI creative tools”. It reported that advertisers new to Advantage+ sales campaigns who turned on its AI targeting saw on average 22% higher return on ad spend. It also says campaigns using its image generation see 11% higher click-through rate and 7.6% higher conversion rate.

**Our read (inference):** these are Meta’s averages, measured by Meta. Useful as a reason to test, not proof it will work for you. And if you use Meta’s AI to make variations, remember: [you’re responsible for what it writes](/blog/meta-ai-ad-edits-your-responsibility).

### 4. Fewer campaigns, more ads in each

**What Meta says:** Meta reports advertisers see on average a 9% lower cost per action with the new streamlined setup for Advantage+ sales campaigns. Its older best-practice advice for Advantage+ shopping campaigns: keep at least 20 diversified ads in the campaign, launch new ads regularly (leaders do it weekly), and put 20–30% of ad budget into testing new ads.

**Our read (inference):** splitting a small budget across eight campaigns with two ads each starves all of them. One or two campaigns with a healthy set of different ads usually gives the system more to learn from. If you spend ₹500 a day, 20 ads isn’t your starting point. Start with 4 to 6 truly different ones and add more as you learn.

### 5. Feed it good conversion data

**What Meta says:** the 2026 post lists the signals Andromeda uses about people, such as recent ad engagement and activity across Meta’s apps. It doesn’t talk about advertiser tracking setups.

**Our read (inference):** the system learns what “good” looks like from the results you report back. If your Pixel misses half your sales, or your Conversions API (Meta’s Conversions API: a way for your website or store to send sales data straight to Meta from your server, not just from the browser) isn’t set up, Meta is learning from a blurry picture. Check that purchases or leads show up in Events Manager, and optimise for the event you actually care about.

**Your account, after Andromeda**

- **If** You run 10 ads that are the same photo with different text **→ then** Replace most of them with 3 to 5 genuinely different ideas.
- **If** Your audience is five stacked interests **→ then** Test a broad audience with only the limits you truly need, like location.
- **If** You have many small campaigns for the same product **→ then** Consider merging them so each one has enough budget and ads to learn.
- **If** Your sales don’t show up properly in Events Manager **→ then** Fix tracking before you change anything else.
- **If** One ad is working well at a cost you can live with **→ then** Leave it running. Add new, different ads beside it. Don’t keep editing the winner.

## Myths and hype to ignore

Andromeda attracts a lot of confident nonsense. Here’s what to tune out.

**What you’ll hear vs what Meta has actually said**

| What you’ll hear | What’s actually true |
| --- | --- |
| Andromeda is a new campaign type you should switch to | It runs behind every Meta ad. There is nothing to turn on. |
| Andromeda is Meta’s ranking system | Meta calls it retrieval, the shortlisting step before ranking. Even some news coverage mixes these up. |
| You need 50 new ads a week now | Meta showed one brand going from 3–4 new ads a week to almost 50. One example, not a rule. |
| Targeting is dead | Meta still checks your targeting rules first. They just matter less as a way to find buyers. |
| Meta says you’ll get 2.4% more sales | Meta says 2.4% more Facebook conversions since 2025, platform-wide, with no baseline or method published. |
| Our agency has a secret Andromeda hack | If it isn’t better ads, cleaner data or a simpler setup, it’s probably a slide deck. |

> **The DumbCMO:** Anyone selling an “Andromeda-proof secret framework” is selling bullsh#t with a GPU on the cover.

## Things your agency won’t say

- 2.4% is Meta’s number, measured by Meta, with no method shared. It’s a direction, not a promise.
- More ads only helps if they’re actually different. Twenty lookalike ads are twenty invoices for one idea.
- A phone, a happy customer and ten minutes can make a genuinely different ad.

## Do this today

- [ ] Open Ads Manager and list your live ads. Circle the ones that are basically the same picture or video.
- [ ] For your main product, write down 3 to 5 different reasons people buy it, or kinds of people who buy it. Make one ad for each.
- [ ] Mix formats: at least one short video, one clear product image, and one with a real customer if you can.
- [ ] Check your audience. Remove interest stacks you can’t explain. Keep real limits like delivery area.
- [ ] Open Events Manager and confirm your purchases or leads are being recorded. Ask your developer about Conversions API if they’re not.
- [ ] Add one or two new, different ads a week, or every two weeks if that’s what you can manage. Leave winners alone.

## Not-so-frequently asked questions

### Do I need to change any setting for Andromeda?

No. Meta announced no new setting. Andromeda already runs on every ad.

### Does it work on Instagram too?

Yes, Meta says retrieval runs when someone opens Instagram or Facebook. Its Q3 2026 figures included 0.24% more conversions on Instagram Reels.

### Is this why my results changed recently?

Maybe, maybe not. Meta changes many parts of its system at once, including ranking. Check your own changes, tracking and season first.

### How many ads should a small business run?

Meta’s 20-ad guidance is for Advantage+ shopping campaigns. With a small budget, a handful of genuinely different ads is a sensible start. That’s our view, not Meta’s rule.

### Should I stop A/B testing?

No. Meta says small tweaks can inform new ideas but shouldn’t replace genuinely different ads.

### Is Andromeda the same as Advantage+?

No. Advantage+ is a set of automated campaign options you choose. Andromeda is the engine underneath that shortlists ads for everyone.

## How this differs from our October Meta roundup

Our [Meta ads October 2026 updates](/blog/meta-ads-october-2026-updates) post covers new ad formats, settings and dates you can act on. This note is about the engine underneath: why different creative and simpler setups matter, whatever formats you use.

## Where to go next

- What changed in Meta ads this month: [Meta ads October 2026 updates](/blog/meta-ads-october-2026-updates)
- If you use Meta’s AI tools: [Meta’s AI can now edit your ads](/blog/meta-ai-ad-edits-your-responsibility)
- Choosing a platform: [Instagram ads or Google ads?](/blog/instagram-ads-or-google-ads)
- Knowing what a customer is worth: [CAC and LTV explained](/blog/cac-and-ltv-explained)
- Everything on ads: [PPC, dumbed down](/ppc)

## Sources

1. Engineering at Meta, [“Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design”](https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/), 28 October 2026: over one billion ads per second, 100x lift, 2.4% Facebook conversions since 2025, about 300x fewer inference operations, “hundreds of millions of new ads”, index matching and co-location, NVIDIA speedups, 10 billion and 100 billion figures, Grace Blackwell.
2. Engineering at Meta, [“Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine”](https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/), 2 December 2024: 10,000x model complexity, +6% recall, +8% ads quality on selected segments.
3. Meta for Business, [“Inside the Three New AI Innovations in Ads Ranking Improving your Ad Performance”](https://www.facebook.com/business/news/three-new-ai-innovations-in-ads-ranking), October 2026: the concierge analogy, “tens of millions of ads” to a shortlist, 10x complexity in Q3 without more compute, 1.37% Facebook and 0.24% Instagram Reels conversions.
4. Meta, [Q3 2026 earnings call transcript](https://s21.q4cdn.com/399680738/files/content_files/META-Q3-2026-Earnings-Call-Transcript.pdf), 21 October 2026: Andromeda improvements “increased conversions by 1.4% on Facebook”.
5. Meta for Business, [“Demystifying Creative Diversification”](https://www.facebook.com/business/news/demystifying-creative-diversification), 16 December 2025: iteration vs diversification, grouping of similar ads, 11% CTR and 7.6% conversion rate with image generation, 22% higher ROAS with Advantage+ AI targeting.
6. Meta for Business, [“The Creative Advantage: Unlocking the Power of Diversification with Meta Andromeda”](https://www.facebook.com/business/news/the-creative-advantage-unlocking-the-power-of-diversification-with-meta-andromeda), 22 April 2025: “from niche targeting to creative diversification”, 9% lower cost per action with streamlined Advantage+ sales campaigns, the 3–4 to almost 50 ads a week example.
7. Meta for Business, [“Three steps to optimize your performance with creative diversification”](https://www.facebook.com/business/news/three-steps-to-optimize-your-performance-with-creative-diversification), holiday 2024: at least 20 diversified ads in Advantage+ shopping campaigns, 20–30% of budget for testing, weekly launches.
8. Social Media Today, [“Meta says its expanded processing capacity is driving ad results”](https://www.socialmediatoday.com/news/meta-says-its-expanded-processing-capacity-is-driving-ad-results/873133/), 28 October 2026 (describes Andromeda as a “ranking system”; Meta calls it retrieval).
9. PPC Land, [“Meta says Andromeda lifted Facebook conversions 2.4% since 2025”](https://ppc.land/meta-says-andromeda-lifted-facebook-conversions-2-4-since-2025/): notes no baseline, test design or Instagram figure for the 2.4% claim.
10. Search Engine Land, [“Inside Meta’s AI-driven advertising system: How Andromeda and GEM work together”](https://searchengineland.com/meta-ai-driven-advertising-system-andromeda-gem-468020), 28 January 2026: practitioner view on broader targeting, simpler account structures and creative-first strategy.

*DumbCMO is a free explainer site. We’re not affiliated with, paid by, or endorsed by Meta or NVIDIA. We don’t run ads for anyone. Numbers above are Meta’s own claims unless we say otherwise.*

## Related notes

- [Meta ads can show discounts you never approved](https://dumbcmo.com/blog/meta-ads-detected-offers-discounts): An ads specialist spotted a Commerce Manager setting, Detected offers, that lets Meta show discounts from your website in your Facebook and Instagram ads, even in copies of ads with promotions off. Meta hasn’t announced it. If you sell products on Meta, check it before copying Black Friday ads.
- [See the fake Google Ads clicks you didn’t pay for](https://dumbcmo.com/blog/google-ads-invalid-traffic-card): Google Ads is rolling out an Invalid traffic card on the Overview page. It shows estimated savings, the fake clicks filtered before billing, and credited cost, refunds for fake clicks found later. You’re not charged for invalid clicks. Check it before buying a click-fraud tool.
- [Google Ads learning phase is now 7 days](https://dumbcmo.com/blog/google-ads-learning-phase-7-days): Google’s help pages now say the Learning status lasts 7 days at most and shows only for a new bid strategy, added or removed campaigns, ad groups or keywords, or a big goal change. Budget, target, ad and targeting edits don’t trigger it. Results still take 1 to 2 conversion cycles to settle.

---

From [DumbCMO](https://dumbcmo.com): The marketing notes you were too embarrassed to take.
