Meta has rebuilt a major part of the system that selects ads across its platforms. Its updated Andromeda retrieval engine can assess more eligible ads in the time it previously needed to consider a much smaller set, according to an engineering update published on October 28, 2026.
The change affects the matching process beneath Facebook and Instagram campaigns, rather than adding a new control to Ads Manager. It also comes alongside changes to Meta’s GEM recommendation model and a new system called GenRecs. Together, these updates show how much of modern ad performance depends on automated decisions advertisers cannot inspect impression by impression.
Key Takeaways
- Meta rebuilt the Andromeda system that retrieves eligible ads before the final ranking stage.
- Meta’s reported performance improvements are company measurements, not independently verified campaign benchmarks.
- Advertisers can respond by improving conversion signals, creative variety and experimentation rather than chasing hidden model settings.
Table of contents
Jump to each section:
- What Meta changed in ad retrieval
- How the three AI systems work together
- What the performance claims actually show
- What advertisers should do next
What Meta changed in ad retrieval
Ad delivery happens in stages. A retrieval system first narrows a vast set of eligible advertisements into a manageable shortlist. A later ranking stage makes finer predictions about which of those candidates is worth showing. If the retrieval stage discards a potentially strong advertisement too early, the later model never evaluates it.
Meta says the new Andromeda design combines model architecture, GPU processing and ad-index matching into a more tightly integrated pipeline. Its engineers describe a mixture-of-experts approach that considers different user interests and campaign signals while avoiding some of the memory and network overhead of the earlier system.
More than 1 billion ads per second. That is Meta’s stated retrieval throughput for the updated production pipeline, described in its October 28 engineering announcement. Meta characterizes it as roughly 100 times the throughput of the 2024 Andromeda system.
The figure describes the internal system’s evaluation throughput, not how many advertisements an individual sees. It also should not be read as a guarantee that every advertiser will see better performance. Its practical significance is that Meta can consider more eligible creative and targeting combinations within a strict delivery-time limit.
How the three AI systems work together
Andromeda is one part of a broader redesign. Meta’s GEM foundation model learns patterns across advertising surfaces, while GenRecs uses concepts inferred from content and behavior to recommend ads even when historical engagement data for a particular creative is limited.
In an interview with AdExchanger, Matt Steiner, Meta’s vice president of ads infrastructure, ranking and AI foundations, said the components are linked: “These three innovations work together.” In broad terms, one system learns patterns, another retrieves candidate ads, and the ranking stage decides what ultimately competes for an impression.
That matters because generating more creative variants does little if the delivery system cannot evaluate them effectively. Conversely, a more capable retrieval engine creates room to use richer signals and potentially discover relevant combinations that older models missed.
What the performance claims actually show
Meta attributes an improvement in Facebook conversions to Andromeda since 2025, and has also reported gains from its wider ranking upgrades. These figures reflect company measurements across its platform, not a controlled comparison that every advertiser can reproduce by altering a campaign.
2.4% higher Facebook conversions. Meta attributes this improvement since 2025 to Andromeda in its October 28 engineering account. The company has not provided a universal advertiser-level uplift guarantee.8.3% more clicks and 15.7% more conversions. AdExchanger reported these Meta-provided figures for Facebook ranking-system improvements in the second quarter of 2026. They cover broader ranking work and should not be conflated with Andromeda alone.
The distinction is important for campaign planning. Model-level improvements can raise aggregate outcomes while individual accounts vary with objectives, creative quality, offer strength, targeting constraints and conversion feedback. The published claims do not settle attribution at the level of a specific business’s return on ad spend.
What advertisers should do next
There is no new Andromeda switch to turn on. The retrieval upgrade operates inside Meta’s ad infrastructure. For practitioners, the sensible response is to keep the inputs that automated systems learn from reliable and diverse, then measure downstream results rather than treating a vendor headline as a forecast.
First, maintain clear conversion events and sensible campaign objectives. A model making more sophisticated predictions still relies on the quality of the feedback it receives. Second, test creative ideas that genuinely differ in message, format or offer instead of multiplying near-identical variations.
Third, use stable measurement windows and compare results against actual business outcomes. Faster retrieval should not become an excuse to stop experiments before they accumulate useful evidence, nor a reason to assume every reported improvement came from the newest model.
Meta’s announcement is ultimately a reminder that ad delivery is increasingly determined by large, interdependent AI systems. Advertisers can influence those systems with clearer signals and better creative. They cannot reliably reverse-engineer the full ranking process from platform-level statistics.
