Meta Has A New Ad Ranking System Built On LLMs, Plus Upgrades To Andromeda And GEM

Meta now scores every eligible ad for every impression at more than a billion ads per second, one of three ad ranking upgrades Meta announced on Wednesday.

The others are an update to GEM, the ads recommendation model it introduced last year, and a new generative recommendations model ( GenRecs) built on LLMs that can predict which ads people will want to see, even with limited engagement data.

According to Meta, improvements to its ranking system drove 8.3% more clicks and a 15.7% lift in conversions on Facebook in Q2.

“These three innovations work together,” said Matt Steiner, VP of ads infrastructure, ranking and AI foundations at Meta.

Each layer feeds the next, he told AdExchanger: the data Meta’s models rely on, the retrieval system that picks which ads move forward and the ranking model that makes the final call.

How ads get chosen

When someone opens Facebook or Instagram, Meta has tens of millions of ads it could show and it has to pick one almost instantly.

First, Andromeda, its retrieval system, narrows the pile to a shortlist of ads that person might like. Then a bigger, more computationally intensive ranking model picks the winner.

Meta’s models learn by tracking what works. Historically, that meant tying performance to identifiers or labels for each advertiser, campaign, ad creative and product. A model might learn, for example, that one particular creative does well with people who engage with posts about certain topics, hobbies or products.

The challenge is that those labels keep changing.

“Ads change all the time, ad campaigns change all the time, campaign structures change all the time,” Steiner said. So it’s more useful for the models to track how performance relates to the underlying concept – say gardening or a particular type of bike pedals, anything really – than how it relates to a specific ad, campaign or creative.

GenRecs swaps those IDs for concepts.

It works like so: LLMs read an ad’s text and images and describe it by its key attributes, such as a product’s brand, color, fit and style, using a set of shared building blocks known as semantic tokens. Because every ad is described in the same terms, the models can see what a new ad has in common with other ads.

With a new product, the model “understands that it’s similar to the types of products that person buys,” Steiner said.

This is meant to ease what’s known as the cold-start problem. A new ad or product has no engagement history, so the models have little to go on and need to test it on people and see what happens. Semantic tokens give the system a head start, because it can draw on what it already knows about similar ads and products.

It doesn’t eliminate the problem, Steiner cautioned – “You still have to learn,” he said – but it shortens the learning phase, which is the stretch of time before a new ad hits its stride.

Retrieval gets an upgrade

GenRecs also powers a new generative retrieval model. Retrieval has traditionally worked by taking every ad targeted to a person and filtering down to the ones they might like.

The new model works differently. Steiner described it as like an LLM, but a small one built for ad retrieval. It looks at the content someone engages with – ads and organic posts – and predicts the ads most likely to be relevant to that person.

Those ads are then added to the ones from Meta’s original retrieval model, which sits in parallel with the new model. This gives the ranking stage “a new source of high-quality ad recommendations,” Steiner said.

Separately, Andromeda got a hardware and software overhaul through Meta’s partnership with Nvidia.

Retrieval runs on GPUs, which are chips that do the heavy computing. Before, information about each ad was stored elsewhere, and getting it to the GPU took time. That left a lot of processing power sitting idle, or “effectively stranded,” Steiner said.

Meta now keeps more of that ad information on the GPUs themselves, so the chips spend less time waiting for data and more time working. Meta also changed the retrieval model so each ad is routed to specialized “expert” models, which makes retrieval smarter.

At the ranking stage, GEM is the foundation model Meta uses to train its other ranking models.

It operates at the scale of an LLM, Steiner said, but it doesn’t pick ads itself. Instead, the smaller models that rank ads on each surface, from Facebook to Instagram Reels to off-Meta conversions, draw on it.

The new version is bigger and reads twice as long a history for every user, which Steiner said helps produce better predictions of what someone will be interested in. According to Meta, this approach has delivered a roughly 19% conversion improvement since last year.

The end of targeting?

Advertisers don’t have to change any settings to benefit from the upgrades, Steiner said.

But the ability to reach people without having to use their past click behavior as a signal is itself a signal that targeting isn’t the performance lever it used to be.

Targeting has traditionally meant building audience segments. Buyers look at performance data, split an audience into groups and test which ones respond best. But Meta’s ranking models now do that work automatically.

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

Which raises the obvious question hanging over agencies and advertisers: If the models handle tasks like audience selection, what’s left for them to do?

Lots of stuff, Steiner said. Machine learning models are good at many things, but they can’t set strategy.

“The more advertisers are willing to use automation levers, the faster they’ll see performance improvements,” he said. “They don’t need to be spending their time trying to outperform the machine at pulling mechanical levers.”