How Meta’s AI Overhaul is Forcing Advertisers to Evolve

If you’ve been running Meta Ads recently, you may have noticed unexplained shifts in performance – spend moving between ad variants, impressions dropping without warning. 

Well, it’s because the playbook media buyers relied on for a decade – intricate audience segmentation, interest stacking, manual budget control – is being rapidly superseded.

The driver? A  rewiring of how Meta decides who sees what – and when.

It’s Bigger Than Andromeda

Most industry commentary pins this shift on Andromeda, Meta’s AI-powered ad retrieval engine. But as Triple Whale’s Bryant Garvin points out, the reality is more complex: Meta didn’t roll out a single algorithm update – they built an entire AI framework of four interlocking systems that work together in ways that make the old ‘hacks’ obsolete.

Those four systems are: GEM (a foundation model that trains all the others), Lattice (the unified decision-making architecture), Andromeda (the ad retrieval engine), and UTIS (a direct user satisfaction survey launched in January 2026). Understanding how they interact matters more than fixating on any one component.

GEM sits at the top – it doesn’t serve ads directly, but teaches all the other models what ‘good’ looks like, learning from ad content and organic engagement across every Meta surface. Andromeda then narrows tens of millions of ads down to a shortlist of candidates worth serving. The result: a 10,000x increase in model capacity for personalisation, with Meta tripling Andromeda’s compute efficiency in 2026.

Critically, when someone discovers a tactic that works, GEM sees it, learns from it and teaches the downstream systems. So, the platform adapts in days or hours. There’s no single lever to exploit anymore.

Creative is the New Targeting

This architecture has flipped the traditional advertising model. Campaign managers no longer dictate targeting – the algorithm does. Meta’s new ad systems incentivise uploading a greater diversity of creative content and showing it to a broad audience; Meta’s AI then determines which messages, formats and angles are most likely to resonate with different sets of users.

The creative itself has become the targeting signal. Andromeda uses computer vision and audio analysis to ‘read’ an ad – aka it takes in the pixels in an image, the pacing of a video, the emotional tone of a script. A video featuring a busy professional talking about saving time will find busy professionals who respond to efficiency messaging. A premium product shot will find users who have historically clicked on high-end content.

This has practical structural implications. Narrow audiences starve the algorithm of data. The modern best practice is broad targeting, simplified campaign structures – often one campaign and one ad set – packed with a diverse library of creatives.

The Volume Bottleneck – and the Real Creative Problem

This creates a significant operational challenge. Meta’s system assigns an ‘Entity ID’ to creatives and measures similarity. Upload ten images that are visually near-identical – same concept, swapped button colour – and Andromeda groups them. Advertisers receive zero additional learning, and accounts can suffer higher CPMs from creative fatigue.

But here’s an important nuance: more creative doesn’t automatically mean better. Marketers often think the new system means they need a huge number of minor variations of the same ad concept – but what actually works best is several fully distinct concepts and stories. One agency head described a recent brief that began as a request for 300 assets and, once properly mapped across four target personas with five concepts each, required 1,000 distinct creative assets.

The demand for genuine creative variety is now a structural feature of how Meta rewards ad spend. A healthy ad set benefits from creative that covers different styles (raw UGC, polished brand video, static, carousel), different angles (feature-led, benefit-led, emotional, social proof) and different hooks designed to stop the scroll for entirely different customer personas.

AI Wariness Is Real – And Worth Acknowledging

The obvious response to this volume demand is AI tooling. But the industry’s actual adoption is more cautious than the hype suggests.

Big brands remain skittish about AI-generated creative, with legal concerns around undisclosed image generation being a primary hesitation. And for good reason. Most brands want to retain control because of the time and effort they’ve invested in building their identity.

Practitioners describe playing ‘Whac-A-Mole’ with Meta’s AI features – constantly identifying what has been turned on by default in order to test it deliberately, rather than being opted in without warning. Meta has since introduced opt-out preferences that persist across campaigns, but the tension between automation and brand control remains live.

That said, the direction of travel is clear. Meta’s Advantage+ campaigns now account for 60–70% of some agencies’ total Meta spend, even as practitioners note the ongoing need for human oversight in areas like placement quality and account-level setup.

The AI Imperative: Smarter, Not Lazier

The solution isn’t to hand everything to Meta’s native tools – it’s to build a modular creative system in which human strategy sets the direction and AI handles scale. This distinction matters.

AI image generation can take a single product shot and place it in multiple contextually distinct settings – a modern kitchen, a busy office, an outdoor environment – without a new photoshoot for every variant. Automated formatting tools can instantly adapt a core design across 9:16, 4:5, and 1:1 ratios, adjusting focal points without manual pixel-pushing. Copy tools can generate multiple hooks from a single value proposition – humorous, urgent, emotional, logical – enabling genuine psychological diversity in testing.

Critically, the system rewards real marketing – unique selling propositions, diverse creative angles and messages that meet people where they’re at as customers. In short, the AI has stopped rewarding lazy tactics. Agencies using AI to produce scale without strategic diversity will see diminishing returns just as quickly as those still working manually.

Turning Creative Pressure Into a Strategic Offering

For creative and media agencies, this shift is both a threat and a genuine opportunity. Meta is attempting to democratise media buying – anyone can launch an Advantage+ campaign and press publish. What brands can’t do themselves is produce and scale strategically diverse creative that actually feeds the algorithm distinct signals.

The strongest agency positioning in this environment is stepping up as a creative diversification partner. That means educating clients that performance is now primarily a function of creative, not targeting; shifting the conversation from ‘five good ads’ to ‘a modular system that tests 40 distinct behavioural signals’; and being transparent that AI is what makes this scale economically viable.

Old Agency vs. The Andromeda-Ready Agency

Conclusion

Meta’s AI overhaul – Andromeda, GEM, Lattice and UTIS working in concert – has made creative the absolute centre of paid social strategy. The old targeting levers are losing relevance; the new performance lever is the quality and genuine diversity of the creative itself.

Agencies slow to adapt from manual versioning will struggle to meet the demand. But those that build AI-assisted creative systems – grounded in real strategic thinking, not just volume for its own sake – will be able to deliver the diverse signals these platforms reward. Where Meta leads, other platforms will follow. The time to build for this new reality is now.