We’ve spent a lot of time talking about how AI is changing the way advertising gets made.
AI is changing how advertising gets understood.
This week, Meta shared details about upgrades to its advertising recommendation systems: Andromeda, GEM, and a newer approach called GenRecs.
The interesting development isn’t simply that Meta can evaluate more ads, faster. It’s that its systems are getting better at understanding what an ad actually represents, rather than relying primarily on historical performance tied to individual ad identifiers.
That could have meaningful implications for creative strategy.
From identifying ads to understanding them
Traditionally, advertising recommendation systems have learned from signals associated with individual campaigns, creatives, and products.
Meta’s GenRecs approach uses semantic representations built from information including an ad’s text and images. Instead of treating every new creative as something entirely unfamiliar, the system can recognize relationships between similar products and concepts.
A newly launched running shoe, for example, can benefit from what the system has already learned about similar footwear.
Meta says this approach helps shorten the learning period for new ads. In its internal February 2026 measurement, GenRecs tokenization, combined with other ranking models, delivered a 0.35% conversion lift.
That’s a specific, company-reported improvement, not proof that every campaign will perform better.
Creative is becoming part of how the platform determines relevance.
More variations aren’t necessarily more ideas
This is where I think things get particularly interesting for creative teams.
We’ve gotten very good at producing variations. Different headlines, colors, backgrounds, calls to action, and aspect ratios.
But if ten variations communicate essentially the same thing, are we really giving the platform ten different ideas to work with?
Consider a mobile gaming campaign.
One ad emphasizes competitive gameplay. Another focuses on solving a puzzle. A third demonstrates character progression and rewards.
Those aren’t simply different visual treatments. They’re different reasons someone might become interested in the game.
That distinction matters if recommendation systems are increasingly able to recognize the concepts behind creative.
Not just, “How many versions can we produce?”
But, “How many meaningful motivations are we actually exploring?”
Where interactive creative fits
For playable and interactive advertising, there’s another layer to consider.
An image can communicate what a game looks like. An interactive experience can help someone discover what playing it feels like.
You might introduce a puzzle through its artwork, but the experience of completing the first move, receiving feedback, or unlocking a reward creates a different kind of connection.
An important distinction: Meta’s published GenRecs explanation does not establish that its semantic models understand every interaction occurring inside an HTML5 playable.
So I wouldn’t assume an algorithm can interpret the entire experience simply because it understands the creative’s visible content.
But that’s exactly why creative teams should look beyond surface-level variation.
We can test different motivations, interactions, and experiences while measuring what people actually do with them.
The opportunity ahead
For me, the practical takeaway is straightforward.
Start with genuinely different creative concepts. Make the value of each one clear. Use interaction data, where available, to understand what resonates. Then connect those signals to meaningful campaign outcomes.
The more intelligently platforms match advertising to people, the more important it becomes to give those systems something genuinely different to match.
AI might make it easier to produce a hundred ads.
The harder, more valuable work is deciding what those hundred ads should actually say, show, and let people experience.
Editorial note: Meta’s conversion result is company-reported and comes from GenRecs Tokenization paired with other models; this article does not claim that Meta’s system understands in-ad HTML5 playable interactions.