Meta ads after Andromeda: why varied creative matters more.

For years, running Meta ads well meant building careful audiences: interests, lookalikes, exclusions. That still has a place, but the centre of gravity has moved. Meta's systems now decide much more of who sees what, and the inputs you control that matter most are the creative you give them and the conversion data you send back.

What Andromeda is

In December 2024 Meta described Andromeda, the machine learning system behind the retrieval stage of its ads pipeline. Retrieval is the first cut: narrowing tens of millions of possible ads down to a few thousand that later models rank for each person. Meta reported a 6 percent improvement in recall and an 8 percent improvement in ads quality on selected segments after rolling it out across Facebook and Instagram, and said it was built to handle the much larger pool of ads that Advantage+ automation and generative AI tools create.

Why creative now does the targeting

If the system can consider far more ads for each person, the ad itself becomes the strongest signal of who it is for. A video built around a specific problem, a testimonial from a particular kind of customer or an offer aimed at first-time buyers each finds its own audience. The practical shift for advertisers is to supply genuinely different ideas, not slight variations of one, which makes content production part of the ad strategy.

• Vary the idea, not the colour: different customer problems, different proof (a demo, a review, a before and after), different formats (short video, carousel, static).
• Fewer, broader campaigns usually give the system more room to learn than many small, overlapping ones.
• Review anything generated automatically. An automatic variation can keep the look of an ad and still weaken the claim.
• Judge results by concept and angle, not only by individual ad, so you learn which ideas to make more of.

Clean conversion data is the other half

The system optimises towards whatever you tell it counts as success. If that signal is patchy, it optimises towards the wrong people. Browser tracking alone misses events, so Meta recommends sending events from your server through the Conversions API as well as the pixel, and removing the double counts.

• Send the same event from browser and server with a matching event name and a unique event ID, which is the deduplication method Meta recommends.
• Deduplication only works within 48 hours of the first event with that ID, so send both promptly.
• Optimise for the event that matters to the business, such as a qualified lead or a purchase, not a cheap proxy like a page view.
• For lead businesses, feed outcomes from your CRM back to Meta so it learns which leads became customers. Workflow automation can send these back without anyone copying them by hand.

Budget follows evidence. With varied creative and trustworthy conversion data in place, the weekly job becomes simple to describe: move money towards the ideas that bring customers at a cost you can live with, and away from the ones that do not. That weekly loop is how we run paid ads on Meta and Google.