LearningLearning

把海外用户增长和 AI 工作流的实践,边做边写下来。Notes from running global user acquisition with an AI workflow — written as the work happens.

背景简单说:七年海外用户增长,腾讯、头部游戏大厂、字节跳动、头部游戏出海公司四段,App 与 PC/Web 两端都完整操盘过。这个站不是简历,是工作笔记:投放该怎么判断、AI 怎么长期接管重复劳动、哪些做法被证伪了——能公开的部分都会放在 Learning 里。Background in one line: seven years in global user acquisition across Tencent, a top-tier games company, ByteDance and a leading games publisher, on both App and PC/Web. This site is not a résumé — it's a working notebook: how to judge paid media, how an AI workflow takes over the repetitive parts for the long run, and which practices turned out wrong. Whatever can be shared goes under Learning.

$1B+
累计操盘媒体消耗Cumulative spend managed
7
年海外 UA,4 家公司Years in UA, 4 companies
3
次归因系统从 0 到 1Attribution systems built from zero
50+
个广告账户经 API 统一在管Ad accounts managed via API
01

背景Background

海外 Google、Meta、TikTok、Bing 与国内主流渠道都完整操盘过,App 和 PC/Web 两端都做。方法上,我把归因系统、投放链路、媒体采买当一个整体来做,而不是三件事。PC/Web 端投放是市面上较稀缺的能力——多数投放人才只做过移动端——也正好是 AI 产品的主要战场。I've run Google, Meta, TikTok, Bing and the major China channels end to end, on both App and PC/Web. My approach treats attribution, the conversion funnel and media buying as one system rather than three jobs. PC/Web acquisition is a scarce skill — most UA people are mobile-only — and it's where AI products live.

02

用户增长是什么,能带来什么What user acquisition is, and what it gets you

用户增长(UA,买量)就是付费在 Google、Meta、TikTok 等平台投广告,把目标用户带进来,变成注册和付费。它和做内容、等口碑的区别只有一条:今天花钱,明天就有数据。一旦算清「花 1 块能回来多少」,它就可以按可控的方式加码。User acquisition — paid growth — means paying Google, Meta or TikTok to put your product in front of the right people, and turning them into signups and paying users. Its one real difference from content and word of mouth: spend today, see data tomorrow. Once you know what a dollar in returns, you can scale it in a controlled way.

  1. 小预算测试Small-budget test用一笔封顶的预算投出去,亏损上限先锁死A hard-capped budget goes out; the downside is fixed first
  2. 把账算清Do the math花 1 块,回来多少注册、多少收入One dollar in — how many users and how much revenue out
  3. 判断 ROI 是否达标Is ROI on target回来的是否 ≥ 花出去的;达标放大,不达标停Do returns cover spend; scale if yes, stop if no

达标 → 放大ON TARGET → SCALE

  • 预算Budget:从每天几百加到几千,收入按比例跟: from hundreds a day to thousands, revenue follows proportionally
  • 渠道Channel:Google 跑通的打法复制到 Meta、Bing、TikTok: a playbook proven on Google replicates to Meta, Bing, TikTok
  • 市场Market:一个国家验证过,复制到更多国家: proven in one country, rolled out to more

同一套验证过的打法可以在预算、渠道、市场三个维度同时放大,这是买量区别于其他获客方式的地方。One proven playbook scales along budget, channel and market at once — that is what sets paid growth apart from other acquisition channels.

不达标 → 停OFF TARGET → STOP

  • 停止加钱Stop feeding it,亏损停在封顶预算内; losses stay under the cap
  • 找原因Find the cause:产品、定价,还是承接的页面和流程: the product, the pricing, or the funnel behind the ad
  • 修好再测Fix, then retest,再用一笔小预算验证 with another small budget

不达标就放大,等于放大亏损。判断什么时候停,比会加预算重要。Scaling an unproven setup scales the losses. Knowing when to stop matters more than knowing how to spend.

03

工作方式:媒体 API 加人的判断How I work: media APIs plus human judgment

付费投广告,老板通常关心三件事:钱花出去有没有赚回来;数据这么多信哪个。下面按这个顺序回答。If you pay for ads, three questions matter: is the money coming back; which numbers can be trusted. Answered in that order.

每天的循环The daily loop

五个环节,每天跑一遍。第 4 步是人;其余由程序完成。Five steps, run daily. Step 4 is a person; the rest is software.

  1. 收需求和素材Take the brief投放计划和素材自动读取,不来回传文件Plans and creatives read in automatically
  2. 程序建广告Ads built by code批量创建和修改,不在后台逐条点击Created and edited programmatically, not by clicking through a console
  3. 三方对账Three-way reconciliation平台报的 × 你后台真实发生的 × 用户进站后每一步,三边对齐才算数Platform claims × your backend × on-site behaviour; numbers count only when all three agree
  4. 人来决定A person decides加预算、减预算、停广告,花钱的决定全部由人做Raise, cut, stop — every spending decision is made by a person
  5. 执行并留痕Execute and log决定由程序执行,每一步有记录;报告自动送达Executed by code, every step logged; reports delivered automatically

典型的一天A typical day

daily-run · 昨天的数据,全渠道yesterday, all channels 广告后台(多账户)、你的业务后台、网站分析,三源拉取完成Ad platforms (all accounts), your backend, site analytics — three sources pulled analyze · 用写成规则的经验rules distilled from experience 每条广告:昨天花了多少 → 带回多少真实用户和收入Every ad: spend in → real users and revenue out 警报:某市场一条广告成本上涨 38%Alert: one ad's cost up 38% in one market → 对照你后台:注册量正常 → 平台数据延迟,不动→ Checked against your backend: signups normal → platform lag, no action 发现:另一市场一条广告连续 7 天盈利且稳定Found: one ad profitable and stable for 7 days propose · 等人确认awaiting approval 建议:该条预算 +20%Proposal: +20% budget on that ad [HUMAN] 已确认approved execute · 全程留痕fully logged 预算已生效;操作写入记录;报告已送达Budget live; action written to the log; report delivered

示意数据。流程、口径与规则为实际生产逻辑。Illustrative numbers. The pipeline, metrics and rules are the real production logic.

六条判断规则Six rules of judgment

分界原则:AI 不独立做花钱的决定;人不手工拉数据。The dividing line: AI never makes a spending decision alone; people never pull data by hand.

这套方式每天怎么保证 AI 记得在管什么、改过什么、上次改动有没有用,见 运行机制How the AI is kept aware of what's live, what changed and whether it worked: see Mechanics.