故事 · 被三个大客户带进沟里
Story · Driven into a ditch by three big accounts
一家 SaaS 公司做新版规划,把调研重点放在了三个营收最大、关系最铁的企业客户身上 —— 毕竟他们最重要、也最愿意聊。
这三家强烈要求更多高级配置项和定制能力,于是产品越做越复杂、越做越重。
一年后数据出来:占用户数 九成的中小客户大批流失,他们的诉求恰恰相反 —— 更简单、开箱即用。
公司听到的,从来只是「最吵的那几个」的声音。问题不在于听错了,而在于样本本身就是歪的:
只在矩阵的一格里取样,自然只能得出那一格的结论。修正它的唯一办法,是先画出客户矩阵,再保证每一格都被听见。
A SaaS company planning its next major release focused its research on the three highest-revenue, closest-relationship enterprise accounts — they mattered most and were happiest to talk.
Those three pushed hard for more advanced configuration and deep customization, so the product grew steadily heavier and more complex.
A year later the numbers came in: the SMB customers — 90% of the user base — were churning in droves, asking for the exact opposite: simpler, works out of the box.
The company had only ever heard "the loudest few." It wasn't that they listened poorly; it was that the sample itself was crooked.
Sample only one cell of the matrix and you only get one cell's answer. The only fix is to draw the matrix first and then guarantee every cell is heard.
1 客户分层矩阵:每格的客户数 / 已抽样数 Customer Segmentation Matrix: Customers per Cell / Samples Taken
点格子加样本2 你的样本,落在了哪几格? Where Did Your Samples Land?
把所有抽到的样本按规模列堆成一条 —— 若整条几乎都是某一种颜色,说明你的结论被那一类客户绑架了。 理想的样本,应大致铺满每一格、与各格的客户分布相称。 All samples stacked into one bar by size column. If the bar is mostly one color, your conclusion is held hostage by that customer type. A good sample spreads across every cell in rough proportion to each cell's customer population.
3 现实里的矩阵抽样 Matrix Sampling in the Real World
抽样代表性:样本的结构要尽量贴近总体的结构,否则再多的访谈,也只是把一格的声音放大。
Sample representativeness: The sample's shape should track the population's shape — otherwise more interviews just amplify one cell.
客户分层:先按真正影响需求的维度(行业、规模、新老、场景)分格,分格的维度选错,矩阵就白搭。
Segmentation: Split along dimensions that genuinely shape needs — industry, size, tenure, use case. Pick the wrong axes and the matrix is useless.
避免幸存者偏差:别只问还在用的客户 —— 已流失的客户往往藏着最关键的问题,他们才是矩阵里最该补的格。
Beware survivorship bias: Don't only ask current users — churned customers usually hold the most critical signal. They're the cell most worth filling in.
配额抽样:给每格定一个最低样本数(配额),确保小众但重要的格子不被声量大的格子淹没。
Quota sampling: Set a minimum sample count (quota) per cell so small-but-important cells aren't drowned out by loud ones.
一句话In One Line
「听客户」听起来天然正确,但真正决定结论的,不是你听得多认真,而是你听的是谁。
人会本能地往「方便、亲近、声量大」的客户那里靠 —— 大客户、老客户、还在用的客户、最爱反馈的客户。
这些便利恰恰制造了系统性偏倚:你以为在描绘全体,其实只放大了一格。
客户矩阵的全部意义,是把这种隐形的偏好摆到台面上:分好格,定好配额,强制让每一格都发声。
代表性不会自动发生,它必须被刻意设计出来。
"Listen to the customer" sounds self-evidently right, but the conclusion isn't decided by how attentively you listen — it's decided by whom you listen to.
People instinctively drift toward the convenient, the close, and the loud — big accounts, long-tenured users, active customers, the chronic feedback-givers.
That convenience is exactly what manufactures systematic bias: you think you're painting the whole picture, but you've magnified a single cell.
The whole point of a customer matrix is to drag that hidden preference into the open — segment the cells, set the quotas, force every cell to speak.
Representativeness never happens on its own; it has to be engineered on purpose.
常见误用Common Mistakes
只问最大 / 最吵 / 关系最近的客户。按矩阵各格配额抽样,让沉默的多数也被听见。
Asking only the biggest / loudest / closest customers. Quota-sample every cell of the matrix so the silent majority is heard.
只调研还在用的客户(幸存者偏差)。把已流失客户也列进矩阵,他们藏着最关键的问题。
Researching only current users (survivorship bias). Put churned customers into the matrix too — they hold the most critical signal.
分层维度随手定,与需求无关。用真正影响需求的维度分格,维度选对矩阵才有意义。
Picking segmentation dimensions arbitrarily, with no link to needs. Segment along dimensions that genuinely shape needs — the right axes make the matrix worth drawing.