Meta Now Scores a Billion Ads Per Second With New LLM Ranking Stack

Meta now scores every eligible ad for every impression at a rate of more than one billion ads per second. The company announced the figure on Wednesday as part of three ad ranking upgrades: that scoring capacity, an update to its GEM ads recommendation model, and a new generative recommendations system called GenRecs built on large language models.

The business impact is already visible. Meta said improvements to its ranking system drove 8.3% more clicks and a 15.7% lift in conversions on Facebook in Q2. A separate GEM-related change has delivered a roughly 19% conversion improvement since last year.

"These three innovations work together," said Matt Steiner, VP of ads infrastructure, ranking and AI foundations at Meta. He told AdExchanger that each layer feeds the next: the data the models rely on, the retrieval system that picks which ads move forward, and the ranking model that makes the final call.

How Meta picks an ad in milliseconds

When someone opens Facebook or Instagram, Meta faces tens of millions of candidate ads and must choose one almost instantly. Andromeda, its retrieval system, narrows the pile to a shortlist. A larger, more computationally intensive ranking model then picks the winner.

Historically, Meta's models learned by tying performance to identifiers for each advertiser, campaign, ad creative and product. Those labels keep shifting. "Ads change all the time, ad campaigns change all the time, campaign structures change all the time," Steiner said.

GenRecs swaps those IDs for concepts. LLMs read an ad's text and images and describe it by key attributes — brand, color, fit, style — using shared building blocks called semantic tokens. Because every ad is described in the same terms, models can see what a new ad has in common with existing ones. With a new product, the model "understands that it's similar to the types of products that person buys," Steiner said.

The approach targets the cold-start problem. A new ad has no engagement history, so models normally test it on users and wait. Semantic tokens give the system a head start by drawing on what it knows about similar ads. It does not eliminate the problem. "You still have to learn," Steiner said, but it shortens the learning phase before a new ad hits its stride.

What changes in retrieval and ranking?

GenRecs also powers a new generative retrieval model. Traditional retrieval filters every ad targeted to a person down to likely matches. The new model works like a small LLM built for ad retrieval: it reads the content a person engages with — ads and organic posts — and predicts the most relevant ads. Those candidates join the output of the original retrieval model, which runs in parallel, giving the ranking stage "a new source of high-quality ad recommendations," Steiner said.

Andromeda separately received a hardware and software overhaul through Meta's partnership with Nvidia. Retrieval runs on GPUs, but ad information previously sat elsewhere, leaving processing power "effectively stranded," Steiner said. Meta now stores more ad data on the GPUs themselves and routes each ad to specialized "expert" models.

At the ranking stage, GEM serves as the foundation model for Meta's other ranking models. It operates at LLM scale but does not pick ads itself; smaller models ranking ads on Facebook, Instagram Reels and off-Meta conversions draw on it. The new version is bigger and reads twice as long a user history, improving prediction quality.

Does targeting still matter?

Advertisers need not change any settings to benefit. But Steiner was blunt about the direction of travel. "This can all be done programmatically with extremely high precision by a machine learning model, and much faster than a human's going to do it with a spreadsheet," he said. "The importance of targeting is diminishing over time as ranking models get better at selecting who an ad would be relevant for."

What does this change for practitioners? Mechanical audience work loses value; strategy does not. "The more advertisers are willing to use automation levers, the faster they'll see performance improvements," Steiner said. "They don't need to be spending their time trying to outperform the machine at pulling mechanical levers."