There is a pervasive narrative circulating among certain performance agencies that succeeding on Meta today requires overwhelming the ad account with sheer volume. The argument usually pushes a sense of urgency, suggesting that brands must drastically ramp up asset production right here, right now, without the hard data to back up the investment. We recently published data proving that a smaller batch of highly distinct creative concepts actually yields a lower CPA. To understand why that data looks the way it does, we have to look past the marketing theory and examine the actual engineering infrastructure of Meta's ad delivery system.
The way Meta processes, retrieves, and ranks an ad has been completely rebuilt over the last three years. By understanding the chronological rollout of Lattice, Sequence Learning, Andromeda, and GEM, e-commerce brands can stop guessing what the algorithm wants and start engineering creative workflows that align with how the machine actually operates.
To understand the current ecosystem, we have to look at the foundational shift that occurred in May 2023 with the rollout of Meta Lattice. Before Lattice, the platform relied on fragmented learning models. An ad served on Instagram Stories learned independently from an ad served on the Facebook Feed.
Lattice collapsed these silos into a single Multi-Domain, Multi-Objective architecture. This meant the algorithm could suddenly use a user's behaviour on Reels to predict their likelihood to convert from an ad in their Feed. This unified ranking system immediately shifted baseline platform efficiency.
Meta's engineering teams documented the initial impact of this unified architecture:
With Lattice acting as the unified ranker, Meta then had to solve the problem of retrieval and context. When a user opens the app, the system has a fraction of a second to select the best possible ad from millions of active campaigns.
In late 2024, Meta introduced two critical updates to manage this. First came Sequence Learning in November. Previously, the algorithm looked at static user snapshots. Sequence Learning allowed the system to process deep chronological timelines of user behaviour, mapping exactly where a user sits in their specific purchase journey (assuming your pixel and API infrastructure are passing back high-fidelity sequence data).
A month later in December 2024, Meta deployed Andromeda. Powered by entirely new superchip infrastructure, Andromeda is the retrieval engine that acts as the absolute first filter in the auction. It evaluates tens of millions of ads instantly to find the best initial candidates to pass up the chain to Lattice.
The deployment of Andromeda resulted in immediate system upgrades:
The final piece of this modern architecture arrived in November 2025 with the Generative Ads Recommendation Model, or GEM. If Andromeda is the filter and Lattice is the final judge, GEM is the central intelligence that bridges them. It processes the vast chronological data from Sequence Learning and applies it to the creative candidates retrieved by Andromeda.
Meta has stated that GEM delivers four times the efficiency of their previous generation of models. It is highly sophisticated at understanding the conceptual differences between creatives and matching them to exact user timelines.
When we map this four-part architecture out, the operational reality for scaling brands becomes clinically clear. It also explains why the urgent advice to flood accounts with sheer volume is fundamentally flawed.
If you upload 50 ads simply to hit an arbitrary volume quota without genuine, foundational conceptual diversity, you are not giving the machine more opportunities to win. You are simply forcing Andromeda to waste retrieval compute filtering through redundancy. Because GEM and Sequence Learning are designed to match distinct conceptual messages to specific moments in a user's timeline, feeding them excess volume dilutes their predictive power and forces your own ads into auction overlap.
The machine is built to categorise and deliver distinct, high-quality concepts. Agencies advising brands to urgently overhaul their creative pipelines for pure volume are ignoring this engineering reality. To leverage Andromeda and Lattice effectively, brands must protect their P&L by investing in true conceptual diversity rather than relying on algorithmic noise.
