What is new on Meta Andromeda 2.0?
On this page 7 sections
Key takeaways
- Andromeda shortlists ads before ranking and the auction even begin.
- The 2.0 update scores every eligible ad, about 100x faster than 2024.
- Your hook, visuals and offer now do most of the targeting work.
- Near-identical ads count as one, so test genuinely different angles.
- Consolidate ad sets, fix tracking and improve offers instead of discounting.
If you have searched for the Meta Andromeda 2.0 update and want a straight answer, here it is: Meta has rebuilt the first filter in its ad system so it scores every eligible ad in real time rather than cutting most of them out early. That changes what works in your account, and it has nothing to do with your targeting settings.
Marketers are calling it "Andromeda 2.0". Meta just calls it the next generation of Andromeda. The lesson for small businesses is the same: your creative now does most of the targeting work. Here is what changed, in plain English, and what it means for your Meta ads management.
What is Meta Andromeda?
Andromeda is the first filter in Meta's ad system: it builds a shortlist of ads that could suit a person before the final auction decides which one they actually see.
Every time someone opens Facebook or Instagram, there are tens of millions of ads they could be shown. Meta can't deeply assess all of them in a split second, so it works in stages:
Retrieval (Andromeda): narrows millions of eligible ads down to a few thousand likely matches.
Ranking: scores that shortlist in depth and predicts who is likely to click or buy.
Auction: combines those predictions with your bid and budget to pick the winning ad.
If your ad never makes the shortlist, your bid doesn't matter. That is why this first stage, launched in December 2024, has such a big effect on results.
Tap a step to see what happens there.
Step 1: Your ad (The creative)
Your creative is the targetingWhat it does
Meta reads the ad itself to work out who it suits. Your hook, visuals, words and offer tell the system far more about your audience than your interest settings do.
What it looks at
- The opening line
- Who appears on screen
- The product
- Tone and audio
What it means for you
Write for one specific person and name them in the first three seconds. Meta appears to treat near-identical ads as one, so ten colour versions of an ad are one test. Ten different angles are ten.
Step 2: Retrieval (Andromeda)
Builds the shortlistWhat it does
Narrows tens of millions of eligible ads down to a few thousand likely matches for this person. Since the 2026 update it scores every eligible ad, instead of cutting most of them out early with rough shortcuts.
What it looks at
- Every eligible ad
- Specialist models
- What the person is doing now
- Over a billion ads a second
What it means for you
If your ad doesn’t make the shortlist, your bid doesn’t matter. Several genuinely different concepts give Andromeda more ways to match your ads with the right people.
Step 3: Ranking (GEM)
Predicts who will actWhat it does
GEM, Meta’s large ranking model, scores the shortlist in depth and predicts who is likely to click or buy. It learns from ad content and behaviour across Facebook and Instagram.
What it looks at
- Chance of a click
- Chance of a sale
- What the ad shows
- Behaviour on Facebook and Instagram
What it means for you
Meta can only optimise for the sales and leads it can see. Set up the Meta Conversions API alongside your pixel so it learns from your real results.
Step 4: Auction (The feed)
Picks the winnerWhat it does
Combines those predictions with your bid and budget to pick the one ad the person actually sees in their feed.
What it looks at
- Predicted results
- Your bid
- Your budget
- The winning ad
What it means for you
Fewer, broader campaigns with more varied ads usually perform better and get through the learning phase sooner. Give changes a few days, because constant tweaks reset learning.
What changed in the Andromeda 2.0 update?
The new version scores every eligible ad for each person, instead of cutting most of them out early with rough shortcuts.
In its October 2026 engineering update, Meta explained that the system now surfaces top candidates at over one billion ads per second, which it says is about 100 times the throughput of the 2024 version. Three changes matter for advertisers:
A team of specialists, not one generalist. The model uses a "mixture of experts" design: many smaller specialist models, plus a router that sends each ad to the few experts best suited to judge it.
More real-time context. It weighs what the person is doing right now, not just their history.
Fewer ads lost early. Because everything gets scored, a new or less obvious ad has a fairer chance of being considered instead of getting zero spend.
Andromeda doesn't work alone. Behind it sits GEM, Meta's large ranking model, which learns from ad content and behaviour across Facebook and Instagram. Think of a shop: Andromeda walks every aisle and pulls the promising products off the shelves, then GEM decides which one to hand the shopper.
| Compared | Andromeda (2024) | Andromeda 2.0 (2026) |
|---|---|---|
| Which ads get scored | Most are cut out early with rough shortcuts | Every eligible ad, for each person |
| How ads are judged | One generalist model | A “mixture of experts”: a router sends each ad to the few specialist models best suited to judge it |
| What it knows about the person | Mostly their history | Their history plus what they are doing right now |
| New or less obvious ads | Often dropped early, so they get zero spend | A fairer chance of being considered |
| Speed | The starting point, launched December 2024 | Over one billion ads a second, about 100 times the 2024 throughput |
| What you switch on | Nothing | Nothing: Ads Manager has no Andromeda setting |
The short answer: Andromeda 2.0 looks at every eligible ad, with more context, far faster. There is nothing to turn on: what changes is what your ads need to give it.
Did the update actually improve results?
Meta says yes, although the gains are smaller than some of the headlines going around.
Meta credits Andromeda with a 2.4% increase in Facebook conversions since 2025. A bigger figure you may have seen, roughly 8% more clicks and 16% more conversions, was credited by Meta for Q2 2026 to GEM and its user understanding models together, not to Andromeda on its own. Meta hasn't published baselines or test designs for any of these numbers, so treat them as a direction, not a promise for your account. There is also nothing to switch on: Ads Manager has no Andromeda setting.
Why is your creative now your targeting?
Andromeda reads the ad itself, so your hook, visuals, words and offer tell Meta who should see it far more than your interest settings do.
The system looks at what is inside the creative: the opening line, who appears on screen, the product, the tone, even the audio. An ad that opens with "Busy parent juggling school pickups?" gets matched to people whose behaviour looks like that, without you building that audience by hand.
Manual targeting still counts as a signal, but stacking narrow interest groups and splitting your budget across a dozen tiny ad sets mostly starves the system of data. Fewer, broader campaigns with more varied ads usually perform better and get through the learning phase sooner.
One catch: Meta appears to group ads that look almost identical and treat them as one. Ten versions of the same video with different background colours are not ten tests. Ten different angles are.
| Compared | The old way (interest-led) | Now (creative-led) |
|---|---|---|
| Who sees your ad | Mostly the interests and demographics you pick | Mostly what the ad shows and says: the hook, who is on screen, the product, the tone and audio |
| How you describe the customer | A demographic, like “women aged 35 to 54” | A person, like “remote working parents trying to get dinner on the table by 6pm” |
| Account structure | Many small ad sets on narrow interest groups | Fewer, broader campaigns with more budget and data in each |
| Creative testing | Small tweaks to one ad, like a new background colour | Genuinely different concepts: near-identical ads appear to be treated as one |
| Learning phase | Budget split thin, so it takes longer | More data in each campaign, so it ends sooner |
| Your interest settings | The main lever | Still a signal, but a much weaker one |
The short answer: Your creative now does most of the targeting. Put your effort into several genuinely different ads, not more audience settings.
How should you set up your ads now?
Focus on three things: different creative for each stage of the buying journey, ads written for a specific person, and offers that protect your margin.
1. Make different ads for different stages
People who have never heard of you need a different ad from people who nearly bought last week.
New audiences: longer problem and solution videos that explain the problem, why your approach is different and what buying from you is like.
Warm audiences: short videos, images or carousels that answer the doubts stopping people from buying. Pull these from enquiries, reviews or abandoned carts. "Is it worth the price?", "How long does it take?" and "What if it doesn't suit me?" are all ads waiting to be made.
Our guide to Meta ads creative testing shows how to run these side by side.
2. Write for a person, not a demographic
"Women aged 35 to 54" tells the system very little. "Remote working parents trying to get dinner on the table by 6pm" tells it a lot. Name that person in the first three seconds of your video and the first line of your copy, and Andromeda can go and find more people like them.
3. Improve your offer instead of discounting harder
Deep discounts can win the click and still lose you money. Offers that lift the order value usually scale better:
Buy more, save more tiers that also cover shipping costs
A free gift or sample pack once the cart passes a set amount
Subscribe and save pricing that turns one-off buyers into regulars
For service businesses, a bundled package or bonus add-on rather than a lower rate
Before you launch an offer, check what return you actually need with our break-even ROAS calculator.
What should you check in your account this week?
Check your structure, your creative variety and your tracking, because those three decide how much Andromeda has to work with.
Consolidate: merge small ad sets chasing similar people so each one has enough budget and data.
Diversify: give each campaign several genuinely different concepts. Many advertisers run 6 to 12; on a small budget, 3 to 5 strong ones is a sensible start. Advantage+ campaigns are built for this.
Stop over-editing: constant tweaks reset learning, so give changes a few days.
Fix your data: the system can only optimise for sales and leads it can see, so set up the Meta Conversions API alongside your pixel.
Conclusion
The Meta Andromeda 2.0 update doesn't make Meta ads harder, but it does change where your effort pays off. The system now looks at every eligible ad and reads your creative closely, so narrow audiences and endless small ad sets matter less than giving it varied, specific, well-priced ads to choose from.
Take one practical step this week: write down the three biggest questions customers ask before they buy, and turn each one into an ad. That alone gives Andromeda more to work with than most accounts have.
