[{"data":1,"prerenderedAt":333},["ShallowReactive",2],{"blog-facebook-ads-split-testing-zh-CN":3,"blog-ad-banners-zh-CN":304},{"id":4,"title":5,"excerpt":6,"content":7,"coverImage":245,"meta":253,"site":257,"status":278,"slug":279,"author":280,"category":290,"publishDate":19,"featured":39,"updatedAt":299,"createdAt":300,"contentHtml":301,"previewUrl":302,"localeSlugs":303},587,"Facebook 广告 A/B 拆分测试：它到底在测什么，结果什么时候只是噪声","拆分测试按人切分受众，这是复制一个广告组做不到的一件事。哪些变量适合测、学习期为什么会吃掉小预算测试，以及让多数广告测试读不出结论的统计功效问题。",{"root":8},{"children":9,"direction":19,"format":15,"indent":13,"type":244,"version":18},[10,22,27,44,52,56,60,64,68,72,85,89,93,97,105,109,134,138,142,146,154,158,162,166,170,194,216,220,226,232,238],{"children":11,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[12],{"detail":13,"format":13,"mode":14,"style":15,"text":16,"type":17,"version":18},0,"normal","","拆分测试真正解决的问题，复制一个广告组解决不了","text",1,null,"heading","h2",{"children":23,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[24],{"detail":13,"format":13,"mode":14,"style":15,"text":25,"type":17,"version":18},"想知道素材 A 是不是比 B 好，最顺手的做法是复制一个广告组、改一处。这个测试在开跑之前就已经被污染了。","paragraph",{"children":28,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[29,31,42],{"detail":13,"format":13,"mode":14,"style":15,"text":30,"type":17,"version":18},"两个瞄准同一批人的广告组会进同一场竞价。Meta 通常是从符合条件的广告组里挑一个去竞价，而不是让它们同时出价——于是你看到的差异里，混进了投放系统\"选谁上场\"的判断，而不只是受众对素材的偏好。更麻烦的是同一个人整周下来两个版本都能看到，所谓\"赢的那版\"可能只是拿到第二次曝光的那版。这和 ",{"children":32,"direction":19,"format":15,"indent":13,"type":35,"version":36,"fields":37,"id":41},[33],{"detail":13,"format":13,"mode":14,"style":15,"text":34,"type":17,"version":18},"Facebook 广告受众重叠","link",3,{"linkType":38,"newTab":39,"url":40},"custom",false,"https://deepclick.com/zh-CN/resources/blog/facebook-ads-audience-overlap/","6a9bd09653c92d00c8572bee",{"detail":13,"format":13,"mode":14,"style":15,"text":43,"type":17,"version":18}," 说的是同一个结构性问题，只不过那篇从成本角度看，这篇从测量角度看。",{"children":45,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[46,48,50],{"detail":13,"format":13,"mode":14,"style":15,"text":47,"type":17,"version":18},"A/B 测试工具专门修的就是\"分配\"这一环：它按",{"detail":13,"format":18,"mode":14,"style":15,"text":49,"type":17,"version":18},"人",{"detail":13,"format":13,"mode":14,"style":15,"text":51,"type":17,"version":18},"切分受众，而不是按曝光切分。每个用户在整个测试期间只被分到一个版本，各分区之间不互相竞价。这就是这个工具存在的全部理由——别的事你手工也能做，随机且互不重叠的分组你做不到。",{"children":53,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[54],{"detail":13,"format":13,"mode":14,"style":15,"text":55,"type":17,"version":18},"什么变量适合测",{"children":57,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[58],{"detail":13,"format":13,"mode":14,"style":15,"text":59,"type":17,"version":18},"工具的设计前提是一次只改一个变量，可测维度是素材、受众、版位、投放优化目标。这个限制不是多余的：如果素材和受众一起改了，结果差异只能告诉你\"这两个组合不一样\"，不能告诉你是哪一半造成的，这个结论也没法迁移到别的广告系列。",{"children":61,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[62],{"detail":13,"format":13,"mode":14,"style":15,"text":63,"type":17,"version":18},"实际操作里要比工具的强制更严一档。把钩子、形式、时长、行动号召全换了的\"新素材\"，是四处改动套了一件衣服。这种测试仍然可以跑——有时你就是想知道新概念作为一个整体打不打得过旧的——但复盘时要如实写清你学到的是什么，因为\"整包赢了\"并不告诉你下一步该改哪里。",{"children":65,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[66],{"detail":13,"format":13,"mode":14,"style":15,"text":67,"type":17,"version":18},"学习期会悄悄吃掉小预算测试",{"children":69,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[70],{"detail":13,"format":13,"mode":14,"style":15,"text":71,"type":17,"version":18},"每个版本都是一个新广告组，每个新广告组都要重新进学习期。在投放稳定之前，单次成效成本既偏高又偏抖，跟它最终会落到的稳态不是一回事。",{"children":73,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[74,76,83],{"detail":13,"format":13,"mode":14,"style":15,"text":75,"type":17,"version":18},"常被引用的门槛是每个广告组每周约 50 次优化事件；具体数字请以产品内当前口径为准，Meta 调整过这个指引。真正要紧的是它逼出来的那道算术：测试预算被摊到各版本上，每个版本都得自己攒够事件。同样一笔钱，作为一个广告组本可以轻松走出学习期，一分为二之后可能两半都卡住——而两个都没跑出来、都还在学习期的广告组互相比较，测的是学习期，不是你的素材。",{"children":77,"direction":19,"format":15,"indent":13,"type":35,"version":36,"fields":80,"id":82},[78],{"detail":13,"format":13,"mode":14,"style":15,"text":79,"type":17,"version":18},"Meta 广告学习期",{"linkType":38,"newTab":39,"url":81},"https://deepclick.com/zh-CN/resources/blog/meta-ads-learning-phase/","6a9bd09653c92d00c8572bef",{"detail":13,"format":13,"mode":14,"style":15,"text":84,"type":17,"version":18}," 讲了什么会重置它、要多久。",{"children":86,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[87],{"detail":13,"format":13,"mode":14,"style":15,"text":88,"type":17,"version":18},"如果诚实的答案是\"这笔钱供不起两边都走出学习期\"，那这个测试就还不该跑。换到量更大的广告系列上跑，或者拉长窗口，或者改测一个攒得更快的指标。",{"children":90,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[91],{"detail":13,"format":13,"mode":14,"style":15,"text":92,"type":17,"version":18},"多数拆分测试真正错的地方：统计功效",{"children":94,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[95],{"detail":13,"format":13,"mode":14,"style":15,"text":96,"type":17,"version":18},"广告测试大多栽在这里，而且是无声地栽——测试给出一个数字，数字看着像答案，没人回头检查它是否有可能是答案。",{"children":98,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[99,101,103],{"detail":13,"format":13,"mode":14,"style":15,"text":100,"type":17,"version":18},"转化是稀有事件。如果每一边只出个位数的转化，那么与你观测结果相容的\"真实转化率\"范围大得吓人。12 比 9 不是 33% 的提升，那是从两个你还分不开的分布里各抽了一次。不太舒服但有用的经验法则是：要检出一个温和的提升——也就是多数素材改动实际能带来的那种量级——需要每边",{"detail":13,"format":18,"mode":14,"style":15,"text":102,"type":17,"version":18},"几百个",{"detail":13,"format":13,"mode":14,"style":15,"text":104,"type":17,"version":18},"转化，不是几十个；而要检出巨大的差异所需样本要少得多。这正是\"两个真正不同的概念\"之间的测试很快能出结论、而\"按钮换个颜色\"永远出不了结论的原因。",{"children":106,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[107],{"detail":13,"format":13,"mode":14,"style":15,"text":108,"type":17,"version":18},"开跑前先定三件事：",{"children":110,"direction":19,"format":15,"indent":13,"type":131,"version":18,"listType":132,"start":18,"tag":133},[111,118,125],{"children":112,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":18},[113,115],{"detail":13,"format":18,"mode":14,"style":15,"text":114,"type":17,"version":18},"值得据此行动的最小差异",{"detail":13,"format":13,"mode":14,"style":15,"text":116,"type":17,"version":18},"。如果 5% 的差别根本不会改变你的投放决策，就别设计一个只能检出 5% 的测试。","listitem",{"children":119,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":124},[120,122],{"detail":13,"format":18,"mode":14,"style":15,"text":121,"type":17,"version":18},"这个差异所需的转化量你出不出得起",{"detail":13,"format":13,"mode":14,"style":15,"text":123,"type":17,"version":18},"。出不起，就要么接受代理指标，要么干脆不跑。",2,{"children":126,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":36},[127,129],{"detail":13,"format":18,"mode":14,"style":15,"text":128,"type":17,"version":18},"停止日期",{"detail":13,"format":13,"mode":14,"style":15,"text":130,"type":17,"version":18},"。在第一次曝光发生之前就写下来。","list","bullet","ul",{"children":135,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[136],{"detail":13,"format":13,"mode":14,"style":15,"text":137,"type":17,"version":18},"Meta 会随结果给一个置信度。置信度低就照它的字面意思理解——工具在告诉你它分不开这两个版本——而不是把它当成对暂时领先那一版的微弱背书。",{"children":139,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[140],{"detail":13,"format":13,"mode":14,"style":15,"text":141,"type":17,"version":18},"为什么不能\"看着不错就收\"",{"children":143,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[144],{"detail":13,"format":13,"mode":14,"style":15,"text":145,"type":17,"version":18},"盯着在跑的测试、一看到某一版拉开就收手，是制造假赢家最有效的办法。测试早期领先者会频繁易主，纯属噪声。只要你允许自己在\"差距看着够说服力\"的任意时刻停手，那么即使拿两个一模一样的素材去测，你迟早也能找到这样一个时刻。",{"children":147,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[148,150,152],{"detail":13,"format":13,"mode":14,"style":15,"text":149,"type":17,"version":18},"防线是流程上的，不是统计上的：提前把结束日期钉死，只在结束时读一次结果。如果确实需要中途盯盘，那就盯",{"detail":13,"format":18,"mode":14,"style":15,"text":151,"type":17,"version":18},"故障",{"detail":13,"format":13,"mode":14,"style":15,"text":153,"type":17,"version":18},"——某一版没花出去、广告被拒、落地页坏了——而不是盯谁赢。",{"children":155,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[156],{"detail":13,"format":13,"mode":14,"style":15,"text":157,"type":17,"version":18},"\"没结论\"本身就是结论",{"children":159,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[160],{"detail":13,"format":13,"mode":14,"style":15,"text":161,"type":17,"version":18},"团队容易把\"无显著差异\"当成白跑一趟。它是个发现：这两个版本之间的差异，小于你出得起钱去测量的差异。这说明该停止在这个维度上继续迭代，转向杠杆更大的地方——报价、受众定义、落地页——而不是再去做第四个标题变体。",{"children":163,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[164],{"detail":13,"format":13,"mode":14,"style":15,"text":165,"type":17,"version":18},"它同时还告诉你一件关于这两个版本的事：随便发哪个都行。既然表现分不出高下，就选制作成本更低、后续更好维护的那个。",{"children":167,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[168],{"detail":13,"format":13,"mode":14,"style":15,"text":169,"type":17,"version":18},"开跑前的检查清单",{"children":171,"direction":19,"format":15,"indent":13,"type":131,"version":18,"listType":132,"start":18,"tag":133},[172,176,180,184,189],{"children":173,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":18},[174],{"detail":13,"format":13,"mode":14,"style":15,"text":175,"type":17,"version":18},"只改了一个变量，而且你能用一句话把它说清楚。",{"children":177,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":124},[178],{"detail":13,"format":13,"mode":14,"style":15,"text":179,"type":17,"version":18},"预算够两边各自独立走出学习期。",{"children":181,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":36},[182],{"detail":13,"format":13,"mode":14,"style":15,"text":183,"type":17,"version":18},"最小差异量与停止日期，两样都事先写下来。",{"children":185,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":188},[186],{"detail":13,"format":13,"mode":14,"style":15,"text":187,"type":17,"version":18},"测试窗口内没有别的广告系列在打同一批受众，免得分组从外部被污染。",4,{"children":190,"direction":19,"format":15,"indent":13,"type":117,"version":18,"value":193},[191],{"detail":13,"format":13,"mode":14,"style":15,"text":192,"type":17,"version":18},"事先想好\"没结论\"时怎么办——趁你还冷静的时候决定。",5,{"children":195,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":13,"textStyle":15},[196,198,205,207,214],{"detail":13,"format":13,"mode":14,"style":15,"text":197,"type":17,"version":18},"上游结构理顺了，能在测试开始前就消掉不少噪声：",{"children":199,"direction":19,"format":15,"indent":13,"type":35,"version":36,"fields":202,"id":204},[200],{"detail":13,"format":13,"mode":14,"style":15,"text":201,"type":17,"version":18},"Facebook 广告账户结构",{"linkType":38,"newTab":39,"url":203},"https://deepclick.com/zh-CN/resources/blog/facebook-ads-account-structure/","6a9bd09653c92d00c8572bf0",{"detail":13,"format":13,"mode":14,"style":15,"text":206,"type":17,"version":18}," 讲一个账户到底需要多少广告系列和广告组，",{"children":208,"direction":19,"format":15,"indent":13,"type":35,"version":36,"fields":211,"id":213},[209],{"detail":13,"format":13,"mode":14,"style":15,"text":210,"type":17,"version":18},"Facebook 广告出价策略",{"linkType":38,"newTab":39,"url":212},"https://deepclick.com/zh-CN/resources/blog/facebook-ads-bidding-strategy/","6a9bd09653c92d00c8572bf1",{"detail":13,"format":13,"mode":14,"style":15,"text":215,"type":17,"version":18}," 讲那些最容易把测试搅浑的投放设置。",{"children":217,"direction":19,"format":15,"indent":13,"type":20,"version":18,"tag":21},[218],{"detail":13,"format":13,"mode":14,"style":15,"text":219,"type":17,"version":18},"常见问题",{"children":221,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":18,"textStyle":15},[222,224],{"detail":13,"format":18,"mode":14,"style":15,"text":223,"type":17,"version":18},"拆分测试该跑多久？",{"detail":13,"format":13,"mode":14,"style":15,"text":225,"type":17,"version":18}," 长到两边都跑出学习期、并攒够你设定的最小差异所需的转化量，且至少跑满一整周，好让\"星期几\"的影响平摊到两边。两个条件不一致时，取更长的那个。",{"children":227,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":18,"textStyle":15},[228,230],{"detail":13,"format":18,"mode":14,"style":15,"text":229,"type":17,"version":18},"能不能拿两个不同的广告目标做拆分测试？",{"detail":13,"format":13,"mode":14,"style":15,"text":231,"type":17,"version":18}," 没什么意义。不同目标优化的是不同事件，投放系统在解两道不同的题，你看到的差异是目标带来的，不是素材带来的。",{"children":233,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":18,"textStyle":15},[234,236],{"detail":13,"format":18,"mode":14,"style":15,"text":235,"type":17,"version":18},"90% 的置信度够不够据此行动？",{"detail":13,"format":13,"mode":14,"style":15,"text":237,"type":17,"version":18}," 完全取决于判断错了的代价。以 90% 的把握把一个素材推到一个广告系列上，是笔划算的赌；据此重建整个账户就不是。置信度是决策的输入，不是决策本身。",{"children":239,"direction":19,"format":15,"indent":13,"type":26,"version":18,"textFormat":18,"textStyle":15},[240,242],{"detail":13,"format":18,"mode":14,"style":15,"text":241,"type":17,"version":18},"预算撑不起一个功效足够的测试怎么办？",{"detail":13,"format":13,"mode":14,"style":15,"text":243,"type":17,"version":18}," 那就去测更大的差异。小预算账户从\"两个真正不同的概念\"里学到的，远多于从\"同一个概念的若干精修版\"里学到的——因为在低量级下，只有大效应才检得出来。","root",{"id":246,"alt":247,"updatedAt":248,"createdAt":248,"url":249,"thumbnailURL":19,"filename":250,"mimeType":251,"filesize":252,"width":19,"height":19},1072,"Abstract split test diagram: one audience stream divided into two equal groups feeding separate A and B panels","2026-09-05T08:11:05.671Z","https://cms-r2.deepclick.com/847-cover-8726f014da73.jpg","847-cover-8726f014da73.jpg","application/octet-stream",106695,{"title":254,"description":255,"image":256},"Facebook 广告 A/B 拆分测试怎么做：设置、样本量与结果解读","Meta 的 A/B 测试如何按人切分受众、哪些变量可以有效测试、学习期为什么会扭曲小规模测试，以及一个结论到底需要多少数据量。",{"id":246,"alt":247,"updatedAt":248,"createdAt":248,"url":249,"thumbnailURL":19,"filename":250,"mimeType":251,"filesize":252,"width":19,"height":19},{"id":18,"key":258,"name":259,"prodHost":260,"testHost":261,"blogPath":262,"docPath":263,"zhPrefix":264,"deployHookTest":265,"deployHookProd":266,"brandAuthor":267,"brandKit":271,"enabled":275,"updatedAt":276,"createdAt":277},"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":124,"name":259,"avatar":268,"updatedAt":269,"createdAt":270},25,"2026-04-22T08:09:35.299Z","2026-04-22T06:42:49.116Z",{"logoLight":19,"logoDark":19,"primaryColor":272,"accentColor":273,"tagline":274,"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-ads-split-testing",{"id":188,"name":281,"avatar":282,"updatedAt":289,"createdAt":289},"Ethan Cole",{"id":283,"alt":284,"updatedAt":285,"createdAt":285,"url":286,"thumbnailURL":19,"filename":287,"mimeType":251,"filesize":288,"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":291,"site":292,"titleZh":294,"titleEn":295,"slug":296,"order":193,"updatedAt":297,"createdAt":298},7,{"id":18,"key":258,"name":259,"prodHost":260,"testHost":261,"blogPath":262,"docPath":263,"zhPrefix":264,"deployHookTest":265,"deployHookProd":266,"brandAuthor":124,"brandKit":293,"enabled":275,"updatedAt":276,"createdAt":277},{"logoLight":19,"logoDark":19,"primaryColor":272,"accentColor":273,"tagline":274,"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","2026-09-05T08:19:40.614Z","2026-09-05T08:11:15.246Z","\u003Cdiv class=\"payload-richtext\">\u003Ch2>拆分测试真正解决的问题，复制一个广告组解决不了\u003C/h2>\u003Cp>想知道素材 A 是不是比 B 好，最顺手的做法是复制一个广告组、改一处。这个测试在开跑之前就已经被污染了。\u003C/p>\u003Cp>两个瞄准同一批人的广告组会进同一场竞价。Meta 通常是从符合条件的广告组里挑一个去竞价，而不是让它们同时出价——于是你看到的差异里，混进了投放系统"选谁上场"的判断，而不只是受众对素材的偏好。更麻烦的是同一个人整周下来两个版本都能看到，所谓"赢的那版"可能只是拿到第二次曝光的那版。这和 \u003Ca href=\"https://deepclick.com/zh-CN/resources/blog/facebook-ads-audience-overlap/\">Facebook 广告受众重叠\u003C/a> 说的是同一个结构性问题，只不过那篇从成本角度看，这篇从测量角度看。\u003C/p>\u003Cp>A/B 测试工具专门修的就是"分配"这一环：它按\u003Cstrong>人\u003C/strong>切分受众，而不是按曝光切分。每个用户在整个测试期间只被分到一个版本，各分区之间不互相竞价。这就是这个工具存在的全部理由——别的事你手工也能做，随机且互不重叠的分组你做不到。\u003C/p>\u003Ch2>什么变量适合测\u003C/h2>\u003Cp>工具的设计前提是一次只改一个变量，可测维度是素材、受众、版位、投放优化目标。这个限制不是多余的：如果素材和受众一起改了，结果差异只能告诉你"这两个组合不一样"，不能告诉你是哪一半造成的，这个结论也没法迁移到别的广告系列。\u003C/p>\u003Cp>实际操作里要比工具的强制更严一档。把钩子、形式、时长、行动号召全换了的"新素材"，是四处改动套了一件衣服。这种测试仍然可以跑——有时你就是想知道新概念作为一个整体打不打得过旧的——但复盘时要如实写清你学到的是什么，因为"整包赢了"并不告诉你下一步该改哪里。\u003C/p>\u003Ch2>学习期会悄悄吃掉小预算测试\u003C/h2>\u003Cp>每个版本都是一个新广告组，每个新广告组都要重新进学习期。在投放稳定之前，单次成效成本既偏高又偏抖，跟它最终会落到的稳态不是一回事。\u003C/p>\u003Cp>常被引用的门槛是每个广告组每周约 50 次优化事件；具体数字请以产品内当前口径为准，Meta 调整过这个指引。真正要紧的是它逼出来的那道算术：测试预算被摊到各版本上，每个版本都得自己攒够事件。同样一笔钱，作为一个广告组本可以轻松走出学习期，一分为二之后可能两半都卡住——而两个都没跑出来、都还在学习期的广告组互相比较，测的是学习期，不是你的素材。\u003Ca href=\"https://deepclick.com/zh-CN/resources/blog/meta-ads-learning-phase/\">Meta 广告学习期\u003C/a> 讲了什么会重置它、要多久。\u003C/p>\u003Cp>如果诚实的答案是"这笔钱供不起两边都走出学习期"，那这个测试就还不该跑。换到量更大的广告系列上跑，或者拉长窗口，或者改测一个攒得更快的指标。\u003C/p>\u003Ch2>多数拆分测试真正错的地方：统计功效\u003C/h2>\u003Cp>广告测试大多栽在这里，而且是无声地栽——测试给出一个数字，数字看着像答案，没人回头检查它是否有可能是答案。\u003C/p>\u003Cp>转化是稀有事件。如果每一边只出个位数的转化，那么与你观测结果相容的"真实转化率"范围大得吓人。12 比 9 不是 33% 的提升，那是从两个你还分不开的分布里各抽了一次。不太舒服但有用的经验法则是：要检出一个温和的提升——也就是多数素材改动实际能带来的那种量级——需要每边\u003Cstrong>几百个\u003C/strong>转化，不是几十个；而要检出巨大的差异所需样本要少得多。这正是"两个真正不同的概念"之间的测试很快能出结论、而"按钮换个颜色"永远出不了结论的原因。\u003C/p>\u003Cp>开跑前先定三件事：\u003C/p>\u003Cul class=\"list-bullet\">\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"1\"\n        >\u003Cstrong>值得据此行动的最小差异\u003C/strong>。如果 5% 的差别根本不会改变你的投放决策，就别设计一个只能检出 5% 的测试。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"2\"\n        >\u003Cstrong>这个差异所需的转化量你出不出得起\u003C/strong>。出不起，就要么接受代理指标，要么干脆不跑。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"3\"\n        >\u003Cstrong>停止日期\u003C/strong>。在第一次曝光发生之前就写下来。\u003C/li>\u003C/ul>\u003Cp>Meta 会随结果给一个置信度。置信度低就照它的字面意思理解——工具在告诉你它分不开这两个版本——而不是把它当成对暂时领先那一版的微弱背书。\u003C/p>\u003Ch2>为什么不能"看着不错就收"\u003C/h2>\u003Cp>盯着在跑的测试、一看到某一版拉开就收手，是制造假赢家最有效的办法。测试早期领先者会频繁易主，纯属噪声。只要你允许自己在"差距看着够说服力"的任意时刻停手，那么即使拿两个一模一样的素材去测，你迟早也能找到这样一个时刻。\u003C/p>\u003Cp>防线是流程上的，不是统计上的：提前把结束日期钉死，只在结束时读一次结果。如果确实需要中途盯盘，那就盯\u003Cstrong>故障\u003C/strong>——某一版没花出去、广告被拒、落地页坏了——而不是盯谁赢。\u003C/p>\u003Ch2>"没结论"本身就是结论\u003C/h2>\u003Cp>团队容易把"无显著差异"当成白跑一趟。它是个发现：这两个版本之间的差异，小于你出得起钱去测量的差异。这说明该停止在这个维度上继续迭代，转向杠杆更大的地方——报价、受众定义、落地页——而不是再去做第四个标题变体。\u003C/p>\u003Cp>它同时还告诉你一件关于这两个版本的事：随便发哪个都行。既然表现分不出高下，就选制作成本更低、后续更好维护的那个。\u003C/p>\u003Ch2>开跑前的检查清单\u003C/h2>\u003Cul class=\"list-bullet\">\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"1\"\n        >只改了一个变量，而且你能用一句话把它说清楚。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"2\"\n        >预算够两边各自独立走出学习期。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"3\"\n        >最小差异量与停止日期，两样都事先写下来。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"4\"\n        >测试窗口内没有别的广告系列在打同一批受众，免得分组从外部被污染。\u003C/li>\u003Cli\n          class=\"\"\n          style=\"\"\n          value=\"5\"\n        >事先想好"没结论"时怎么办——趁你还冷静的时候决定。\u003C/li>\u003C/ul>\u003Cp>上游结构理顺了，能在测试开始前就消掉不少噪声：\u003Ca href=\"https://deepclick.com/zh-CN/resources/blog/facebook-ads-account-structure/\">Facebook 广告账户结构\u003C/a> 讲一个账户到底需要多少广告系列和广告组，\u003Ca href=\"https://deepclick.com/zh-CN/resources/blog/facebook-ads-bidding-strategy/\">Facebook 广告出价策略\u003C/a> 讲那些最容易把测试搅浑的投放设置。\u003C/p>\u003Ch2>常见问题\u003C/h2>\u003Cp>\u003Cstrong>拆分测试该跑多久？\u003C/strong> 长到两边都跑出学习期、并攒够你设定的最小差异所需的转化量，且至少跑满一整周，好让"星期几"的影响平摊到两边。两个条件不一致时，取更长的那个。\u003C/p>\u003Cp>\u003Cstrong>能不能拿两个不同的广告目标做拆分测试？\u003C/strong> 没什么意义。不同目标优化的是不同事件，投放系统在解两道不同的题，你看到的差异是目标带来的，不是素材带来的。\u003C/p>\u003Cp>\u003Cstrong>90% 的置信度够不够据此行动？\u003C/strong> 完全取决于判断错了的代价。以 90% 的把握把一个素材推到一个广告系列上，是笔划算的赌；据此重建整个账户就不是。置信度是决策的输入，不是决策本身。\u003C/p>\u003Cp>\u003Cstrong>预算撑不起一个功效足够的测试怎么办？\u003C/strong> 那就去测更大的差异。小预算账户从"两个真正不同的概念"里学到的，远多于从"同一个概念的若干精修版"里学到的——因为在低量级下，只有大效应才检得出来。\u003C/p>\u003C/div>","https://deepclick.com/zh-CN/resources/blog/facebook-ads-split-testing",{"zh-CN":279,"en":279},[305],{"id":36,"title":306,"site":307,"image":309,"mobileImage":319,"targetUrl":328,"enabled":275,"postDateFrom":329,"postDateTo":330,"updatedAt":331,"createdAt":332},"2026.10.20 Jakarta summit (deepclick)",{"id":18,"key":258,"name":259,"prodHost":260,"testHost":261,"blogPath":262,"docPath":263,"zhPrefix":264,"deployHookTest":265,"deployHookProd":266,"brandAuthor":124,"brandKit":308,"enabled":275,"updatedAt":276,"createdAt":277},{"logoLight":19,"logoDark":19,"primaryColor":272,"accentColor":273,"tagline":274,"wechatName":19,"xhsHandle":19,"wechatQr":19},{"id":310,"alt":311,"updatedAt":312,"createdAt":312,"url":313,"thumbnailURL":19,"filename":314,"mimeType":315,"filesize":316,"width":317,"height":318},1069,"TrafficTalking Jakarta summit 2026.10.20 (zh)","2026-09-04T10:44:19.548Z","https://cms-r2.deepclick.com/traffictalking-jakarta-banner-zh-aa032503f69b.png","traffictalking-jakarta-banner-zh-aa032503f69b.png","image/png",1128551,2000,377,{"id":320,"alt":321,"updatedAt":322,"createdAt":322,"url":323,"thumbnailURL":19,"filename":324,"mimeType":315,"filesize":325,"width":326,"height":327},1071,"TrafficTalking Jakarta summit 2026.10.20 (mobile-zh)","2026-09-04T10:44:29.391Z","https://cms-r2.deepclick.com/traffictalking-jakarta-banner-mobile-zh-372fe26c1a2d.png","traffictalking-jakarta-banner-mobile-zh-372fe26c1a2d.png",649961,1200,400,"https://t.me/rrrun_r","2026-09-03T00:00:00.000Z","2026-10-04T00:00:00.000Z","2026-09-04T10:45:05.578Z","2026-09-04T10:45:03.031Z",1788753210131]