Back to Details

ads-performance-analytics

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Ads Performance Analytics

付费媒体绩效分析

A data-team-mentor's playbook for interpreting paid media dashboards without fooling yourself.
The dashboard is the moment of truth for paid media decisions. The numbers on it determine whether you scale, hold, or kill. They also expose every platform's self-attribution bias, every modeled-conversion shortcut, every cross-platform double-count. Most "scale this campaign" decisions trace back to misreading the dashboard.
This skill is the discipline that prevents misreading. It assumes the campaign was strategically sound (see 
paid-media-strategy
). It assumes the creative was tested properly (see 
ads-creative-development
). The hard part is knowing what each number actually means, what it does not, and how to reconcile platform-reported metrics with the truth in your warehouse.
When to use this skill: any time you are about to scale, kill, or rebudget a campaign based on platform metrics; reconciling platform reports with revenue data; evaluating an agency's reporting; or building a paid media dashboard that will not lie to you.
这是一份数据团队导师编写的指南，教你如何正确解读付费媒体仪表盘，避免自我误导。
仪表盘是付费媒体决策的关键时刻，上面的数字决定了你要扩大投放、维持现状还是终止活动。同时，它也暴露了每个平台的自我归因偏见、每一个建模转化的捷径、每一次跨平台的重复统计。大多数“扩大该活动”的决策，根源都是对仪表盘的误读。
本技能就是防止误读的准则。它假设活动本身具备战略合理性（详见
paid-media-strategy
），创意也经过了恰当测试（详见
ads-creative-development
）。最难的部分是理解每个数字的实际含义、它不代表什么，以及如何将平台报告的指标与仓库中的真实数据进行对账。
适用场景：当你准备基于平台指标扩大、终止或重新分配广告预算时；当你核对平台报告与收入数据时；当你评估代理商的报告时；或是当你构建一份不会误导你的付费媒体仪表盘时。

What this skill is for

本技能的适用范围

This skill spans paid media result interpretation. It does not cover paid media strategy (use 
paid-media-strategy
), creative production (use 
ads-creative-development
), or platform-specific tooling (covered in the integrations microsites). Pair this skill with the relevant integrations microsite for platform-specific MCP commands and example prompts.
The audience is a marketer, growth analyst, agency analyst, or founder evaluating paid media reports. The voice is patient and clinical. There is no "trust the platform's number" or "ignore the platform entirely." Both are wrong. The discipline is knowing which numbers from which platform mean what, and what to reconcile against to make the actual decision.
本技能涵盖付费媒体结果解读，但不包含付费媒体策略（请使用
paid-media-strategy
）、创意制作（请使用
ads-creative-development
）或平台特定工具（相关内容在集成微站点中）。将本技能与对应的集成微站点结合使用，可获取平台特定的MCP命令和示例提示。
受众为营销人员、增长分析师、代理商分析师或评估付费媒体报告的创始人。内容风格耐心且严谨，既不主张“完全信任平台数据”，也不建议“彻底忽略平台数据”——这两种做法都是错误的。核心准则是了解不同平台的哪些数据代表什么，以及需要与哪些数据对账才能做出正确决策。

The result panel: what every paid media platform should expose

结果面板：每个付费媒体平台都应展示的内容

A trustworthy result panel exposes nine things. Anything missing is a signal to treat reported numbers with extra skepticism.
  1. Spend, impressions, clicks. Table-stakes metrics. Should match across platforms within rounding.
  2. Conversions with definition and window visible. Not just a count; the definition of what counts as a conversion and the attribution window applied. Without this, the count is unreadable.
  3. Attribution breakdown. Last-click vs view-through vs modeled. The mix of how the conversions were credited.
  4. Frequency. Impressions per unique user. The fatigue early-warning system.
  5. Audience saturation. Where the platform exposes it. A flat audience-saturation curve means there is room to scale; a steep curve means efficiency is dropping.
  6. Time series. Daily breakdown to spot novelty effects, fatigue, day-of-week patterns, and exogenous variance.
  7. Cost metrics in clear currency. CPC, CPM, CPA, ROAS with the math defined and the currency labeled. Do not assume USD.
  8. Conversion path data. Touchpoints before conversion, where available. Tells you whether a campaign is a closer or an opener.
  9. Filters, segments, and exports. Without these, the panel is a brochure, not a tool.
Platforms hide what makes their reporting look weakest. Google PMax hides keyword-level and placement-level data. Meta hides the modeled-conversion share. LinkedIn hides cross-device click paths. Treat hidden metrics as the place to dig.
一个可靠的结果面板应展示以下9项内容。任何一项缺失，都意味着你需要对报告数据保持额外的怀疑。
  1. 支出、曝光量、点击量：基础指标，不同平台间的数据应在合理误差范围内一致。
  2. 带定义和窗口期的转化数据：不能只展示转化数量，还需明确转化的定义和应用的归因窗口期。没有这些信息，转化数据毫无意义。
  3. 归因细分：最后点击、浏览归因、建模归因的占比，即转化信用的分配方式。
  4. 触达频次：每位独立用户的曝光次数，是疲劳预警系统。
  5. 受众饱和度：平台展示的受众饱和曲线。平缓曲线意味着仍有投放空间；陡峭曲线则表示投放效率正在下降。
  6. 时间序列：按日细分的数据，用于识别新奇效应、疲劳效应、周度规律和外部变量影响。
  7. 明确币种的成本指标：CPC、CPM、CPA、ROAS需明确计算方式和币种，不要默认是美元。
  8. 转化路径数据：可获取的转化前触点数据，用于判断活动是促成转化的收尾环节还是引流环节。
  9. 筛选、细分和导出功能：没有这些功能，面板只是一份宣传册，而非实用工具。
平台会隐藏那些让其报告显得薄弱的数据。Google PMax隐藏关键词和展示位置数据；Meta隐藏建模转化的占比；LinkedIn隐藏跨设备点击路径。对于隐藏的指标，要深入挖掘。

Platform-reported vs reality

平台报告数据 vs 真实情况

Every platform's dashboard is optimized to make the platform look effective. This is not a moral failing; it is a structural incentive. Platforms with rosier reporting attract more spend.
Conversion windows. Meta defaults to 7-day click plus 1-day view. Google defaults to 30-day click plus 1-day view. Different windows, same activity, different reported numbers. If you compare Google's 30-day-click count against Meta's 7-day-click count, you are comparing different definitions and pretending they are the same.
View-through attribution. Counted by Meta and Google for users who saw but did not click. Often half the reported "conversions" are view-through. Treat view-through as a signal of awareness contribution, not as a direct response measurement. The user might have converted from organic search anyway.
Modeled conversions. When iOS users opt out of tracking, Meta and others statistically model what the conversion would have been. Modeled numbers are educated guesses, not measurements. They are useful for direction; they are not reliable for precision.
Self-attribution bias. Every platform's pixel fires on conversion and the platform claims credit. If you ran Meta, Google, and TikTok in the same week, all three platforms report your conversions as theirs. Sum-of-platforms is always greater than 100% of actual conversions.
The discipline. Never report platform-reported numbers as fact in board decks. Always reconcile against the single source of truth (warehouse, GA4, or unified analytics platform). Detail in 
references/platform-reporting-quirks.md
.
每个平台的仪表盘都经过优化，旨在让平台看起来更有效。这并非道德问题，而是结构性激励使然——报告数据越乐观，吸引的广告支出就越多。
转化窗口期：Meta默认7天点击+1天浏览归因；Google默认30天点击+1天浏览归因。相同的活动，不同的窗口期，报告数据也会不同。如果你拿Google的30天点击转化数和Meta的7天点击转化数对比，本质是在对比不同定义的数据，却假装它们是同一标准。
浏览归因：Meta和Google会统计看到广告但未点击的用户的转化，通常报告的“转化”中有一半是浏览归因。要将浏览归因视为品牌认知贡献的信号，而非直接响应指标——用户可能本来就会通过自然搜索完成转化。
建模转化：当iOS用户选择退出追踪时，Meta等平台会通过统计建模估算转化情况。建模数据是有根据的猜测，而非实际测量值。它们可用于判断趋势，但无法保证精度。
自我归因偏见：每个平台的像素在转化时触发，平台就会宣称功劳。如果你同时投放Meta、Google和TikTok，三个平台都会将你的转化算到自己头上。各平台数据相加，总会超过实际转化的100%。
准则：永远不要在董事会汇报中把平台报告数据当作事实。始终以单一可信来源（仓库、GA4或统一分析平台）的数据为准。详情见
references/platform-reporting-quirks.md
。

Attribution models in practice

实际应用中的归因模型

Six models and one anti-model. None is right. They are all approximations. The discipline is picking one, committing, and reading the others as sanity checks.
Last-click. Simple, reproducible, undercredits awareness. The conversion is fully credited to the last click before the conversion event. Easy to compute; easy to compare across channels; bad for understanding upper-funnel contribution.
First-click. Opposite bias. Fully credits the first touchpoint, undercredits closing channels. Useful as a sanity check against last-click; rarely the right primary view.
Linear. Equal credit across all touchpoints. Gives every channel something. Defensible; not informative. Most useful for board reporting where avoiding "Google gets 70% so we cut Meta" politics matters more than precision.
Time-decay. More credit to recent touchpoints. Reflects the intuition that recent ads are more influential. Hard to argue against; hard to verify.
U-shaped (position-based). Heavy on first and last (40% each), light on middle (20% distributed). Honors both opener and closer roles. The default in many MTA tools.
Data-driven attribution (DDA, Google). Machine-learning model that distributes credit based on observed conversion paths. Opaque; hard to audit. The closest to "right" for digital channels but a black box.
Marketing mix modeling (MMM). Regression-based, top-down. Uses spend and revenue time series across channels to estimate channel contributions. Requires 2+ years of data. The strongest defense against platform self-attribution because it does not rely on platform-reported conversions at all.
The anti-model: trusting platform-reported attribution. Each platform's "DDA" or "attributed conversions" is the platform's self-attribution. Sum across platforms exceeds reality. Use platform attribution for in-flight optimization within the platform; use a unified attribution model for cross-channel decisions.
Practical guidance.
  • Early-stage. Use last-click plus a single guardrail metric (warehouse-attributed CAC). Sophisticated attribution requires data volume you do not have.
  • Mid-stage. Data-driven attribution from Google plus GA4, with explicit awareness vs closing channel labeling.
  • Mature. MMM as the canonical incremental reference. MTA for in-flight optimization. Last-click for channel-level decisions where ambiguity is acceptable.
Detail and a decision matrix in 
references/attribution-model-comparison.md
.
六种模型和一种反模型。没有一种是绝对正确的，它们都是近似值。准则是选择一种模型并坚持使用，将其他模型作为 sanity check（合理性验证）。
最后点击归因：简单、可复现，但低估了品牌认知的贡献。转化的全部信用归于转化事件前的最后一次点击。计算简单，便于跨渠道对比，但不利于理解漏斗上层的贡献。
首次点击归因：与最后点击归因相反，将全部信用归于第一个触点，低估了收尾渠道的作用。可作为最后点击归因的合理性验证，但很少作为主要参考模型。
线性归因：所有触点平分信用。每个渠道都能获得一定认可，有说服力但缺乏信息量。最适合用于董事会汇报，此时避免“Google占70%所以砍掉Meta”这类争议比精度更重要。
时间衰减归因：越近期的触点获得越多信用，反映了“近期广告影响力更大”的直觉。难以反驳，但也难以验证。
U型（基于位置）归因：首尾触点各占40%，中间触点共占20%。兼顾引流和收尾角色，是许多MTA工具的默认模型。
数据驱动归因（DDA，Google）：基于观察到的转化路径分配信用的机器学习模型。不透明，难以审计。是最接近数字渠道“正确”结果的模型，但属于黑箱。
营销组合模型（MMM）：基于回归的自上而下模型，利用跨渠道的支出和收入时间序列估算渠道贡献。需要2年以上的数据。是对抗平台自我归因的最强手段，因为它完全不依赖平台报告的转化数据。
反模型：信任平台报告的归因：每个平台的“DDA”或“归因转化”都是平台的自我归因。跨平台数据相加会超过真实值。平台归因仅适用于平台内的实时优化；跨渠道决策需使用统一归因模型。
实用指南：
  • 早期阶段：使用最后点击归因，搭配一个单一的保障指标（仓库归因的CAC）。复杂归因需要的数据量你暂时没有。
  • 中期阶段：使用Google的数据驱动归因搭配GA4，明确标注品牌认知渠道和收尾渠道。
  • 成熟阶段：将MMM作为标准的增量参考，MTA用于实时优化，最后点击归因用于可接受模糊性的渠道级决策。
详情和决策矩阵见
references/attribution-model-comparison.md
。

Multi-platform reconciliation

多平台对账

The trap. Google says you spent $50K with 800 conversions. Meta says $30K with 600. LinkedIn says $20K with 200. Total reported equals 1,600 conversions. Your warehouse says 950. Where did 650 go?
The answer. Nowhere. They never existed. Each platform claimed conversions other platforms also claimed.
The reconciliation pattern.
  • Trust the warehouse for total conversions and total revenue.
  • Trust platforms for relative ranking within platform (which campaign won, which audience won).
  • Never trust platform sums.
  • Compute blended CAC as (total ad spend across platforms) divided by (total new customers from warehouse). Not from platform reports.
The board-deck pattern. Report warehouse-attributed conversion counts, never platform-summed. Report blended CAC, not channel-by-channel CAC unless explicitly noted as platform-self-attributed. Detail and templates in 
references/dashboard-reconciliation-patterns.md
.
常见陷阱：Google显示你花费5万美元，获得800次转化；Meta显示花费3万美元，获得600次转化；LinkedIn显示花费2万美元，获得200次转化。平台报告的总转化数为1600次，但你的仓库数据显示只有950次。那650次去哪了？
答案：它们从未存在过。每个平台都声称了其他平台也声称过的转化。
对账模式：
  • 以仓库数据为准，获取总转化数和总收入。
  • 信任平台内部的相对排名（哪个活动表现更好，哪个受众群体效果更佳）。
  • 永远不要信任平台数据的总和。
  • 计算混合CAC：（跨平台总广告支出）÷（仓库统计的新客户总数），不要使用平台报告的数据。
董事会汇报模式：汇报仓库归因的转化数，永远不要用平台数据相加的结果。汇报混合CAC，而非按渠道划分的CAC（除非明确标注是平台自我归因的数据）。详情和模板见
references/dashboard-reconciliation-patterns.md
。

ROAS vs LTV: the time horizon trap

ROAS vs LTV：时间陷阱

ROAS is short-term. Revenue from purchases attributed to a campaign in a fixed window, often 7 to 30 days. LTV is long-term. Total customer lifetime revenue.
Decisions made on ROAS can be wrong if LTV varies by channel. A worked example.
Meta drives 2.5x ROAS at $40 CAC with $80 LTV. The 7-day-click revenue covers 1.5x payback over the customer lifetime.
Google drives 1.8x ROAS at $60 CAC with $200 LTV. The 7-day-click revenue covers 3.3x payback over the customer lifetime.
Google looks worse on ROAS, better on LTV-adjusted return. Allocating budget to Meta because the ROAS is higher is the wrong move.
The fix. Cohort-based LTV by acquisition channel, updated quarterly. Compare channels on payback period or LTV-CAC ratio, not raw ROAS. The 2x ROAS heuristic is a dangerous shortcut. Same ROAS at different LTVs equals different actual returns.
The trap that compounds. Performance teams optimize for short-term ROAS because the metric refreshes weekly. Brand and high-LTV channels get cut because their short-term ROAS is lower. Six months later, the brand pipeline has eroded and short-term ROAS itself drops because the cheap channels are saturated. The metric that drove the decision was the wrong horizon.
ROAS是短期指标，指固定窗口期（通常7至30天）内活动带来的购买收入。LTV是长期指标，指客户生命周期内的总收入。
如果不同渠道的LTV存在差异，基于ROAS做出的决策可能是错误的。举个例子：
Meta的ROAS为2.5倍，CAC为40美元，LTV为80美元。7天点击收入在客户生命周期内可覆盖1.5倍的成本回收。
Google的ROAS为1.8倍，CAC为60美元，LTV为200美元。7天点击收入在客户生命周期内可覆盖3.3倍的成本回收。
从ROAS看Google表现更差，但从LTV调整后的回报看，Google表现更好。因为ROAS更高就把预算分配给Meta是错误的决策。
解决方案：按获客渠道进行基于 cohort（群组）的LTV分析，每季度更新。对比渠道的回收期或LTV-CAC比率，而非原始ROAS。“2倍ROAS”的经验法则是危险的捷径——相同的ROAS，不同的LTV意味着实际回报完全不同。
更严重的陷阱：绩效团队会因为短期ROAS指标每周更新而优化短期ROAS，品牌和高LTV渠道会因为短期ROAS较低被砍掉。六个月后，品牌渠道的流量枯竭，短期ROAS本身也会因为低成本渠道饱和而下降。驱动决策的指标选错了时间维度。

Cohort analysis vs daily metrics

群组分析 vs 每日指标

Daily metrics tell you what happened today. Cohort analysis tells you whether today's customers are different from last month's.
Three cohort cuts that matter for paid media.
By acquisition month. Are users acquired in March retaining better than users acquired in January? A declining LTV over rolling acquisition cohorts means recent acquisition is lower quality; the daily metrics will show this two to three months later when the retention starts hurting.
By acquisition channel. Are users from Meta retaining better than users from Google? Channel-level cohort divergence is the data behind the LTV-vs-ROAS argument. Meta might drive volume at lower LTV; Google might drive lower volume at higher LTV. The cohort tells the story the daily ROAS hides.
By acquisition campaign. Campaign-level cohort signals. Useful for diagnosing why a campaign that "works" in week 1 produces no recurring revenue.
The signal to act on. Declining cohort LTV over two consecutive months is the alarm. Pause the channel or campaign before the daily metrics force you to. Detail in 
references/cohort-analysis-templates.md
.
每日指标告诉你今天发生了什么。群组分析告诉你今天的客户和上个月的客户是否存在差异。
付费媒体领域有三种重要的群组划分方式：
按获客月份划分：3月获客的用户留存率是否高于1月获客的用户？滚动获客群组的LTV下降，意味着近期获客质量降低——每日指标会在2到3个月后，当留存率开始影响数据时才会体现这一点。
按获客渠道划分：Meta获客的用户留存率是否高于Google获客的用户？渠道层面的群组差异是LTV vs ROAS争论的数据支撑。Meta可能带来高流量但低LTV，Google可能带来低流量但高LTV。群组分析能揭示每日ROAS隐藏的真相。
按获客活动划分：活动层面的群组信号，用于诊断为什么某个活动在第一周“表现良好”，却没有产生复购收入。
需要采取行动的信号：连续两个月群组LTV下降是警报，要在每日指标被迫让你采取行动前暂停该渠道或活动。详情见
references/cohort-analysis-templates.md
。

Statistical noise in performance metrics

绩效指标中的统计噪声

Most "the campaign improved 15% week over week" stories are noise. Real performance changes are 30% or more in metrics that vary 10 to 20% naturally. Below that threshold, you are looking at variance and calling it signal.
Sources of noise in paid media metrics.
  • Day-of-week effects. B2C tends to weekend dips. B2B tends to weekend gains. A "Monday morning is better" hypothesis often dissolves when day-of-week is normalized.
  • Holiday and seasonal effects. Q4 dwarfs most "optimization" effects. A campaign launched in Q4 looks great because of seasonality, not strategy.
  • Weather, news, competitor activity. Real exogenous variance. Last week's news cycle can shift CPMs across an entire vertical.
  • Pixel fire and reporting delay. Conversions reported on a 7-day click window arrive incrementally. Reading the panel on Monday for last week's performance undercounts.
The fix. Pre-commit to test duration before drawing conclusions. Use the experimentation discipline from 
experimentation-analytics
for any directional change you want to claim is real. The signal-to-noise problem in paid media metrics is the same as the signal-to-noise problem in product experiments; the framework transfers.
This is where 
experimentation-analytics
bridges in. The statistical patterns are the same; the application is different. Read both for the full picture.
大多数“活动周环比提升15%”的说法都是噪声。真实的绩效变化通常在30%以上，而指标本身自然波动范围在10%到20%之间。低于这个阈值，你看到的只是方差，却误以为是信号。
付费媒体指标噪声的来源：
  • 周度效应：B2C业务通常周末数据下滑，B2B业务通常周末数据上升。“周一表现更好”的假设，在排除周度效应后往往不成立。
  • 节假日和季节性效应：第四季度的季节性影响远大于大多数“优化”效果。在第四季度推出的活动看起来表现良好，往往是因为季节性，而非策略有效。
  • 天气、新闻、竞争对手活动：真实的外部变量影响。上周的新闻周期可能会影响整个行业的CPM。
  • 像素触发和报告延迟：7天点击窗口期的转化数据会逐步上报。周一查看上周的绩效数据会低估转化数。
解决方案：在得出结论前预先确定测试时长。对于任何你声称是真实的方向性变化，使用
experimentation-analytics
中的实验准则。付费媒体指标的信噪比问题与产品实验的信噪比问题相同，框架可以通用。
这正是
experimentation-analytics
的衔接点。统计模式相同，应用场景不同。同时阅读两者可获取完整视角。

Incrementality testing

增量测试

The honest test. If we had not run this ad, would we still have gotten the conversion? The number above zero is the incremental contribution.
Most paid media is 30 to 70% incremental, not 100%. Some is zero. Branded search bidding is often 5 to 20% incremental (most converters would have found you organically). Retargeting is often 20 to 40% incremental (many of those users were going to convert anyway). Prospecting is often 50 to 90% incremental.
Four methods.
Geo holdout. Hold one region out from the campaign. Measure the difference in conversions between the holdout region and the matched test region. The cleanest causal test for paid media at scale.
Ghost bidding (Google). Google's own incrementality tool. Bids on a holdout share of impressions but does not actually serve the ad. Reports incremental conversions. Honest signal; some teams find the math opaque.
Conversion lift studies (Meta). Splits audiences into test and control. Test sees the ad; control does not. Reports incremental lift. The cleanest within-Meta test.
PSA tests. Serve some users a public service announcement instead of your ad. Compare conversion rates. Useful for legacy brands with deep budget.
Incremental rate ranges by channel type are in 
references/incrementality-testing-playbook.md
. The discipline is to run incrementality tests at least quarterly on the highest-spend channels. Without them, you are optimizing against platform-reported attribution that systematically overcounts.
最诚实的测试：如果我们没有投放这个广告，还能获得这次转化吗？大于零的数值就是增量贡献。
大多数付费媒体的增量在30%到70%之间，而非100%。有些甚至为零。品牌搜索出价的增量通常在5%到20%之间（大多数转化用户本来就会通过自然搜索找到你）；再营销的增量通常在20%到40%之间（许多用户本来就会完成转化）；拉新活动的增量通常在50%到90%之间。
四种方法：
Geo holdout（地理对照组）：将一个地区排除在活动之外，测量对照组和匹配测试组之间的转化差异。这是大规模付费媒体最可靠的因果测试方法。
Ghost bidding（幽灵出价，Google）：Google自带的增量工具。对一部分曝光出价但不实际投放广告，报告增量转化。信号真实，但部分团队认为计算方式不透明。
Conversion lift studies（转化提升研究，Meta）：将受众分为测试组和对照组，测试组看到广告，对照组看不到，报告增量提升。这是Meta内部最可靠的测试方法。
PSA测试：向部分用户投放公益广告而非你的广告，对比转化率。适用于预算充足的成熟品牌。
不同渠道类型的增量率范围见
references/incrementality-testing-playbook.md
。准则是至少每季度对支出最高的渠道进行增量测试。没有这些测试，你就是在针对系统性高估的平台归因数据进行优化。

Geo experiments and holdouts

地理实验与对照组

For paid media specifically, geo-based testing is the most reliable causal method.
Geo holdout. Turn off paid media in one region. Measure baseline organic conversions. The difference between expected and actual conversions in the holdout region is the incremental paid contribution.
Geo lift. Scale spend in one region by 2x. See if conversions scale linearly. A linear scale means the channel has headroom. A sublinear scale means saturation; further spend is diminishing returns.
Switchback. Alternate weeks of campaign on and off. Compare on-weeks to off-weeks. Useful when geo splitting is not feasible.
Pre-and-post analysis. Launch in a region; measure 30 days before vs 30 days after. Weak design because external factors confound the comparison. Use only when no other test is available.
The right setup. Matched markets (similar demographics, similar baseline conversion rates). Statistical power calculation upfront (how big a difference can the test actually detect). Pre-committed analysis window (so you do not stop early when the data looks good or wait too long when it looks bad).
The trap. Calling a geo test successful because of timing-correlated revenue lift. A campaign launched in October will see "lift" because Q4 is starting; without a control region, the lift is not attributable to the campaign.
对于付费媒体而言，基于地理的测试是最可靠的因果方法。
Geo holdout（地理对照组）：在一个地区关闭付费媒体投放，测量基线自然转化数。对照组的预期转化数与实际转化数的差值，就是付费媒体的增量贡献。
Geo lift（地理提升）：将一个地区的支出翻倍，看转化数是否线性增长。线性增长意味着该渠道仍有空间；次线性增长意味着饱和，进一步支出会导致收益递减。
Switchback（交替测试）：每周交替开启和关闭活动，对比开启周和关闭周的数据。适用于无法进行地理拆分的场景。
前后分析：在一个地区启动活动，对比启动前30天和启动后30天的数据。设计薄弱，因为外部因素会干扰对比。仅在没有其他测试方法时使用。
正确的设置：匹配市场（相似的人口结构、相似的基线转化率）。预先计算统计功效（测试能检测到多大的差异）。预先确定分析窗口期（避免在数据看起来好时提前停止，或在数据看起来差时过度等待）。
陷阱：因为时间相关的收入增长就宣称地理测试成功。10月启动的活动会看到“增长”，因为第四季度开始了；没有对照组的话，增长无法归因于活动。

Platform self-attribution bias

平台自我归因偏见

A specific failure mode worth its own section.
The mechanism. Platform's pixel fires on conversion. Platform claims credit. The user might have converted from any channel; the platform that loaded the pixel last gets the credit on the platform's own dashboard.
Why platforms reward this design. More credit on the platform dashboard equals better-looking ROAS equals more advertiser spend. Platforms have no incentive to underreport their own contribution.
Detection patterns. When platform-reported conversions exceed warehouse-attributed conversions for the same channel by more than 30%, you have heavy double-counting. When sum-of-platform-reports exceeds total conversions in the warehouse, you have cross-platform double-counting.
The fix. Warehouse as canonical for board reporting. Platform reporting as in-flight signal only. Incrementality tests at least quarterly to keep the channel-attribution honest.
A worked example. A retargeting campaign in Meta showed 3.5x ROAS for six months. The team scaled spend from $20K to $80K per month. A geo holdout test revealed that 65% of the "conversions" would have happened anyway from organic. Real ROAS adjusted for incrementality was 1.2x, not 3.5x. The campaign got cut and warehouse-attributed CAC dropped 18% in the next quarter.
这是一个值得单独讨论的错误模式。
机制：平台的像素在转化时触发，平台就会宣称功劳。用户可能通过任何渠道完成转化，但在平台自己的仪表盘中，最后加载像素的平台会获得信用。
平台为何采用这种设计：平台仪表盘上的信用越多，ROAS看起来越好，广告主的支出就越多。平台没有动力低估自己的贡献。
检测模式：当平台报告的转化数比仓库归因的同一渠道转化数高出30%以上时，存在严重的重复统计；当各平台报告数据之和超过仓库中的总转化数时，存在跨平台重复统计。
解决方案：董事会汇报以仓库数据为标准，平台报告仅作为实时信号。至少每季度进行增量测试，确保渠道归因真实。
实例：Meta的再营销活动连续六个月显示3.5倍ROAS，团队将支出从每月2万美元增加到8万美元。Geo holdout测试显示，65%的“转化”本来就会通过自然渠道完成。考虑增量后的真实ROAS是1.2倍，而非3.5倍。该活动被砍掉后，仓库归因的CAC在接下来的季度下降了18%。

Common interpretation failures

常见的解读错误

Twelve patterns recur in paid media reporting work. Detail in 
references/common-interpretation-failures.md
.
  • "ROAS dropped 20% week over week, kill the campaign." Could be noise. Pre-commit a test window before acting on weekly variance.
  • "Meta says 500 conversions, my warehouse says 200, who is right?" Both are wrong; warehouse is closer to truth, Meta self-attributes. Reconcile, do not pick a winner.
  • "We turned off Google PMax and conversions did not drop." PMax was harvesting branded search you would have gotten free. Audit branded queries inside PMax.
  • "The new campaign hit 5x ROAS in week 1." Likely retargeting hot leads. Check the audience composition before declaring victory.
  • "We A/B tested and one creative wins by 12%." Within margin of platform noise. Not significant.
  • "Our LTV calculation says this channel is profitable." Check cohort age. Recent cohorts may not have hit LTV yet; the calculation is a projection, not a measurement.
  • "The platform says high frequency is fine because conversions are still happening." Fatigue masked by free organic conversions. The campaign is taking credit for conversions that would have happened anyway.
  • "Last-click attribution shows Meta at 60% credit." Last-click bias. First-click view of the same data shows different. Pick a model and stick.
  • "We scaled spend 5x and conversions only doubled." Saturation. The channel found its ceiling; the marginal CAC at the new spend level is much higher.
  • "Brand campaigns underperform on direct ROAS." They do not have to. Brand impact shows up in other channels' efficiency. Measure brand against brand-search lift, not direct ROAS.
  • "ROAS held steady but profit dropped." The mix shifted toward lower-margin products. Channel-level ROAS hides product-mix effects.
  • "Agency reported a 4x ROAS month." Whose number? Platform-reported, warehouse-attributed, or model-adjusted? The unit of measurement matters more than the magnitude.
付费媒体报告工作中反复出现12种错误模式。详情见
references/common-interpretation-failures.md
。
  • “ROAS周环比下降20%，砍掉这个活动。”可能只是噪声。在根据周度方差采取行动前，预先确定测试窗口期。
  • “Meta显示500次转化，我的仓库显示200次，谁是对的？”两者都不对；仓库数据更接近真相，Meta是自我归因。要对账，而非选边站。
  • “我们关闭了Google PMax，转化数没有下降。”PMax一直在抢占你本来可以免费获得的品牌搜索流量。检查PMax中的品牌关键词。
  • “新活动第一周ROAS达到5倍。”可能是针对高意向用户的再营销。在宣布成功前检查受众构成。
  • “我们做了A/B测试，一个创意胜出12%。”在平台噪声范围内，不具备显著性。
  • “我们的LTV计算显示这个渠道盈利。”检查群组时长。近期群组可能还未达到LTV，计算结果是预测值，而非实际测量值。
  • “平台显示高触达频次没问题，因为转化还在发生。”疲劳效应被免费自然转化掩盖了，活动在抢本来就会发生的转化的功劳。
  • “最后点击归因显示Meta占60%的信用。”这是最后点击偏见。同一数据的首次点击视角会显示不同结果。选择一个模型并坚持使用。
  • “我们将支出增加5倍，转化数只翻倍。”饱和了，该渠道已达上限，新支出水平下的边际CAC高得多。
  • “品牌活动的直接ROAS表现不佳。”这并非必须，品牌影响体现在其他渠道的效率提升上。要衡量品牌搜索提升，而非直接ROAS。
  • “ROAS保持稳定，但利润下降了。”产品组合转向低利润率产品。渠道级ROAS掩盖了产品组合的影响。
  • “代理商报告当月ROAS为4倍。”是谁的数据？平台报告的、仓库归因的还是模型调整的？测量标准比数值大小更重要。

The framework: 12 considerations for trustworthy paid media interpretation

框架：可信付费媒体解读的12项考量

When reading a paid media dashboard about to inform a decision, walk these 12 considerations. Skipping any of them is how teams ship the wrong call.
  1. Result panel completeness. What is the platform showing vs hiding.
  2. Platform-reported vs reality. View-through, modeled conversions, conversion windows.
  3. Attribution model. Pick one and read the others as sanity checks.
  4. Multi-platform reconciliation. Sum-of-platforms is always inflated.
  5. ROAS vs LTV horizon. Short-term metric, long-term impact.
  6. Cohort vs daily. Cohort tells the quality story; daily tells the volume story.
  7. Statistical noise. Weekly variance, day-of-week, seasonal, exogenous.
  8. Incrementality. What would have happened without the spend.
  9. Geo and holdout testing. The honest causal test.
  10. Self-attribution bias. Platforms claim credit they do not deserve.
  11. Decision rule. Pre-committed scale up, hold, or pull back.
  12. Single source of truth. Warehouse over platform reporting for board metrics.
The output of the framework is one of three answers. Scale (the campaign is incremental and unit economics work). Hold (data is ambiguous; gather more before deciding). Kill (the campaign is not incremental enough to justify the spend).
当你准备根据付费媒体仪表盘做出决策时，逐一检查这12项考量。跳过任何一项都可能导致错误决策。
  1. 结果面板完整性：平台展示了什么，隐藏了什么。
  2. 平台报告数据 vs 真实情况：浏览归因、建模转化、转化窗口期。
  3. 归因模型：选择一种模型，将其他模型作为合理性验证。
  4. 多平台对账：平台数据总和总会被高估。
  5. ROAS vs LTV时间维度：短期指标，长期影响。
  6. 群组分析 vs 每日指标：群组分析揭示质量真相，每日指标展示流量情况。
  7. 统计噪声：周度方差、周度效应、季节性、外部变量。
  8. 增量性：没有这笔支出会发生什么。
  9. 地理与对照组测试：最诚实的因果测试。
  10. 自我归因偏见：平台会宣称不属于自己的功劳。
  11. 决策规则：预先确定扩大、维持或缩减的标准。
  12. 单一可信来源：董事会指标以仓库数据为准，而非平台报告。
该框架的输出是三个答案之一：扩大（活动具备增量性，单位经济可行）、维持（数据模糊，需收集更多信息再决策）、终止（活动增量不足，不值得投入）。

Reference files

参考文件

  • references/metric-definitions-glossary.md
     - CTR, CPC, CPM, CPA, ROAS, LTV, AOV, frequency, reach, impressions, conversion window, view-through, modeled conversion, blended CAC, MER.
  • references/attribution-model-comparison.md
     - Last-click, first-click, linear, time-decay, U-shaped, DDA, MMM. Decision matrix by business stage.
  • references/platform-reporting-quirks.md
     - Google PMax black box, Meta iOS impact and Conversions API, LinkedIn 30-day click defaults, TikTok video-completion attribution, programmatic viewability gates.
  • references/incrementality-testing-playbook.md
     - Geo holdout, ghost bidding, conversion lift, PSA tests, switchback designs. Setup, duration, analysis pattern, expected incremental rates.
  • references/dashboard-reconciliation-patterns.md
     - Warehouse as canonical, platform as in-flight signal, blended CAC formula, board-deck patterns, reconciliation cadence.
  • references/cohort-analysis-templates.md
     - By acquisition month, channel, and campaign. Retention curves, when to act on cohort signals.
  • references/common-interpretation-failures.md
     - Twelve failure patterns with symptom, root cause, fix, prevention.
  • references/metric-definitions-glossary.md
     - CTR、CPC、CPM、CPA、ROAS、LTV、AOV、触达频次、覆盖人数、曝光量、转化窗口期、浏览归因、建模转化、混合CAC、MER。
  • references/attribution-model-comparison.md
     - 最后点击、首次点击、线性、时间衰减、U型、DDA、MMM。按业务阶段划分的决策矩阵。
  • references/platform-reporting-quirks.md
     - Google PMax黑箱、Meta iOS影响与Conversions API、LinkedIn 30天点击默认设置、TikTok视频完成归因、程序化可见性门槛。
  • references/incrementality-testing-playbook.md
     - Geo holdout、幽灵出价、转化提升研究、PSA测试、交替测试设计。设置、时长、分析模式、预期增量率。
  • references/dashboard-reconciliation-patterns.md
     - 仓库作为标准来源、平台作为实时信号、混合CAC公式、董事会汇报模式、对账频率。
  • references/cohort-analysis-templates.md
     - 按获客月份、渠道、活动划分的群组分析。留存曲线、何时根据群组信号采取行动。
  • references/common-interpretation-failures.md
     - 12种错误模式，包含症状、根源、解决方案、预防措施。

Closing: the courage to call it incremental zero

结语：敢于承认增量为零

Most paid media spend is not 100% incremental. Some channels are 70% incremental. Some are 30%. Some are zero, paying for conversions you would have gotten anyway.
The discipline of accepting that channels can be incremental zero, and pulling spend accordingly, is the single highest-impact skill in paid media analytics. Most accounts have at least one campaign that looks profitable in the platform but is incremental zero in the warehouse. Branded paid search at $4 CPC when the same users find you at position one organically. Retargeting at $0.30 CPC for users who already added items to cart. PMax cannibalizing free brand traffic.
The discipline of finding those campaigns and killing them is the work. The platform will not tell you. The platform's incentive is the opposite. The warehouse, paired with quarterly incrementality tests, is the only honest source.
When in doubt about whether a campaign is incremental, run a geo holdout. The two-week test is cheaper than a quarter of unincremental spend. The team that does not run incrementality tests is optimizing against numbers that are systematically wrong, and the size of the error is exactly the size of the budget waste.
大多数付费媒体支出并非100%增量。有些渠道增量为70%，有些为30%，有些则为零——你在为本来就会获得的转化付费。
接受渠道可能增量为零，并相应缩减支出，是付费媒体分析中影响最大的技能。大多数账户至少有一个在平台报告中看起来盈利，但在仓库数据中增量为零的活动。比如，当用户本来就能在自然搜索结果第一页找到你时，却以4美元CPC投放品牌搜索；再比如，为已经加购的用户以0.3美元CPC投放再营销；或是PMax抢占免费品牌流量。
找到这些活动并终止它们，就是核心工作。平台不会告诉你这些，平台的激励机制恰恰相反。只有仓库数据搭配季度增量测试，才是唯一诚实的来源。
如果不确定一个活动是否具备增量性，就进行Geo holdout测试。两周的测试比一个季度的无增量支出更便宜。不进行增量测试的团队，是在针对系统性错误的数据进行优化，错误的规模就是预算浪费的规模。