How Meta's ad system chooses which ads reach ranking at all

Meta's engineering posts describe a step that comes before ranking. Retrieval decides which ads are considered, and an ad that is not retrieved never reaches the later stages.

Two systems appear in Meta's descriptions. Andromeda handles retrieval. GEM, the Generative Ads Recommendation Model, supplies predictions to the models that rank and serve ads. Meta published updates on both in 2026.

Retrieval comes first, ranking second

Meta describes its ad system as multi-stage. Retrieval is the first stage.

Later stages use larger ranking models. Meta says they predict value to people and to the advertiser, and they choose the final ads shown. In its October 28, 2026 post, Meta describes ranking as the more computationally expensive stage, scoring a small set of candidates in finer detail.

The stages differ in volume. Meta put retrieval at three orders of magnitude more ads than the stages after it. A heavy model cannot run on every ad, so the first stage has to be fast and the later stages can afford to be thorough.

What Andromeda changed in 2024 and 2026

The December 2024 post introduced Andromeda as the retrieval engine behind Advantage+ automation. Meta reported a 6% improvement in recall and an 8% improvement in ad quality on selected segments. It ran on NVIDIA Grace Hopper Superchips and Meta's own MTIA hardware.

The October 2026 post describes the latest iteration of Andromeda. A mixture-of-experts design replaced the single user tower with an ensemble of user experts. Index matching, scoring and top-K selection moved onto one GPU server, and the scoring steps were fused into a single kernel.

Meta says it and NVIDIA launched a joint collaboration in early 2026. The reported results are about 300 times fewer inference-time FLOPs than the 2024 retrieval model and network traffic cut by about 100 times. The full pipeline now scores every eligible ad at over 1 billion ads per second on a single GPU server, which Meta calls a 100-fold lift on the 2024 system.

Meta gives a reason for the scale. Generative AI creative tools have helped advertisers produce hundreds of millions of new ads, which enlarges the candidate pool. Meta says the larger pool is searched within the same latency budget.

For results, the post states that Andromeda has delivered a 2.4% increase in conversions on Facebook since 2025.

What GEM predicts

GEM sits behind the models that serve ads. Meta's November 2025 post called it the largest foundation model for recommendation systems in the industry. It was trained on ad content and user engagement data from both ads and organic interactions.

The October 13, 2026 post says GEM is now a single central model replacing separate per-surface models for placements such as Facebook, Instagram and Reels. It jointly predicts click-through rate across all surfaces and conversion rate across the main advertiser objectives. Meta lists off-site conversions and app installs among them.

Meta says the model learns from data across all of its ads platforms, and that 3.6 billion people use at least one of its apps each day. Training runs on several thousand of the latest-generation GPUs.

How GEM reaches the models that serve ads

Meta says running GEM inside a live ad request is not affordable. Meta uses it as a teacher for hundreds of vertical models that serve ads in real time. The 2025 post describes transfer through knowledge distillation, representation learning and parameter sharing.

The 2026 post adds two components. LoopFM compresses GEM's internal representations and combines them with sequential historical memory. Meta says it doubles the intelligence transfer together with the knowledge distillation used in the previous version, with no additional compute cost or latency.

Solaris handles freshness. It precomputes GEM embeddings for user-ad pairs that upcoming requests are likely to need, so downstream models get near-real-time signals. When an exact embedding is missing, it falls back to aggregated or similar-user embeddings.

A heavy model cannot run on every ad, so the first stage has to be fast and the later stages can afford to be thorough.

The reported gains, and who reported them

Every performance figure in these posts is Meta's own.

The November 2025 post reported that GEM's launch lifted ad conversions by 5% on Instagram and 3% on Facebook Feed in the second quarter. The October 2026 post reports an 18.9% conversion improvement and a 12% click-through improvement across Meta's ad surfaces and objectives.

The same post gives further conversion uplift figures for targeted users over the past year: 5.8% on Facebook, 4.1% on Instagram and 1.6% off-site.

The 2024 Andromeda post also cited a 22% increase in return on ad spend for advertisers who had not previously used Advantage+ creative and turned on its AI-driven targeting features.

What the descriptions mean for account setup

The posts do not give advertisers a settings checklist. They do say what the systems predict and what they score.

The conversion objective is one. GEM predicts conversion rate for the major optimization goals, from off-site conversions to mobile app installs.

Creative volume is another. Meta ties the growth of the candidate pool to generative creative tools, and says retrieval now scores every eligible ad.

Measurement is the third. Meta's gain figures are platform-reported, and the Advantage+ and GEM numbers cover different populations and time frames.