blog31 | 10 | 2025

Meta’s Andromeda AI: Evolution, Learnings, and Optimization

Author

Romina Schwarz

Index

BLOG

Meta Ads It’s no longer optimized like it used to be.

In recent months, Meta completed the consolidation of Andromeda, its new artificial intelligence system for ad delivery.
It didn’t come with big announcements or visible alerts in Ads Manager, but its impact is real and is already being felt in campaigns.

Many marketing teams began to notice that:
Some campaigns became more unstable in the short term.
Microsegmentation stopped performing as well as it used to.
Creativity began to explain a large part of the results.

None of this is a coincidence.

Andromeda changed the optimization logic of Meta Ads: today the system decides which ad to show to each person in real time, prioritizing behavioral signals and creative quality.

Adapting is not optional. But it’s not complicated either if you understand the underlying change.

IN THIS ARTICLE

What is Andromeda?

Meta introduced Andromeda as its new internal machine learning engine for ad delivery: a system that processes millions (or tens of millions) of ad candidates, narrows them down to thousands, and then decides which one to show each person in real time.

Among its technical innovations: deep neural network architectures, specialized hardware (such as NVIDIA’s Grace Hopper Superchip or Meta’s own MTIA), and a hierarchical indexing logic to handle the massive volume of possible creatives.

Additionally, it operates in a layer prior to the final ad ranking, filtering millions of initial options down to only those most likely to perform, thereby improving the efficiency of the delivery system. This represents a direct update to the old retrieval engine, which could no longer handle the current scale of formats, ads, and user signals.

Meta also reported that Andromeda runs on internal high-performance infrastructure, integrating proprietary hardware with advanced GPUs to accelerate machine learning and reduce latency in ad delivery.

The effects already reported:

Improvements in ad quality (~8%),
Improvements in system recall (~6%),
ROAS growth (~22% in some cases with Advantage+), according to industry sources.

What changed with Andromeda (in a nutshell)?

Until recently, Meta Ads primarily operated on:

Manually defined audiences
Fragmented structures
Stricter optimization rules

With Andromeda, the focus is different:

AI predicts, at the individual level, which ad is most likely to generate a result.
Creativity gains weight with the audience.
The system learns best when it has fewer structures and more data volume.

In practice, this translates into less manual control and more automated intelligence.

Step by step: how to adapt your campaigns with Andromeda

1. Simplify the campaign structure.

The first step is to let go of unnecessary complexity.

Working with fewer campaigns, broad audiences, and a consolidated budget allows Andromeda to gather more signals to learn from and optimize.

Highly fragmented structures, which used to work, now tend to work against you.

Sometimes, simplifying is the greatest advance.

2. Treat creativity as the new segmentation.

With Andromeda, ads become the true engine of performance.

This implies:

Activate multiple creatives per ad set
Vary messages, formats, and approaches
Think of each ad as a distinct story

It’s no longer enough to change some text or a color.

Creativity is what allows the system to find the right person.

3. Use automatic locations.

Allowing Meta to optimize placements is key in this new scenario.

Activating Advantage+ placements and designing assets specifically for vertical format helps AI allocate inventory where it’s most likely to deliver results, without imposing manual assumptions.

4. Feed the algorithm with more signals.

Optimizing only for the final event can limit learning, especially in low-volume campaigns.

Including intermediate events, such as engagement, content views, or partial leads, gives the system more information to learn quickly and optimize better.

More signals, better decisions.

5. Give learning time.

One of the most common mistakes is adjusting campaigns reactively.

Andromeda needs stability to learn.

Constant changes in budget, structure, or creative assets often slow down that process.

Evaluating in 7- to 10-day windows and adjusting one variable at a time usually yields better results than daily optimization.

How to analyze whether the changes actually worked

Hacer los ajustes correctos es solo una parte del trabajo.
La otra, igual de importante, es leer bien los datos.

1. Compare equivalent periods.

To assess the real impact, it’s crucial to compare:
30 days before vs. 30 days after
Campaigns with the same objective

Avoiding day-to-day analysis helps understand the system’s actual behavior.

2. Look beyond the CPA

With Andromeda, it’s not all about a single metric.

It is important to analyze:

Conversion trends
Cost per result along with volume
CTR and engagement per ad
Average frequency
Distribution of results by creative

There may be more daily variation, but better cumulative performance.

How do Bunker Analytics and ADA AI help interpret this new scenario?

In an increasingly automated environment, the real value lies in understanding what’s happening, not just in executing.

That’s where Bunker Analytics and ADA AI make a difference.

Bunker Analytics: a clear vision for making decisions

Bunker Analytics centralizes Meta Ads data in an environment designed for strategic analysis.

This allows you to compare periods, identify trends, and understand performance by campaign, ad, and creative, without relying on the rigid views of Ads Manager.

Less time assembling reports.

More time understanding results.