Skip to content
Social Rocket

Meta Andromeda evaluates one billion ads per second – what advertisers get from it

Meta reports 100 times more throughput for Andromeda and 2,4 % additional Facebook conversions since 2025. What the figures actually measure – and why rising ad prices put the benefit into perspective.

From Social Rocket editorial team

This text was created with AI assistance and reviewed editorially before publication. All sources are linked below.

Official illustration of the Meta Andromeda AI for ad pre-selection
Meta

On 28 October 2026, Meta presented a new architecture for its ad selection system Andromeda. According to Meta's engineering department, the infrastructure evaluates more than one billion ad candidates per second in production – roughly 100 times as many as the system from 2024. The promise for advertisers is more precise matching. Whether this makes a specific campaign profitable cannot be derived from the technical record alone.

One billion candidates are not one billion ad contacts

Andromeda takes over the pre-selection: from a very large number of fundamentally suitable ads, a few candidates for the more complex ranking are identified in milliseconds. According to Meta, the new combination of specialised AI models and GPU optimisation can consider more creative assets without the selection taking significantly longer. The billion therefore refers to internal model evaluations, not deliveries or visible impressions, clicks or new customers.

This is particularly relevant because generative creative tools increase the number of distinct assets. More variants bring little if a system already sorts out suitable ads in the first selection stage. Meta complements this method with semantic representations of ads: the GenRecs system is meant to classify content with little prior interaction history better too.

Which success figures are actually available

Meta puts the conversion increase achieved through Andromeda since 2025 at 2,4 percent on Facebook. The third-quarter figures also cite 1,4 percent more Facebook conversions from newer Andromeda improvements. These percentages refer to different periods or changes and must not simply be added together.

These are performance changes reported by the platform operator – not an independently confirmed revenue increase for every advertiser. The report from Social Media Today mentions further Meta-internal tests on ranking, Business Assistant and generative recommendations; these figures too are not a general success forecast.

Better matching alongside higher ad prices

The economic cross-check is in Meta's official quarterly report from 21 October: the number of ad impressions delivered across the Family of Apps rose by 10 percent compared with the same quarter last year, while the average price per ad rose by 14 percent. Both figures apply platform-wide and are not a German CPM.

For German media teams, this means more automated relevance can coincide with higher buying prices. What remains decisive is your own qualified conversion value, returns, the share of new customers and additional sales, rather than platform attribution alone.

In practice, it is worth testing several clearly distinguishable creatives with stable conversion signals and evaluating the results by audience, placement and actual contribution margin. Our analyses of the performance of individual Meta creatives, the learning phase in Ads Manager and the AI assistant for campaigns fit alongside this. A controlled lift test is more meaningful than transferring a global model benchmark to your own ad account.