Meta upgrades ad ranking with LLM tags, GEM and Andromeda

So a new ad no longer starts from zero: the system can match its labels to products that already have results. Matt Steiner, who runs Meta’s ads infrastructure, says this shortens the learning phase, the first stretch when the system is still finding out who responds to an ad, without ending it. He also says advertisers need not change any setting.

Andromeda now runs over the whole pool of eligible ads, more than a billion a second, on one GPU server, according to an engineering post PPC Land summarised. Meta credits it with 2.4 per cent more Facebook conversions since 2025 and gives no baseline. GEM, the central ranking model, now looks twice as far back through each person’s activity. Steiner’s reading: “The importance of targeting is diminishing over time.”

Meta’s Andromeda post credits its creative tools with hundreds of millions of new ads. Our census of AI ad figures holds Meta’s own counts of their users, and today adds F8: nearly 10 million advertisers in the third quarter, the figure Meta’s finance chief gave on 21 October.

So when one of those advertisers uses Meta’s tools to make a new version of a product ad already running, the new ad shares its product labels with an ad that has results, the case Steiner says learns faster.