Digital Marketing

Meta's Advertising AI Is Far Bigger Than Andromeda. Here's What's Really Happening.

AI in Meta Ads Algorithm

Meta's Advertising AI Is Far Bigger Than Andromeda. Here's What's Really Happening.

Most advertisers have heard about Meta Andromeda by now.

It's been discussed across marketing communities, advertising agencies and digital marketing circles as one of the biggest changes to Facebook and Instagram advertising.

But here's something that most discussions miss.

Andromeda is only one piece of a much larger AI infrastructure that Meta has been building.

Behind every advertisement you see on Instagram or Facebook, there are sophisticated machine learning systems trying to understand what you're interested in, what you're likely to engage with, what you might purchase and which advertisement is worth showing you.

And Meta is increasingly moving beyond traditional recommendation algorithms.

It is building advertising systems powered by foundation models, generative recommendation architectures, behavioural sequence learning and increasingly sophisticated prediction engines.

In October 2026, Meta published several major engineering announcements revealing just how far this infrastructure has evolved.

Among them were GEM, GenRecs Tokenization and the next generation of Andromeda.

These developments have significant implications for advertisers, digital marketing agencies and businesses spending money on Meta Ads.

So what exactly is happening behind the scenes?

Let's look at the technologies, how they work together and, most importantly, what advertisers should do differently.

First, how does Meta actually decide which advertisement you see?

Imagine opening Instagram.

There are millions of advertisements that could potentially be shown to you.

Meta cannot simply display advertisements randomly. It needs to determine which advertisements are eligible, identify the most relevant candidates, evaluate their predicted performance and decide which ones should enter the auction and be delivered.

This involves several stages.

At a simplified level, the process looks like this:

  1. User understanding: Meta processes signals that help it understand your interests, behaviour and intent.
  2. Advertisement retrieval: It identifies a relevant set of advertisements from an enormous candidate pool.
  3. Advertisement ranking: It predicts which advertisements are most likely to generate valuable outcomes.
  4. Auction and delivery: Eligible advertisements compete based on their overall predicted value and other auction factors.
  5. Feedback: Subsequent interactions provide additional signals for future predictions.

This is where the different AI technologies come into play.

Andromeda helps retrieve relevant advertisements. GEM provides powerful shared recommendation intelligence. GenRecs improves how advertisements and user preferences are represented and matched. Adaptive Ranking makes sophisticated predictions computationally practical.

These are not interchangeable technologies. They solve different problems within a larger recommendation infrastructure.

Let's examine them individually.

1. GEM: The foundation model behind Meta's advertising recommendations

GEM, short for Generative Ads Recommendation Model, is arguably one of the most important developments in Meta's advertising infrastructure.

Meta introduced GEM in November 2025 and announced a substantially expanded architecture in October 2026.

Traditional recommendation systems often rely on multiple models trained for specific platforms, objectives or advertising surfaces.

For example, one model might specialize in predicting clicks while another focuses on conversions.

GEM takes a different approach.

It acts as a large, shared recommendation foundation model that learns from advertising signals across Meta's ecosystem.

It can learn relationships involving:

  • User interaction histories.
  • Advertisement engagement.
  • Product and content characteristics.
  • Different advertising surfaces.
  • Conversion behaviour.
  • Relationships between awareness, consideration and purchase.

One particularly interesting development is its ability to transfer knowledge to downstream recommendation models.

Think of it this way.

A smaller recommendation model might be responsible for making a specific advertising prediction. Instead of learning everything independently, it can benefit from representations and knowledge learned by a much larger foundation model.

This allows improvements in the central model to benefit multiple downstream systems.

What's new about GEM in 2026?

In its October 13, 2026 engineering announcement, Meta described a new generation of GEM that moves towards a central model across Facebook, Instagram and different advertising objectives.

Meta reported an 18.9% conversion improvement and a 12% CTR improvement across its advertising surfaces and objectives for the expanded system.

These are Meta-reported system-level results, not a promise that every advertiser will experience the same improvement.

Why should advertisers care?

Because Meta is getting better at learning relationships across different types of user interactions and advertising outcomes.

An advertisement that generates meaningful engagement can contribute to a much richer understanding of user preferences than a system relying on isolated campaign-level signals.

Official technical announcement: Meta's Generative Ads Model (GEM).

2. Andromeda: The system that finds relevant advertisements

Andromeda is probably the most widely discussed Meta advertising AI technology.

But what does it actually do?

Imagine Meta has a billion eligible advertisement candidates.

It needs to identify a relatively small set of potentially relevant advertisements for an individual user.

That's where Andromeda comes in.

It is Meta's personalized advertisement retrieval engine.

Its responsibilities include:

  • Retrieving relevant advertisement candidates.
  • Processing enormous candidate pools.
  • Understanding relationships between users and advertisements.
  • Reducing the candidate set before more expensive ranking operations.

Andromeda uses deep neural networks, hierarchical indexing and GPU-optimized computation.

The October 2026 Andromeda upgrade

On October 28, 2026, Meta published details about its next-generation Andromeda architecture.

The system introduces Mixture-of-Experts (MoE) techniques into advertisement retrieval.

Rather than relying on one homogeneous user representation, it uses multiple specialized user experts to capture different interests and contextual signals.

It also incorporates:

  • Sequential advertising engagement events.
  • User cohort representations.
  • Language-model-derived semantic features.
  • Cross-surface engagement signals.
  • Context-aware advertisement representations.

Meta reports that the new system can surface top candidates at over one billion advertisements per second in production, representing a 100-fold throughput increase compared with its 2024 system.

Meta also reported a 2.4% conversion increase on Facebook since 2025 from Andromeda.

Why should advertisers care?

Andromeda makes advertisement retrieval more sophisticated.

It is increasingly important to provide genuinely different creative concepts rather than simply producing dozens of nearly identical advertisements.

The more varied your advertising concepts, the more opportunities you create to communicate different products, benefits and problems to potential customers.

Technical announcement: Meta Andromeda: Next-Generation Ads Retrieval.

3. GenRecs: Meta's move towards understanding what advertisements actually mean

This is one of the most interesting developments in Meta's advertising ecosystem.

GenRecs stands for Generative Recommenders.

To understand why it matters, consider a simple example.

Imagine a company launches a completely new running shoe.

It has:

  • No historical advertisement engagement.
  • No previous campaign conversions.
  • No meaningful product interaction history.

How does Meta determine which people might be interested in it?

Traditional recommendation systems often rely heavily on historical interactions and item identifiers.

GenRecs introduces a more semantic approach.

Instead of treating every advertisement as an isolated identifier, it helps the recommendation system understand what the advertisement represents.

For example, an advertisement could contain information about:

  • Product category.
  • Brand.
  • Product characteristics.
  • Intended use.
  • Benefits.
  • Visual content.
  • Audience relevance.

This allows recommendation models to learn relationships between similar products and user preferences.

A new product can potentially benefit from knowledge the model already has about related products, even when the new product has little historical engagement.

GenRecs Tokenization

In October 2026, Meta also published details about GenRecs Tokenization.

The technology converts semantic representations of advertisements, products and other items into discrete tokens.

The process involves three main stages:

  1. Encoding content into a semantic embedding space.
  2. Decomposing the embeddings into separate subspaces.
  3. Quantizing those representations into discrete tokens.

Think of it as giving the recommendation system a more meaningful vocabulary for understanding products and advertisements.

Meta reported a 0.35% conversion improvement across advertising surfaces when GenRecs Tokenization was paired with GEM and Adaptive Ranking.

Why should advertisers care?

The content of your advertisement matters beyond its visual appearance.

Your messaging, product descriptions, imagery and positioning collectively communicate what you're offering.

Clear creative concepts can help recommendation systems understand the relevance of an advertisement.

Technical announcement: GenRecs Tokenization: Transforming Ads Recommendation at Meta.

4. Adaptive Ranking: Giving different users different levels of AI computation

Once Meta retrieves potential advertisements, it needs to rank them.

But ranking billions of potential user-advertisement interactions with extremely sophisticated models is computationally expensive.

This creates a problem.

How can Meta use increasingly powerful AI models while maintaining the speed required to serve advertisements to billions of people?

Its answer includes the Adaptive Ranking Model.

Traditional ranking approaches often use relatively fixed computational configurations.

Adaptive Ranking moves towards dynamically allocating model complexity based on the context of an advertising request.

For example, different requests may benefit from different levels of computational sophistication.

The system can use information about a person's context and intent to determine how much computation is appropriate.

This allows Meta to use more sophisticated recommendation models without imposing the same computational cost on every request.

Meta described this architecture in its March 31, 2026 engineering publication.

Why should advertisers care?

The ranking process is becoming increasingly sophisticated.

Your advertisement is not simply competing for attention based on an interest category.

The system is evaluating how relevant an advertisement might be to an individual person and how likely that person is to perform a valuable action.

Technical announcement: Meta Adaptive Ranking Model.

5. Sequence Learning: Understanding what people do, not just what they like

This technology has particularly interesting implications for targeting.

Traditional advertising systems often relied on aggregated user characteristics.

For example:

  • Someone frequently interacts with fitness content.
  • Someone has clicked several electronics advertisements.
  • Someone follows multiple marketing-related pages.

But these signals don't necessarily reveal what someone currently intends to do.

Consider two people.

Person A:

  1. Watches a video about digital marketing.
  2. Searches for SEO services.
  3. Visits a digital marketing agency website.
  4. Interacts with a case study about increasing organic traffic.

Person B:

  1. Watches a video about SEO.
  2. Watches a cooking video.
  3. Watches travel content.
  4. Likes a random social media post.

Both people may have similar broad interests.

But their behaviour tells a very different story.

Sequence Learning allows recommendation systems to consider the order and context of interactions rather than relying exclusively on aggregated characteristics.

Meta's work in this area includes transformer-based user modelling and systems that combine longer behavioural histories with fresh signals.

This allows models to understand how user interests evolve over time.

Why should advertisers care?

It helps explain why broad targeting can sometimes outperform extremely narrow interest-based targeting.

A person's recent behaviour and changing intent may reveal much more than a static interest category.

It doesn't mean demographic and interest targeting are useless. It means they are only part of a much richer recommendation process.

6. Meta Lattice: Connecting different advertising objectives

Meta Lattice is another important part of the evolution of its recommendation infrastructure.

Historically, advertising systems often relied on different models for different objectives and surfaces.

For example:

  • Predicting advertisement clicks.
  • Predicting video views.
  • Predicting purchases.
  • Optimizing Facebook Feed.
  • Optimizing Instagram placements.

Lattice was designed to consolidate and share knowledge across multiple advertising objectives and surfaces.

This is particularly useful because advertising outcomes are connected.

Someone might watch an advertisement today, click it tomorrow and purchase a product several days later.

A recommendation system needs to understand these relationships.

Meta introduced Lattice in 2023 and subsequently continued developing its unified modelling approach.

Why should advertisers care?

Advertising outcomes should not be viewed as isolated events.

A campaign designed to generate awareness can influence future consideration and purchases.

The infrastructure behind Meta's advertising recommendations increasingly needs to understand these relationships.

7. Intelligent Delivery: AI deciding where computing power should go

One of the less obvious challenges in Meta's advertising infrastructure is computational efficiency.

Its recommendation models are becoming increasingly sophisticated.

But Meta cannot spend unlimited computing resources evaluating every possible advertisement for every user.

Intelligent Delivery addresses this problem.

It helps determine where additional computational resources are likely to improve advertisement ranking.

This involves decisions about:

  • Which requests need more sophisticated evaluation.
  • Which models should receive additional resources.
  • How computational workloads should be distributed.
  • Where additional ranking capacity could improve results.

The interesting part is that advertising optimization is no longer just about improving the recommendation model.

It is also about optimizing the infrastructure that serves those models.

For advertisers, this is an indirect improvement. It isn't a campaign setting you can manually enable.

8. Hierarchical Interest Representation: Connecting users, products and advertisers

Meta has also explored Hierarchical Interest Representation, or HIR.

This is an important research direction involving the relationships between users, advertisers, products and interests.

Imagine someone regularly engages with:

  • Hiking content.
  • Outdoor equipment.
  • Adventure travel.
  • Running shoes.
  • Fitness products.

Even if that person has never interacted with a particular outdoor equipment company, there may be meaningful relationships between their interests and the company's products.

Graph-based learning and semantic representations can help recommendation systems identify these relationships.

HIR explores ways to combine different information sources into richer representations.

It is important to distinguish this from a confirmed universal advertising feature. Meta's published work describes an emerging research direction rather than a setting advertisers can directly control.

9. Rank More Ads and Serve Better Models: Improving recommendation quality

Meta has also described two approaches to improving recommendation quality through additional computational capacity.

Rank More Ads (RMA)

The idea is relatively straightforward.

When additional computational capacity is available, the system can evaluate more advertising candidates.

More candidates can provide additional opportunities to identify relevant advertisements.

Serve Better Models (SBM)

Rather than evaluating every request using exactly the same model configuration, Meta can direct more requests towards more sophisticated ranking models when resources permit.

Both approaches demonstrate an important direction in Meta's infrastructure: using available computational resources to improve the quality of advertising recommendations.

These are backend infrastructure mechanisms, not advertiser-facing optimization switches.

10. Solaris: Making foundation-model intelligence available during live advertising

A large foundation model such as GEM contains an enormous amount of learned information.

But running that entire model for every advertisement candidate during every live request would be extremely expensive.

Solaris addresses part of this problem through precomputation.

It helps precompute user-advertisement representations asynchronously so that downstream recommendation systems can use richer information during live serving.

In simple terms:

Instead of repeatedly calculating everything from scratch, Meta can prepare useful representations in advance.

This helps make large-scale recommendation models more practical.

It is another example of how Meta is optimizing not just the intelligence of its AI models, but also the infrastructure required to deploy them.


So how do all these technologies work together?

Let's put everything into one simplified picture.

TechnologyPrimary responsibilityGEMShared recommendation foundation modelAndromedaAdvertisement retrievalGenRecsSemantic representation and generative recommendationsAdaptive RankingContext-aware ranking and computationSequence LearningUnderstanding behavioural historyLatticeShared learning across advertising objectivesIntelligent DeliveryComputational resource allocationHIRResearch into hierarchical interest representationsRMAEvaluating additional advertisement candidatesSolarisPrecomputed representations for downstream serving

These technologies do not operate as ten independent algorithms.

They represent different parts of a larger, evolving advertising recommendation ecosystem.

Meta has not publicly disclosed every internal component or every production integration. This is a map of its publicly documented technologies, not a complete internal architecture.

What does this mean for advertisers? 6 things you should change.

Understanding these technologies is interesting.

But what matters more is how they change the way you run advertising campaigns.

Here are six practical implications.

1. Stop producing 20 versions of the same advertisement

Changing the background colour, replacing a photograph or slightly modifying a headline doesn't necessarily create a meaningful new advertising concept.

Instead, develop different creative angles.

For an SEO agency, these might include:

  • Financial loss: How much revenue are you losing because customers cannot find your business on Google?
  • Competition: Your competitors are getting the customers who should be yours.
  • Authority: What would happen if your business ranked first for its most valuable keywords?
  • Pain point: Getting website traffic but no actual enquiries?
  • Curiosity: We analysed why businesses struggle to generate leads through organic search.
  • Education: Three SEO mistakes that are costing businesses customers.

Each concept communicates a different problem, motivation or value proposition.

2. Test broad targeting instead of relying entirely on interests

Meta's recommendation infrastructure increasingly incorporates behavioural and contextual information.

This makes broad targeting worth testing.

Instead of automatically separating campaigns into dozens of interest-based audiences, test a broader eligible audience and allow the recommendation system to identify potential customers.

However, broad targeting is not automatically superior in every situation.

Compare qualified leads, acquisition costs and actual revenue rather than judging performance by cheap clicks.

3. Make your creative messaging extremely clear

If you're advertising a product or service, explain:

  • What you're offering.
  • Who it's for.
  • What problem it solves.
  • Why someone should care.
  • What makes it different.

Don't expect an algorithm to compensate for confusing messaging.

A clear advertisement gives both the viewer and the recommendation system a better understanding of what is being promoted.

4. Feed Meta meaningful conversion signals

This is one of the most important practical recommendations.

If your objective is to generate qualified leads, optimizing exclusively for form submissions can create a problem.

You may end up with hundreds of inexpensive leads that never become customers.

Build a feedback loop:

Advertisement → Landing page → Lead → Qualification → Sales conversation → Customer.

Use Meta Pixel, Conversions API and supported CRM conversion-event integrations to communicate meaningful outcomes.

This helps Meta optimize against deeper-funnel events when sufficient data and the relevant campaign configuration are available.

Remember that sending CRM events does not guarantee every foundation model directly trains on your individual customer records. It provides optimization signals through Meta's available advertising systems.

5. Don't neglect landing pages

A sophisticated recommendation engine cannot compensate for a poor landing page.

Your website still needs:

  • Fast loading.
  • Clear positioning.
  • A compelling offer.
  • Relevant proof.
  • Trust signals.
  • Simple enquiry forms.
  • Strong calls to action.

The algorithm can help identify potential customers.

Your website must convince them.

6. Build a creative testing system

Instead of randomly producing advertisements, build a structured experimentation process.

For example:

VariableWhat to testProblemFinancial loss, poor leads, low visibilityHookQuestion, statistic, unexpected claimFormatStatic, carousel, videoMessagingEducational, emotional, directOfferFree audit, consultation, case studyVisualProduct-focused, founder-led, infographic

Change variables systematically and track the impact on qualified enquiries and sales.

Creative diversity is valuable, but it should not become uncontrolled campaign fragmentation.


The biggest shift: Advertising is moving from audience selection towards intelligent matching

For years, advertisers spent considerable time selecting:

  • Age groups.
  • Gender.
  • Interests.
  • Behaviours.
  • Demographics.
  • Custom audiences.

These options still matter.

But the direction of Meta's AI development suggests that the competitive advantage is increasingly moving towards understanding customers and communicating the right message through the right creative.

Think about it.

You can manually select an audience interested in entrepreneurship.

But that doesn't necessarily mean everyone in that audience wants your service.

A person who owns a business and is actively looking for SEO services is fundamentally different from someone who simply likes entrepreneurship content.

The newer recommendation infrastructure is designed to understand much richer relationships between users, advertisements and outcomes.

This is why creative strategy, customer psychology, positioning and conversion measurement are becoming increasingly important.

The future of Meta advertising isn't simply about finding the right audience. It's about creating advertisements that the recommendation system can intelligently match with the right people.

Andromeda is an important part of that change.

But it is certainly not the whole story.


Frequently asked questions

What is Meta GEM in advertising?

GEM (Generative Ads Recommendation Model) is Meta's large recommendation foundation model. It learns from advertising signals across different surfaces and objectives and transfers knowledge to downstream recommendation systems.

Is Andromeda the same as GEM?

No. Andromeda is primarily responsible for advertisement retrieval, while GEM is a shared recommendation foundation model. They address different stages and aspects of the advertising recommendation infrastructure.

What is GenRecs in Meta Ads?

GenRecs stands for Generative Recommenders. It uses semantic representations and generative recommendation approaches to improve how Meta understands advertisements, products and user preferences.

Does Meta AI automatically target the right audience?

Meta's recommendation systems use multiple behavioural, contextual and advertising signals to predict relevance. However, advertisers still need to configure campaign objectives, eligibility, budgets, conversion events and other settings appropriately.

Should advertisers stop using interest targeting?

No. Interest targeting can still be useful. But it should be tested against broader targeting approaches rather than treated as the only way to find customers.

Does creative diversity improve Meta Ads performance?

Meaningful creative diversity creates different advertising messages, propositions and visual concepts for the recommendation system to evaluate. It is worth testing, although there is no universal number of creatives that guarantees better performance.

How can I optimize Meta Ads for the newer AI algorithms?

Start with differentiated creative concepts, clear messaging, appropriate conversion objectives, accurate tracking, strong landing pages and systematic testing. Evaluate results using qualified leads and revenue rather than impressions or clicks alone.


Final thoughts

Meta is investing heavily in recommendation infrastructure that understands users, advertisements and their relationships at a much deeper level.

GEM, Andromeda, GenRecs and Adaptive Ranking are particularly important developments.

For advertisers, this means one thing:

The old approach of finding a collection of interests, uploading a few advertisements and hoping for cheap leads is becoming less of a competitive advantage.

The agencies that understand consumer behaviour, develop compelling creative concepts, build strong offers and measure actual business outcomes will be in a much better position.

The algorithm is getting smarter. Your advertising strategy needs to catch up.


Technical sources and further reading

The following are primary engineering publications from Meta:

  1. Meta's Generative Ads Model (GEM), October 2026
  2. Next-Generation Andromeda, October 2026
  3. GenRecs Tokenization, October 2026
  4. Adaptive Ranking Model, March 2026

Research note: This article explains technologies documented in Meta's public engineering publications. Performance figures are attributed to Meta and should not be interpreted as guaranteed results for individual advertisers. Some architectural details have been simplified for clarity.

Share:
V
Vishnu Viswanath
Team at BlackBox Learning · Published November 16, 2026
Previous
Could Jensen Huang Join Elon Musk's Terafab? What We Know

Comments (0)

No comments yet. Be the first!