Skip to content

+ Education · Noviembre 2026

Meta Andromeda and how the AI behind your ads works

What Meta Andromeda is, how it selects the ads you see on Facebook and Instagram, how it relates to Advantage+ and how we apply it when building campaigns in Puerto Rico.

Conceptual representation: thousands of ads pass through glass filters until only a few remain, one highlighted in red

Article

The essentials about Meta Andromeda.

Analysis reference date: February 2025. Editorial review of content and sources: November 2026.

  • What it is: Meta Andromeda is the machine learning system Meta uses for ad retrieval, the first stage of its ad system.
  • What it does: from tens of millions of eligible ads, it picks a few thousand relevant candidates for each person in milliseconds.
  • Why it matters: Advantage+ automation and generative AI multiplied the number of ads in circulation, and Meta needed a system able to evaluate them all.
  • What changes for advertisers: real variety in messages and creative, plus reliable measurement, matter more than manual micro-targeting.
  • What changes for agencies: the work shifts toward strategy, concepts, production, measurement and analysis of business results.

In this article:

+

What Meta Andromeda is.

Every time someone opens Facebook or Instagram and an ad slot appears, Meta has to decide what to show. The problem is scale: at that moment there may be tens of millions of ads that person could, in principle, see.

Meta solves that decision in several stages. The first is called retrieval, and that is where Andromeda works. According to Meta's engineering team, this stage narrows tens of millions of candidates to a few thousand relevant ads. Then larger, more sophisticated ranking models estimate the value for the person and the advertiser and determine the final set of ads shown. Engineering at Meta, 2024

It is important to place its role correctly. Andromeda is not a tool the advertiser configures or a campaign type. It is internal infrastructure. What the advertiser controls, such as the objective, audience, budget and ads, is set in Ads Manager, and many of those decisions can be automated with Advantage+. Andromeda is the piece that lets the system consider all that volume of ads in time.

Meta publicly introduced Andromeda on December 2, 2024, describing it as deployed across Facebook and Instagram. Engineering at Meta, 2024

+

How it works, explained simply.

Think of a huge library with millions of books and a librarian who has one second to recommend something. They cannot read them all. First they go straight to the aisles that make sense for you and pull a manageable stack. Then, more carefully, they compare that stack and pick the best. Andromeda is that first step: getting quickly to the right stack.

Illustrative example. Imagine a person in Puerto Rico thinking about renovating their home. The system holds ads for paint, hardware stores, contractors, furniture and personal loans, plus thousands of ads unrelated to them, like mobile games or pet insurance. In the retrieval stage, the system keeps a group of candidates it considers relevant to that person. This example is hypothetical: Meta does not publicly document which specific signals it uses for each person, so we do not attribute any.

Then other stages come in. Ranking models calculate, for each candidate, the likelihood that the person will engage or convert and the value for the advertiser. Finally, the auction combines those predictions with the bid and ad quality to decide what appears. PPC Land, 2026

The practical consequence is clear: if an ad does not pass the first stage, it never gets to compete in the next ones.

+

The technology behind Andromeda.

Deep neural networks

A deep neural network is a model that learns patterns from many examples. In Andromeda, its job is to find relationships between people's interests and what ads offer, including relationships that are not obvious. Meta explains it designed a network with greater compute complexity to learn higher-order interactions between people and ads data. Engineering at Meta, 2024

Hierarchical index

Reviewing millions of ads one by one would be too slow. So Andromeda organizes ads into a multi-layer index, like a catalog with sections, subsections and shelves, and only goes deeper into the most promising branches. Meta says the index and the model are trained together, which improves precision compared with more common methods. Engineering at Meta, 2024

Specialized infrastructure

All of this happens in milliseconds, so hardware matters. Andromeda was designed for the NVIDIA Grace Hopper chip and Meta's own accelerators (MTIA), storing precomputed ad information in the chip's memory to respond faster. Engineering at Meta, 2024

What the numbers say

  • +6% recall: according to Meta, the system retrieves 6% more of the ads that were actually relevant. It measures the quality of the initial selection, not sales.
  • +8% ads quality: Meta reported this on selected segments, not across the whole system.
  • 10,000x model capacity: refers to the size of the model retrieval can use, thanks to a compute cost that grows less than proportionally.
  • 10x inference efficiency: from the system's ability to adjust model complexity to available resources.

All are technical figures Meta published about its own system. They are not guaranteed results for any particular campaign. Engineering at Meta, 2024

+

Andromeda, Advantage and generative AI.

Advantage+ is Meta's suite of AI solutions advertisers can use to automate audience, placements, budget and creative, either in full sales, app or leads campaigns or in individual features. Meta Business Help Center

With Advantage+ creative, Meta can adjust images, videos and text, and some enhancements use generative AI, such as adding animation or generating backgrounds. Meta Business Help Center

Here is the connection: each creative variant is one more ad the system must evaluate. Meta explained that automation and generative AI tools would significantly grow the number of creatives, and that Andromeda was designed precisely to handle that growth. Engineering at Meta, 2024

In the same article, Meta shared three figures worth reading separately:

  • 22% higher ROAS: advertisers who did not use Advantage+ creative and turned on its AI-driven targeting features. This is a result of an Advantage+ tool, not Andromeda.
  • +7% conversions: this is a Meta estimate for businesses using image generation.
  • More than 15 million ads in a month, created by over a million advertisers with generative AI tools. It measures adoption, not performance.

In short: Advantage+ is what the advertiser uses; Andromeda is what lets the system evaluate the volume that automation produces.

+

How it affects businesses that advertise.

What is documented

  • Initial selection considers a much larger universe of ads with more capable models, which Meta associates with more personalized ads. Engineering at Meta, 2024
  • In October 2026, Meta reported that, since 2025, Andromeda has delivered a 2.4% increase in conversions on Facebook. Engineering at Meta, 2026
  • Industry analysts describe Andromeda and the GEM ranking model as now largely determining how ads are selected and delivered, with creative acting as a key signal. Search Engine Land, 2026

Our editorial recommendations

The following is Alterno's judgment, not a Meta rule. We treat these as hypotheses to test for each business:

  • Real variety, not copies. Ten ads saying the same thing in another color will likely behave like one. It is more useful to test different angles: price, benefit, testimonial, demonstration.
  • Simpler structures. Concentrating budget in fewer ad sets can give the system more data to learn from. Whether it fits depends on budget, offer and sales cycle.
  • Broad targeting with good signals. It can work well when measurement is properly set up; with little conversion data, narrowing may be preferable.
  • Measurement first. Pixel, Conversions API and correct events are the foundation. If the system optimizes toward the wrong signal, it will find many people doing the wrong thing.

+

The role of agencies.

Automation does many formerly manual tasks well, like allocating budget or finding audiences. What it cannot do is understand a business. That is still human work:

  • Knowing the consumer: what worries them, what holds them back and what convinces them.
  • Positioning and offer: what makes the brand different and why someone should choose it today.
  • Creative concepts and production: distinct, well-executed ideas in the right formats.
  • Measurement: well-defined events, a fast and clear website, and forms that produce quality leads.
  • Business analysis: clicks or cost per lead are not enough; you need to see how many become sales.

Automation changes how work is distributed, but it demands more rigor in evaluating what goes into the system and what comes out. That is the core of our Digital Advertising and Media service, complemented by Dynamic Retargeting and AI + Automation.

+

How we would apply it at Alterno.

To make it concrete, we describe a proposed methodology with a hypothetical example: a paint and sealant brand in Puerto Rico. It is not a real case and has no obtained results.

1. A concrete objective. Before producing anything, we define what success means: in-store sales, contractor leads or quote requests. Each objective changes the campaign and what we measure.

2. Distinct concepts. We propose four angles that speak to different motivations:

  • Roof protection: the sealant against sun, rain and hurricane season.
  • Ease of application: for people who want to do it themselves over a weekend.
  • Transforming spaces: the before and after of a living room or a facade.
  • Contractor testimonials: professionals explaining why they trust the product.

3. Formats by message. Transformation works best in short video or carousel; ease, in a brief tutorial; protection, in a clear image with a product fact. We fit the number of pieces to the budget and available material, without producing variants nobody will be able to evaluate.

4. Production. Our team develops the copy, design and audiovisual production for each concept, with the brand's identity.

5. Measurement. We review the pixel, Conversions API, website events and form before launch, so the system optimizes toward real leads or sales.

6. Evaluation. We analyze which angles produce business results, not just engagement, and use those learnings for the next round of creative.

+

What we reviewed in November 2026.

The original Andromeda announcement is from December 2, 2024. In this review we reread the sources and found a later development:

  • On October 28, 2026, Meta published the next generation of Andromeda. According to Meta, the system now scores every eligible ad with near ranking-level performance and processes over one billion ads per second per server, 100 times the 2024 system, within the same response time. Engineering at Meta, 2026
  • In the same article, Meta stated that, since 2025, Andromeda has delivered a 2.4% increase in conversions on Facebook. Engineering at Meta, 2026

The rest of the article is a review of content and sources: it confirms and contextualizes existing documentation.

+

Frequently asked questions.

What is Meta Andromeda?

It is the machine learning system Meta uses in the first stage of its ad system to choose, from tens of millions of eligible ads, a few thousand relevant candidates for each person.

How does it work inside Meta Ads?

Andromeda makes the initial selection. Then larger ranking models estimate the value of each candidate, and the auction, combining those predictions with the bid and ad quality, decides what is shown.

How is it related to Advantage+?

Advantage+ is the automation suite advertisers use for audience, placements, budget and creative. Andromeda is Meta's internal infrastructure that can evaluate the volume of ads that automation generates.

Do I have to turn on Andromeda?

No. Meta does not present it as an Ads Manager option: it runs inside its delivery system on Facebook and Instagram for all advertisers.

How does it affect ad creation?

Because the system can evaluate many more ads, it pays to give it variants that are truly different in message, format and angle. Volume alone does not guarantee better results; it has to be tested for each business.

What role does an agency play in automated campaigns?

Defining positioning, the offer and creative concepts, producing quality assets, setting up measurement correctly and analyzing whether results turn into sales or qualified leads.

+

References.

+

Let's talk about your digital advertising strategy.

If you invest in Facebook and Instagram and want your ads to make better use of Meta's automation, we can help with creative strategy, production and measurement.

Let's talk about your digital advertising strategy or request a proposal +.

Let's talk

Want to turn these ideas about artificial intelligence and digital media into a strategy for your brand?

Tell us about your goals, and we can develop a proposal built around your business needs.

REQUEST A PROPOSAL +

Conversation +

Comentarios.

0 published

Your perspective helps this conversation grow. Share a comment, a question or your experience related to this article, we're listening.

or with your email

Create an account or log in to comment with your real name.

Be the first to comment.

Keep reading

More from the blog.

← Back to the blog

+ Ideas that keep moving

Get the next article.

Creativity, marketing, digital advertising, technology and artificial intelligence, straight to your inbox.

Let’s work together

Ready to talk about your brand?