PLG Data Analyst Cheatsheet详解 ↗
目标:拿到一个陌生 PLG 产品时,不从“哪个数字最低”开始,而是建立一条从获客 → 价值 → 留存 → 付费 → 扩张的可验证证据链。
0. 总模型详解 ↗
PLG Business Model
Acquire → Activate / Retain → Monetize → Expand
↘ Growth Loops ↻
Data Analyst Model
Business Question
→ Metric Definition
→ Data Grain
→ Event Model
→ Funnel / Cohort / Segment
→ Bias Check
→ Validation
→ Hypothesis
→ Experiment
→ Decision
指标不是事实;指标是对用户价值的一种测量假设。
1. 拿到陌生产品:标准分析 SOP详解 ↗
- 先问业务问题:不要接受“看看 PLG 数据”“看看 retention”“做个 dashboard”作为问题本身。
- 确定 Grain:User / Workspace / Account / Organization / Subscription / Revenue。
- 定义 Value Event:产品定位 → 核心价值 → 用户真正获得价值的时刻 → Activation Metric → 下游验证。
- 拆 Metric Definition:Numerator / Denominator / Window / Eligibility / Cohort Anchor / Segment / Attribution。
- 检查 Event Model:谁、做了什么、何时、上下文、价值、商业状态是否都能追踪。
- 沿价值链分析:Acquire → Activate → Retain → Monetize → Expand。
- 检查 Bias:Mix Shift / Selection / Confounding / Survivorship / Leakage / Attribution / Simpson’s Paradox。
- 找到 Predictor 后不要直接当 Cause:Predictor → Causal Hypothesis → Experiment。
一句话:先把这个指标的 SQL 口径说清楚。
2. Funnel Analysis详解 ↗
回答:用户沿着有顺序的路径,掉在哪里?
Visit → Signup → Setup → Value Event → Retained → PQL → Paid
同时看:
- Conversion Rate
- Absolute Loss
- Time Between Steps
优先级原则:
先判断当前 PLG 主矛盾在哪一层,再看这一层内部哪段绝对损失最大。
适合:路径流失、Onboarding、注册→激活、商业化漏斗。
不适合:长期用户为什么留下——这个问题切 Cohort。
3. Cohort / Retention详解 ↗
Retention 真正测的是:第一次价值发生以后,这个价值有没有重复发生。
Activation = 第一次价值
Retention = 重复价值
一个 Retention 定义必须有:
- Cohort Anchor
- Return Event
- Return Window
常见 Anchor:
- Signup Cohort:看获客质量 / 全生命周期
- Activation Cohort:看获得真实价值后的 retention
- Paid Cohort:看付费后的 retention / expansion / NRR
Retention 类型:
- N-Day Retention
- Windowed Retention
- Rolling Retention
Natural Frequency 要匹配产品频率:
- Daily:D1 / D7 / D30
- Weekly:W1 / W4 / W8
- Monthly:M1 / M3 / M6
Retention Curve 重点看:
- Early Drop
- Long Decline
- Plateau
真激活验证:
P(Retention | Activated)
vs
P(Retention | Non-activated)
注意:predictive ≠ causal。
4. TTV / Time-to-Event详解 ↗
TTV = Product Friction + Natural Waiting
把:
Signup
→ Integration Start
→ Integration Complete
→ First Input
→ First Result
→ Value Event
分段计算。
常见 Time-to-X:
- Time to Activation
- Time to First Paid
- Time to Expansion
- Time to Churn
行为数据常偏斜,优先看 Median / P25 / P75 / P90,少依赖 Mean。
5. Segmentation详解 ↗
目的不是切得细,而是:找到不同的业务机制。
优先切:
- Business mechanism:ICP / Use case / Persona / Account size
- Acquisition mechanism:Channel / Campaign / Landing page
- Product path:Activation path / Workflow / Feature adoption
- Commercial motion:Self-serve / Sales-assisted / Plan
- Environment:Country / Device / OS
好的 segmentation 标准:
切完以后,业务结论变了。
6. Mix Shift / Bias详解 ↗
Mix Shift详解 ↗
Overall Change
=
Within-segment behavioral change
+
Segment composition change
Simpson’s Paradox详解 ↗
每个 segment 都变好,但 overall 反而变差,常由用户结构变化导致。
Selection Bias详解 ↗
问:为什么这些人会进入这组?
Confounding详解 ↗
第三个变量同时影响 X 和 Y。 第一步先做 Stratification。
Survivorship Bias详解 ↗
只看留下的人,会高估很多长期行为的作用。 解决:固定 early observation window。
Look-ahead / Leakage详解 ↗
Observation Window → Scoring Point → Outcome Window
不能偷看未来。
Attribution Bias详解 ↗
B2B PLG 至少区分:
- Source
- Product / Activation Influence
- Close
典型:
Source = SEO
Activation = Product-led
Close = Sales
7. Acquisition Quality详解 ↗
不要停在:
Visit → Signup
至少追到:
Visit
→ ICP Signup
→ Activated
→ Retained
→ Paid
更成熟的效率指标可能是:
- CAC / Activated User
- CAC / Retained User
- CAC / Retained Paid Account
8. PQL / Predictive Analysis详解 ↗
核心问题:
在需要采取商业动作的那个时点,哪些已经发生的产品行为,可以预测未来付费 / 扩张?
Base Rate详解 ↗
P(Paid)
Lift详解 ↗
Lift(signal)
=
P(Paid | Signal)
/
P(Paid)
Coverage详解 ↗
这个 signal 覆盖多少真实商业机会?
Precision详解 ↗
TP / (TP + FP)
Recall详解 ↗
TP / (TP + FN)
Threshold 取决于 intervention cost:
- 自动 CTA:可容忍低 precision
- Sales contact:需要更高 precision
PQL signals 可分:
- Value Realization
- Engagement
- Collaboration / Expansion
- Commercial Intent
- Firmographic
成熟 B2B:
PQL = Product Intent × Commercial Potential
Observation Window 示例:
Day 0–7 Behavior
→ Day 7 Score
→ Predict Day 8–30 Paid
第一版可以 rule-based scoring,先别急着上复杂 ML。
9. Monetization Analysis详解 ↗
商业化漏斗:
Activated
→ Retained / Repeated Value
→ PQL
→ Paywall / Upgrade Trigger
→ Checkout
→ Paid
同时保留:
- Signup → Paid
- Activated → Paid
- PQL → Paid
Free Tier详解 ↗
免费必须换回某种增长价值:
- Activation
- Habit
- Referral
- Team invite
- PQL
- Paid conversion
Usage Distribution详解 ↗
不要只看平均,至少看: Median / P75 / P90 / P95 / P99
Usage × Commercial Outcome详解 ↗
比较 usage bucket 对:
- Paid rate
- Retention
- ARPA
- Expansion
- Cost
10. Value Metric / Pricing Metric / Cost详解 ↗
Value Metric详解 ↗
问题:用户获得的价值沿哪个维度增长?
候选: Tasks / Seats / Workflows / Storage / Automation Runs / Time Saved
Pricing Metric详解 ↗
不一定等于 Value Metric。
真实价值 = 节省人工小时
↓
难测 / 难预测 / 易争议
↓
Pricing Proxy = Task / Workflow / Credit
好的 Pricing Metric:
- Value aligned
- Measurable
- Predictable
Cost详解 ↗
Value 是主轴,Cost 是 guardrail。
Value alignment
→ Simplicity / Predictability
→ Cost guardrail
11. Paywall / Pricing Cohort详解 ↗
Paywall Analysis详解 ↗
不要只看 Paywall → Paid,还要看:
Who hit paywall?
→ Activated?
→ Usage before paywall
→ Repeated paywall hits
→ Checkout
→ Paid
→ Post-paid retention
重点比较:
Paywall before Activation
vs
Paywall after Activation
Pricing Cohort详解 ↗
价格调整后比较:
- Activation
- Paid conversion
- ARPA
- Retention
- Expansion
- Gross Margin
避免只优化单笔收入。
12. Revenue / Expansion Analysis详解 ↗
Revenue Decomposition详解 ↗
New Revenue
+ Expansion Revenue
- Contraction
- Churn
= Net Revenue Change
NRR详解 ↗
NRR
=
(Starting Revenue - Churn - Contraction + Expansion)
/
Starting Revenue
但:NRR 是结果,不是 insight。
继续拆 Expansion Driver:
- Seat
- Usage
- Workflow
- Module
- Org
- Capacity
- Compound Value
至少同时看:
- Expansion Amount
- Expansion Distribution
- Expansion Driver
Concentration
看:
- % Accounts Expanded
- Median Expansion
- P90 Expansion
- Top Account Contribution
Expansion Cohort详解 ↗
M0 → M3 → M6 → M12
Expansion Velocity详解 ↗
看 Paid → First Expansion,以及不同 expansion axis 的时间。
13. Growth Loop Analysis详解 ↗
例如 Invite Loop:
Activated Users
→ Invite Rate
→ Invites / Inviter
→ Acceptance
→ Signup
→ Activation
→ Retention
→ Invite Again
近似:
New Activated Users
=
Activated Users
× Invite Rate
× Invites/User
× Acceptance
× Activation
目标不是“做分享”,而是找到 loop bottleneck。
14. Experimentation / A/B Test详解 ↗
从 Observational 到 Causal:
Predictor
→ Causal Hypothesis
→ Randomized Experiment
→ Decision
实验卡至少提前定义:
- Business Problem
- Causal Hypothesis
- Eligible Population
- Randomization Unit
- Treatment
- Primary Metric
- Guardrails
- Baseline + MDE
- Sample / Observation Window
- Success / Failure / Gray Zone
- Kill Condition
- Owner
- Next Action
重点:
- Primary Metric 只设一个主判断
- Guardrail 防止上游/下游被伤害
- Randomization Unit 匹配 User / Workspace / Account
- MDE 连接 Business Significance 与 Sample Size
- 不只看 p-value,要看 Effect Size + Confidence Interval + Business Impact
- No significance ≠ No effect
- 先查 Data Quality / SRM 再做统计
- 前后对比不是严格 A/B
- ITT 默认按最初 assignment 分析
最终输出:
- What happened
- Effect size
- Uncertainty
- Guardrails
- Key segments
- Decision
- What we learned
决策:
Ship / Kill / Iterate / Continue
15. Dashboard / Analysis / Experiment详解 ↗
Dashboard = Monitoring
发生了什么?
Analysis = Diagnosis
为什么发生?
Experiment = Causal Validation
改 X 会不会真的改变 Y?
成熟工作流:
Monitor
→ Diagnose
→ Hypothesis
→ Experiment
→ Decision
16. 高频判断口诀
别人说:注册率 12%
问:这些人 Activated 了吗?
别人说:Activation 58%
问:定义是什么?能预测 Retention 吗?
别人说:Retention 掉了
问:Same people got worse,还是 user mix changed?
别人说:这个行为和付费高度相关
问:Selection / Confounding / Leakage 有吗?
别人说:这个渠道 conversion 高
问:转到 Signup,还是 Retained Paid?
别人说:PQL 很多
问:Base Rate / Lift / Coverage / Precision / Recall 呢?
别人说:NRR 120%
问:谁贡献的?哪个 Expansion Axis?集中度呢?
别人说:实验赢了
问:Effect Size、CI、Guardrail、核心 Segment 呢?
17. 最终一页记忆版
PLG DATA ANALYST
1. DEFINE
Business Question
→ Grain
→ Value Event
→ Numerator / Denominator / Window / Eligibility
2. TRACE VALUE
Acquire
→ Activate
→ Retain
→ Monetize
→ Expand
3. CHOOSE METHOD
掉在哪里? → Funnel
后来怎么样? → Cohort
为什么总体变化? → Segmentation
什么行为预测商业结果? → PQL / Lift
多久发生? → Time-to-Event
收入为什么变化? → Revenue Decomposition
Loop 卡哪里? → Loop Funnel
X 是否导致 Y? → Experiment
4. CHECK BIAS
Mix Shift
Selection Bias
Confounding
Survivorship
Leakage
Simpson’s Paradox
Attribution
5. VALIDATE
Activation → Retention
Usage / PQL → Paid
Paid → Expansion
Product Signal → Future Commercial Outcome
6. EXPERIMENT
Hypothesis
→ Primary Metric
→ Guardrail
→ MDE / Sample
→ Randomize
→ Effect + CI
→ Ship / Kill / Iterate
最终目标:
不是“找最低的数字”,
而是建立一条能从获客一路解释到价值、留存、付费和扩展的证据链。