Meta advertising has reached a turning point. Not with a big announcement or an overnight switch-off, but through a fundamental change in how ads are selected, shown and scaled across the platform.
In the early days of Meta advertising, performance was heavily influenced by how precise campaigns were setup and adjusted over time. Success often depended on how much time you could spend refining targeting rules, structuring accounts, optimising ads and closely managing budget shifts.
Over the years, we’ve seen Meta introduce advancements that streamline setup and improve delivery, but effective performance has always required a level of strategic oversight. In that era, the system rewarded those who could most effectively "out-tinker" the available controls.
However, we’ve seen that the traditional operating model has shifted significantly. As privacy regulations evolved and signal loss became a reality, the deterministic targeting we once relied on lost its edge.
Rather than simply adjusting the existing framework, Meta has fundamentally rebuilt its advertising infrastructure around an AI-first core with ad delivery now a fully AI-driven system. This is a complete pivot from manual control to a system driven by machine learning and automated optimisation.
Meta's new AI-driven advertising system is built around two major components:
Together, they’re reshaping how ads are selected, ranked and scaled across Facebook and Instagram.
Let’s break down what these changes mean, why it matters and how marketers should respond.
Meta’s latest update moves ad delivery into a fully AI-driven system, where:
All in real time.
Previously, Meta’s ad delivery worked in stages:
Now, Meta’s AI can:
In short: the system has got smarter, quicker and less dependent on rigid inputs.
Andromeda is Meta’s machine learning (ML) system designed for retrieval in ad recommendation, essentially the system that decides which ads are even worth considering for a user at any given moment.
Instead of pulling from a defined audience, Andromeda begins with analysing past performance engagement, ad copy, ad creative and ad formats. By doing this, Andromeda can predict which ads a user will find most interesting, helping advertisers meet their campaign objectives, whether increasing brand awareness or acquiring new customers.
The first step in Meta's new multi-stage ad recommendation system is retrieval, where it scans millions of potential ads and narrows them down by selecting a few thousand relevant ad candidates. The next step, a more sophisticated ranking model, predicts user and advertiser value to determine the final sets of ads to be shown to the user.
Essentially, we’re moving away from narrow targeting and toward creative-led growth. When we give the system more creative assets to work with and a broader campaign structure, it has a much better chance of finding the combinations that actually drive conversions.
If Andromeda is the engine, GEM (Meta’s Generative Ads Recommendation Model) is the central intelligence directing it. Think of it as a massive pattern-recognition machine. It’s constantly looking at how people interact with organic posts, which ad sequences they actually enjoy and which messaging styles lead to a sale.
It synthesises all those behaviours, what people click, how they browse, and what they buy, to create a "prediction map." GEM then feeds these insights directly into Andromeda, helping the system predict exactly which ad will resonate with which person, and precisely when it should be shown.
GEM's goal is to deliver increased ad performance and advertiser ROI by enhancing other ad recommendation models' ability to serve relevant ads.
Since its rollout in mid-2025, the impact has been significant. By the end of last year, Meta reported that GEM is four times more efficient at driving performance gains compared to the old ranking models. Essentially, the system isn't just guessing anymore; it’s learning at a scale that current optimisation simply can’t match.
This update changes how success on Meta is achieved. Instead of ongoing adjustments, the focus now moves toward a creative-first strategy and a simplified account structure. It’s about giving the system the stability it needs to learn, which means favouring patience over frequent changes.
Targeting still matters, but creative is doing more of the work than ever.
We recommend:
Over-segmentation now works against the system. Meta’s AI performs best when:
Fewer campaigns. Fewer ad sets. More focus on what people actually see.
The system adapts quickly, but only when it has enough data. That means:
Short-term tinkering can reset learning more than it helps.
Budgets are now moved dynamically toward what’s working. Your job isn’t to force spend, it’s to:
Meta’s latest update allows marketers to spend less time managing ads and more time improving what people actually see. Instead of spending hours adjusting audiences, bids and structures, the focus can now shift to shaping stronger ideas, better creative and clearer strategy.
One hero ad isn’t enough anymore. Winning brands are:
Variety fuels performance.
GEM understands nuance better than ever. Clear, natural language consistently outperforms:
If it wouldn’t sound right out loud, rethink it.
Meta’s AI doesn’t need micromanaging; it needs guidance. Set strong foundations:
Then let the system do what it’s designed to do.
With more automation, vanity metrics matter less. Focus on:
The platform is optimising, your reporting should evolve too.
Meta’s move to AI-first advertising can feel daunting, but it also creates new opportunities for brands willing to adapt.
At STM AGENCY, our performance and paid social specialists are already working with Meta’s AI-led approach, helping brands adapt their creative, structure and strategy to what actually drives results. If you need help navigating Meta's AI-first innovation, get in touch to see how we can help.