Counterintuitive Growth: Add Friction, Kill A/B Tests
TL;DR: Two ideas from Anthropic’s growth team that broke my priors: adding friction to onboarding increases conversion when your product has a discoverability problem, and AI-first companies should flip the optimization ratio — 70% big bets, 30% micro. The playbook changes when the ceiling won’t sit still.
Anthropic’s growth team deliberately adds steps to Claude’s onboarding. More questions. More clicks before you see the chat box. Conversion goes up.
That sounds wrong. It’s not.
Picked this up from Amol Avasare (Anthropic’s Head of Growth) on Lenny’s Podcast. Two ideas stuck — mostly because they violate muscle memory I’ve built over years of product work.
Friction as Feature
The standard playbook: nuke every step between signup and value. Fewer clicks, faster activation, ship it. We’ve all internalized this.
Claude has the opposite problem. You open it, you see a text box, you type “summarize this.” Done. Except it can also write production code, simulate a debate partner, analyze papers, build entire apps. The user thinks it’s a fancy autocomplete. It’s not. The gap between capability and discoverability is enormous.
So they added an onboarding quiz. “What do you do? What are you trying to accomplish?” Two questions, personalized suggestions, conversion climbs.
Not original. Notion does the same thing — blank page syndrome is brutal when your product can be a wiki, a tracker, a database, and a notebook simultaneously. Ask what the user wants, pre-load a template, problem solved.
MasterClass went further: quiz before the paywall. “What interests you?” Recommend courses. Purchase rate up.
The pattern: capability » discoverability → guided friction beats a clean funnel.
“Guided” is the key word. This isn’t a tutorial talking at you. It’s the product asking about you. Good salesperson behavior — don’t pitch features, ask what they need.
Here’s the part people miss though. This only works when the gap is real. A calculator app adding an onboarding quiz is just insulting. You need honest self-assessment: can a new user figure out your product’s value in 30 seconds? If yes, get out of the way. If no, guide them.
Most founders think they have a discoverability problem. Most don’t. They have a value problem. Adding friction to a product nobody wants just makes it worse. Be honest.
Big Bets > Button Colors
Traditional growth: 70% micro-optimization (A/B test the CTA, tweak copy, move the button 3px left), 30% big bets. Anthropic inverts it. 70% big bets, 30% micro.
The logic is almost too simple. If your product’s value grows 30-50% over two years — a grocery delivery app, say — micro-optimizations can capture most of that. The ceiling is visible. Squeeze every percent. Makes sense.
But if your product might be 100x better in two years? AI-first products where each model generation is a step function? Optimizing the current funnel is rearranging deck chairs. Your A/B test winner from January might be meaningless by June because the underlying model changed everything.
Their growth team built a Chrome extension as a big bet. It became the foundation for Claude’s Cowork feature. That’s not something you get from a 2% conversion lift experiment. That’s something you get from asking “what if agents could live in your browser?” and then actually building it.
I think this framework is genuinely useful. But — and Amol said this too — it only applies when core value is AI-driven. Sprinkled some AI on top? Autocomplete here, recommendation there? Your ceiling isn’t moving that fast. Stick to the traditional split. Be honest about which category you’re in.
The uncomfortable part: micro-optimizations give you signal in 1-2 weeks. Big bets take months. In a world where models upgrade quarterly, your big bet might be obsolete before it ships. You need strong directional judgment and fast execution.
Most teams have neither. That’s the real problem nobody talks about.
So What
Both ideas share a thread: the rules change when capability growth is exponential.
Standard growth playbooks assume a stable product. AI products aren’t stable. They’re on a curve that makes last quarter look ancient.
Two questions if you’re building AI stuff right now:
-
Is there a real gap between what your product can do and what users discover? If yes, friction is your friend. If no, stop adding onboarding quizzes and go build something worth discovering.
-
Is your ceiling moving faster than your optimization cycle? If yes, stop testing button colors. Make bigger bets.
Neither answer is obvious. Both require honesty about where you actually are — not where your pitch deck says you are.
一句话总结： Anthropic 增长团队的两个反常识操作：产品能力远超用户认知时，加步骤反而提转化；AI 产品该把 70% 资源砸在大赌注上，别磨按钮颜色。天花板在跑的时候，老方法论不好使。
Anthropic 故意在 Claude 的 onboarding 里加步骤。更多问题、更多点击，转化率反而涨了。
听着离谱。但它是对的。
这是从 Lenny’s Podcast 上 Anthropic 增长负责人 Amol Avasare 那听来的。两个观点让我卡住了——因为跟我过去做产品攒的直觉完全反着来。
摩擦变功能
老规矩：注册到核心体验之间，步骤越少越好。少点击、快激活、别废话。都刻在DNA里了。
Claude 的问题正好反过来。打开一看——聊天框。输”帮我总结一下”。完事。但它其实能写生产级代码、模拟辩论对手、分析论文、搭应用。用户以为是个高级自动补全。不是。能力和认知之间的 gap 大到离谱。
于是加了个 onboarding 问卷。”你干什么的？想用来做什么？”两个问题，给个性化推荐，转化率上去了。
这招不新鲜。Notion 一样的——空白页综合症，产品能当 wiki、项目管理、数据库、笔记本，用户反而懵了。问清楚要干嘛，预装模板，立刻激活。
MasterClass 更猛，付费墙前面加测试。”你对什么感兴趣？”推荐课程，付费率涨。
规律：能力远大于可发现性 → 引导式摩擦赢过干净漏斗。
关键词”引导式”。不是 tutorial 单方面教你——那种东西用户只想跳过。是产品在了解你。好销售不上来就讲功能，先问你要什么。
但这里有个大前提人们经常忽略。只有 gap 真实存在才有效。计算器 app 加 onboarding 问卷纯属有病。你得诚实：新用户 30 秒能不能搞明白你的产品能干什么？能，那就闪开。不能，才引导。
大部分创始人觉得自己有 discoverability 问题。其实没有。他们有 value 问题。产品本身没人要，加摩擦只会更差。别骗自己。
大赌注优先，别改按钮颜色
传统增长：70% 微优化（A/B 测按钮颜色、改文案、挪位置）+ 30% 大赌注。Anthropic 反过来。70 大 30 小。
逻辑简单到几乎不需要解释。产品价值两年后涨 30-50%——比如买菜 app——微优化能吃大部分。天花板看得见，每个百分点都值得榨。
但产品价值两年后可能 100 倍？AI 原生产品，每代模型都是台阶式跳跃？那优化当前漏斗就是在泰坦尼克号上摆椅子。一月份跑赢的 A/B 测试六月可能就废了，因为底层模型把一切都改了。
他们增长团队做了个大赌注：自己写 Chrome 扩展。后来成了 Claude Cowork 的基础。这种东西靠 2% 转化提升的小实验永远做不出来。这是”如果 agent 能住在浏览器里会怎样”这种问题的产物。
我觉得这个框架确实有用。但 Amol 自己也说了——只适用于核心价值 AI 驱动的产品。就是加了点 AI 调味？这里自动补全一下、那里推荐一下？天花板没在动，老老实实传统比例。对自己在哪个象限要诚实。
让人难受的 tradeoff：微优化一两周出信号，大赌注要几个月。模型一个季度升级一次的世界里，大赌注可能还没上线就过时了。需要极强的方向判断力加极快的执行速度。
大部分团队两样都没有。这才是没人说的真问题。
所以呢
两个观点一根线：能力增长指数级的时候，规则变了。
传统方法论假设产品稳定。AI 产品不稳定，在一条让上季度看起来像古董的曲线上。
做 AI 产品的话，两个问题值得问自己：
-
产品能做的和用户能发现的之间有没有真实 gap？有就加引导。没有就别搞 onboarding 问卷了，去做点值得被发现的东西。
-
天花板跑得比优化周期快吗？是的话别测按钮颜色了，去下大注。
两个答案都不显而易见。都需要对自己产品的真实位置诚实——不是 pitch deck 里写的那个位置。