Meta updated its ad ranking system — what is Adaptive Ranking Model and how it works

Meta continues to rebuild its advertising system around artificial intelligence. In March 2026, the company introduced a new ad ranking system called the Adaptive Ranking Model, which is already being used on Instagram.

Adheart – Meta Ads Ad Intelligence Tool to uncover your competitors' top-performing ads and trending products

The key idea is that Meta has started using models at a similar scale and complexity to modern AI models to select which ads to show. This means the system can analyze far more signals and better understand user interests and intent — which directly affects ad performance.

This is one of the most important changes to Meta’s ad algorithms in recent years because it changes the fundamental way ads are selected and ranked.

Meta Ads attribution changes: what counts as a click and updates to video engagement

Meta is updating its ad measurement model: click attribution will now include only link clicks, while social interactions move to a separate category

Adheart BlogAdheart Team

Let’s break down what this means.

How ad selection works in Meta

Every time a user opens Instagram or Facebook, the system must decide very quickly:

• which ads to show,
• in what order,
• from which advertiser,
• with what bid,

— all in less than a second.

The process roughly looks like this: user opens the feed → the system analyzes behavior (clicks, purchases, interactions, interests) → a list of potential ads is generated → the system ranks these ads by probability of click or conversion → the auction happens → the ad is shown.

The ad ranking stage is exactly the part that was completely rebuilt in the new Adaptive Ranking Model.

Why Meta had to change the system

Meta says they ran into a fundamental problem called the inference trilemma — a trade-off between model complexity, response speed, and computing cost.

The more complex the model, the better it understands users and selects ads. But more complex models require more computing power, higher infrastructure costs, and slower response times.

For an advertising system, this is critical because:

• ads must be selected in milliseconds;
• the system runs for billions of users;
• computing costs must stay low or ads become unprofitable.

So Meta needed a way to use very complex models without losing speed or massively increasing infrastructure costs.

What is the Adaptive Ranking Model

The Adaptive Ranking Model solves this problem.

The main idea is to move away from a “one model for everything” approach and instead use different models depending on the situation.

Previously, the system used roughly the same model for all requests. Now Meta uses dynamic request routing, meaning the system decides which model to use for each specific user and situation.

Celebrity ads: what celeb bait is and how It works

Learn what celeb bait is in advertising, how it works, which industries use it and how to find these ad creatives in Meta Ads

Adheart BlogAdheart Team

The system analyzes the context: user behavior, complexity of the situation, number of signals, probability of conversion, — and then decides whether to use a simpler or more complex model.

This is called intelligent request routing.

In simple terms:

• if the situation is simple → use a faster, lighter model;
• if the situation is complex → use a more powerful and accurate model.

This allows Meta to use LLM-level model complexity but only run those heavy models when necessary.

As a result, Meta can simultaneously improve ad selection accuracy, keep the system fast,

and control computing costs.

The shift to LLM-scale models

The biggest news in 2026 is that Meta managed to scale its ad recommendation system to LLM-scale models.

This means ad ranking models now:

• have far more parameters;
• use longer sequences of user behavior;
• better understand user intent;
• can work with a huge number of signals.

At the same time, the system still works in real time and selects ads in milliseconds.

Meta says their system can now work with models at the trillion-parameter scale, which was previously impossible for real-time recommendation systems.

Three key technology changes

Meta describes three main technological directions that made this possible.

1. Inference-efficient model scaling

The system moved to a request-centric architecture — now complex computations are done once per user request, not separately for each ad.

Previously, the system calculated signals for each user–ad pair. Now complex user signals are computed once and reused for all ad candidates.

This significantly reduced computing costs.

2. Model / system co-design

Meta started designing models together with infrastructure and hardware.

This means model architecture, GPUs, memory, and infrastructure are now designed as one system, which allows resources to be used much more efficiently and enables very large models to run faster.

3. New Infrastructure for extremely large models

Meta built infrastructure that allows them to:

• distribute models across multiple GPUs;
• work with very large embedding tables;
• scale models to trillions of parameters;
• automatically scale infrastructure depending on load.

This is essentially a new AI infrastructure layer for Meta’s systems.

Results

According to Meta, after launching the Adaptive Ranking Model on Instagram in Q4 2025:

• +3% ad conversions
• +5% CTR

At Meta’s scale, these are huge numbers — even a 1% change in the ad system can mean billions of dollars.

Adheart – Meta Ads Ad Intelligence Tool to uncover your competitors' top-performing ads and trending products

What this means for advertisers

The main takeaway is that ad algorithms are becoming more complex and more behavior-driven, and the system is moving away from manual targeting toward signals and creatives.

Meta is gradually moving from this model targeting → audiences → interests to the model user behavior → signals → algorithm → creative.

Key trends:

• algorithms understand user intent, not just interests;
• the system decides who sees the ad — manual targeting becomes less important;
• behavioral data (clicks, views, interactions) becomes more important than interests;
• creatives and offers become the key success factors;
• the algorithm increasingly optimizes campaigns automatically.

Meta is moving toward a system where the advertiser:

• uploads creatives,
• sets the budget,
• selects the optimization event,
• and the algorithm does the rest.

This means that in Meta advertising, creatives, offers, behavioral signals, conversion data, and optimization events are becoming more important, while detailed manual targeting and interest audiences are becoming less important.

Conclusion

The Adaptive Ranking Model is essentially a new generation of Meta’s advertising system.

Meta is starting to use LLM-scale models not only for chatbots and AI assistants, but also for ad selection, content ranking, recommendations, and feed personalization.

This means advertising in Meta will become:

• more personalized,
• more behavior-driven,
• more dependent on creatives,
• and even more controlled by algorithms rather than manual targeting settings.

For advertisers, this means one thing: understanding algorithms and working with creatives is becoming more important than manual audience targeting.

In a world where Meta’s algorithms are becoming more complex and increasingly driven by behavioral signals and creatives, advertisers need to understand which creatives and approaches are currently working in different niches and GEOs.

That’s exactly why tools like Adheart are used — a Meta Ads intelligence platform that allows advertisers to analyze competitors’ ads, find long-running creatives, study funnels, offers, and advertising strategies across different markets.

When the algorithm becomes smarter, the advertiser’s main advantage is not targeting — but creatives, offers, and market understanding.