何时打开:你要出海 / 做海外推广 / 想搭多账号矩阵 / 在出海基础设施上踩坑。本 wiki 把 3 派视角融合(王凯主导 / 老黄独立产品 / 老 D 国内适配)。
一句话核心:海外启动慢 10 倍但 PMF 后订阅率高 3-5 倍 + 单用户 ARPU 高 2-3 倍(王凯);海外平台养号 = 信用积累 8-12 周,没有捷径;多账号矩阵化 = 单条成本压到趋近 0 → 用产能换概率。
0. 3 派的"出海"
| 视角 | 来源 | 模式 | 一句话立场 |
|---|---|---|---|
| 矩阵化出海派 | 王凯 出海推广 SOP 与海外多账号矩阵实战 / 王凯 6 平台海外运营 SOP / 王凯多 Agent 浏览器矩阵架构(v1.1 实操指南) | 6 平台 SOP + 多浏览器矩阵 + 4 重隔离 | "国内卷,海外是必须要去的高地" |
| 独立产品出海派 | newtype · Vibe Coding 独立产品实战 / newtype AI 商业、行业与投资观察 | newtype.pro + Substack + 北美用户群 | "Pro User 一定多工具 + 2026 三件事(银行卡+美股+AI 全访问)" |
| 国内适配派 | dontbesilent AI 自媒体课程 / 13 节系统化方法论 | rap.dev + 小红书 MCP + 国内不科学上网 | "所有 SOP 在不科学上网情况下也跑得通" |
1. 跨作者共识(5 条)
1.1 推荐流量 > SEO 流量(创新型产品冷启动)
王凯王凯 出海推广 SOP 与海外多账号矩阵实战 §1 + §7:新产品 90% 精力做推荐流量,不做 SEO。
| 类型 | 流量策略 |
|---|---|
| 创新型产品(如 AiPathly) | 必须先做推荐流量教育用户 |
| 已知品类(如电商工具) | SEO 直接竞争 + 投流 |
Why AI 编程 / AI 求职报告不能直接做 SEO?
- 搜索是滞后指标
- 用户不知道"AI 求职报告"这个东西
- 必须先通过 TikTok / KOL 推广教育市场
- SEO 是防御性策略,防止他人收割
1.2 共通 AI 写作铁律(王凯独家 7 条自检)
王凯 6 平台海外运营 SOP §2:
核心规则:"Claude 只做翻译,不做润色。要求 Claude 不发挥,按你的意思口语化翻译,越短越好、越正式越死。"
翻译提示词模板:
你是我的 Reddit 翻译助手。规则:
1. 我给你中文,你直接翻译成英文,不要发挥、不要润色、不要补内容
2. 用 Reddit 用户日常口语,避免任何正式书面语
3. 缩写优先(gonna, wanna, kinda, tbh, imo, ymmv, ngl)
4. 句子要短,平均 < 12 个英文单词
5. 不要 "In conclusion" "Furthermore" 这类衔接词
6. 不要 emdash(—),用普通逗号或句号
7. 第一人称强烈,多用 I, my, me
8. 偶尔可以打错字、漏标点,更像真人
7 条自检清单:
- 句子平均长度 < 12 词?
- 没有 emdash?
- 没有 "In conclusion / Furthermore"?
- 没有完美语法?
- 第一人称 > 3 次?
- 段落 < 4 行?
- 像 Slack 消息不像论文?
1.3 共通基础设施 4 件套(王凯独家)
1. 海外手机号(英国卡 45 元 + 10 欧元激活,2026-05 推荐)
2. 海外银行卡(Wise / Payoneer / PayPal)
3. 多套账号(Google / Apple / Claude / Stripe)
4. 住宅代理 + 浏览器指纹(矩阵必备)
老黄印证(2026 三件事):海外银行卡 + 美股 + AI 全访问(见 newtype AI 商业、行业与投资观察 §8)。
1.4 矩阵化是新范式(共识 + 规模差 50 倍)
| 派 | 矩阵规模 |
|---|---|
| 王凯 | 50 频道 / 每天 500 条 Shorts(10 台电脑 × 5 浏览器 × 1 Google + 1 频道) |
| 老黄 | 4-5 平台精耕(newtype.pro + YouTube + Substack + X) |
| 老 D | 1-2 个核心平台(抖音 + 小红书) |
1.5 Agent 自主注册账号(2026 跑通)
王凯王凯 出海推广 SOP 与海外多账号矩阵实战 §3:
- AgentMail(YC S25)Agent 专用邮箱
- Chrome MCP 控制真实浏览器
- reCAPTCHA 通过(checkbox 点击就过)
- HN 新账号限制(发 Show HN 拒,只能评论养 karma)
结论:平台一视同仁,要遵守规则养号。
2. 显式分歧(3 条 —— 不抹平)
2.1 分歧 1:出海 vs 留国内?
| 视角 | 立场 |
|---|---|
| 王凯(强出海派) | "国内卷,海外是必须要去的高地" —— PMF 后订阅率 3-5 倍 / ARPU 2-3 倍 / 但启动成本 5 倍 |
| 老黄(海外独立产品) | 北美用户群 —— Stripe / 报税合规 / Substack 订阅经济 |
| 老 D(国内派) | "所有 SOP 在不科学上网情况下也跑得通" —— DeepSeek + Monica + rap.dev + 抖音 200 万播放 |
真分歧:身份决定路径(详见 AI 自媒体与内容工程化(跨作者主题综合) §2.1)。
2.2 分歧 2:矩阵概率赌 vs 精耕账号资产?
| 视角 | 立场 | 数量 |
|---|---|---|
| 王凯(矩阵主义) | 多 Agent 多浏览器多账号 | 5+ 浏览器 / 50 频道 |
| 老黄(精耕) | 跨平台分发但每个平台精耕 | 4-5 个平台 |
| 老 D(单平台) | 单平台精耕 1 个号 | 1-2 个 |
真分歧:
- 王凯做"产能 × 概率"(50 频道总会出 1-2 个爆款)
- 老黄做"账号资产 × 复利"(1 年积累 = 不可复制资产)
- 老 D 做"个人 IP 信任飞轮"
2.3 分歧 3:海外投流值不值?
王凯王凯 出海推广 SOP 与海外多账号矩阵实战 §9.2:踩坑教训:
- 海外 Google Ads / Facebook Ads / TikTok Ads
- 王凯团队试过 200 美金海外 Chrome 插件投流 → 0 个用户
- "小微 team 没有海外投流能力"
对策:走"流量验证"链路:
- 先制作广告素材
- 投给"不存在"的产品页面(占位)
- 收集获客成本
- 决定是否做产品
3. 王凯 6 平台 SOP 速查(独家)
| 平台 | 冷启动 | 核心策略 | 红线 |
|---|---|---|---|
| 8 周 | niche subreddits / 1:9 黄金比例 / 4 类打法(调研/素材/产品理念/舆情) | shadowban / 同 IP 多号 / emoji 大写组合 | |
| HackerNews | 8 周养号 | 新号纯评论 / 新号不能发 Show HN / 8 周后才能 / 4 种高赞范式 | 求朋友投票 / 频繁刷新 / 同产品 Re-Show |
| Twitter / X | 90 天真人号 | 50% 项目 / 30% 行业 / 20% 个人 / 回 50-200 转发未爆推 | 同 IP 必死 / 7 天 shadow review / 跨号互动协同检测 |
| YouTube Shorts | 90 天单频道 | 5 主题 × 5 语种 × 2 风格 = 50 频道矩阵 / 前 3 秒决定一切 | 完播率 / 矩阵反作弊靠指纹 |
| ProductHunt | 反冲榜 12 小时 | 一次性推广 / 真实用户运营 | 国内拉票不计数 |
| Dev.to | cross-post 矩阵 | Long form articles + SEO 友好 | 与 Medium / Substack 配合 |
3.1 HN Show HN 模板
Title: Show HN: [Product] – [one-line value prop, max 80 chars]
I'm [your role]. Built [product] because [specific personal pain].
What it does:
- [specific capability 1]
- [specific capability 2]
- [specific capability 3]
Tech stack: [key tech]. Architecture decisions worth noting:
[1-2 design tradeoffs that other devs will appreciate]
Pricing: [be transparent — free tier / paid / open source]
Honest about limitations: [admit 1-2 known gaps]
Would love feedback on [specific question].
Source code: [github link if applicable]
Demo: [link]
3.2 Twitter 招聘推模板(王凯实测)
Hiring: [Role] for [project]
[1 sentence about what we're building]
What we need:
- [skill 1]
- [skill 2]
- [working style: remote/async/etc]
What you get:
- [comp range — be specific]
- [equity if applicable]
- [interesting tech challenge]
Reply or DM. Retweets appreciated.
前提:招聘推有效 = 真人号 1 年以上声誉 × 关注者中目标人群浓度。
3.3 YouTube 50 频道矩阵差异化
5 主题 × 5 语种 × 2 风格 = 50 频道:
| 维度 | 示例 |
|---|---|
| 主题 | AI 工具评测 / 行情解读 / 历史故事 / 编程教程 / 心理学 |
| 语种 | 英语 / 西班牙语 / 葡萄牙语 / 印地语 / 阿拉伯语 |
| 风格 | 严肃讲解 / 搞笑反转 / 数据可视化 / 真人出镜 |
每频道独立:频道名 + logo + banner + 简介 + 配音音色 + 缩略图风格 + 发布时段。
3.4 Hook 模板(前 3 秒决定一切)
- 反差型:
"Everyone says [common belief]. They're wrong. Here's why." - 数据型:
"This single change made [metric] go [direction] [number]." - 问题型:
"Why does [common phenomenon] actually happen?" - 演示型:
"Watch what happens when [action]."
4. 王凯多 Agent 浏览器矩阵架构(独家工程化)
王凯多 Agent 浏览器矩阵架构(v1.1 实操指南) v1.1 完整实操指南:
4.1 核心定义
身份 (Identity) = 一个浏览器 app + 一套登录态 + 一个住宅 IP + 一个 cdp-proxy 进程
↓
工种 (Role) = 一个 Claude Code agent(独立工作目录),跟同身份其他 agent 共享浏览器实例
4.2 端口对分配(按身份维度 v1.1)
| 身份 | 浏览器 app | proxy 端口 | chrome 端口 | 住宅代理出口 |
|---|---|---|---|---|
| identity-A | Chrome | 9401 | 9222 | US-East |
| identity-B | Arc | 9402 | 9223 | US-West |
| identity-C | Brave | 9403 | 9224 | UK |
| identity-D | Edge | 9404 | 9225 | DE |
| identity-E | Vivaldi | 9405 | 9226 | JP |
4.3 4 重隔离
- 浏览器隔离(不同 app)
- 账号隔离(独立 Google / 独立用户数据目录)
- IP 隔离(住宅代理出口)
- 指纹隔离(Multilogin / GoLogin / AdsPower)
4.4 浏览器指纹工具选型
| 工具 | 价格 | 何时用 |
|---|---|---|
| GoLogin | 免费 3 profile / $24/mo | 起步 / ≤ 3 身份 |
| AdsPower | 免费 5 profile / $5.4/mo | 国内出海团队 |
| Multilogin | $99/mo 起 | TikTok / X 严打期 / ≥ 5 身份 |
| BitBrowser | (未给价) | 备选 |
4.5 住宅代理选型
| 提供商 | 价格 | 备注 |
|---|---|---|
| Bright Data | $8.4/GB | 老牌,IP 池最大 |
| Soax | $7/GB | 性价比 ok |
| IPRoyal | $7/GB | 起步价低 |
| Smartproxy | $7/GB | 入门友好 |
机房代理(Datacenter)绝对禁用 —— X / TikTok / Reddit 在机房 IP 上注册即封。
4.6 地理一致性铁律
- identity-A = 美国东海岸用户 → US-East 住宅 IP → 时区
America/New_York→ 平台账号都用 EST 时间发布 - 地理矛盾是风控最容易抓的点
4.7 CDP 安全(必读)
- 默认监听 127.0.0.1
- ❌ 绝不改成
0.0.0.0 - ❌ 绝不
docker -p 9222:9222端口映射 - ❌ 矩阵浏览器禁止登录私人重要账号(私人 Gmail / 工作邮箱 / 银行)
5. 出海基础设施清单(王凯独家)
graph TD
A[出海基础设施] --> B[支付层]
A --> C[账号层]
A --> D[流量层]
A --> E[审核层]
B --> B1[Stripe / Lemon Squeezy / Gumroad]
B --> B2[Wise / Payoneer / PayPal]
B --> B3[Apple / Google Play 海外账号]
C --> C1[英国卡 45 元 + 10 欧元激活]
C --> C2[AgentMail Agent 专用邮箱]
C --> C3[多套 Google / Apple / Claude 账号]
D --> D1[Reddit / Twitter / YouTube]
D --> D2[HackerNews / ProductHunt]
D --> D3[Dev.to / Medium / Substack]
E --> E1[Google OAuth 安全审核]
E --> E2[Apple Developer 99 美金/年]
E --> E3[Stripe Atlas 公司 1 千美金起]
5.1 出海隐形成本
| 维度 | 踩坑 |
|---|---|
| 数据安全审核 | 接 Google Workspace API → 类似车检 / Quanta Quest 卡在这 |
| 投流能力门槛 | 200 美金海外 Chrome 插件投流 → 0 用户 |
| 信用体系 | HN / Reddit 都要养号 / 至少 3-6 个月 |
| 支付门槛 | Stripe Atlas 1 千美金 / Apple Developer 99 美金 / Wise 往 ZA Bank 拦截 |
5.2 海外推广避坑清单
| 避坑 | 原因 |
|---|---|
| ❌ 直接软广发 Reddit / HN | 一律拒收 |
| ❌ 中文翻译直接发海外 | 文化差异严重 |
| ❌ 国内拉票 ProductHunt | 不计数,白浪费 |
| ❌ 海外投流没本地化素材 | ROI 永远跑不正 |
| ❌ 单一账号大规模 | 平台一封全损失 |
| ❌ 用 Google Voice / Skype 数字账号 | 被识别封号 |
| ❌ 用大陆卡走 ZA Bank 汇款 | 拦截率高 |
6. 老黄独立产品出海(独家)
6.1 个体商业经历
- 美国公司
- Apple Developer 99 美金/年
- Substack 200 订阅
- newtype.pro 月度 Newsletter
- Pro User 多工具(订阅 Claude / ChatGPT / Cursor / Gemini)
6.2 2026 三件事
newtype AI 商业、行业与投资观察 §8(老黄给的"2026 该做什么"):
- 海外银行卡
- 美股
- AI 全访问(订阅一线模型)
6.3 别和巨头比"输入框"
"像 ChatWise 这类 Chatbot,如果不能找到一条细缝让自己生存下去,那最终都会被大厂碾过去。这是 AI 创业最忌讳的事。"
老黄做 Prompt House 不是"输入框",而是"Prompt 管理 + MCP 协议调度"(原子化中间层)。
7. 老 D 国内一键发布(独家)
dontbesilent AI 自媒体课程 / 13 节系统化方法论 §L11:
7.1 rap.dev + 小红书 MCP 工作流
1. 在 RAP 输入:"帮我把这条推文发布到小红书"+ 链接
2. RAP 调用:
a. Playwright 操控无头浏览器截图(确保 3:4 比例)
b. 字数多 → 左右填白边(Python Pillow)
c. 字数少 → 拉远拍 → 裁剪
3. Claude 4.5 Sonnet 给小红书写标题(自动 ≤ 20 字 / 从原文找金句)
4. 小红书 MCP server 自动登录发布
5. 还可以定时发布("1 小时后发")
7.2 国内 vs 海外节奏对比
| 平台 | 国内特征 | 海外特征 |
|---|---|---|
| 节奏 | 节奏更快但内容质量门槛更高 | 信用积累 8-12 周,没有捷径 |
| 反作弊 | 更宽容(老 D 没强调养号) | 极严(同 IP 必死) |
| 付费意愿 | 持续下降 | 持续高 |
| 审核 | 内容审核(政策敏感) | 数据安全 + 隐私合规 |
8. 怎么选 —— 按身份的出海路径
| 你是谁 | 主用方法 | 起点 |
|---|---|---|
| 小微 AI team 出海 | 王凯矩阵化 | 6 平台 SOP + 多浏览器矩阵 + 出海基础设施 4 件套 |
| 1 人非技术独立产品 | 老黄独立产品 | newtype.pro + Substack + Pro User + 2026 三件事 |
| 国内不科学上网素人 | 老 D 国内适配 | rap.dev + DeepSeek + 小红书 MCP + 抖音 |
| 想做 50 频道矩阵 | 王凯多浏览器矩阵 | 4 重隔离 + 指纹工具 + 住宅代理 + CDP 安全 |
| 关注海外内容工厂 | 王凯 YouTube 50 频道 | 5 主题 × 5 语种 × 2 风格 + Hook 模板 |
通用纪律:
- 新产品 90% 精力做推荐流量,SEO 是防御
- 海外平台养号 = 信用积累 8-12 周,没有捷径
- Claude 只做翻译不做润色(海外)/ 保留个人风格不消灭 AI 味(国内)
- 同一份稿子绝对不能两个平台都发
- 小微 team 没有海外投流能力,走流量验证链路
- 机房代理绝对禁用,必须住宅代理 + 地理一致性
- CDP 端口绝不暴露 0.0.0.0
9. 引用清单
主要原始 author wiki:
- 王凯 出海推广 SOP 与海外多账号矩阵实战 —— 226 帖出海实战 / 出海基础设施清单 / 创新型 vs 已知品类 / 隐形成本 / 避坑清单
- 王凯 6 平台海外运营 SOP —— 6 平台 SOP 全 / AI 写作 7 条自检 / HN Show HN 模板 / Twitter 招聘模板 / YouTube 50 频道
- 王凯多 Agent 浏览器矩阵架构(v1.1 实操指南) —— v1.1 多浏览器矩阵 / 4 重隔离 / 端口对 / 指纹工具 / 住宅代理 / CDP 安全
- 王凯 AI 时代创业方法论与商业判断框架 —— "出海要本地化" / 多项目并行
- newtype · Vibe Coding 独立产品实战 —— 老黄独立产品 / 北美用户群 / Pro User
- newtype AI 商业、行业与投资观察 §8 —— 2026 三件事(海外银行卡 + 美股 + AI 全访问)
- dontbesilent AI 自媒体课程 / 13 节系统化方法论 §L11 —— rap.dev + 小红书 MCP / 国内适配
相关主题 wiki(本系列):
- AI 自媒体与内容工程化(跨作者主题综合) —— AI 自媒体与内容工程化(Batch 1)
- Claude Code 工作流(跨作者主题综合) —— Claude Code 工作流(Batch 1)
- 创业方法论(从 0 到 1)(跨作者主题综合) —— 创业方法论(Batch 1)
History
- 2026-05-18:Topic-dimension wiki 首发。从 7 份 author wiki 综合改写。显式标注 3 条分歧:出海 vs 留国内 / 矩阵概率赌 vs 精耕 / 海外投流值不值
来源与关联资料
2026-10-18 增补：从渠道矩阵转向“人群、价值与反馈”驱动的增长
本节综合42章经不同年份采访，补充前文以渠道与产能为主的视角。前文的海外付费倍数、平台养号时长、账号与IP判断、固定渠道占比均保留为原作者当时经验，不是本次重新核验的平台规则或普遍增长定律。 下列收入、留存、付费和ROI也只在各自案例、时点与口径中理解；不以一个市场的自报数字替另一个产品作预算。本文是跨期知识提炼，不构成自动发帖、招募联系或投放授权。
一、每阶段先找一个主要约束，再选渠道与组织
增长不是从“该开几个账号”开始，而是先问：目标用户还不知道产品、不相信产品、没体验到价值、不愿付费，还是体验过后不再回来？这些问题对应不同动作。没有价值交付时扩大曝光，可能只增加失败体验；已有稳定付费但无人知晓时继续堆功能，也未必改变结果。OpusClip前增长负责人谢君陶在2025年采访中将冷启动、转化、留存和洞察分别拆开：早期找对人，已有注册流量后学习付费点，存量扩大后补反馈闭环，数据基础设施按复杂度投入。其意义是定位约束，不是要求所有公司同时建立完整增长部门。节目原文
ACE Studio在2025年的HF0经历提供另一种组织方式：用有期限的单目标冲刺，把周报从“做了多少功能”转为“哪项假设带来结果”。但它入营前已有约10万美元月收入，属于已有价值基础的团队；驻留环境、定价、渠道、归因和回流同时变化，不能认为孵化器能把零起点团队普遍放大十倍。可借鉴的是固定周期内减少注意力分散，每个试验仍保留退款、质量、留存和成本护栏；不能复制长工时和高压作为必需条件。节目原文
高途2024年的内部实践说明优先级还取决于控制范围与结果可见性：节约辅导教师时间未必兑现为同等现金，公域内容获客又受平台和竞争影响，自有线索中的响应、答疑和转化更容易跟到结果。零线索新产品不能照抄“转化优先”，已有业务也不应拿AI覆盖率代替成交。阶段只选一件主要事，是为了能判断原因，而不是把其他责任取消。节目原文
二、渠道×人群×内容×信任，是一个联合假设
同一功能投给不同人，可能不是同一门生意。Answer.AI在2024年解释，面向高中生选择TikTok，不是平台天然便宜，而是目标人群与内容习惯匹配；代理熟悉买量操作，不必然能做适合具体产品的素材。印度用户不理解放大镜图标的经历则说明，界面习惯也会阻碍漏斗，低转化不能直接归结为没有需求。应将语言、教育制度、设备与生活场景纳入研究，而不是把中文话术翻译后视为本地化完成。节目原文
ACE Studio在2024年跨期访谈里，起初联系百余YouTuber少有人回应，后来在NAMM展会让音乐人真实试用、建立关系，才获得推广回应。它的专业歌声产品需要用户相信能进入工作流，现场演示、行业身份与口碑改变了触达条件。团队口述海外新增付费率更高、美元收入占比提高，样本同时换了人群、渠道、触达方式和信任条件，因此不能抽象成“海外用户天然更愿付费”。节目中的ROI=0或100没有定义、费用范围与归因窗，保留为感受性描述，不用作已审计收益倍数。节目原文
Opus早期把Podcasters作为清晰人群，在其围绕KOC学习剪辑的网络中找长期伙伴。关键是伙伴本来就在用、愿意共建与示范，经济安排则让关系可持续；不是让任何粉丝多的人按佣金发一次广告。少量真实伙伴能同时产生产品反馈和可信传播。它与规模化KOL漏斗不是互相否定：前者帮助确定价值与语言，后者在机制较清晰后降低重复协作成本。扩量前需区分自然使用者、付费合作、分销者和普通消费者，避免把它们混成一个“达人用户”标签。
可执行的渠道试验应写成一整句：“在某个地区/语言/人群中，通过某种触点，展示某项真实交付，让用户相信并完成某个动作。”若只填平台名与CPM，缺少人群、内容、信任，测试结果很难解释，更不能无差别扩到多账号矩阵。
三、人工招募买到的应是理解与有效供给，不只是样本数
FlowGPT在2023年11月录制、2024年发布的采访中，自述前两周私聊五六百人、累计几千人。其Discord、Reddit、Twitter招募针对愿意制作和消费作品的社区成员，不只是购买一次访问。对UGC平台，用户自己比团队更懂玩法，早期人工交流还在建立供给。人数是嘉宾回顾而非可复现门槛；小团队应更关心触达后谁真的发布、作品是否有人消费、是否继续更新。节目原文
Mercor在2025年从广义招聘愿景切到AI Labs急需的领域专家，因为付款方有明确预算与交付压力。专业论坛、线下活动、主动招募和推荐共同补供给；推荐占注册的重要部分，不宜引用成“超过一半”，因为该比例由主持人提问而未被精确确认。更重要的是注册后还有身份、履历、技能、工时、质量、付款和争议处理。以sign-up计增长，会漏掉平台真正需要的“有效专家完成合格工作”。长期信任来自履约，不只是找到联系方式。节目原文
用户研究也必须验证招募来源。Answer.AI曾收到第三方声称的高活跃用户，查UID却发现是刚注册，前几天访谈和需求理解因此失去依据。样本真实性先于访谈技巧：确认使用时间、核心动作、付费状态、地区与语言，保留筛选条件，并记录未回复和流失用户。不能只访谈愿意赞美产品的社区活跃者，再将结论代表全体。
Monica在2023年收购ChatGPT for Google后，没有强迫其用户直接升级，而是用部分初始样本帮助验证，同时建设自己的获客能力。问卷中“产品消失会非常遗憾”的比例自述由35%升到55%，但样本量、选择规则和离开者未披露，不能把40%经验线视为统一PMF认证。收购来的用户、采访愿意说的话、行为留存与真实付费，是不同证据。节目原文
四、免费留存、真实付费、持续毛利，必须分别验收
“团队喜欢自己的原型→有人免费用并回来→有人愿意持续付费”是三个假设。Answer.AI认为商业化必须早验证，推理成本也使纯免费扩量更难长期维持。但不同品类不应用同一个次日留存或付费率：AiPPT的低频非专业用户可能只在需求时回来；ACE面对专业制作工作流；捏Ta的持续需求可能是养角色而非创作漫画。低频不自动等于失败，高频也不自动等于支付价值。节目原文 节目原文
FlowGPT还提醒，娱乐消费要把单作品/角色留存、品类留存、平台留存分开。用户换了一个Bot但仍每周回来找内容，不应直接算平台完全失去价值；反过来，平台活跃不证明某段角色关系稳固。UGC平台还需看创作者留存和复创，防止消费侧增长掩盖供给萎缩。
Opus的付费点实验强调先定位用户体验到什么价值：字幕模板被更明显展示后，转化意外改善，说明用户可能比团队预想更重视这项结果。Arcade公开产品演示时的付费拦截是嘉宾自己的购买体验，不是对所有用户最优的证明。定价、权益、提示时机与文案要一并记录，保护已购用户；每次变化同时看首次付费、后续使用、取消、退款和反馈。AI成本核算还要包括免费试用消耗，不能只用付费用户的毛利。
ACE在2025年称一次提价让日收入很快翻倍，随后测试不同价格、试用与月卡。这是一个短期现金杠杆，不证明长期续费不受损。月收入80万美元简单年化约960万美元，也不自动成为全年实收或可续订ARR。Mercor的1亿美元run rate同样不是已累计收入。增长报告必须把现金实收、收入、GMV、run rate、ARR和利润写清，不让标题替代定义。
五、增长质量还包括生态外溢与用户承接能力
于北川回顾2018—2020年抖音时，强调早期年轻人群、黑色UI调性、精选供给和不过早导入泛流量共同建立定位；随后引入新垂类、给探索曝光，让推荐有机会学习，再吸引新人群。这是内部经验解释，缺少隔离各因素的实验，但机制提醒我们：拉新人不只是获取一个ID，还要有内容、进入动机与可学习的反馈承接。节目原文
分发容器会同时改变观看效率和作者回报。单列、双列不能只比一次观看时长，还要看普通作者的首个反馈、复创、供给多样性和用户自主选择。作者同时服务实验内外消费者，普通用户级A/B会有网络外溢，短期消费优势未必代表长期生态更好；需要明确分群、长观察窗口与潜在污染，而不是因为难测就免检。
同一历史还有代理指标失真：增加好友与互动的OKR可被红包拉高，却没有自然关系场景；临近春节活动改目标，可能既未建立社交也未做好拉新。这些例子提示，关注、分享、消息、生成消耗都只是过程，必须回到用户为何回来。捏Ta的体力不足投诉可能表明需求增强，但还需验证真实购买、复购与供给影响，不能把抱怨或消耗直接当商业化成功。
内容传播也可能伤害承接：低价流量让注册暴增，但用户不匹配、客服跟不上、作者得不到反馈，会使短期CAC好看而后续价值变差。增长质量应包含错误人群比例、任务失败、投诉、退款、反感与供给流失；不能只用“有曝光”解释投入合理。
六、把实验结果接回产品和经营，才算完成增长闭环
增长闭环至少有五步：提出当前约束→写明机制与指标→有限预算试验→看结果和护栏→决定保留、修订或停止。Mercor描述的一周快决策，前提是先说清未来一周应该看到什么、没看到怎么办；速度不是略过预期、责任和失败方案。复杂或不可逆行动需要更高门槛，不能照搬可逆试验的节奏。
Opus将客服问题连接到产品优先级、修复、客服回访：社区愿望、付费故障与路线图可以分渠道收集，但最后必须有人负责闭环。只自动分类或发了邮件不等于修好了。续费还需把主动取消与Past Due支付失败分开；其重试成功率20%—30%到40%—50%是未给口径的经验，不是增长预算。实际应看可重试对象、追回金额、授权、重复扣款与投诉，尊重用户已经取消的选择。
高途节目提供重要口径纠正：官方章节称成单率三四倍，正文具体谈的是每条线索GMV的“单效”三四倍，另说转化率约提高一倍，且只是小范围试验。经营回读应优先具体定义，保留来源、实验比例、基线和窗口；不能将章节标题扩成全公司营收翻倍。覆盖更多线索、回复更多消息，仍只是过程数。
七、迁移到产品团队的实践：先做一张能回读的增长实验卡（编辑推论）
对AI陪伴/角色与创作产品，可先固定一个地区和语言、一种目标用户、一段核心体验、一条渠道，不在同一周同时改价格、入口、内容和推荐。卡片至少记录：目标人群及真实性、现有替代方案、首次价值动作、渠道与素材版本、成本包含项、实验分母与时间窗、回访及付费、退款和质量护栏、作者反馈，以及停止/扩量条件。定性原话要回到对应样本，行为结果要回到事件和付费流水，避免把问卷、演示、后台保存、上线与用户真实体验混为一层。
创作者增长先看“被招来→完成作品→首个有效消费反馈→继续创作”；消费者增长看“看卡→开始互动→得到预期体验→再次回来”；商业化再看真实购买、续费和毛利。角色级、品类级和平台级分别报告，关系体验受损不能由短期付费抵消。渠道跑出好结果后，再逐步增加人群与规模，每次检查原有信任和供给机制是否仍成立。矩阵是已验证机制的放大器，不能代替机制本身。
English version
2026-10-18 Addition: Growth Driven by Audiences, Value, and Feedback, Beyond a Channel Matrix
This section synthesizes 42 Zhangjing interviews from different years, adding to the preceding discussion's emphasis on channels and production capacity. The earlier claims about overseas payment multiples, time spent establishing platform accounts, account and IP judgments, and fixed channel shares remain the original author's experience at the time. They are not platform rules reverified for this update or universal laws of growth. The revenue, retention, payment, and ROI figures below must likewise be understood within their respective cases, dates, and definitions. One market's self-reported figures should not become another product's budget assumptions. This is a synthesis across episodes; it does not authorize automated posting, recruitment outreach, or advertising expenditure.
1. Identify One Main Constraint at Each Stage, Then Choose Channels and Organization
Growth starts with a question, rather than with how many accounts to open: Do target users not know the product, not trust it, fail to experience its value, decline to pay, or stop returning after trying it? Each problem calls for a different action. Expanding exposure before delivering value may simply create more failed experiences. Adding features to a product that already has stable paying users but little awareness may also leave the outcome unchanged. In a 2025 interview, former OpusClip growth lead Xie Juntao (谢君陶) separated cold start, conversion, retention, and insight: find the right people early; learn what drives payment once registration traffic exists; build feedback loops as the installed base grows; and invest in data infrastructure according to complexity. The point is to locate the constraint, not to require every company to build a complete growth department at once. Original episode
ACE Studio's 2025 experience at HF0 offers another organizational approach: use a time-limited sprint around a single goal, shifting weekly reports from “how many features we built” to “which assumption produced a result.” But the team already had approximately US$100,000 in monthly revenue before entering the program, so it had an existing foundation of value. Its working environment, pricing, channels, attribution, and returning traffic changed together. This does not show that an accelerator can generally multiply a team starting from zero by ten. The transferable practice is reducing divided attention over a fixed period, while retaining refund, quality, retention, and cost guardrails for every experiment. Long hours and high pressure should not be copied as necessary conditions. Original episode
Gaotu's internal work in 2024 shows that priorities also depend on what a team can control and whether outcomes are observable. Saving tutors' time does not necessarily turn into equivalent cash savings; acquiring customers through public-platform content is affected by platforms and competition; and responses, answers, and conversion within an existing lead pool are easier to follow through to outcomes. A new product with no leads cannot simply copy a “conversion first” strategy, while an established business should not substitute AI coverage for completed sales. Choosing one main priority at a stage helps identify causes; it does not remove other responsibilities. Original episode
2. Channel × Audience × Content × Trust Is a Joint Hypothesis
The same feature offered to different people may amount to a different business. Answer.AI explained in 2024 that it chose TikTok for high school students because the audience and content habits matched, not because the platform was inherently cheap. An agency's familiarity with paid acquisition does not necessarily mean it can produce creative suited to a particular product. Its experience with users in India who did not understand a magnifying-glass icon also shows how interface conventions can obstruct the funnel. Low conversion cannot immediately be attributed to a lack of demand. Research should include language, education systems, devices, and everyday contexts, rather than treating translated Chinese copy as completed localization. Original episode
In a 2024 interview included in this synthesis across episodes, ACE Studio described initially contacting more than a hundred YouTubers with little response. Responses to promotion followed only after it let musicians genuinely try the product and built relationships at the NAMM trade show. Its professional singing-voice product required users to believe it could fit into their workflow; live demonstrations, industry standing, and word of mouth changed the conditions for reaching them. The team reported a higher rate of new users becoming paying users overseas and an increased share of revenue in US dollars. But the sample changed in audience, channel, contact method, and trust conditions simultaneously. It cannot therefore be generalized into “overseas users are inherently more willing to pay.” The episode's ROI figures of 0 or 100 have no stated definition, cost scope, or attribution window. They are retained as impressions, not as audited return multiples. Original episode
Early on, Opus identified podcasters as a specific audience and found long-term partners in the networks where those podcasters learned editing from key opinion consumers (KOCs). What mattered was that the partners already used the product and wanted to help build and demonstrate it; financial arrangements made those relationships sustainable. This was not simply a one-off commission-based advertisement from anyone with a large following. A few genuine partners can produce both product feedback and credible distribution. This approach and a scaled key opinion leader (KOL) funnel are compatible: the former helps establish value and language, while the latter reduces the cost of repeated collaboration once the mechanism is clearer. Before scaling, distinguish organic users, paid partners, distributors, and ordinary consumers instead of combining them into one “influencer user” label.
An actionable channel experiment should form a complete sentence: “For a particular region, language, and audience, use a particular touchpoint to demonstrate a real deliverable, so users trust it and complete a particular action.” A test specified only by platform name and CPM lacks audience, content, and trust. Its results will be difficult to interpret, and cannot justify indiscriminate expansion across an account matrix.
3. Manual Recruitment Should Buy Understanding and Effective Supply, Not Just Sample Size
In an interview recorded in November 2023 and released in 2024, FlowGPT reported privately messaging five or six hundred people during its first two weeks, and several thousand in total. Its recruitment on Discord, Reddit, and Twitter targeted community members willing to create and consume works, rather than merely purchasing a visit. On a user-generated content platform, users themselves may understand the possibilities better than the team; early manual conversations also build supply. These counts are the guest's retrospective account, not a reproducible threshold. A small team should care more about who actually publishes after being contacted, whether anyone consumes those works, and whether creators keep updating them. Original episode
In 2025, Mercor moved from a broad recruitment vision toward domain experts urgently needed by AI labs, whose buyers had clear budgets and delivery pressure. Professional forums, offline events, proactive recruitment, and referrals together supplied experts. Referrals were an important part of registrations, but should not be quoted as “more than half”: the host raised that proportion in a question and it was not precisely confirmed. More importantly, registration was followed by identity, résumé, skill, hours, quality, payment, and dispute management. Measuring growth by sign-ups misses what the platform actually needs: effective experts completing qualified work. Long-term trust comes from fulfilling commitments, not merely finding contact details. Original episode
User research must also verify recruitment sources. Answer.AI once received participants whom a third party described as highly active users, only to discover from their UIDs that they had just registered. The interviews and understanding of demand developed over the preceding days consequently lost their foundation. Sample authenticity comes before interview technique: confirm usage history, core actions, payment status, region, and language; preserve screening conditions; and record nonrespondents and churned users. Interviews only with active community members who are willing to praise the product cannot represent everyone.
After acquiring ChatGPT for Google in 2023, Monica did not force its users to upgrade directly. It used part of the initial sample to help validate the product while building its own acquisition capabilities. It reported that the share of survey respondents who would be “very disappointed” if the product disappeared rose from 35% to 55%. But sample size, selection rules, and people who had left were not disclosed; a 40% rule of thumb cannot be treated as a universal PMF certification. Acquired users, what people are willing to say in interviews, behavioral retention, and actual payment are different forms of evidence. Original episode
4. Retention Among Free Users, Actual Payment, and Sustained Gross Profit Need Separate Validation
“The team likes its own prototype → people use it for free and return → people are willing to keep paying” contains three hypotheses. Answer.AI argued that monetization must be tested early, while inference costs make purely free scaling harder to sustain. But different categories should not share one next-day retention or paid-conversion target: AiPPT's infrequent, nonprofessional users may return only when a need arises; ACE serves professional production workflows; and NieTa's recurring demand may be nurturing characters rather than creating comics. Infrequent use does not automatically mean failure, and frequent use does not automatically demonstrate value worth paying for. Original episode Original episode
FlowGPT also highlights the need to separate retention for an individual work or character, category retention, and platform retention in entertainment consumption. A user who switches to another bot but still returns weekly to find content should not be counted as a complete loss of platform value. Conversely, platform activity does not prove that a particular character relationship is stable. UGC platforms must also track creator retention and continued creation, so growth on the consumption side does not conceal shrinking supply.
Opus's experiments with payment prompts emphasized first identifying the value users actually experience. Making subtitle templates more visible unexpectedly improved conversion, suggesting users may have valued that result more than the team anticipated. The paywall encountered during Arcade's public product demonstration was the guest's own purchasing experience, not proof of the best approach for all users. Pricing, entitlements, prompt timing, and copy should be recorded together, while protecting existing purchasers. Each change should be evaluated against first payments, subsequent use, cancellations, refunds, and feedback. AI cost accounting must include consumption during free trials, not just the gross margin of paying users.
In 2025, ACE said one price increase quickly doubled daily revenue, followed by tests of different prices, trials, and monthly plans. That is a short-term cash lever, not proof that long-term renewals were unharmed. Simply annualizing US$800,000 in monthly revenue gives approximately US$9.6 million; it does not automatically become cash collected over a full year or renewable ARR. Mercor's US$100 million run rate likewise does not mean that amount had already been earned in cumulative revenue. Growth reports must clearly distinguish cash collected, revenue, GMV, run rate, ARR, and profit, rather than allowing a headline to substitute for definitions.
5. Growth Quality Includes Ecosystem Spillovers and the Capacity to Serve Acquired Users
Reflecting on Douyin from 2018 to 2020, Yu Beichuan (于北川) emphasized how its early young audience, black UI aesthetic, curated supply, and restraint in importing broad, untargeted traffic established its positioning. It subsequently introduced new verticals and allocated exploratory exposure so recommendations could learn, then attracted new audiences. This is an explanation based on internal experience, without experiments isolating each factor. But the mechanism reminds us that acquisition means more than obtaining an ID: content, a reason to enter, and feedback the system can learn from must be there to receive the user. Original episode
The distribution interface changes both viewing efficiency and creator returns. Comparing single-column and two-column layouts requires more than a single measure of viewing time. It should include ordinary creators' first feedback, continued creation, supply diversity, and user choice. Creators serve consumers both inside and outside an experiment, so ordinary user-level A/B tests can have network spillovers. A short-term consumption advantage may not mean a better long-term ecosystem. Clear segmentation, longer observation windows, and potential contamination need to be addressed; measurement difficulty does not exempt a mechanism from evaluation.
The same history contains examples of distorted proxy metrics. Red packets could raise OKRs for adding friends and interacting without creating organic relationship contexts. Changing an activity's objective near Chinese New Year might leave it delivering neither social relationships nor effective acquisition. These examples suggest that follows, shares, messages, and generation consumption are only process measures; evaluation must return to why users come back. Complaints about insufficient energy in NieTa may indicate stronger demand, but actual purchases, repeat purchases, and effects on supply still need validation. Complaints or consumption alone do not establish monetization success.
Distribution can also damage the experience awaiting users. Cheap traffic may cause registrations to surge while bringing mismatched users, overwhelming customer support, and leaving creators without feedback. Short-term CAC then looks good as subsequent value deteriorates. Growth quality should include the share of poorly matched users, task failures, complaints, refunds, negative reactions, and supply churn. “It generated exposure” is not enough to justify expenditure.
6. Growth Comes Full Circle Only When Experiment Results Feed Back into Product and Business Decisions
A growth loop has at least five steps: identify the current constraint → state the mechanism and metrics → experiment with a limited budget → examine outcomes and guardrails → decide whether to retain, revise, or stop. Mercor's description of fast decisions within a week depends on first stating what should become visible during that week and what to do if it does not. Speed does not mean omitting expectations, accountability, and failure plans. Complex or irreversible actions require higher thresholds and should not copy the pace of reversible experiments.
Opus connected customer-support issues to product priorities, fixes, and support follow-up. Community wishes, paid-service failures, and roadmap input can be collected through different channels, but someone must ultimately own closure. Automatic classification or sending an email does not mean the problem is fixed. Renewals also require separating voluntary cancellations from Past Due payment failures. Its reported improvement in retry success rates from 20%–30% to 40%–50% is experience with no stated measurement definition, not a growth-budget assumption. In practice, teams should examine eligible retries, recovered amounts, authorization, duplicate charges, and complaints, while respecting users who have already canceled.
The Gaotu episode provides an important correction to metric definitions: the official chapter heading referred to a three- or fourfold sales-conversion rate, while the detailed transcript discussed a three- or fourfold increase in “per-lead efficiency,” meaning GMV per lead. It separately described conversion approximately doubling, and only in a small-scale experiment. Business readback should prioritize specific definitions and retain the source, experimental share, baseline, and window. A chapter heading must not be inflated into a company-wide doubling of revenue. Covering more leads and answering more messages remain process counts.
7. Applying This to Product Teams: Start with a Growth Experiment Card Whose Results Can Be Verified (Editorial Proposal)
For AI companion, character, and creative products, start by fixing one region and language, one target user type, one segment of the core experience, and one channel. Avoid changing pricing, entry points, content, and recommendations in the same week. The card should record at least: the target audience and its authenticity; existing alternatives; the first action that delivers value; channel and creative versions; costs included; the experimental denominator and time window; returning use and payment; refund and quality guardrails; creator feedback; and conditions for stopping or scaling. Qualitative quotations should trace back to the corresponding participants, and behavioral outcomes to events and payment records. Surveys, demonstrations, saved admin settings, deployment, and users' actual experience must remain distinct layers of evidence.
For creator growth, first track “recruited → completes a work → receives the first meaningful consumption feedback → continues creating.” For consumer growth, track “views a card → starts interacting → receives the expected experience → returns.” Monetization then tracks actual purchases, renewals, and gross profit. Report character, category, and platform results separately. Short-term payment gains cannot offset damage to the relationship experience. Once a channel produces good results, expand audiences and scale gradually, checking each time whether the original mechanisms of trust and supply still hold. A matrix amplifies a validated mechanism; it cannot replace the mechanism itself.