WayToGlobal Insights · AI Buying

WayToGlobal Insights · AI 买量

Meta Rebuilt Ad Ranking on LLMs: What GenRecs, Andromeda and GEM Mean for DTC Advertisers

Meta 把广告排序换成了大模型:GenRecs、Andromeda、GEM 三件套,DTC 投放怎么变

2026-10-31 · 8 min read · checked against primary sources

2026-10-31 · 约 8 分钟读完 · 逐条核对原始来源

All insights全部文章

30-second read

  • Meta now scores 1B+ ads per second through three layers: GenRecs reads new ads as semantic tokens (shorter cold start), Andromeda retrieves on-GPU, GEM reads twice the user history (~19% conversion lift since last year).
  • Targeting is depreciating: Meta's own VP says ranking models increasingly pick the audience. Value moves to creative supply, event plumbing and strategy.
  • Three moves now: replace audience-pack time with a creative matrix, audit event return completeness, and track learning-phase length as your acceptance metric.
The three-layer ranking stack, from AdExchanger's Oct 28 report
The three-layer ranking stack, from AdExchanger's Oct 28 report

Meta swapped out the engine under its ads system. AdExchanger's October 28 report lays out the three layers: GenRecs, a new LLM-based ranking system; a hardware-and-software overhaul of the retrieval layer called Andromeda; and a bigger, longer-memory upgrade to GEM, the foundation model behind ranking. Advertisers need to change nothing to benefit — but the flip side of "nothing" is that the manual targeting craft you built in Ads Manager is being taken over by the system.

What each layer got

GenRecs: understanding ads by semantics. The old ranking tracked performance history tied to IDs — campaigns, creatives, products. New ads have no history, so cold start was slow. GenRecs instead uses an LLM to describe ads through shared semantic tokens — attributes like brand, color, fit and style — so the model understands a new ad by its concepts. It doesn't eliminate cold start, but it shortens the learning phase.

Andromeda: a hardware-level refresh of retrieval. Retrieval narrows tens of millions of candidate ads down to the few hundred worth bidding. Through Meta's Nvidia partnership, ad data now lives more on the GPUs themselves, cutting idle compute, and ads route to specialized "expert" models for smarter first-pass filtering.

GEM: bigger base model, twice the user history. GEM doesn't pick ads itself; it trains the surface-level ranking models. The new version is larger and reads twice as long a user history, improving interest predictions. Meta credits the GEM approach with roughly 19% conversion improvement accumulated since last year.

The scale: per impression, Meta filters tens of millions of candidates down to one, scoring over 1 billion ads every second. Q2 results on Facebook: 8.3% more clicks and a 15.7% conversion lift.

Targeting depreciates; the operator's value moves

Matt Steiner, Meta's VP of ads infrastructure, ranking and AI foundations, put it plainly: "The importance of targeting is diminishing over time as ranking models get better at selecting who an ad would be relevant for." The core agency craft — audience packs, interest stacks, lookalike tuning — is becoming an internal function of the system.

This is not Meta-specific; it is the shared direction of every large-model ranking platform. For DTC operators, value moves to three places. Creative supply: the system reads ads by semantics, so creative variety and concept coverage matter more than ever — the more concepts you feed it, the wider it can explore. Data plumbing: the ranking model eats what you send it, and the completeness of purchase/ROAS events now decides learning-phase speed and survival. Strategy: budget allocation, channel mix, margin constraints — the model doesn't manage your P&L.

Three changes to make now

First, stop stacking audience packs. Move that time into a creative matrix: multiple expressions of each selling point, multiple formats, multiple scenarios — give the semantic tokens material. Second, audit event return completeness; the new system's demand for clean inputs is only higher, and gaps cost more. Third, change your acceptance metric to learning-phase length: whether new creatives cold-start faster than last quarter is the most direct read on whether you're capturing this upgrade. For what a second engine looks like under the same logic, see our AppLovin AXON guide — a fully automated ranking system built on the same premise.


Sources: AdExchanger, "Meta Has A New Ad Ranking System Built On LLMs, Plus Upgrades To Andromeda And GEM" (Oct 28, 2026; Matt Steiner interview and Q2 figures); Meta Q2 earnings figures (click and conversion lifts). Platform-mechanics explainer — no proprietary campaign data.

Series — buying in the AI era: Google AI Mode ads entry · the pre-BFCM incrementality test.

Want your Meta account tuned for the new ranking stack?

Leave your category and current Meta performance — we'll come back with a concrete read on your creative matrix and event plumbing for the new ranking stack. Managed Meta buying is our main line.

30 秒看完

  • Meta 每秒对超 10 亿条广告打分,三层分工:GenRecs 按语义 token 认识新广告(冷启动更短)、Andromeda 在 GPU 上做检索初筛、GEM 读取翻倍的用户历史(累计约 19% 转化提升)。
  • 定向在贬值——Meta 自己的副总裁说排序模型越来越擅长替你选人。投手的价值挪向素材供给、数据回传和策略层。
  • 马上能改的三件事:把堆人群包的时间换成素材矩阵、审计事件回传完整性、把「学习期长短」当验收指标。
三层排序流水线(据 AdExchanger 2026-10-28 报道)
三层排序流水线(据 AdExchanger 2026-10-28 报道)

Meta 的广告系统完成了一次底层更换。10 月 28 日 AdExchanger 的报道把三层架构讲透了:新的 GenRecs 排序系统、检索层 Andromeda 的软硬件翻新、排序基础模型 GEM 的升级——三层各管一段,层层喂给下一层。广告主不需要改任何设置就能受益,但「受益」的另一面是:你过去在后台练就的手工定向功夫,正在被系统接管。

三层各换了什么

GenRecs:用「语义」认识广告。 老排序追踪的是 ID(campaign、素材、商品)的历史表现,新广告没有历史,冷启动慢。GenRecs 改用大模型把广告拆成语义 token——品牌、颜色、版型、风格这类属性——让模型按「概念」理解一条新广告。它不能消灭冷启动,但把学习期缩短了。

Andromeda:检索层的硬翻新。 检索是从几千万候选广告里捞出「值得参与竞价的几百条」。Meta 与 Nvidia 合作把广告数据更多驻留在 GPU 上,减少算力空转,并把广告路由给专门的「专家」模型做更聪明的初筛。

GEM:排序基础模型变大、看用户历史变长。 GEM 本身不挑广告,它是训练各场景排序模型的地基。新版本模型更大、读取的用户历史长度翻倍,兴趣预测更准。Meta 称 GEM 这套路线自去年以来累计带来约 19% 的转化提升。

规模口径:每次展示,Meta 要从数千万候选里筛出一条,每秒对超过 10 亿条广告打分。Q2 实测(Facebook 侧):点击 +8.3%、转化 +15.7%。

定向贬值,投手的价值挪到哪

Meta 广告基础设施副总裁 Matt Steiner 把话说得很直白:「随着排序模型越来越擅长挑出广告对谁相关,定向的重要性在随时间下降。」过去 agency 的核心手艺——人群包、兴趣分层、lookalike 调优——正在变成系统内部的事。

这不是 Meta 独家的趋势,而是所有「大模型排序」平台的共同方向。对 DTC 投手,价值挪向三处:素材供给(系统按语义理解广告,素材的多样性和「概念覆盖」比以往更重要——你给它的概念越全,它能探索的人群越宽);数据回传(排序模型吃得越准,学习期越短,购买/ROAS 事件的完整性直接决定生死);策略层(预算分配、渠道组合、毛利约束——模型不管你的 P&L)。

实操上要改的三件事

第一,别再堆人群包了。把原来花在精细定向上的时间,转移到素材矩阵:同一卖点的多种表达、多版式、多场景,让语义 token 有足够材料。第二,检查事件回传的完整性——新系统对「干净输入」的要求只高不低,回传缺口的代价被放大。第三,验收口径换成「学习期长短」:新素材冷启动是否比上个季度快,是判断你有没有吃到这波升级的最直观指标。至于 Meta 之外的第二个增长渠道怎么配,我们写过一篇 AppLovin 的进场指南——AXON 同样是全自动排序引擎,逻辑互为印证。


出处:AdExchanger《Meta Has A New Ad Ranking System Built On LLMs, Plus Upgrades To Andromeda And GEM》(2026-10-28,Matt Steiner 访谈与 Q2 数据);Meta Q2 财报口径(点击与转化提升)。本篇为平台机制解读,不含本司投放数据。

系列——AI 时代买量:Google AI Mode 进场 · 黑五前增量测试。

想让账户跟上新排序系统?

留下品类和当前 Meta 表现,我们回来聊素材矩阵和数据回传怎么配——Meta 托管投放是我们的主业。