Meta’s Billion-Ads-a-Second Upgrade Shifts the Ad Targeting Balance

Meta now scores every eligible ad for every impression at more than 1 billion ads per second, the company said Wednesday, as it announced three connected changes to ad ranking.

The package includes GenRecs, a new generative recommendations model built on large language models; an updated GEM foundation model that Meta introduced last year; and a hardware and software overhaul of Andromeda with Nvidia.

Meta said its ranking improvements drove 8.3% more clicks and a 15.7% lift in Facebook conversions in Q2. The announcement did not specify the comparison baseline or break out the figures by advertiser, campaign or industry, so the gains remain company-reported results, not a guarantee for every campaign.

Meta’s VP of ads infrastructure, ranking and AI foundations, Matt Steiner, said the work forms a pipeline in which each layer feeds the next: the data models use, retrieval that selects candidates, and ranking that makes the final call. Andromeda narrows the field, while a more compute-intensive model chooses among the survivors.

When someone opens Facebook or Instagram, Meta may have tens of millions of ads it could show that person and must select one almost instantly. Andromeda retrieves a shortlist. A larger ranking model then makes the final choice.

Historically, Meta linked performance to identifiers for advertisers, campaigns, ad creatives and products. Such labels become less useful because ads, campaigns and campaign structures keep changing. GenRecs swaps those identifiers for concepts such as gardening or a particular bike pedal. The system can then connect a new creative to known interests instead of waiting for that exact ad, campaign or product to build history.

LLMs read an ad’s text and images, then describe attributes such as brand, color, fit and style with shared units called semantic tokens. This common vocabulary lets Meta compare a new product with similar products a person buys. It also addresses the cold-start problem, which occurs when a new ad or product has little engagement history. Semantic tokens let the system borrow from known ads and products, shortening the learning phase before a new ad hits its stride. Steiner cautioned: “You still have to learn.” The system still requires testing.

GenRecs also powers a smaller LLM-style generative retrieval model. Traditional retrieval starts with every ad targeted to a person, then filters. The new model reads the ads and organic posts someone engages with, predicts relevant ads and adds them to the original retrieval system’s output. Both systems run in parallel, giving ranking another source of high-quality recommendations.

Meta’s Andromeda upgrade keeps more ad information on the GPUs that perform retrieval. Previously, the company stored that information elsewhere, leaving chips idle while they waited for data. The Nvidia partnership also routes each ad toward specialized “expert” models, which Meta says makes retrieval smarter.

At the ranking stage, GEM trains smaller models serving Facebook, Instagram Reels and off-Meta conversions. It operates at the scale of a large language model but does not select ads itself. Meta’s updated version is larger and reads twice as long a user history. Steiner said that broader context improves interest predictions, while Meta attributes roughly 19% conversion improvement since last year to the approach.

What does this change for advertisers?

Advertisers do not need to change settings to benefit, according to Steiner. Their audience job changes. Models can select people based on semantic similarity and broader engagement with ads and organic posts, rather than requiring past click behavior or manually built segments. Steiner said machine learning can do that work with more precision and speed than spreadsheet segmentation.

“The importance of targeting is diminishing over time as ranking models get better at selecting who an ad would be relevant for,” Steiner said.

For budgets, Meta announced no bid, CPM or cost-per-conversion result, so the 15.7% lift cannot be treated as cheaper media. For reach, semantic similarity may help a new ad find relevant buyers sooner, but Meta did not report an audience-size increase. The evidence supports a relevance and conversion claim, not proof of incremental reach.

Strategically, advertisers should spend less time trying to beat the machine at segmenting and more time on strategy. Steiner expects performance gains to come faster as advertisers use automation rather than trying to outperform models at mechanical levers. The next test is whether Meta’s gains persist beyond its own Q2 results and whether advertisers can retain useful audience control without relying on a history of clicks.