A System Change Hiding in Plain Sight

If you run Meta ads, you are already living in a new world, even if nobody told you.

Every time you scroll Facebook, Instagram, or Reels, Meta’s AI is working behind the scenes to decide which ads you see. Most advertisers do not realize this, but the last eighteen months have brought the biggest shift to that system in more than ten years.

This is not a small update to Ads Manager.

Meta rebuilt the core engine that powers ad recommendations and moved from “classic” machine learning to foundation-scale models inspired by large language models.

The result:

• Ads feel more relevant
• Journeys are more connected
• The system is much better at understanding creative
  
  

That is good news if you know how to feed it. Not so great if you are still running the same three ad concepts from last summer.

Let’s walk through what changed and what it means for your ad account.

‍

The Five AI Building Blocks Behind Meta’s New Ranking System

To keep things simple, think of Meta’s new stack as a very smart brain with five key parts:

1. GEM – the super brain that learns from everything
2. InterFormer (inside GEM) – the layer that understands context and connections
3. Sequence Learning – the long-term memory that reads behavior in order
4. Andromeda – the concierge that picks which creatives to consider
5. Lattice – the giant library that lets surfaces learn from each other
   
   

These all work together. You do not “turn them on” in Ads Manager. They are the environment you are already operating in.

The question is how to work with that environment instead of against it.

‍

GEM: Meta’s New Super Brain For Ads

GEM is Meta’s largest and most advanced ads model. It is built at a similar scale to modern language models, but instead of predicting the next word in a sentence, it predicts:

• Which ad
• For which person
• In which context
• Is most likely to drive value
  
  

GEM learns from both paid and organic behavior

This is one of the biggest shifts and something most advertisers do not think about.

Older systems leaned heavily on “ad-only” signals. GEM learns from everything people naturally do across Meta’s apps, including:

• Organic posts they view
• Creators they interact with
• Videos they watch to the end
• Ads they click
• Ads they convert on
• The order all of that happens in
  
  

So if someone:

• Watches a creator’s tutorial
• Saves a product Reel
• Clicks a brand’s carousel a few days later
• Then finally purchases from an ad
  
  

GEM is learning from that entire journey, not just the final purchase event.

That richer signal is a big part of why Meta is seeing:

• up to 5% more conversions on Instagram
• up to 3% more conversions on Facebook Feed
  
  

And it is a big reason why “creative is the new targeting” is not just a tagline anymore. The system deeply understands what people actually interact with.

‍

Inside GEM: How InterFormer Makes It Smart

InterFormer is one of the core architectural innovations inside GEM that makes this whole thing work.

If GEM is the brain, InterFormer is the part that understands the story instead of just bullet points.

The problem with older systems

In the old world, recommendation models did something like this:

• Compress a person’s history into a small vector
• Throw away the detailed order of events
• Try to predict what happens next from that compressed snapshot
  
  

That means the system had:

• Limited memory
• Limited context
• A weak understanding of sequences
• Poor sensitivity to creative nuance
  
  

InterFormer changes that.

What InterFormer actually does

InterFormer lets GEM do two things at the same time:

1. Understand sequences
   It looks at the order of actions across surfaces:
   
   
   • Which Reels someone watched
   • Which posts they saved
   • Which ads they clicked
   • What they purchased
   • In what order
     
     
2. Understand cross-feature relationships
   It connects:
   
   
   • Who the person is
   • What creative they saw
   • Which format it was in
   • What platform they were on
   • What time it happened
     
     

Instead of flattening behavior into a small summary, InterFormer preserves the full sequence and then learns which parts actually matter.

That is how GEM can understand that “watching three creator reviews then clicking an ad” is different from “clicking one ad cold.”

Multi-domain learning: context across surfaces

Here is where this gets interesting for actual media buying.

Traditionally, systems either:

Treated Facebook, Instagram, and other surfaces as totally separate worlds

OR

Treated them all the same and ignored that people behave differently in each place

Both approaches are flawed.

GEM, using InterFormer, does something more advanced:

• It learns from cross-surface behavior
• But it optimizes differently for each environment
  
  

So if GEM sees that someone loves short-form product demos on Instagram, it can use that signal to improve which ads they see in Facebook Feed.

At the same time, it knows that Feed behavior is different from Reels, so it does not blindly copy-paste rules from one surface to another. It adjusts the ranking logic to fit the objective and environment.

That is why Meta keeps saying performance is improving “across surfaces” even though you don’t see a new toggle in Ads Manager.

‍

Why InterFormer Makes Creative Diversity Non-Negotiable

Here is where this directly hits your account.

Because InterFormer keeps the full behavioral story and understands relationships across surfaces, it is extremely good at spotting when your creative is basically the same idea in a new outfit.

It can see when you are running the same:

• Hook with slightly different wording
• Creator in the same setting over and over
• Product shot with a new headline
• Design system in 40 variations
  
  

If your ad library is repetitive, GEM has very little to learn from.
If your ad library is diverse, GEM gets rich, distinct signals to work with.

That is why creative diversity has turned into a real performance lever, not just a branding preference.

Because InterFormer understands:

• Which ads someone engaged with organically
• Which creator videos they watched
• Which posts they saved
• Which ads they ignored
• How that changed over time
  
  

It can also decide:

• What angle they should see next
• What format is more likely to land
• When an ad is effectively “done” for that person
  
  

This is the intelligence layer that ties everything together.

‍

Sequence Learning: Meta’s Long-Term Memory

Sequence Learning sits alongside GEM as another key shift. Think of it as Meta moving from looking at a single screenshot of your behavior to watching the full movie.

Older models would say:

• “This person clicked a ski ad. Show them more ski ads.”
  
  

The new system looks at the sequence around that click.

What Sequence Learning does

• Tracks the events before and after a conversion
• Learns how interests evolve over time
• Builds realistic purchase journeys, not just isolated events
• Predicts what someone should see next in the sequence
  
  

In practice, this has already led to 3–4% more conversions in test segments.

A simple example

Old world:

• You book a ski resort
• Meta keeps showing you ski resort ads
  
  

New world:

• You book a ski resort
• Meta starts showing you ski jackets, gear, luggage, and related offers
  
  

The system is no longer basing everything on a single action. It is optimizing for the journey.

What that means for you

You cannot just run “buy now” ads and call it a day.

You need:

• Awareness creative
• Consideration and education creative
• Post-purchase and cross-sell creative
  
  

If you only produce bottom-of-funnel ads, you are under-feeding the very system that is now built to understand sequences.

‍

Andromeda: The Personal Concierge Behind Your Creatives

While GEM and Sequence Learning decide what type of ad someone should see, Andromeda helps decide which specific creatives get considered.

Andromeda is basically the retrieval engine. It runs on Meta’s custom MTIA chips plus NVIDIA Grace Hopper to handle far more complex models than Meta could previously run at this scale.

What Andromeda does

At a high level:

• It looks at tens of millions of ads in the pool
• Uses a much more complex model to score relevance
• Narrows that down to a few thousand candidates in real time
• Sends those candidates into the ranking system
  
  

The key point for advertisers:

Andromeda works best when you give it options.

It thrives on:

• Multiple creative formats
• Different hooks and angles
• UGC alongside product and lifestyle shots
  
  

If you only have a handful of similar creatives in the account, you are not letting Andromeda do its job.

‍

Meta Lattice: The Giant Library Behind The Scenes

Lattice is Meta’s move from “lots of small, specialized models” to “one large model that can share learnings across objectives and surfaces.”

Before Lattice:

• You might have separate ranking models for different placements or goals
  
  

With Lattice:

• A single large architecture can learn across many of these, then apply what it learns in smarter ways
  
  

That is why Meta has reported:

• Up to 12% higher ad quality
• Up to 6% more conversions in some cases
  
  

The takeaway here is straightforward:

When you give Meta more surfaces, more placements, and more room to explore, Lattice can actually use that to your advantage.

This is part of why Advantage+ Shopping and broad strategies are performing so well in many accounts.

‍

The Big Picture: What This Actually Changes For Advertisers

If we pull this all together, here is the practical reality:

You are no longer just “targeting audiences” in the way you used to. You are feeding a foundation model.

That model:

• Learns from paid and organic behavior
• Understands sequences across time and surfaces
• Evaluates creative as a core signal, not an afterthought
• Can see when your creative is repetitive or low-variety
• Optimizes for the journey, not isolated last-click actions
  
  

So the old habits that used to kind of work, like:

• Spinning minor creative variations
• Over-optimizing narrow audiences
• Clinging to one “hero” ad for months at a time
  
  

Are increasingly misaligned with how Meta actually works after this GEM update.

‍

Action Plan: How Brands Should Adapt In 2025

Here is how I would approach this from a practical standpoint.

1. Build real creative diversification pipelines

Not “20 captions on one image.”
Different concepts and creative sources.

You want diversity across:

• Hooks
• Angles
• Personas
• Visual styles
• Formats (static, UGC, founder-led, testimonial, product demo, etc.)
• Different agecnies or creative strategies bringing net new view points and ideas to the table
  
  

If you have 100 ads but only 10 actual concepts, you have a diversity problem.

2. Shorten your creative iteration cycles

Think weeks, not quarters.

• Drop new creative every 1–2 weeks in active accounts
• Retire or rotate out concepts before they fully die
• Treat ad creative as a living portfolio, not a one-time project
  
  

GEM and Andromeda can only work with what you give them.

3. Scale UGC and creator-led content

Scaling UGC production is not a trend piece here. It is a structural advantage.

Creator content:

• Introduces new faces and styles
• Helps break creative similarity
• Often maps better to how people consume Reels and Stories
  
  

Whitelisting from creator handles can also act as an audience unlock when rolling reach starts to stall. Personally, I believe the demand for UGC will rise in 2026 as brands realize it’s one of the few ways to maximize true creative diversity at scale, in Meta’s view. 

4. Lean into Advantage+ and broad strategies

You do not need to micromanage every audience segment.

Let Lattice and GEM:

• Figure out who to show what
• Across surfaces and placements
• While you focus on feeding better creative
  
  

This does not mean never using structure. It means respecting what Meta is good at now and not fighting the system on every campaign.

5. Watch for creative sameness

If your ad library looks “on brand” but visually identical, you will hit a ceiling.

You want to avoid creatives with the same:

• Backdrop in every video
• UGC creator in 15 similar scripts
• Studio look and feel in every shot
  
  

Think in themes and angles, not just visual polish.

6. Map your creative to the full journey

Use the creative bucket system to build our content across the full purchaser journey.

Trigger Ads — 10–20%

• You’ve got the problem. Here’s the fix.
• Purpose: Solve an urgent pain point right now.

‍

Exploration Ads — 25%

‍

• Curious? Let us explain.
• Purpose: Educate consumers in a casual, compelling way.

‍

Evaluation Ads — 25%

‍

• Why we’re different (and better).
• Purpose: Build trust, comparison, and differentiation.

‍

Offer/Purchase Ads — 30–40%

‍

• Take the next step.
• Purpose: Move educated buyers to purchase.
• Offer Types: Bundles, limited runs, bonus gifts, or consultation-based.

‍

Sequence Learning is doing this work behind the scenes. Your job is to give it the right building blocks.

‍

Welcome To The Creative-First Era Of Meta Ads

Meta’s new AI systems are not something you can ignore and hope they go away. This is the direction the platform is committed to.

The advertisers who win in this environment are not the ones with the “best hack” or the most complicated naming conventions. They are the ones who:

• Take creative diversity seriously
• Design for real customer journeys
• Give the system room to explore
• Measure performance with that reality in mind
  
  

You cannot scale by showing the same people the same creative forever, no matter how efficient it looks today.

You are now feeding a foundation model.
The more thoughtful, diverse, and journey-aware your creative inputs, the better that model will perform for you.

The brands that adapt will win. The brands that resist will get left behind.

If you want support building a creative diversification pipeline that aligns with how GEM and InterFormer actually rank ads, Brighter Click can help.