1. CAC 要把人算进去。销售和市场的工资、外包、工具费全摊进去，再除以同期新客户数。只算广告费的 CAC 常常低了一半以上——你会据此认为渠道很划算，然后越投越亏。
2. 分渠道算，别只看总数。总 CAC 是被最好的那个渠道拉平的平均值。分开算之后，通常有一两个渠道是明确亏的，而你一直在给它加预算。
3. LTV 用毛利、看中位数。用收入算等于把交付成本当成利润；用均值算，几个超大客户会把它拉到一个没有任何真实客户代表的位置。中位数才是「典型客户」。
4. 别用还没到期的客户估留存。只签了三个月的客户群，你算不出「平均留存 24 个月」——那是假设，不是数据。诚实的写法是：「按目前月流失率外推，约 24 个月，尚未验证。」
EN Half the numbers on your dashboard have never once changed what you did — that's not a metric, it's decoration. Only three earn their place, and all three answer one question: if I spend another unit of money acquiring a customer, do I get it back, and how fast? CAC is every cost of acquisition including salaries, divided by new customers in the same period. LTV is monthly gross profit times months retained — gross profit, never revenue. Payback is CAC divided by monthly gross profit. Watch payback hardest, because it's the only one that requires no guess about the future: LTV asks you to predict how many years someone stays, payback only uses money already in the bank. Four ways people get these wrong: leaving headcount out of CAC, which usually understates it by more than half; computing one blended CAC instead of one per channel, so the good channel keeps subsidising the loss-making one you keep funding; using revenue and the mean for LTV, when a handful of huge accounts drag the average to a place no real customer occupies — the median is your typical customer; and estimating retention from cohorts that haven't aged yet, which is an assumption wearing a number's clothes.
机制：这三个数决定的是「加钱能不能放大」。回本周期决定你烧钱的速度上限（钱要多久转一圈回来），LTV/CAC 决定这门生意本身值不值得放大。而你之所以需要它们、不能靠感觉，是因为销售恰好是那种直觉学不会的环境。
诚实边界：客户只有几十个时，LTV 基本是幻觉——留存曲线还没长出来，你算的是自己的假设乘以一个大数。这时只看回本周期和单笔毛利，够用。另外这三个数管的是「这门生意能不能规模化」，管不了「眼前这一单要不要追」——后者是资格判定那件事，别指望财务指标替你做单笔判断。
1. 下结论前先问「这需要多少单」。算不出来就用上图的量级感：几十单只能看方向不能看小数点，几百单才谈得上比较。想清楚这一条，你会自动少做一大半错误决策。
2. 数量不够就往上游看过程指标。成交是最稀疏的那一层，所以噪声最大。往上一层：约到的会面数、二次会面率、见到拍板人的比例、报价后七天内的回复率——这些量级大十倍、噪声小、而且发生得更早，改了也更快知道结果。
3. 先定口径和时间窗，再看数。「按线索进入的月份算，还是按成交的月份算」必须提前说死。看到结果再挑窗口、再定口径，你能让任何东西显得有效——这不叫分析，叫给自己讲故事。
周会上的一句话：「这个提升是真的还是噪声？我们这个月一共多少单——够不够看出这么大的差别？」
EN "We closed 28% this month against 20% last month, so the new script works." At twenty-something deals that sentence says nothing at all. The wobble in a close rate is computable by hand: the standard error is roughly the square root of p(1−p)/n. At p=20% and n=25 that's about eight percentage points, so a 95% interval spans roughly ±16 — the true rate behind your observed 20% could be anywhere from 4% to 36%. Worse, regression to the mean guarantees that a terrible month is usually followed by an ordinary one, so anything you did at the bottom will look like it worked: the new script, the morning stand-up, a different coffee mug. Three rules protect you. Ask how many deals a claim would need before you make it; a few dozen can show direction, never a decimal point. When volume is short, move upstream to process metrics — meetings booked, second-meeting rate, share of deals where you met the real decision-maker — an order of magnitude more numerous, far less noisy, and earlier. And fix your definitions and time window before you look, because choosing the window afterwards can make anything look effective.
机制：成交率是从一串「成 / 不成」里估出来的比例，样本越小估计越抖。抖动本身不含任何信息，但人天生会给每一次抖动配一个原因——尤其是配上自己刚做的那件事。
边界：小样本不等于没信息。丢单里客户说的原话、反复卡在同一个障碍上——这类定性信号几例就有价值，它不能拿来算比例，却足以指出下一步该改什么。真正的错误是拿五单的百分比当结论、再据此改掉整套流程。
1. 分档统计：五档各有多少单、实际成了多少。只要每档有十来单，形状就看得出来了。
2. 看两个偏差：概率——每一档实际值都低于标注值，说明你系统性乐观，按上图折算即可；日期——统计实际成交日比预计晚了多少天，取中位数，以后所有预计日期一律加上它。这两条几乎人人中招，而且改起来不花一分钱。
3. 永远给两个数，别混成一个：承诺数＝只算 90% 那一档（你敢背的），预测数＝各档按校准后的概率加权求和。老板问「这季度多少」，答：「承诺 80 万，预测 130 万，历史上我这个预测偏乐观约一成。」
要改概率时的一句话：「他今天把财务拉进了会——这是新信息，所以我把这单从 50% 提到 70%。」没有这种句子，改动就只是心情。
EN "I'd say this one's eighty percent there" is a feeling, not a forecast, because nothing about it can later be judged wrong. Replace it with a rule: every deal carries two written values — a close probability from a fixed ladder (10/30/50/70/90) and an expected close date — changeable only on new information, and only with one line saying what the information was. At quarter end, pull every deal you marked 70% and count: how many actually closed? If four in ten did, then your "70%" means 40% in the real world, and from now on you discount it accordingly. Your forecast becomes accurate immediately, and you learned no new skill — you simply stopped fooling yourself. Two biases show up in nearly everyone: probabilities run optimistic across every rung, and dates run early, so take the median slippage in days and add it to every future date. Always report two numbers rather than one blended figure: the commit (only the 90% rung, what you'll stand behind) and the forecast (all rungs weighted by your calibrated probabilities).
机制：学习来自「预期与结果之差」。如果你没有事先把预期固定成一个数，事后大脑会自动把预期改写成已经发生的结果，误差被抹成零——看起来什么都没错，于是什么也学不到。写下来，只是为了让误差还存在。
边界：档位别超过五档，17% 这种精度只会制造精确的假象。更要紧的一条：别把预测和考核绑在一起——一旦报的数直接决定奖惩，人报的就不再是他真信的数，而是安全的数。这和提成的道理完全一样：你考核什么，就会拿到什么的表演版。想两者都要，就把预测和承诺分成两个数，只对承诺问责。
问客户（一封邮件就够，不辩解、不再推销，只求学到东西）：
1. 「最后是在哪两个方案之间做的选择？」——问的是替代方案，不是让他给理由。
2. 「决定的那一刻，最关键的一条是什么？」——问那一刻，比问「为什么不选我们」诚实得多。
3. 「我们这边哪一步让你觉得最费劲？」——这条最常挖出真正能改的东西。
内部规矩一：复盘会先读当初系统里写的下一步和标的概率，再开始讨论。顺序反了，讨论就变成给已知结果编一套合理的解释。
内部规矩二：每季度挑两三单做事前验尸——「假设三个月后这单黄了，各自写下最可能的原因」。写完你多半会发现同一个原因反复出现：从没见过真正拍板的人。
EN Review only your losses and you'll learn plenty about how not to lose and nothing about why you win — and worse, the bad decisions inside a won deal get laundered by the good outcome and then copied across the team as "what works". Two biases stand in the way. Hindsight: once you know the result you sincerely feel the signs were there all along, and you can no longer recall how genuinely unsure you were. Outcome bias: you judge the quality of a decision by how it turned out. There's exactly one defence — compare against what you wrote down at the time, never against memory. That's the second use of the probability and the next-step date from the previous move. For losses, one email with three questions beats an hour of internal theorising: which two options were you finally choosing between (ask about the alternative, not for a reason), what was the single deciding thing at that moment, and which part of dealing with us was the most effort. And run an occasional pre-mortem: assume this deal is dead in three months, everyone writes the most likely cause. The same cause usually surfaces — nobody ever met the person who actually decides.
机制：结果一旦已知，它会反过来重写你对当时的记忆与判断，于是复盘从「检查决策」滑向「给已知结论找证据」。对抗它不靠自律，靠当时留下的记录——记录是唯一不会被结果篡改的东西。
边界：丢单回访的答案要打折——客户常给一个体面的理由，「预算不够」是最省事的挡箭牌。所以问替代方案和过程，别问「为什么不选我们」。另外，复盘一旦影响考核就立刻失效：人只会带来能过审的版本。想让复盘有用，就得让说真话在这个房间里是安全的。
六十分钟，把「感觉」换成四个能查的数。