Meta Explains How Its AI Narrows Millions of Ads to a Shortlist

Meta’s October 28 engineering disclosure gives marketers a clearer view of what happens before an ad reaches the final selection stage. For a small team paying for Meta ads, it also draws a useful boundary around claims that producing more creative will automatically improve results.

The company says its production Andromeda system evaluates more than a billion ads a second on one server using a graphics chip. That measures processing speed. It doesn’t count impressions or promise an advertiser more sales.

In the disclosure, Haiyu Lu and 11 coauthors say the system maintains that speed:

while scoring every eligible ad in the candidate pool.

An ad needs to survive that first cut before the final ranking models can consider it.

The shortlist comes before the winner

Meta’s original Andromeda explanation, published in December 2024, describes a two-stage problem. Retrieval reduces tens of millions of possible ads to a few thousand candidates. More demanding ranking models then assess the survivors and determine what someone sees.

Think of a selection process with a shortlist. Improving how carefully the finalists are compared won’t help an appropriate candidate that was discarded earlier. The first selection has its own effect on what the later decision can achieve.

That makes retrieval worth understanding even when the advertiser never touches its controls. A team looking only at the final result sees the combined effect of several decisions inside the platform. Naming one system doesn’t isolate its contribution to a campaign’s performance.

The 2024 account already connected this problem to the growing supply of ads from Advantage+ automation and generative creative tools. Better generation creates a matching problem: the platform needs to consider more possibilities within the short time available to choose an ad.

Our earlier coverage of ChatGPT’s creative-enhancement permissions concerned who can change the advertisement. Andromeda concerns an earlier selection question about which eligible advertisements advance. Marketers need to keep both questions visible when discussing AI with their media team.

What the new disclosure establishes

The newer design keeps eligibility matching, scoring, and shortlist selection together on that chip, reducing movement of data between machines. Meta says this preserves advertisers’ targeting rules. The October post describes a production system, without dating its deployment or announcing an advertiser setting to enable.

It also attributes a 2.4% increase in Facebook conversions to Andromeda since 2025. That company-reported figure spans a longer period than this disclosure. It doesn’t isolate the new architecture’s effect or establish what any particular account will gain.

A platform’s engineering benchmark answers a question about its capacity. An advertiser’s creative test answers whether a particular message earns worthwhile business at an acceptable cost. Moving between those questions in a budget meeting requires evidence the benchmark doesn’t supply.

Meta has also described separate improvements elsewhere in its recommendation system. Its November 2025 GEM account explains how a larger model learns from ad content and people’s interactions, then passes learning to other recommendation models. Those systems give further reasons to resist assigning every movement in an account to Andromeda. Their reported gains aren’t numbers an advertiser can simply add together.

An older warning about creative volume still applies

Jon Loomer, founder of the Meta-advertising education business Jon Loomer Digital, wrote about their own experience in November 2025. After trying 25 and 50 ads per ad set, they wrote:

I didn’t see any clear advantage to it.

Loomer described returning to creative variety at a smaller scale. That was a practitioner’s experience, without a controlled comparison that would establish a universal ad count. They sell advertising education and consultations. The observation predates this week’s disclosure and supplies no independent test of the new design.

It does identify a cost the platform’s throughput figure leaves out: someone’s time producing and reviewing those additional ads. A system capable of considering more versions doesn’t tell a small team which versions deserve that time.

For marketers, a sensible test still needs a clear question. A different customer problem or a different way to demonstrate the product provides something to investigate. An arbitrary production quota provides an output count. Our earlier analysis of audience perceptions of AI creative addresses another reason to judge the resulting advertisement carefully.

Keep the working campaign as a comparison when testing a distinct creative idea. Judge the extra production by the account’s results before funding another batch around a platform benchmark.