Why Meta performance has been so volatile: GEM, Andromeda and creative-level reporting
Creative-level reporting for dynamic ads
Meta now offers creative-level reporting for dynamic ads in Ads Manager (announced Oct 2026): commonthreadco.com/blogs/coachs-corner/meta-just-unlocked-creative-level-reporting-in-ads-manager-what-ecommerce-brands-must-check-right-now
We read this as a strong signal that Meta wants advertisers to use dynamic creative more often, and it likely ties into the GEM and Andromeda changes below.
GEM moved to a single model
Meta’s generative ads recommendation model (GEM) went from a separate model per platform (2025) to one model for Facebook, Instagram and Threads. It was announced in July 2026; it’s unclear whether the rollout started then or later: commonthreadco.com/blogs/coachs-corner/metas-generative-ads-model-is-delivering-18-9-more-conversions-what-every-ecommerce-brand-must-do-before-q4
Between publishing an ad and that ad showing, Meta runs three steps:
- Retrieval (Andromeda): picks a few hundred ads out of the millions someone could see.
- Ranking (GEM): ranks those few hundred and sends a handful of candidates to the auction.
- Auction: decides what shows based on available inventory and bids.
Collapsing GEM from one model per platform into one for all three likely caused a lot of volatility.
Andromeda overhaul
On Oct 28, 2026, Meta Engineering announced a full overhaul of Andromeda: engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design
- The first version (2024) handled tens of millions of ads per second.
- The previous model filtered 2.4 billion ads per second; the new architecture filters 10 billion.
- The overhaul was driven by Advantage+ creative and the millions of new ad variations it produces.
What it means
GEM’s move to a single model, most likely rolled out after July based on what we see in the data, explains part of the volatility. The Andromeda overhaul, landing around now, is likely another part of why Meta performance has been rough. Put both together with an election year and most of the volatility is explained.