[{"data":1,"prerenderedAt":373},["ShallowReactive",2],{"blog-facebook-lookalike-audience-zh-CN":3},{"id":4,"title":5,"excerpt":6,"content":7,"coverImage":314,"meta":322,"site":326,"status":347,"slug":348,"author":349,"category":359,"publishDate":19,"featured":43,"updatedAt":368,"createdAt":369,"contentHtml":370,"previewUrl":371,"localeSlugs":372},561,"Facebook 相似受众：源名单质量、体量门槛，以及为什么建不出来（2026）","Facebook 相似受众到底是从什么长出来的、决定它能不能建起来的按国家体量门槛、创建失败的真实原因，以及怎么读它的表现才不会陷入循环论证。",{"root":8},{"children":9,"direction":19,"format":15,"indent":13,"type":313,"version":18},[10,21,25,31,48,52,58,64,68,72,84,88,92,98,102,106,112,122,128,143,157,161,165,177,181,206,210,214,220,226,232,245,249,292,296,300],{"children":11,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[12],{"detail":13,"format":13,"mode":14,"style":15,"text":16,"type":17,"version":18},0,"normal","","Facebook 相似受众（lookalike audience）做的事是：你把一份已经有价值的名单交出去——买过的人、订阅的人、高价值用户——让 Meta 去找更多\"像他们\"的账号。机制一句话说得清，却极容易用错。大多数被归咎于\"算法\"的结果，其实追溯回去都落在三个更早的决定上：源名单里放了谁、名单有多大、百分比开到了多宽。","text",1,null,"paragraph",{"children":22,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[23],{"detail":13,"format":13,"mode":14,"style":15,"text":24,"type":17,"version":18},"这篇讲清楚：模型到底在匹配什么、决定受众能不能建起来的体量门槛、创建失败的真实原因，以及怎么读相似受众的表现才不会自欺。",{"children":26,"direction":19,"format":15,"indent":13,"type":29,"version":18,"tag":30},[27],{"detail":13,"format":13,"mode":14,"style":15,"text":28,"type":17,"version":18},"相似受众到底是从什么长出来的","heading","h2",{"children":32,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[33,35,46],{"detail":13,"format":13,"mode":14,"style":15,"text":34,"type":17,"version":18},"输入是一份源受众——由客户名单、像素或转化 API 事件、App 行为、或主页与视频互动建出来的",{"children":36,"direction":19,"format":15,"indent":13,"type":39,"version":40,"fields":41,"id":45},[37],{"detail":13,"format":13,"mode":14,"style":15,"text":38,"type":17,"version":18},"自定义受众","link",3,{"linkType":42,"newTab":43,"url":44},"custom",false,"https://deepclick.com/resources/blog/facebook-custom-audience-guide/","6a952a3be37d2600c8168d8f",{"detail":13,"format":13,"mode":14,"style":15,"text":47,"type":17,"version":18},"。Meta 给源里的人做画像，再在你选定的国家里，按\"与这个画像的相似度\"给其他所有人打分。",{"children":49,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[50],{"detail":13,"format":13,"mode":14,"style":15,"text":51,"type":17,"version":18},"由此推出两件决定了大半结果的事：",{"children":53,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[54,56],{"detail":13,"format":18,"mode":14,"style":15,"text":55,"type":17,"version":18},"模型会连你源名单里的偏差一起继承，包括你自己都没意识到的那部分。",{"detail":13,"format":13,"mode":14,"style":15,"text":57,"type":17,"version":18}," 如果源是\"最近 180 天买过的人\"，而其中 70% 来自某次折扣活动、吸引来的是薅羊毛的人，那相似受众会忠实地帮你找来更多薅羊毛的人。模型没有\"你想要哪种客户\"这个概念，它只看得见你递过去的那份名单。",{"children":59,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[60,62],{"detail":13,"format":18,"mode":14,"style":15,"text":61,"type":17,"version":18},"相似度是对着源算的，不是对着你的生意算的。",{"detail":13,"format":13,"mode":14,"style":15,"text":63,"type":17,"version":18}," 用邮件订阅者建的相似受众，找到的是\"像订阅者的人\"——也就是容易注册各种东西的人。这和\"容易付钱的人\"是两个人群。选源就是在选结果，而这个决定发生在花第一分钱之前。",{"children":65,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[66],{"detail":13,"format":13,"mode":14,"style":15,"text":67,"type":17,"version":18},"落到操作上：源要建在离钱最近的行为上。用购买事件建的相似受众，和用全站访客建的相似受众，不是同一件事的两个版本。",{"children":69,"direction":19,"format":15,"indent":13,"type":29,"version":18,"tag":30},[70],{"detail":13,"format":13,"mode":14,"style":15,"text":71,"type":17,"version":18},"体量门槛，以及贴着门槛会发生什么",{"children":73,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[74,76,78,80,82],{"detail":13,"format":13,"mode":14,"style":15,"text":75,"type":17,"version":18},"Meta 的硬性要求是源受众里",{"detail":13,"format":18,"mode":14,"style":15,"text":77,"type":17,"version":18},"单一国家至少 100 人",{"detail":13,"format":13,"mode":14,"style":15,"text":79,"type":17,"version":18},"才允许建相似受众。那是系统受理请求的下限，",{"detail":13,"format":18,"mode":14,"style":15,"text":81,"type":17,"version":18},"不是结果开始可靠的那条线",{"detail":13,"format":13,"mode":14,"style":15,"text":83,"type":17,"version":18},"。",{"children":85,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[86],{"detail":13,"format":13,"mode":14,"style":15,"text":87,"type":17,"version":18},"真正可用的区间要高得多。源只有几百人时，模型能拿来画像的信号极少，建出来的受众很不稳定：过一个月成员略有变化再重建一次，人群就会明显不同。源在几千人量级时，多次重建之间稳定得多——这一点很重要，因为随着客户基数变化，你迟早要重建它。",{"children":89,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[90],{"detail":13,"format":13,"mode":14,"style":15,"text":91,"type":17,"version":18},"但也不是越大越好，\"直接用我最大的那份名单\"这个直觉恰恰错在这里。五万条良莠不齐的联系人，画出来的像是一团模糊的平均值；三千个真正的高价值客户，画出来的像才具体。当体量和纯度必须二选一时，通常纯度赢——前提是你离 100 这条底线留足了余量。",{"children":93,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[94,96],{"detail":13,"format":18,"mode":14,"style":15,"text":95,"type":17,"version":18},"百分比是与选源相互独立的另一个决定。",{"detail":13,"format":13,"mode":14,"style":15,"text":97,"type":17,"version":18}," 1%–10% 那个滑块设定的是\"按相似度排序后，取该国人口的多少比例\"。在大国，1% 已经是几百万人。往宽了拉并不会\"提升\"受众质量，只是把相似度更低的人依次追加到排序尾部。默认从 1% 起步，只有在投放确实受量级限制时才放宽。",{"children":99,"direction":19,"format":15,"indent":13,"type":29,"version":18,"tag":30},[100],{"detail":13,"format":13,"mode":14,"style":15,"text":101,"type":17,"version":18},"相似受众建不出来的原因",{"children":103,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[104],{"detail":13,"format":13,"mode":14,"style":15,"text":105,"type":17,"version":18},"创建报错很常见，原因大多平淡无奇：",{"children":107,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[108,110],{"detail":13,"format":18,"mode":14,"style":15,"text":109,"type":17,"version":18},"源还没填充完。",{"detail":13,"format":13,"mode":14,"style":15,"text":111,"type":17,"version":18}," 客户名单类的自定义受众要等匹配跑完，事件类的要等合格事件攒够。显示\"正在填充中\"或体量低于门槛的源，此时还不能用，解法是等。",{"children":113,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[114,116,118,120],{"detail":13,"format":18,"mode":14,"style":15,"text":115,"type":17,"version":18},"所选国家里不足 100 人。",{"detail":13,"format":13,"mode":14,"style":15,"text":117,"type":17,"version":18}," 一份五千人的名单分散在四十个国家，你选的那个国家里可能不到 100 人。门槛是",{"detail":13,"format":18,"mode":14,"style":15,"text":119,"type":17,"version":18},"按国家算",{"detail":13,"format":13,"mode":14,"style":15,"text":121,"type":17,"version":18},"的，而界面上显示的体量是总数——这两个数字对不上的频率，比多数人预期的高得多。",{"children":123,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[124,126],{"detail":13,"format":18,"mode":14,"style":15,"text":125,"type":17,"version":18},"源本身就是相似受众。",{"detail":13,"format":13,"mode":14,"style":15,"text":127,"type":17,"version":18}," 不能用相似受众再建相似受众，得回到最初那份自定义受众。",{"children":129,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[130,132,134,141],{"detail":13,"format":18,"mode":14,"style":15,"text":131,"type":17,"version":18},"源已失去资格。",{"detail":13,"format":13,"mode":14,"style":15,"text":133,"type":17,"version":18}," 某些事件类型建的受众、或底层数据源后来不可用了，都会让它不再合格。如果一个一直能用的源突然不能用了，先查底层的像素、",{"children":135,"direction":19,"format":15,"indent":13,"type":39,"version":40,"fields":138,"id":140},[136],{"detail":13,"format":13,"mode":14,"style":15,"text":137,"type":17,"version":18},"转化 API 接入",{"linkType":42,"newTab":43,"url":139},"https://deepclick.com/resources/blog/meta-conversions-api-setup/","6a952a3be37d2600c8168d90",{"detail":13,"format":13,"mode":14,"style":15,"text":142,"type":17,"version":18},"或 App 事件流是否还在正常上报——故障点往往在受众的上游。",{"children":144,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[145,147,149,156],{"detail":13,"format":18,"mode":14,"style":15,"text":146,"type":17,"version":18},"账户本身有限制。",{"detail":13,"format":13,"mode":14,"style":15,"text":148,"type":17,"version":18}," 广告账户被限制或在审核中时，受众创建是会被一并挡住的功能之一。那是另一个问题穿着\"受众\"的外衣；账户处在这种状态时，先",{"children":150,"direction":19,"format":15,"indent":13,"type":39,"version":40,"fields":153,"id":155},[151],{"detail":13,"format":13,"mode":14,"style":15,"text":152,"type":17,"version":18},"把账户问题解决掉",{"linkType":42,"newTab":43,"url":154},"https://deepclick.com/resources/blog/facebook-ad-account-disabled-appeal/","6a952a3be37d2600c8168d91",{"detail":13,"format":13,"mode":14,"style":15,"text":83,"type":17,"version":18},{"children":158,"direction":19,"format":15,"indent":13,"type":29,"version":18,"tag":30},[159],{"detail":13,"format":13,"mode":14,"style":15,"text":160,"type":17,"version":18},"自动化改变的是相似受众的位置，不是它的作用",{"children":162,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[163],{"detail":13,"format":13,"mode":14,"style":15,"text":164,"type":17,"version":18},"Meta 的自动化受众产品越来越把你填的定向当作\"建议\"而非\"边界\"——当系统预测别处效果更好时，它会跑到你指定的受众之外去探索。于是反复有人问：相似受众是不是要没了。",{"children":166,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[167,169,171,173,175],{"detail":13,"format":13,"mode":14,"style":15,"text":168,"type":17,"version":18},"比较有用的读法是：相似受众正在从",{"detail":13,"format":18,"mode":14,"style":15,"text":170,"type":17,"version":18},"围栏",{"detail":13,"format":13,"mode":14,"style":15,"text":172,"type":17,"version":18},"变成",{"detail":13,"format":18,"mode":14,"style":15,"text":174,"type":17,"version":18},"输入信号",{"detail":13,"format":13,"mode":14,"style":15,"text":176,"type":17,"version":18},"。在老的购买流程里，选了 1% 相似受众，就意味着广告只投给这些人；在新的自动化流程里，同一份受众是一个很强的起点提示，系统可以往外走。",{"children":178,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[179],{"detail":13,"format":13,"mode":14,"style":15,"text":180,"type":17,"version":18},"实际影响：",{"children":182,"direction":19,"format":15,"indent":13,"type":203,"version":18,"listType":204,"start":18,"tag":205},[183,190,197],{"children":184,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":18},[185,187],{"detail":13,"format":18,"mode":14,"style":15,"text":186,"type":17,"version":18},"相似受众之间互相 A/B 测，会变得没那么干净。",{"detail":13,"format":13,"mode":14,"style":15,"text":188,"type":17,"version":18}," 如果系统在两个受众之外都做了探索，那么实际触达人群的差异，会小于你在后台选择上的差异。","listitem",{"children":191,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":196},[192,194],{"detail":13,"format":18,"mode":14,"style":15,"text":193,"type":17,"version":18},"源的质量更重要了，不是更不重要。",{"detail":13,"format":13,"mode":14,"style":15,"text":195,"type":17,"version":18}," 提示也还是提示。用真实购买者建的源，仍然比用全部流量建的源，把系统指向更好的方向。",2,{"children":198,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":40},[199,201],{"detail":13,"format":18,"mode":14,"style":15,"text":200,"type":17,"version":18},"\"泛投 vs 相似受众\"不再是干净的二选一。",{"detail":13,"format":13,"mode":14,"style":15,"text":202,"type":17,"version":18}," 老实的做法是：在可比的花费下，加上相似受众这个信号，每个结果的成本有没有变化——当成一次测试去跑，别预设任何一边的答案。","list","bullet","ul",{"children":207,"direction":19,"format":15,"indent":13,"type":29,"version":18,"tag":30},[208],{"detail":13,"format":13,"mode":14,"style":15,"text":209,"type":17,"version":18},"怎么读表现才不会自欺",{"children":211,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[212],{"detail":13,"format":13,"mode":14,"style":15,"text":213,"type":17,"version":18},"三个习惯能挡掉大部分错误结论：",{"children":215,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[216,218],{"detail":13,"format":18,"mode":14,"style":15,"text":217,"type":17,"version":18},"看每个结果的成本，别看受众层的比率。",{"detail":13,"format":13,"mode":14,"style":15,"text":219,"type":17,"version":18}," 相似受众的点击率通常会好过泛投，因为它本来就是在挑\"像已经跟你互动过的人\"的人——这近乎循环论证。真正要看的是它有没有用更低的成本换来结果。",{"children":221,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[222,224],{"detail":13,"format":18,"mode":14,"style":15,"text":223,"type":17,"version":18},"给源留出刷新的时间。",{"detail":13,"format":13,"mode":14,"style":15,"text":225,"type":17,"version":18}," 事件类源建的相似受众会随源更新。用\"最近 30 天购买者\"建的受众是个移动靶，拿它第 1 周和第 6 周的表现对比，比的是两个不同的人群。",{"children":227,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":18,"textStyle":15},[228,230],{"detail":13,"format":18,"mode":14,"style":15,"text":229,"type":17,"version":18},"下\"A 比 B 好\"的结论前先查重叠。",{"detail":13,"format":13,"mode":14,"style":15,"text":231,"type":17,"version":18}," 两个来源相近的相似受众可能高度重叠。真重叠了，同时跑就不是测试，而是同一个受众挂了两个名字、把预算劈成两半。",{"children":233,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[234,236,243],{"detail":13,"format":13,"mode":14,"style":15,"text":235,"type":17,"version":18},"归因也该保持同样的怀疑：如果相似受众的结果好得或差得离谱，先确认参与对比的转化数据是完整的，再去重做受众。服务端事件是否去重、",{"children":237,"direction":19,"format":15,"indent":13,"type":39,"version":40,"fields":240,"id":242},[238],{"detail":13,"format":13,"mode":14,"style":15,"text":239,"type":17,"version":18},"链接跟踪是否一致",{"linkType":42,"newTab":43,"url":241},"https://deepclick.com/resources/blog/link-tracking-for-ad-campaigns-2026/","6a952a3be37d2600c8168d92",{"detail":13,"format":13,"mode":14,"style":15,"text":244,"type":17,"version":18},"，决定了你拿来比较的那些数字有没有意义。",{"children":246,"direction":19,"format":15,"indent":13,"type":29,"version":18,"tag":30},[247],{"detail":13,"format":13,"mode":14,"style":15,"text":248,"type":17,"version":18},"一个能避开常见错误的搭建顺序",{"children":250,"direction":19,"format":15,"indent":13,"type":203,"version":18,"listType":290,"start":18,"tag":291},[251,257,263,269,276,283],{"children":252,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":18},[253,255],{"detail":13,"format":18,"mode":14,"style":15,"text":254,"type":17,"version":18},"挑离收入最近的那个行为",{"detail":13,"format":13,"mode":14,"style":15,"text":256,"type":17,"version":18},"建自定义受众。购买优于加购，加购优于浏览。",{"children":258,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":196},[259,261],{"detail":13,"format":18,"mode":14,"style":15,"text":260,"type":17,"version":18},"看按国家拆的人数",{"detail":13,"format":13,"mode":14,"style":15,"text":262,"type":17,"version":18},"，不是总数。确认目标国家过 100 且有明显余量。",{"children":264,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":40},[265,267],{"detail":13,"format":18,"mode":14,"style":15,"text":266,"type":17,"version":18},"先按 1%、单个国家建",{"detail":13,"format":13,"mode":14,"style":15,"text":268,"type":17,"version":18},"。一个相似受众只对一个国家，读数比多国混在一起干净得多。",{"children":270,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":275},[271,273],{"detail":13,"format":18,"mode":14,"style":15,"text":272,"type":17,"version":18},"等它建完再挂预算。",{"detail":13,"format":13,"mode":14,"style":15,"text":274,"type":17,"version":18}," 还在填充的受众会跑不出量，而你会把这误读成定向失败。",4,{"children":277,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":282},[278,280],{"detail":13,"format":18,"mode":14,"style":15,"text":279,"type":17,"version":18},"一次只改一个变量。",{"detail":13,"format":13,"mode":14,"style":15,"text":281,"type":17,"version":18}," 换源或换百分比，别同时换。",5,{"children":284,"direction":19,"format":15,"indent":13,"type":189,"version":18,"value":289},[285,287],{"detail":13,"format":18,"mode":14,"style":15,"text":286,"type":17,"version":18},"按周期重建。",{"detail":13,"format":13,"mode":14,"style":15,"text":288,"type":17,"version":18}," 客户基数会变，用去年买家建的相似受众，画的是去年的生意。",6,"number","ol",{"children":293,"direction":19,"format":15,"indent":13,"type":29,"version":18,"tag":30},[294],{"detail":13,"format":13,"mode":14,"style":15,"text":295,"type":17,"version":18},"这件事落在哪一层",{"children":297,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[298],{"detail":13,"format":13,"mode":14,"style":15,"text":299,"type":17,"version":18},"相似受众是一个分发决定，但它依赖一个度量决定。源名单的质量上限，等于填充它的那些事件的质量，也就是说像素与服务端事件的质量位于你建的每一个受众的上游。事件不全或重复，源就是错的，而基于一个错误的源精心配置出来的相似受众，只是更快地触达了错的人。",{"children":301,"direction":19,"format":15,"indent":13,"type":20,"version":18,"textFormat":13,"textStyle":15},[302,304,311],{"detail":13,"format":13,"mode":14,"style":15,"text":303,"type":17,"version":18},"DeepClick 做的是下面那一层——",{"children":305,"direction":19,"format":15,"indent":13,"type":39,"version":40,"fields":308,"id":310},[306],{"detail":13,"format":13,"mode":14,"style":15,"text":307,"type":17,"version":18},"面向 Meta 与 TikTok 投放的链接与投放基础设施",{"linkType":42,"newTab":43,"url":309},"https://deepclick.com/solutions/meta-tiktok-advertisers","6a952a3be37d2600c8168d93",{"detail":13,"format":13,"mode":14,"style":15,"text":312,"type":17,"version":18},"，跟踪一致性与落地页交付，决定了你用来建受众的那些事件，究竟有没有如实反映真实发生过的事。","root",{"id":315,"alt":316,"updatedAt":317,"createdAt":317,"url":318,"thumbnailURL":19,"filename":319,"mimeType":320,"filesize":321,"width":19,"height":19},1042,"Concentric rings of dots expanding outward from a dense seed cluster, illustrating how a Facebook lookalike audience ranks people by similarity to a source audience","2026-08-31T07:15:48.336Z","https://cms-r2.deepclick.com/cover-739-d842bf808a95.jpg","cover-739-d842bf808a95.jpg","application/octet-stream",151770,{"title":323,"description":324,"image":325},"Facebook 相似受众怎么建：体量门槛与失败排查（2026）","相似受众会连源名单里的偏差一起继承。本文讲清按国家 100 人的门槛、创建失败的五个真实原因、自动化改变了什么，以及怎么判断表现。",{"id":315,"alt":316,"updatedAt":317,"createdAt":317,"url":318,"thumbnailURL":19,"filename":319,"mimeType":320,"filesize":321,"width":19,"height":19},{"id":18,"key":327,"name":328,"prodHost":329,"testHost":330,"blogPath":331,"docPath":332,"zhPrefix":333,"deployHookTest":334,"deployHookProd":335,"brandAuthor":336,"brandKit":340,"enabled":344,"updatedAt":345,"createdAt":346},"deepclick","DeepClick","https://deepclick.com","https://www-test-deepclick.qiliangjia.one","/resources/blog/{slug}","/docs/{slug}","/zh-CN","https://api.cloudflare.com/client/v4/pages/webhooks/deploy_hooks/60a9adef-153b-4c89-8d07-7118e91e9522","https://api.cloudflare.com/client/v4/pages/webhooks/deploy_hooks/05323321-f694-4ce8-a5af-173c507b8bae",{"id":196,"name":328,"avatar":337,"updatedAt":338,"createdAt":339},25,"2026-04-22T08:09:35.299Z","2026-04-22T06:42:49.116Z",{"logoLight":19,"logoDark":19,"primaryColor":341,"accentColor":342,"tagline":343,"wechatName":19,"xhsHandle":19,"wechatQr":19},"#1a73e8","#ff6600","一次点击，多重价值",true,"2026-08-25T06:40:57.663Z","2026-07-14T06:26:38.962Z","published","facebook-lookalike-audience",{"id":275,"name":350,"avatar":351,"updatedAt":358,"createdAt":358},"Ethan Cole",{"id":352,"alt":353,"updatedAt":354,"createdAt":354,"url":355,"thumbnailURL":19,"filename":356,"mimeType":320,"filesize":357,"width":19,"height":19},922,"Ethan Cole — growth & ad-tech editor","2026-07-27T07:17:05.290Z","https://cms-r2.deepclick.com/gpt_1785136218181_0-d1b8139c1927.png","gpt_1785136218181_0-d1b8139c1927.png",1965798,"2026-07-27T07:17:07.920Z",{"id":360,"site":361,"titleZh":363,"titleEn":364,"slug":365,"order":282,"updatedAt":366,"createdAt":367},7,{"id":18,"key":327,"name":328,"prodHost":329,"testHost":330,"blogPath":331,"docPath":332,"zhPrefix":333,"deployHookTest":334,"deployHookProd":335,"brandAuthor":196,"brandKit":362,"enabled":344,"updatedAt":345,"createdAt":346},{"logoLight":19,"logoDark":19,"primaryColor":341,"accentColor":342,"tagline":343,"wechatName":19,"xhsHandle":19,"wechatQr":19},"技术导航","Tech Guides","tech-guides","2026-04-27T08:37:10.576Z","2026-04-23T02:59:13.436Z","2026-08-31T07:16:15.121Z","2026-08-31T07:16:04.323Z","\u003Cdiv class=\"payload-richtext\">\u003Cp>Facebook 相似受众（lookalike audience）做的事是：你把一份已经有价值的名单交出去——买过的人、订阅的人、高价值用户——让 Meta 去找更多"像他们"的账号。机制一句话说得清，却极容易用错。大多数被归咎于"算法"的结果，其实追溯回去都落在三个更早的决定上：源名单里放了谁、名单有多大、百分比开到了多宽。\u003C/p>\u003Cp>这篇讲清楚：模型到底在匹配什么、决定受众能不能建起来的体量门槛、创建失败的真实原因，以及怎么读相似受众的表现才不会自欺。\u003C/p>\u003Ch2>相似受众到底是从什么长出来的\u003C/h2>\u003Cp>输入是一份源受众——由客户名单、像素或转化 API 事件、App 行为、或主页与视频互动建出来的\u003Ca href=\"https://deepclick.com/resources/blog/facebook-custom-audience-guide/\">自定义受众\u003C/a>。Meta 给源里的人做画像，再在你选定的国家里，按"与这个画像的相似度"给其他所有人打分。\u003C/p>\u003Cp>由此推出两件决定了大半结果的事：\u003C/p>\u003Cp>\u003Cstrong>模型会连你源名单里的偏差一起继承，包括你自己都没意识到的那部分。\u003C/strong> 如果源是"最近 180 天买过的人"，而其中 70% 来自某次折扣活动、吸引来的是薅羊毛的人，那相似受众会忠实地帮你找来更多薅羊毛的人。模型没有"你想要哪种客户"这个概念，它只看得见你递过去的那份名单。\u003C/p>\u003Cp>\u003Cstrong>相似度是对着源算的，不是对着你的生意算的。\u003C/strong> 用邮件订阅者建的相似受众，找到的是"像订阅者的人"——也就是容易注册各种东西的人。这和"容易付钱的人"是两个人群。选源就是在选结果，而这个决定发生在花第一分钱之前。\u003C/p>\u003Cp>落到操作上：源要建在离钱最近的行为上。用购买事件建的相似受众，和用全站访客建的相似受众，不是同一件事的两个版本。\u003C/p>\u003Ch2>体量门槛，以及贴着门槛会发生什么\u003C/h2>\u003Cp>Meta 的硬性要求是源受众里\u003Cstrong>单一国家至少 100 人\u003C/strong>才允许建相似受众。那是系统受理请求的下限，\u003Cstrong>不是结果开始可靠的那条线\u003C/strong>。\u003C/p>\u003Cp>真正可用的区间要高得多。源只有几百人时，模型能拿来画像的信号极少，建出来的受众很不稳定：过一个月成员略有变化再重建一次，人群就会明显不同。源在几千人量级时，多次重建之间稳定得多——这一点很重要，因为随着客户基数变化，你迟早要重建它。\u003C/p>\u003Cp>但也不是越大越好，"直接用我最大的那份名单"这个直觉恰恰错在这里。五万条良莠不齐的联系人，画出来的像是一团模糊的平均值；三千个真正的高价值客户，画出来的像才具体。当体量和纯度必须二选一时，通常纯度赢——前提是你离 100 这条底线留足了余量。\u003C/p>\u003Cp>\u003Cstrong>百分比是与选源相互独立的另一个决定。\u003C/strong> 1%–10% 那个滑块设定的是"按相似度排序后，取该国人口的多少比例"。在大国，1% 已经是几百万人。往宽了拉并不会"提升"受众质量，只是把相似度更低的人依次追加到排序尾部。默认从 1% 起步，只有在投放确实受量级限制时才放宽。\u003C/p>\u003Ch2>相似受众建不出来的原因\u003C/h2>\u003Cp>创建报错很常见，原因大多平淡无奇：\u003C/p>\u003Cp>\u003Cstrong>源还没填充完。\u003C/strong> 客户名单类的自定义受众要等匹配跑完，事件类的要等合格事件攒够。显示"正在填充中"或体量低于门槛的源，此时还不能用，解法是等。\u003C/p>\u003Cp>\u003Cstrong>所选国家里不足 100 人。\u003C/strong> 一份五千人的名单分散在四十个国家，你选的那个国家里可能不到 100 人。门槛是\u003Cstrong>按国家算\u003C/strong>的，而界面上显示的体量是总数——这两个数字对不上的频率，比多数人预期的高得多。\u003C/p>\u003Cp>\u003Cstrong>源本身就是相似受众。\u003C/strong> 不能用相似受众再建相似受众，得回到最初那份自定义受众。\u003C/p>\u003Cp>\u003Cstrong>源已失去资格。\u003C/strong> 某些事件类型建的受众、或底层数据源后来不可用了，都会让它不再合格。如果一个一直能用的源突然不能用了，先查底层的像素、\u003Ca href=\"https://deepclick.com/resources/blog/meta-conversions-api-setup/\">转化 API 接入\u003C/a>或 App 事件流是否还在正常上报——故障点往往在受众的上游。\u003C/p>\u003Cp>\u003Cstrong>账户本身有限制。\u003C/strong> 广告账户被限制或在审核中时，受众创建是会被一并挡住的功能之一。那是另一个问题穿着"受众"的外衣；账户处在这种状态时，先\u003Ca href=\"https://deepclick.com/resources/blog/facebook-ad-account-disabled-appeal/\">把账户问题解决掉\u003C/a>。\u003C/p>\u003Ch2>自动化改变的是相似受众的位置，不是它的作用\u003C/h2>\u003Cp>Meta 的自动化受众产品越来越把你填的定向当作"建议"而非"边界"——当系统预测别处效果更好时，它会跑到你指定的受众之外去探索。于是反复有人问：相似受众是不是要没了。\u003C/p>\u003Cp>比较有用的读法是：相似受众正在从\u003Cstrong>围栏\u003C/strong>变成\u003Cstrong>输入信号\u003C/strong>。在老的购买流程里，选了 1% 相似受众，就意味着广告只投给这些人；在新的自动化流程里，同一份受众是一个很强的起点提示，系统可以往外走。\u003C/p>\u003Cp>实际影响：\u003C/p>\u003Cul class=\"list-bullet\">\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"1\"\n        >\u003Cstrong>相似受众之间互相 A/B 测，会变得没那么干净。\u003C/strong> 如果系统在两个受众之外都做了探索，那么实际触达人群的差异，会小于你在后台选择上的差异。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"2\"\n        >\u003Cstrong>源的质量更重要了，不是更不重要。\u003C/strong> 提示也还是提示。用真实购买者建的源，仍然比用全部流量建的源，把系统指向更好的方向。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"3\"\n        >\u003Cstrong>"泛投 vs 相似受众"不再是干净的二选一。\u003C/strong> 老实的做法是：在可比的花费下，加上相似受众这个信号，每个结果的成本有没有变化——当成一次测试去跑，别预设任何一边的答案。\u003C/li>\u003C/ul>\u003Ch2>怎么读表现才不会自欺\u003C/h2>\u003Cp>三个习惯能挡掉大部分错误结论：\u003C/p>\u003Cp>\u003Cstrong>看每个结果的成本，别看受众层的比率。\u003C/strong> 相似受众的点击率通常会好过泛投，因为它本来就是在挑"像已经跟你互动过的人"的人——这近乎循环论证。真正要看的是它有没有用更低的成本换来结果。\u003C/p>\u003Cp>\u003Cstrong>给源留出刷新的时间。\u003C/strong> 事件类源建的相似受众会随源更新。用"最近 30 天购买者"建的受众是个移动靶，拿它第 1 周和第 6 周的表现对比，比的是两个不同的人群。\u003C/p>\u003Cp>\u003Cstrong>下"A 比 B 好"的结论前先查重叠。\u003C/strong> 两个来源相近的相似受众可能高度重叠。真重叠了，同时跑就不是测试，而是同一个受众挂了两个名字、把预算劈成两半。\u003C/p>\u003Cp>归因也该保持同样的怀疑：如果相似受众的结果好得或差得离谱，先确认参与对比的转化数据是完整的，再去重做受众。服务端事件是否去重、\u003Ca href=\"https://deepclick.com/resources/blog/link-tracking-for-ad-campaigns-2026/\">链接跟踪是否一致\u003C/a>，决定了你拿来比较的那些数字有没有意义。\u003C/p>\u003Ch2>一个能避开常见错误的搭建顺序\u003C/h2>\u003Col class=\"list-number\">\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"1\"\n        >\u003Cstrong>挑离收入最近的那个行为\u003C/strong>建自定义受众。购买优于加购，加购优于浏览。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"2\"\n        >\u003Cstrong>看按国家拆的人数\u003C/strong>，不是总数。确认目标国家过 100 且有明显余量。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"3\"\n        >\u003Cstrong>先按 1%、单个国家建\u003C/strong>。一个相似受众只对一个国家，读数比多国混在一起干净得多。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"4\"\n        >\u003Cstrong>等它建完再挂预算。\u003C/strong> 还在填充的受众会跑不出量，而你会把这误读成定向失败。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"5\"\n        >\u003Cstrong>一次只改一个变量。\u003C/strong> 换源或换百分比，别同时换。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"6\"\n        >\u003Cstrong>按周期重建。\u003C/strong> 客户基数会变，用去年买家建的相似受众，画的是去年的生意。\u003C/li>\u003C/ol>\u003Ch2>这件事落在哪一层\u003C/h2>\u003Cp>相似受众是一个分发决定，但它依赖一个度量决定。源名单的质量上限，等于填充它的那些事件的质量，也就是说像素与服务端事件的质量位于你建的每一个受众的上游。事件不全或重复，源就是错的，而基于一个错误的源精心配置出来的相似受众，只是更快地触达了错的人。\u003C/p>\u003Cp>DeepClick 做的是下面那一层——\u003Ca href=\"https://deepclick.com/solutions/meta-tiktok-advertisers\">面向 Meta 与 TikTok 投放的链接与投放基础设施\u003C/a>，跟踪一致性与落地页交付，决定了你用来建受众的那些事件，究竟有没有如实反映真实发生过的事。\u003C/p>\u003C/div>","https://deepclick.com/zh-CN/resources/blog/facebook-lookalike-audience",{"zh-CN":348,"en":348},1788167894864]