What Is Meta Andromeda? How It Works and What It Means for Advertisers

What is Meta Andromeda and what does it do?

Meta Andromeda is the AI system that runs the retrieval stage of Meta's ad recommendation pipeline. It narrows tens of millions of candidate ads down to a few thousand that might suit one person. Ranking models then make the final pick. Meta introduced it on its engineering blog in December 2024.

That definition sounds technical. For an advertiser, however, the takeaway is simple: your ad has to enter a candidate pool before anyone decides who sees it.

Also, the way into that pool no longer depends only on audience settings. The creative itself helps the system understand who the ad suits.

In this guide, we explain Meta Andromeda in two layers. First, we report what Meta has officially published. Then we share how we read the likely impact on ad accounts, based on our field experience. We keep the two apart on purpose, because a lot of speculation circulates around Andromeda.

Where does Meta Andromeda sit in the ad system?

Meta's ad delivery pipeline has two broad stages. The first is retrieval, which finds candidate ads. The second is ranking, which orders those candidates.

First, in the retrieval stage, the system selects a few thousand ads from tens of millions of candidates. According to Meta's post, this stage handles roughly three orders of magnitude more ads than later stages. So this is where the funnel is widest.

Then, in the ranking stage, larger and more detailed models take over. They predict the value an ad offers to the person and to the advertiser, then choose what to show.

For example, think of a library. Retrieval is the librarian who puts ten books on your desk out of thousands. Ranking decides which of those ten you read first.

Therefore, keep this in mind: Andromeda is not the system that shows your ad. It is the system that lets your ad into the race. An ad that never enters the race never gets shown, however good it is.

When did Meta announce Andromeda and what figures did it share?

Meta announced Andromeda on 2 December 2024 on the Engineering at Meta blog. The post describes it as a next-generation personalized ads retrieval engine and says it strengthens Advantage+ automation.

Here are the main figures from the post:

• A 6% recall improvement in the retrieval system, meaning it finds more of the relevant ad candidates.
• An 8% ad quality improvement on selected segments.
• More than a 3x improvement in end-to-end inference queries per second.
• Also, a 100x improvement in feature extraction latency compared with earlier CPU components.
• Finally, a 10x efficiency gain from model elasticity features.

However, a warning is needed here. These numbers come from Meta's own internal measurements. They do not mean you will see 6% or 8% in your account.

The same post also cites a 22% ROAS lift and a 7% conversion lift. However, Meta ties those figures to Advantage+ creative tools and image generation features, not to Andromeda alone. Please do not mix the two.

Source: Engineering at Meta, Andromeda post.

What hardware does Meta Andromeda run on?

According to Meta, Andromeda runs on the NVIDIA Grace Hopper Superchip. The company also says it plans to add its own MTIA chip (Meta Training and Inference Accelerator) in the future.

Put simply, an advertiser has no direct use for hardware details. Still, one point matters: the retrieval stage has to score a huge number of ads within milliseconds. As capacity grows, the number of ads that can enter the pool grows too, at least in theory.

Meta's post links this capacity to exponential growth in ad creatives. It also says that more than one million advertisers created over 15 million ads per month with generative AI tools.

In short, the hardware story prepares for a world with far more ads. In that world, your ads need to be distinguishable. We come back to this point in later sections.

How does Meta Andromeda relate to Advantage+?

Meta's post positions Andromeda as infrastructure behind Advantage+ automation. Advantage+ automates audience, budget and creative combinations. Andromeda makes it possible to scan the wide candidate pool that this automation needs.

So the two are not the same thing. For example, Advantage+ is the product family you see and configure. Andromeda is the engine behind the scenes.

In practice, you can read it like this: when you loosen targeting limits in an Advantage+ campaign, the system selects candidates from a wider pool. As a result, your creative variety decides how well you use the opportunity.

To go deeper on the Advantage+ side, read our guide on what Meta Advantage+ Audience is. For catalog campaigns, our Advantage+ catalog ads guide will help.

What does Meta Andromeda change inside your ad account?

First, know that Meta Andromeda adds no setting to your account. There is no "turn on Andromeda" switch in the dashboard. Its effect shows up indirectly, in how your ads enter the pool and get selected.

From our field observations, you may notice three things:

• Accounts with creative variety reach wider audiences more easily.
• Ads that look very alike compete with each other for the same budget.
• Narrow targeting and many small ad sets make it harder for the system to learn.

These three points are not official Meta guarantees. They are trends we have seen repeatedly in accounts we manage. So test them in your own account with small experiments.

What matters is that optimization shifts from "who should I show this to?" toward "what should I show?" Understanding that shift is the shortest path to understanding Andromeda.

Why has creative diversity become so important?

The retrieval stage works from a very large pool. Therefore, the system infers who should see an ad largely from the ad's content. The clearer and more distinct your visuals, copy, video and message are, the easier it is for the system to match the ad with the right person.

Take an example. A shoe brand wants to reach three people: a runner, an office worker and someone shopping for a gift. If you show all three the same photo and the same sentence, you give the system one signal.

But if you prepare three ads, one about performance, one about comfort and one about unboxing a gift, you give three different signals. As a result, each ad has a chance to find its own audience.

Meta's official post does not state a rule like "creative diversity works this way." It does say the system is preparing for growth in ad volume. So we treat diversity as a strong field hypothesis, not as an official requirement. For testing logic, see our Meta ads creative testing guide.

What is the difference between real diversity and cosmetic variation?

Put simply, real diversity changes the story, the promise, the target person or the format of an ad. Cosmetic variation changes a colour, a font or a few words. The first carries a new message, while the second repeats the same message.

In practice, you will see the difference like this:

• Real diversity: a different problem, a different objection, a different use case.
• Real diversity: different formats such as customer quotes, product demos, comparisons or founder stories.
• Cosmetic variation: the same photo with different filters.
• Cosmetic variation: the same headline rewritten with synonyms.

One claim circulates widely online: Andromeda groups similar ads under one identity and makes them compete. That account comes from third-party agencies. Meta's engineering post does not describe such a mechanism.

Still, the logic leads to a defensible conclusion: fifty copies of the same message do not create fifty new opportunities. So we prefer producing concepts instead of copies. For a related view, read our post on banner blindness and attention-grabbing creative.

How many different creative concepts does an account need?

Meta gives no official number on this, so treat any claim of "exactly this many" with suspicion. We can share a starting range from field experience, but it is not a guarantee.

In a mid-size e-commerce or service account, we see 3 to 6 clearly distinct concepts per campaign as a reasonable start. For each concept, you can try two or three formats, such as a static image, a short video and a carousel.

Example calculation: imagine a cosmetics brand with a monthly ad budget of 60,000 in local currency. In one campaign, you prepare 4 concepts: ingredient education, a customer quote, a before and after story, and a gift set. Adding 2 formats per concept gives 8 ads.

If four of those eight ads copied the same idea, you would effectively test with four ads. This is an invented example to show the logic. It is not a real client result.

Does targeting lose its importance in the Andromeda era?

No, targeting does not lose its importance. Its role changes. You still set the budget, geography, legal limits and exclusions. What changes is that the system takes on more responsibility for finding the right person within a broad audience.

Therefore, very narrow, multi-layered interest stacks often become unnecessary. Moreover, narrow audiences give the system little data and lengthen the learning phase.

In new accounts especially, we often start broad and separate the creatives instead. Treat this as a starting point, not a principle. In some sectors, narrow targeting still makes sense.

For example, a 10 km radius for a local service is already a natural filter. Here, "broad" means adding no interest restriction inside that radius. We explain audience logic in detail in our Advantage+ Audience guide.

How do you simplify account structure for Andromeda?

Simplifying means reducing the number of campaigns and ad sets so that data gathers in a few places. Each ad set has to learn on its own, so many small sets spread both budget and data thin.

You can follow these steps:

1. Set up one or two main campaigns by goal, for example sales and leads.
2. Keep few ad sets in each campaign and merge similar audiences.
3. Put ads with clearly separate concepts inside those sets.
4. Each week, pause the weakest ads and add a new concept in their place.
5. Avoid frequent budget changes, because big edits can reset learning.

If you are unsure about budget structure, see our comparison of ABO and CBO for Meta ads. If the account layout overwhelms you, our team can help through social media management.

How do the learning phase and budget interact with Meta Andromeda?

The learning phase is the period in which Meta collects enough data to predict how an ad set will perform. Meta describes this process in its help pages. We found no official statement that Andromeda changes this logic.

Two points still matter. First, too many small ad sets make it hard for each one to collect enough events. Second, constantly changing creatives delays stabilization.

So add new concepts gradually. Adding one or two new concepts per week is healthier than swapping ten ads in one day.

If you see a "learning limited" warning, review the number of sets and your budget first. More creative variety alone will not fix that warning. For details, read our guide to the Meta ads learning phase.

Why does measurement quality matter even more with Meta Andromeda?

The more ads a system evaluates, the more the quality of the feedback signal matters. If your conversion events are missing or wrong, Meta cannot learn which ad truly drives sales. It may then shift budget to the wrong creatives.

Before you launch, we suggest checking these points:

• Do your pixel and Conversions API events run without double counting?
• Is your event match quality good enough?
• Do conversion values pass through correctly?
• Does the attribution window suit your business model?

We have a guide for each of these. Our posts on setting up the Meta Pixel and improving event match quality cover the basics.

In short, investing in creative variety while measurement is broken is like lighting a candle in the dark and ignoring the wall. Make sure the numbers are reliable first.

What myths surround Meta Andromeda?

Whenever a new system is announced, myths appear fast. The table below collects the three or four we hear most often, with the more accurate picture beside each one.

One row carries a "not an official rule" note. We added it on purpose, because much of the information in this area comes from agency observation.

When you read a claim, ask which Meta page states it. If there is no answer, test it as a hypothesis and do not apply it as fact.

How does the old Meta Ads approach differ from the new one?

The old approach centred on audience layers. However, the new approach puts creative strategy, measurement quality and a simple account structure first. The table below summarizes the shift.

Also, this table is not a rulebook. Every business has a different product, margin and sales cycle. Still, in our accounts this shift has tended to improve results; do not read that as a guarantee.

How does Meta Andromeda affect small-budget businesses?

Also, a small budget means little data, which slows the system's learning. So the biggest risk for small businesses is spreading the budget over too many ads and sets.

If your budget is limited, we suggest this path:

• Start with one campaign and a few ad sets.
• Prepare three or four strong concepts, not twenty weak ads.
• For each concept, first try the cheapest format you can produce, such as a phone video.
• Make no major changes for two weeks.

You can also use our ROAS calculator to see your own numbers. Switching ads on and off without knowing your break-even ROAS is a bigger mistake than lacking creative variety.

So settle the basic maths first, then increase variety. Reversing the order wastes budget.

What should a creative brief answer?

In practice, a good brief does not leave diversity to chance. A few questions answered before production help your team make ads that truly differ from each other. We keep the brief to one page.

Ask these questions in the brief:

• Which single problem or wish does this ad answer?
• What will the viewer see in the first three seconds, and why will they stop?
• Which objection does it break, such as price, trust or time?
• How does this concept differ from the concepts already live?
• Which signal will you use to judge success?

The fourth question matters most. If you ask "is it different?" before production, you cut the risk of making lookalike ads from the start.

Also, feed the brief with customer reviews, sales calls and support tickets. Real customer language often grabs more attention than agency language. So collecting your own customers' sentences is the cheapest way to find concepts.

Which mistakes burn budget in the Meta Andromeda era?

Old habits can get expensive with new systems. We share the mistakes we see most often as field observations. They are not official Meta warnings.

• Uploading the same video dozens of times with different headlines.
• Pausing a new concept after three days because it "did not work".
• Raising budget before fixing pixel errors.
• Opening a separate campaign and ad set for every product.
• Focusing on changing audiences instead of creative.
• Picking a winning ad without calculating break-even ROAS.

If you make two or three of these at once, the problem is the account setup, not Andromeda. For example, pausing a concept after three days denies the system a chance to learn.

So wait for enough spend and conversion signal before you judge a concept. How long you wait depends on your average order value and sales cycle.

Is there a parallel between Meta Andromeda and Google's automation?

Yes, in direction there is. Google also hands control to the system through broad match, AI Max and automated bidding in search campaigns. On Meta's side, Advantage+ and Andromeda represent a similar direction.

On both platforms, the advertiser's role moves from turning dials to improving inputs. Those inputs are creative, measurement data and business goals. In short, understanding what the system looks at matters more than dashboard settings.

We covered the similar shift on Google's side in our Google Ads AI Max guide. If you are deciding which platform to lead with, our Google Ads vs Meta Ads for ecommerce comparison can help.

Still, a parallel does not mean the same solution. Google works with intent-driven search queries, while Meta works with an interest and discovery feed. That is why creative usually plays a bigger role on Meta.

Who in your team should handle what for Meta Andromeda?

This change does not concern only the media buyer. Creative production, data measurement and strategy must meet at one table. Otherwise, the need for variety gets stuck in the production line.

We suggest a simple split:

• Media buyer: keeps the account structure simple, watches spend distribution, separates winners from losers.
• Creative producer: makes different concepts and formats from the brief.
• Analytics owner: checks that pixel, Conversions API and conversion values are correct.
• Brand owner: makes sure the message fits the brand tone and legal rules.

In a small business, one person may hold all four roles. However, if you do not think of the roles separately, measurement or brand control usually gets skipped. So writing the roles down helps whatever your headcount.

If you want support on brand consistency, see our brand identity service. A strong identity guide keeps a shared voice while you produce variety.

How do you measure the effect of Meta Andromeda?

You cannot measure Andromeda as a separate metric. Instead, you look at the overall health of your account: costs, audience breadth, how spend spreads across ads, and how new concepts perform.

Ask these questions every week:

• Does most of the budget go to one ad, or is the spread balanced?
• Do new concepts get spend, or do they get no delivery?
• Is frequency rising quickly?
• When costs rise, are there signs of creative fatigue?

The answers show whether your account lacks creative variety. For example, if most spend sits on one ad, the account weakens fast once that ad tires.

Instead, do not trust unsourced numbers. Most tables of the type "this many creatives give this much ROAS" rest on no measurable source.

What does a 30-day action plan look like?

The plan below helps you move in small steps instead of adapting to Meta Andromeda in one go. Adjust it to your own account.

Week 1, foundation: Check measurement. Verify the pixel, Conversions API and event match quality. Calculate your break-even ROAS.

Week 2, structure: Simplify the number of campaigns and sets. Merge similar audiences.

Week 3, concepts: Pull 4 to 6 concepts out of customer objections and produce them in different formats.

Week 4, review: Look at spend distribution. Keep the winner, pause the weak ones and add a new concept.

Put simply, following this plan guarantees nothing. However, it aligns the account with how the system works. If you want to review your account layout together, write to us through our social media management page.

What should you do about Meta Andromeda?

Meta Andromeda is infrastructure that strengthens Meta's ad candidate retrieval and supports Advantage+ automation. That is the part you can verify from an official source. The importance of creative variety is a strong field reading, but it is not an official rule.

In practice, do three things. First, harden your measurement. Second, simplify your account. Third, produce genuinely different creative concepts on a regular basis.

Do not get swept up in the noise. Every time a new system name appears, agencies shouting "do this now" is a familiar marketing reflex. We state only what we can verify as fact and present the rest as hypothesis.

Set up small tests in your own account, read the results and update your decisions with data.

One last point: Meta keeps developing this system, and details may change over time. So any commentary you read today may age within a few months.

Therefore, the safest approach is to follow official announcements and run measurable experiments in your own account. For example, testing one new concept group each month and noting the results makes you independent of general commentary.

In the accounts we work on with our team, we follow the same method: write the hypothesis, test it with a small budget, record the result and decide the next step from data. This discipline keeps its value whichever algorithm name is in the news.