一句话：一家美国手游厂商全球停投广告 24 天，自然新增安装掉了 20%–30%。这些用户不是广告直接带来的，而是广告把游戏推上了应用商店榜单，榜单又带来了自然量。所以只看“付费带来的安装数”会系统性低估买量，而按平台榜单运作的产品，连“留 10% 用户不看广告”这种标准增量实验都可能测不准。
研究者拿到了一家美国手游厂商（论文里叫 GameSpace）近两年的投放数据：500 天、85 家广告平台、6 款游戏、iOS 与 Android 双端，期间产生超过 1000 亿次广告曝光。数据由一家代投公司提供，论文明确说明“日历日期因隐私原因隐去”，也就是说外界无法核对具体是哪款游戏、哪段时间。
在这 500 天里，厂商做了一件在行业里很少见的事：全球全平台停投，从第 75 天到第 98 天，共 24 天。不是停某个渠道，是搜索、社交、展示、应用内全部停掉。
停投之后，付费安装几乎归零（这是意料之中）。意料之外的是自然安装——用事件研究法（event study）估计，自然安装下降了 20%–30%，随带宽（7 天 / 11 天 / 23 天）在 −0.305 到 −0.359 之间。用固定效应面板模型换算成金额：每 100 美元投放对应 32.3 个付费安装和 2.2 个自然安装。两种方法算出来的量级基本一致。
同时，这 6 款游戏在应用商店分类榜上的排名恶化了 2.3 到 3.3 倍（对数排名口径）。
关键证据不是“自然量掉了”，而是掉在哪里。
研究者把榜单排名数据并进模型后做中介检验（mediation analysis），分两步：
第二步是整篇论文里信息量最大的一句话。如果投放真的通过排名影响自然量，那么“控制住排名”就等于掐断了下游那一段，上游的相关性自然随之消失。作者据此判断：排名是中介变量（mediator），不是又一个同时变化的伴随现象。
把它写成一条能讲给同事听的链：
投放增加 → 付费安装增加（安装速度上升）
→ 商店把"安装速度快的应用"往榜单上抬
→ 榜单是可被所有用户看见的公共位置
→ 自然浏览带来的安装增加
→ 自然安装又进一步抬高榜单位置最后一环是它和自我强化的循环接上的地方，也是它“反直觉”的地方。
这篇论文的对照对象是搜索广告领域几个著名的关停实验。eBay 停掉品牌词搜索广告后，大部分流量只是从付费位置挪到了自然位置，广告基本是在抢自然流量的功劳（over-attribution）；后续在 Edmunds、以及 Golden 与 Horton 的研究里，付费搜索对自然流量的替代率在 37% 到 63% 之间。
同一个“停投”手法，搬到移动应用上，结论反过来了。区别在于用户是怎么找到你的：
所以“买量到底有没有用”这个问题问错了。 应该问：我的用户主要通过“有意图的搜索”还是“无意图的浏览”找到我。前者会高估买量，后者会低估买量。
论文还给了两个边界：
论文里有一句话，比前面所有数字都更值得抄进笔记本：
因为排名是平台层决定的，局部的投放变动会外溢到全局——这违反了 SUTVA（个体处理稳定性假设）。于是，“给 10% 用户停投”这类标准增量测试，可能因为压掉了对照组本可以获得的榜单提升而系统性低估投放效果。
换句话说：在一个由平台公共榜单分发的环境里，你的对照组不是干净的对照组。你为了测量效果而按住的那 10%，本身就在压制整个产品被看见的机会。这不是执行得不好，是这套设计在这个环境里测不准。
这条推论同样出自论文作者，本卡只能转述。但它是可检验的。
论文给的是机制，你的产品有没有被这条机制支配，要自己判。要看的不是某一个数字的高低，而是三条曲线之间的相对运动。
假设你在 12 月做了一次投放冲刺，预算翻了三倍，事后复盘发现：付费安装涨了很多，自然安装也涨了 40%，榜单位置从第 80 名升到第 45 名。你会怎么解读？
如果只看“自然量涨了”，你会得出结论：平台在给我免费流量，多投很划算。 但这条推理里少了一个环节——自然量涨，也可能只是因为付费安装多了、总安装盘子大了，自然量按固定比例跟着涨，与榜单无关。
要分清，把三列排在一起看：
| 观察 | 说明什么 |
|---|---|
| 投放涨，排名明显上升，自然量跟着上升 | 链条生效。投放同时在买曝光位，削减预算的代价会延迟出现在自然量上 |
| 投放涨，排名基本不动，自然量按比例跟着涨 | 链条没生效。你买到的是付费安装本身，自然量的增长是盘子变大带来的，别把它读成“榜单红利” |
| 投放涨，排名上升，但自然量没动 | 榜单位置没有转化成自然量。可能是品类自然流量太小，也可能是榜单位次的变化幅度还不足以改变曝光 |
这三种情况对应三种完全不同的动作。 第一种：削减预算要按“付费 + 自然”的总账算，只算付费会算亏。第二种：投放量不是杠杆，素材与渠道结构才是，应该把精力放在提高单次安装的效率上。第三种：先别急着加预算，先去确认这个平台的榜单到底有没有自然流量。
这就是本卡标题里“广告真正买的是被看见”的落地方式：判断的不是买量有没有用，而是你的钱花在了链条的哪一环。 花在“付费安装”这一环，优化对象是单价与转化；花在“榜单位置”这一环，优化对象是投放的节奏与集中度——同样一笔钱，集中在一周内砸出去和摊平到三个月，对安装速度的影响完全不同，而榜单看的是速度。
当你的自然量占新增比例不低、且产品的发现入口是商店榜单或平台推荐位时，不要把“付费安装数”当作买量效果的全部，因为投放同时在买一个公共曝光位；此时砍预算的代价会以自然量的形式延迟出现。
反过来，当你已经删掉了买量的最后一环（不再争夺榜单位置，例如投放量太小、或只在私域里投放）时，这条链断开，此时按付费安装数评估买量基本准确。
迁移性上，本卡只站得住一半：链条的前半段（投放→安装速度→榜单）是平台机制，跨品类通用；后半段（榜单→自然量）依赖该平台的自然流量规模，不同品类差别很大。休闲和超休闲比中重度更依赖榜单发现，也更容易被这条机制左右。
- [Advertising Spillovers in Mobile Apps: Evidence from Ad Shutoffs and Store Rankings](https://arxiv.org/abs/2504.16151) —— 独立研究（学术论文），v1 提交于 2025-04-22，v2 修订于 2026-07-10。提供本卡全部数据：停投窗口、事件研究估计、面板模型系数、榜单中介检验
- [论文 HTML 全文 v2](https://arxiv.org/html/2504.16151v2) —— 同上，正文方法段与文献综述段（搜索广告替代率 37%–99.5% 的对照来源）
- [CEDEC2021「あの施策は売上にいくら貢献したのか？」傾向スコア分析による統計的因果推論の適用事例紹介](https://cedec.cesa.or.jp/2021/session/detail/s60648834255dc.html) —— 从业者会议演讲（WFS，日本），提供"手游里 A/B 测试常常无法实施"这一反面场景，用于本卡"什么时候不成立"一节证据等级 A，置信度中高。 理由：
当前推断链上最脆弱的一环：从“控制排名后相关性消失”推到“排名是中介变量”，这一步依赖模型设定正确。作者做了稳健性检验（对数排名、负对数、倒数、以及把榜外日编码为 401 名），结论不变；但中介分析本身不能排除“排名与自然量同时被第三个因素驱动”的可能——例如平台在同期调整了该品类的推荐权重。
需要什么数据才能确认：同一厂商在其他平台（例如不设公共榜单的渠道）的关停对照，或第三方复现同一实验设计。
**
One line: A US mobile studio paused all ad spend for 24 days, and organic installs dropped 20–30%. Those users were not delivered by the ads; the ads pushed the game up the store charts, and the charts delivered the users. Judging paid UA by attributed installs alone therefore understates it — and on store-chart-driven products, even a standard incrementality holdout can be biased.
The authors obtained almost two years of campaign data from a large US mobile game developer (called GameSpace in the paper): 500 days, 85 ad publishers, six games, iOS and Android, more than 100 billion impressions served. The data came through an ad-tech company that managed the spend. The paper states that calendar dates are withheld for privacy, so nobody outside can verify which game this was or when.
Within those 500 days the developer did something rare: a global shutoff across every app and platform, from day 75 to day 98 — 24 days. Not one channel; search, social, display and in-app all went dark.
Paid installs collapsed, as expected. Organic installs did not stay put. Event-study estimates put the drop at 20–30%, ranging from −0.305 to −0.359 across 7-, 11- and 23-day bandwidths. A fixed-effects panel model converts the effect into dollars: every $100 of spend was associated with 32.3 paid installs and 2.2 organic installs. Two different methods, same order of magnitude.
Meanwhile the six games' category chart positions deteriorated 2.3–3.3× (log-rank terms).
The important evidence is not that organic fell. It is where the fall showed up in the model.
The authors merged daily category-rank data and ran a mediation test in two steps:
Step 2 is the most informative sentence in the paper. If spend moves organic installs through ranking, then controlling for ranking cuts the chain at its downstream link, and the upstream correlation vanishes with it. The authors read this as ranking being a mediator, not merely a co-moving variable.
Written as a chain you can explain to a colleague:
More spend → paid installs rise (install velocity increases)
→ the store ranks apps with rising install velocity higher
→ chart position is a public slot visible to all users
→ browsing users install organically
→ organic installs push the rank higher stillThat last link closes a self-reinforcing loop, and it is where the finding stops being intuitive.
The paper's benchmark is a set of famous search-advertising shutoffs. When eBay paused branded search ads, most of the lost paid traffic simply shifted to organic. Paid search was largely taking credit for organic traffic that would have arrived anyway — an over-attribution problem. Replications at Edmunds and later work by Golden and Horton put paid-search substitution of organic traffic somewhere between 37% and 63%.
The same method, moved to mobile apps, reverses the sign. The difference is how users find you:
So "does paid UA work" is the wrong question. The right one is whether your users find you through intent-driven search or intent-free browsing. The first overstates paid UA; the second understates it.
Two boundaries the paper draws itself:
One passage in the paper deserves to be copied into a notebook more than any of the numbers:
Because ranking is decided at the platform level, a localized change in spend spills over globally, violating SUTVA. Standard incrementality tests — withholding spend from a holdout — may therefore suffer systematic downward bias by suppressing the counterfactual ranking lift that would otherwise drive organic installs.
In plain terms: in an environment distributed by a public platform chart, your holdout is not clean. The 10% you hold back to measure the effect is itself suppressing the whole product's chance of being seen. That is not sloppy execution; it is a design that cannot measure this environment correctly.
This inference is the authors' own. This card relays it. It is, however, testable.
The paper gives you the mechanism. Whether your product is governed by it is your call. The test is not the level of any single number but the relative movement of three series.
Suppose you ran a spend push in December, tripled the budget, and the retrospective shows paid installs way up, organic installs up 40%, and chart position moved from #80 to #45. How do you read it?
Look only at "organic went up" and the conclusion writes itself: the platform is giving me free traffic, spend more. But that reasoning skips a step. Organic could also have risen simply because paid installs grew the total install base, lifting organic proportionally — with the chart playing no part.
To separate the two, put all three series side by side:
| What you observe | What it means |
|---|---|
| Spend up, rank clearly improves, organic follows | The chain is live. Spend is buying an exposure slot, and cutting budget will show up later as lost organic |
| Spend up, rank barely moves, organic rises proportionally | The chain is not live. You are buying paid installs; the organic lift is the larger base, not a chart dividend |
| Spend up, rank improves, organic does not move | Chart position is not converting into organic installs — either the category has too little organic traffic, or the rank move is too small to change exposure |
Those three cases call for three different actions. Case one: evaluate budget cuts on paid plus organic, or you will mis-count the cost. Case two: volume is not the lever, creative and channel mix are — put effort into cost per install instead. Case three: do not add budget yet; first confirm the platform's chart actually carries organic traffic.
This is what the headline means by "ads really buy being seen." The question is not whether paid UA works, but which link of the chain your money lands on. Money landing on paid installs is optimized through unit cost and conversion. Money landing on chart position is optimized through pacing and concentration — the same budget dumped into one week and spread across three months produce entirely different install velocity, and the chart ranks velocity.
When organic installs are a meaningful share of your new users and discovery runs through store charts or platform recommendation slots, do not treat attributed paid installs as the whole return on paid UA — because the spend is simultaneously buying a public exposure slot, and cutting budget will show up later as lost organic installs.
Conversely, when the last link of the chain is already absent — spend too small to move a chart, or confined to private channels — the chain is broken, and evaluating paid UA by attributed installs is broadly accurate.
Transferability is only half-supportable. The first half of the chain (spend → install velocity → chart position) is a platform mechanism and travels across genres. The second half (chart → organic installs) depends on how much organic traffic the platform has, and varies a lot by genre. Casual and hypercasual titles lean on chart discovery far more than mid-core or hardcore ones, and are correspondingly more exposed to this loop.
Evidence tier A, confidence medium-high. Reasoning:
The weakest link in the current inference chain: moving from "the correlation disappears once rank is controlled" to "rank is the mediator" depends on correct model specification. The authors ran robustness checks (log rank, negative log, reciprocal, and coding off-chart days as rank 401) with unchanged conclusions — but mediation analysis cannot rule out a third factor driving both rank and organic installs, such as the platform adjusting category weights in the same period.
What data would settle it: a comparable shutdown on a channel without a public chart for the same developer, or independent replication of the same design.
Terminology follows references/glossary.md. This card is one file; there is no separate .en.md.