MetaMeta

Inside the Three New AI innovations in Ads Ranking Improving your Ad Performance

2026年10月28日·3 分钟阅读

摘要

文章聚焦Meta广告排序系统的最新AI创新。Meta表示,其系统正在重建和演进,以决定哪些广告进入候选、如何更准确预测表现,以及新广告多快能起量;这些能力覆盖更多广告、跨App/版位/目标的统一预测,并利用LLMs理解广告内容、结合用户历史反馈。相关优化自动应用于广告主账户,无需额外手动调参。2026年Q2,Meta广告排序系统改进在Facebook带来广告点击量提升8.3%、转化提升15.7%,属于其更广泛广告投放系统的一部分。

作为排序流程的第一步,Meta Andromeda承担广告召回:把数千万级广告池收敛为候选短名单。2024年版Andromeda强调深度,通过软硬件与机器学习协同设计,并基于MTIA与NVIDIA Grace Hopper Superchip,使召回模型复杂度提升10,000倍。新一代Andromeda与NVIDIA再次协同设计,转向更聪明的广度:不再用单一通用模型处理所有广告,而是将每个广告路由至专门的“专家”模型,把接近排序级的判断力带入召回阶段。生产环境中,系统对每个符合条件的广告打分,并以超过10亿广告/秒的速度输出头部候选,形成召回阶段模型的新吞吐基准。

召回短名单质量的提升最终反映在点击和转化上。自2025年以来,Meta Andromeda在Facebook带来2.4%的转化增长;Q3其复杂度提升10倍,但未增加计算成本或硬件需求,并进一步带来Facebook 1.37%、Instagram Reels 0.24%的转化提升。文章用“私人导购”类比:过去只检查最熟悉、最热门货架,现在能在同样时间内阅读每个货架上的每个商品标签,带回更符合用户偏好的商品组合。对应到广告投放,就是更多广告、创意变体和商品会在同一瞬间被评估,提升合适广告触达合适人群的概率。

对广告主而言,这类系统级优化意味着投放策略可更重视创意多样性与商品/SKU覆盖:更多素材与商品有机会进入候选池,由系统在召回与排序中完成匹配,从而提升用户获取效率与增量转化。由于优化自动生效,广告主无需为每一轮模型迭代手动调参,但仍应保持充足的创意供给和准确的商品信息,以便专家模型和LLM更好理解广告内容并复用历史反馈。整体看,Andromeda的演进体现了广告排序从“单一大模型”走向“专家化召回+规模化吞吐”的趋势,为广告变现效率与投放ROAS的长期提升提供基础设施。

分析师点评

值得关注的是,这次迭代的共同方向是:把“排序级”的判断力前移到召回阶段,并用专家路由(expert routing)替代单一通用大模型。这与 Google、字节等平台近两年在统一排序、多任务建模上的演进路径趋同,说明头部平台的竞争焦点已从“可用的定向手段”转向“模型容量与推理效率”。关键信号在于 10 倍复杂度下算力成本未增——规模化不再靠堆硬件,而靠软硬协同与稀疏化架构,这决定了中小广告主能否以同等成本获得更强系统能力。

对 UA 与变现团队而言,素材与商品目录的覆盖广度正成为系统能否“读到每一层货架”的前提;投放优化中,对单条素材的精雕细琢,其相对价值正在被整体供给的丰富度稀释。此外,把跨版位、跨目标的预测统一起来,也意味着把预算按版位单独分割的做法,其参考意义在下降。

猜你喜欢

TikTokTikTok

TikTok Advertising Blog: Find Product Updates, Insights & Tips

TikTok for Business is promoting tiered first-30-day ad credit incentives for new self-serve SMB advertisers: spend $100 get $100, $500 get $500, or $1,500 get $1,500 plus 1-to-1 expert onboarding support, with a headline entry offer of 50% off first-month spend up to $200 and matching credits up to $6,000. Credits are applied within 7 days of hitting the highest threshold or after the 30-day accumulation window. Eligibility excludes agency, Business Center, and TikTok Shop accounts, and limits one ad account per new TikTok For Business account. For ad ops teams, these terms dictate budget pacing, account structure, and credit-expiry planning.

just paying·2026年9月21日·6 分钟阅读阅读文章 →
TikTokTikTok

TikTok Advertising Blog: Find Product Updates, Insights & Tips

This TikTok for Business landing page promotes a limited-period new-advertiser incentive program offering up to USD 6,000 in matched ad credits. Three self-serve tiers are presented: 50% off first 30-day spend (up to $200), dollar-for-dollar match up to $500, and match up to $6,000. A separate headline offer advertises Spend $100/Get $100, Spend $500/Get $500, and Spend $1,500/Get $1,500 plus 1-to-1 expert onboarding support. Registration requires email or phone sign-up with 2-step verification. Eligibility is restricted to new SMB self-serve accounts, excluding agency, Business Center, and TikTok Shop accounts; credits post within 7 days of hitting the highest threshold or after the 30-day window.

Dan Madden·2026年10月7日·3 分钟阅读阅读文章 →
AppsFlyerAppsFlyer

Why attribution fails: How to fix it

Attribution fails because platforms self-report, data is siloed, identity breaks across surfaces, and traditional models answer the wrong question. The same purchase is often claimed by three or four platforms, inflating attributed revenue 2–4x and distorting ROAS and budget decisions. For ad ops leaders, the fix is not another dashboard but an independent, deduplicated measurement layer with cross-surface customer ID, full coverage of mobile, web, CTV, and retail media, plus incrementality testing. Clean attribution is also the prerequisite for reliable AI bidding and budget automation; broken inputs scale misallocation faster.

Rachel Siegman·2026年10月26日·16 分钟阅读阅读文章 →
TikTokTikTok

TikTok Advertising Blog: Find Product Updates, Insights & Tips

TikTok is offering new advertisers up to $6,000 in ad credits through a tiered spend incentive ($100/$500/$1500) that includes 1-to-1 expert support at the top tier. However, eligibility is restricted to new SMB self-serve accounts, and credits expire. Alongside the offer, TikTok has rolled out several ad tech innovations—Symphony AI creative suite, Streaming Ads, Agentic Hub, Market Scope, and new MMM data—that provide actionable opportunities for testing and scaling performance. Ad ops teams should review eligibility criteria carefully and consider leveraging these tools to maximize ROI during the promotional window.

MCP from TikTok and·2026年8月3日·3 分钟阅读阅读文章 →
MolocoMoloco

Inside Moloco's CARA: the AI that compounds

The open internet presents unique challenges for performance advertising: fragmented identity, closed first-price auctions, and non-stationary supply. Moloco's CARA compound architecture tackles this with six integrated technical domains—Campaign Automation, Supply, Ad Recommendations, Bidding, Creative, and Signals—running on a unified ML infrastructure. Key insights for ad ops: the system continuously learns from every interaction, uses knowledge distillation to serve real-time predictions under 10ms latency, and validates improvements through rigorous live experiments. In 2025, 65 validated model updates reduced CPA by 17% and improved ROAS by 27%. The key takeaway: compound AI architectures that connect prediction, bidding, creative, and data can unlock measurable performance gains beyond walled gardens.

Tal Shaked·2026年8月7日·14 分钟阅读阅读文章 →
MetaMeta

Meet the 2026 Meta Agency Award Winners

Meta's 2026 award winners show ad ops teams winning with automation, first-party signals, and creator/business messaging. Dior used Moment Maker plus Reminder Ads to scale a livestream, then Lead Ads, driving 267% more reminder signups at 24% lower cost. Its holiday rebuild consolidated campaigns into Advantage+ Shopping with CRM data, lifting ROAS 59% and cutting CPA 41%. Feastables used creative testing to find clean-label messaging, yielding a 23.9-point ad recall lift. American Eagle scaled organic creator posts via Partnership Ads; Boost Mobile used Click-to-WhatsApp plus incrementality testing; Winn-Dixie fed store conversions via Conversions API; NAR tested Advantage+ Audience; Compartés scaled with Advantage suite. Let data pick winners, then scale when efficiency holds.

2026年9月21日·9 分钟阅读阅读文章 →
AppsFlyerAppsFlyer

Cross-channel marketing analytics: how it works in 2026

Cross-channel marketing analytics isn't about putting Meta, Google, and TikTok numbers side by side—they often double-count the same customer journey. Fragmented identity is the real culprit; without a first-party Customer User ID, attribution measures platform credit, not customer value. The article explains that deduplicating conversions across mobile, web, and CTV can lift attributed revenue by 30–60% and improve ROAS by 20%. It walks through attribution models, warns against platform-native analytics, and advises using an independent MMP for true cross-channel measurement. Ad ops takeaway: fix identity resolution first, because AI-driven optimization and budget allocation depend on trustworthy, deduplicated data.

Rachel Siegman·2026年9月3日·14 分钟阅读阅读文章 →
MetaMeta

Advertising Week New York 2026: New AI Capabilities to Guide Campaigns and Reach Customers

Meta’s Advertising Week New York updates center on AI-driven campaign automation and faster creative production. Ads Creative Studio is expanding with performance insights. Meta AI business assistant gains creative insights, deeper business memory, and agentic abilities to plan, create, optimize, and schedule tasks, with advertiser permissions and approvals. AI-enabled audience discovery arrives in Detailed Targeting, and Customer Lifecycle Strategy helps prioritize new customers. Generative AI video expands for static assets, catalog product video, and smart recommendations. Direct checkout tests launch in Meta Business Agent on Messenger, Meta AI, and Muse. For ad ops, these tools promise efficiency, but governance and testing remain critical.

2026年10月6日·5 分钟阅读阅读文章 →

更多来自 Meta

Peak Season 2026: You Don’t Beat The Auction. You brief it.

旺季投放的本质不是盯着后台调参,而是给竞价系统写一份清晰的"brief"(简报)——优化目标、出价强度、归因口径三条线决定了系统去找哪一类人;去年 42% 的消费者声称要少花钱,实际支出却增加了 4.2%,说明"谨慎型计划者"与"果断型买家...

2026年10月21日·6 分钟阅读

The Team of One That Doesn't Look Like It

文章以经营12年的女装品牌JaeLuxe Shoetique为例,说明单人创始团队如何借助Muse for Small Business把网站、邮件自动化、Meta广告投放和财务分析等分散工作整合为可执行的“虚拟团队”,从而清理积压并备战Q...

2026年10月20日·4 分钟阅读

Creative diversification at scale: how Nest builds for the way Meta technologies work now

Meta 广告投放系统的算法进化让创意多样性从"素材供给"变成直接的成效杠杆,Nest Commerce 以 AI 创意方案 Nest AI Creative 将单一 brief 批量转化为平台原生变体,把创意生产从成本中心重构为效率驱动引...

2026年10月19日·3 分钟阅读

Performance Spotlight: How Agencies Are Testing for Incrementality

文章指出,广告投放的核心原则之一,是在看效果时必须区分“本来就会发生的转化”与广告真正带来的增量,因此增量测试(incrementality testing)正从一次性项目变为代理商的标准作业流程——通过将广告从随机选定的对照组中撤出、再对...

2026年10月19日·2 分钟阅读

Meta Brand Building Summit: New Ways to Help Marketers Build Unforgettable Brands on Meta

Meta 在 Brand Building Summit 上集中发布并测试多项品牌广告新品,围绕短视频形态与文化热点时刻,帮助广告主与代理商抢占用户注意力并沉淀长期品牌记忆。测试数据显示,Ads Super Burst 带来广告回忆度中位数...

2026年10月15日·6 分钟阅读

The Second Brain on a Sixth-Generation Farm

文章以特拉华州第六代家族农场 Bennett Orchards 为例,说明在无冷链缓冲、全年收入押注 3—4 月几个霜冻夜的极端波动业务中,Muse for Small Business 通过接入分钟级温度记录仪与多源天气数据、自动生成简报...

2026年10月8日·4 分钟阅读