Meta Rebuilds Ad Ranking on LLMs, Upgrades Andromeda and GEM
The Development
Meta Platforms announced three upgrades to its ad ranking stack on Wednesday: a new ranking system built on large language models, plus improvements to Andromeda and GEM, the retrieval and modelling layers already powering its delivery. Per the announcement, the new system scores every eligible ad for every impression at more than a billion ads per second. Previously, ranking pipelines relied on staged filtering, narrowing a large candidate pool to a shortlist before final scoring. Scoring the full eligible set removes that narrowing step and changes how ads compete for each impression. The announcement positions the three components as a single system, with Andromeda handling retrieval, GEM supplying learned signals, and the LLM-based ranker making final selection decisions.
Our Take
This is Meta finishing the move from advertiser-directed delivery to model-directed delivery. Once every eligible ad is scored for every impression, manual levers such as narrow audiences, segmented ad sets, and bid micromanagement lose most of their explanatory power. What remains is the quality and diversity of your creative and the integrity of the signals you feed back. Advertisers who still structure accounts around audience hypotheses are optimising a layer the system now treats as input, not instruction. The strategic read: Meta is building a closed-loop machine where the marginal advertiser contribution is creative volume and first-party signal. Expect results to diverge sharply between accounts that supply both and accounts that supply neither.
What Changed
Meta can now evaluate the entire eligible ad inventory against each impression in real time using LLM-based ranking, rather than pre-filtering candidates. Selection becomes a comprehensive, per-impression judgment, so an ad's chance of winning depends on its own relevance and quality, not on surviving an upstream cut.
Marketing Impact
Media buying teams should expect account structure to matter less and creative throughput to matter more. Creative operations become the primary performance lever, since LLM-based ranking differentiates ads on substance. Measurement and CRM teams own the second lever: conversion signal quality now directly governs ranking outcomes.
Competitive Implication
In-house creative teams running high-volume generative pipelines and brands with clean server-side conversion data gain structural advantage, because the ranker rewards variety and signal depth. Agencies selling audience-engineering and manual optimisation, and brands with thin creative libraries, become vulnerable as their levers lose influence.
Strategic Outlook
Expect Meta to push further consolidation into automated campaign types through Q1 2027, with less advertiser control and more reliance on creative inputs. Google and Amazon will answer with comparable full-inventory ranking claims. Performance gaps between creative-rich and creative-poor accounts will widen measurably within two quarters.
The Exploit
Action Item
Heads of paid social should audit conversion signal quality and creative variety on Meta before Q1 2027 budgets lock, then shift test spend from audience segmentation to producing at least ten distinct creative concepts per campaign.