InfiniSynapse
What is user retention?什么是用户留存？
User retention is a product's ability to bring an eligible group of users back to a meaningful value-producing action over a defined period. A defensible measure states who entered the cohort, which action proves return, when that return counts, and whether the rule is exact-period, bracketed, or on-or-after.
用户留存是产品让一组符合条件的用户在明确时间内回来完成有意义价值动作的能力。可靠的衡量必须说明谁进入同期群、哪个动作代表回访、回访在何时计数，以及采用精确周期、区间还是“该时间点及之后”的规则。
Retention is valuable because acquisition counts can rise while the product fails to create repeat value. It is not, however, a single universal KPI. A collaboration tool might define weekly return as completing a shared task; a tax product may have an annual cadence; a marketplace must decide whether buyer and seller retention are separate. The definition must follow the product's natural value cycle.
留存之所以重要，是因为即使获客数量上升，产品仍可能无法创造重复价值。但留存并非一个通用 KPI。协作工具可以把每周完成共享任务定义为回访，报税产品可能具有年度节奏，市场平台则需要分别定义买家和卖家留存。口径必须服从产品的自然价值周期。
When user retention is the right question—and when it is not何时应研究用户留存，何时不应
Evaluate onboarding quality, repeated product value, cohort health, feature adoption, re-engagement, lifecycle friction, and whether an intervention changes return behavior.
评估引导质量、重复产品价值、同期群健康度、功能采用、重新参与、生命周期摩擦，以及干预是否改变回访行为。
Explain motivation, prove a feature caused an outcome, measure revenue durability, or judge infrequent products whose value cycle is longer than the observation window.
解释用户动机、证明功能导致结果、衡量收入稳定性，或评估价值周期长于观察窗口的低频产品。
Retention analysis becomes important when active users decline, repeat use weakens, a cohort curve flattens, or a gap appears between sign-up and ongoing value. The analysis localizes the loss; interviews, usability sessions, support records, and experiments help explain and test it. If the main question is how to compute a percentage, use the retention rate formula guide. If it is how different start groups behave, use the cohort analysis workflow.
当活跃用户下降、重复使用不足、同期群曲线趋平，或注册与持续价值之间出现断层时，就需要开展留存分析。留存分析用于定位损失；访谈、可用性测试、支持记录和实验用于解释并检验原因。如果主要问题是如何计算比例，请阅读留存率公式指南；如果问题是不同起始群体之后如何表现，请使用同期群分析流程。
Prepare the data and metric contract before analysis分析前准备数据与指标契约
Start with event records containing a stable pseudonymous user or account identifier, timestamp, event name, product version, and only the properties needed for the question. Add acquisition source, plan, device, geography, consent state, or experiment assignment when relevant. Document identity merging, deleted accounts, internal users, bots, late events, time zone, and the date after which instrumentation is trustworthy.
首先准备事件记录，其中包含稳定的化名用户或账户标识、时间戳、事件名称、产品版本，以及回答问题所需的最少属性。必要时加入获客来源、套餐、设备、地区、同意状态或实验分组。记录身份合并、删除账户、内部用户、机器人、迟到事件、时区，以及埋点开始可靠的日期。
| Decision决策项 | Example示例 | Why it matters重要性 |
|---|---|---|
| Cohort entry同期群进入条件 | First completed project, not sign-up首次完成项目，而不是注册 | Separates exposure from receiving initial value区分接触产品与获得初始价值 |
| Return event回访事件 | Completed another collaborative task再次完成协作任务 | Measures repeat value rather than any login衡量重复价值，而非任意登录 |
| Interval rule时间规则 | Week 4 bracket, days 22–28第 4 周区间，即第 22–28 天 | Prevents exact-day and rolling retention from being mixed避免混用精确日与滚动留存 |
| Eligibility资格条件 | Human users with complete observation window具有完整观察窗口的真实用户 | Removes bots and immature recent cohorts排除机器人与尚未成熟的新同期群 |
Privacy boundary: collect only what the analytical decision requires. Retention work does not justify exposing direct identifiers or sensitive free text. Apply access controls, aggregation thresholds, deletion rules, and the consent obligations relevant to your product and market.
隐私边界：只收集分析决策真正需要的数据。留存分析并不能成为暴露直接身份信息或敏感自由文本的理由。应实施访问控制、聚合阈值、删除规则，以及与你的产品和市场相关的同意义务。
How to measure user retention without distorting it如何衡量用户留存并避免失真
For a fixed cohort, user retention is the original cohort members who complete the defined return event in an interval divided by the eligible original cohort size, multiplied by 100. Preserve numerator and denominator counts beside every percentage. A cell based on 20 users should not be interpreted like one based on 20,000.
对于固定同期群，用户留存率等于在某个时间区间完成既定回访事件的原始同期群成员数，除以符合条件的原始同期群人数，再乘以 100。每个百分比旁都应保留分子和分母。基于 20 人的单元格不能与基于 20,000 人的单元格同样解释。
| Method方法 | Question answered回答的问题 | Best fit适用情况 | Main caution主要注意事项 |
|---|---|---|---|
| Exact N-day精确第 N 日 | Did the user return on that exact age?用户是否在精确队列年龄回访？ | Daily-habit products每日习惯型产品 | Misses nearby legitimate use可能漏掉相邻日期的合理使用 |
| Bracket retention区间留存 | Did the user return within a period?用户是否在一个区间内回访？ | Weekly or monthly workflows每周或每月工作流 | Bracket boundaries must remain stable区间边界必须保持稳定 |
| Rolling retention滚动留存 | Did the user return on or after age N?用户是否在第 N 日或之后回访？ | Irregular return cadence不规则回访节奏 | Naturally reads higher than exact retention数值自然高于精确留存 |
| Unbounded return无界回访 | Did the user return at any later time?用户是否在之后任何时间回访？ | Long or uncertain cycles较长或不确定周期 | Observation length changes the result观察时长会改变结果 |
The official Google Analytics cohort exploration documentation separates cohort inclusion from return criteria. Adobe's official retention analysis documentation likewise describes start events, return events, duration, retention, and churn. These first-party definitions support a general rule: write the measurement contract before viewing the curve.
Google Analytics 的同期群探索官方文档将同期群纳入条件与回访条件分开定义。Adobe 的留存分析官方文档也分别说明起始事件、回访事件、时长、留存与流失。这些第一方定义共同支持一条通用规则：先写指标契约，再查看曲线。
A repeatable user retention analysis workflow可重复执行的用户留存分析流程
- Define repeated value定义重复价值Name the action that means the user received value again. Avoid using “any activity” unless any activity truly represents the product outcome.明确哪个动作表示用户再次获得价值。除非任意活动确实代表产品结果，否则不要使用“任何活动”。
- Build comparable cohorts建立可比同期群Group users by a shared start event and calendar grain. Exclude incomplete observation windows and show the cohort size.按共同起始事件和日历粒度分组，排除观察窗口不完整的群体，并展示同期群人数。
- Locate the loss stage定位损失阶段Separate activation failure, early-value failure, habit or cadence failure, mature-use friction, and involuntary loss such as billing or access problems.区分激活失败、早期价值失败、习惯或节奏失败、成熟使用摩擦，以及账单或访问问题等非自愿流失。
- Segment with a prior hypothesis基于先验假设进行细分Compare plan, channel, device, use case, role, geography, or version only when the segment existed before the outcome and has enough observations.只有当套餐、渠道、设备、用例、角色、地区或版本在结果前已经存在且样本足够时，才进行比较。
- Triangulate causes交叉验证原因Join event patterns with support contacts, surveys, interviews, usability evidence, release notes, outages, and billing changes. Treat observed associations as hypotheses, not causes.把事件模式与支持记录、调查、访谈、可用性证据、发布说明、故障和账单变化结合。把观察到的关联视为假设，而不是原因。
- Test and validate实验并验证Choose one intervention, define the primary retention outcome and guardrails before launch, wait for cohorts to mature, and reproduce the result from source counts.选择一项干预，在上线前定义主要留存结果与护栏指标，等待同期群成熟，并从源数据计数复现结果。
Example: diagnosing weak Week 4 retention示例：诊断较弱的第 4 周留存
Hypothetical example: a team-workflow product defines cohort entry as creating a first shared project and the return event as completing a task with another member. A mature January cohort has 1,200 eligible users; 396 return in days 22–28, so Week 4 bracket retention is 33%. This number is an example, not an InfiniSynapse customer result or benchmark.
假设示例：某团队工作流产品把创建首个共享项目定义为同期群进入条件，把与另一名成员共同完成任务定义为回访事件。成熟的一月同期群有 1,200 名合格用户，其中 396 人在第 22–28 天回访，因此第 4 周区间留存为 33%。这一数字仅为示例，不是 InfiniSynapse 客户结果或行业基准。
The team first verifies identity merges, event completeness, and cohort maturity. Segmentation shows lower retention among users whose first project has no collaborator. Interviews reveal that invitations are unclear after project creation. The team therefore tests a contextual invitation step only for solo-created projects. The experiment's primary outcome is Week 4 bracket retention; guardrails include project completion, invite errors, notification opt-outs, and support contacts. It does not claim success until both groups reach the complete Week 4 window and source totals reconcile.
团队先检查身份合并、事件完整性和同期群成熟度。细分结果显示，首个项目没有协作者的用户留存更低。访谈发现，项目创建后的邀请流程不清楚。因此团队只针对单人创建项目的用户测试情境化邀请步骤。实验主要结果是第 4 周区间留存，护栏指标包括项目完成率、邀请错误、通知退订和支持联系量。只有两组都完成第 4 周观察窗口，且源数据总数对账一致后，才判断实验结果。
How to improve user retention by lifecycle failure如何按生命周期失败类型改善用户留存
| Failure pattern失败模式 | Evidence to inspect需要检查的证据 | Candidate action候选措施 | Guardrail护栏指标 |
|---|---|---|---|
| Users never reach first value用户从未达到首次价值 | Onboarding funnel, setup errors, time to value引导漏斗、设置错误、价值实现时间 | Remove unnecessary setup; guide the next meaningful action移除不必要设置，引导下一个有意义动作 | Downstream task success, not completion alone后续任务成功，而不只是完成引导 |
| Value is received once价值只获得一次 | Repeat-action paths, natural cadence, user interviews重复动作路径、自然节奏、用户访谈 | Make saved state, collaboration, or the next use case discoverable让保存状态、协作或下一个用例更易发现 | Notification fatigue and shallow activity通知疲劳与浅层活动 |
| Retention drops after a release版本发布后留存下降 | Version cohorts, errors, performance, changed paths版本同期群、错误、性能、路径变化 | Fix regressions or restore a displaced workflow修复回归问题或恢复被打断的工作流 | Crash rate, latency, task completion崩溃率、延迟、任务完成率 |
| Mature users become inactive成熟用户变得不活跃 | Feature changes, role turnover, billing, support themes功能变化、角色更替、账单、支持主题 | Repair missing value or offer relevant re-entry; avoid spam修复缺失价值或提供相关回归入口，避免骚扰 | Opt-outs, complaints, involuntary churn退订、投诉、非自愿流失 |
Common user retention strategies—better onboarding, clearer activation, reliable performance, relevant lifecycle messaging, support, collaboration, personalization, and win-back flows—are not interchangeable. Choose one only after finding the failure it could plausibly address. Dark patterns, excessive notifications, hidden cancellation, and rewards that create empty activity may move a chart temporarily while harming trust.
常见用户留存策略包括更好的引导、更清晰的激活、可靠性能、相关生命周期消息、支持、协作、个性化和召回流程，但它们不能互换。只有在找到其可能解决的失败模式后，才应选择具体措施。暗黑模式、过度通知、隐藏取消入口，以及制造空洞活动的奖励，也许能短期改变图表，却会伤害信任。
Analyze user retention across your own data sources跨自有数据源分析用户留存
Prepare a read-only connection or export with stable user or account IDs, cohort-entry timestamps, return events, relevant segment fields, and a written metric definition. InfiniSynapse can analyze supported databases, warehouses, files, and multi-source data through natural-language questions. It is not presented as a dedicated retention dashboard or experimentation platform: validate population rules, source totals, query logic, and any causal claim.
请准备只读连接或导出文件，其中包含稳定的用户或账户 ID、同期群进入时间、回访事件、相关细分字段和书面指标定义。InfiniSynapse 可以通过自然语言问题分析受支持的数据库、数据仓库、文件和多源数据。这里不把它描述为专用留存仪表板或实验平台；你仍需验证人群规则、源数据总数、查询逻辑和任何因果结论。
Open the InfiniSynapse data analysis app打开 InfiniSynapse 数据分析应用Common user retention mistakes and limitations常见用户留存错误与局限
- Using login as value: a return session can reflect confusion, a password reset, or an unresolved task. Prefer a product-specific value event.
- 把登录当成价值：回访会话可能代表困惑、密码重置或未解决任务。应优先使用产品特定的价值事件。
- Comparing immature cohorts: a cohort created two weeks ago cannot have complete Week 4 retention. Hide or label incomplete cells.
- 比较未成熟同期群：两周前创建的同期群不可能拥有完整的第 4 周留存。应隐藏或标记不完整单元格。
- Changing definitions silently: a new return event, identity rule, time zone, or interval creates a different metric. Version the contract and backfill only when reproducible.
- 静默更改定义：新的回访事件、身份规则、时区或区间会产生不同指标。应为契约版本化，只在可复现时回填。
- Assuming correlation is cause: retained users may adopt a feature because they were already motivated. Use experiments or credible causal designs for causal claims.
- 把相关性当成因果：留存用户可能因为本来就更有动机而采用某功能。因果结论需要实验或可信的因果设计。
- Chasing a universal benchmark: cadence, audience, maturity, pricing, acquisition mix, and definitions differ. Compare like with like and disclose the denominator.
- 追逐通用基准：产品节奏、受众、成熟度、定价、获客组合和定义各不相同。应进行同类比较并披露分母。
- Ignoring survivorship and missing data: blocked tracking, account deletion, cross-device identity, outages, and seasonality can change who remains observable.
- 忽略幸存者偏差与缺失数据：追踪拦截、账户删除、跨设备身份、故障和季节性都会改变仍可观察的人群。
Validate retention results before making a decision决策前验证留存结果
- The cohort entry, return event, interval, eligibility, identity, and time zone are written and versioned.同期群进入条件、回访事件、区间、资格、身份与时区均已书面记录并版本化。
- Each percentage includes raw numerator and denominator counts; small groups are flagged.每个百分比都包含原始分子与分母，小样本群体已标记。
- Only cohorts with complete observation windows are compared at the same age.只比较观察窗口完整且年龄相同的同期群。
- Event totals reconcile with an independent source and instrumentation changes are annotated.事件总数与独立来源对账一致，埋点变更已有标注。
- Segments existed before the outcome, have sufficient size, and were not selected only because they looked favorable.细分在结果前已存在、样本量足够，且不是因为结果看起来有利才选择。
- Experiment allocation, exposure, primary outcome, guardrails, and stopping rules were defined before reading results.实验分配、曝光、主要结果、护栏指标与停止规则均在查看结果前定义。
Retention should connect to its diagnostic neighbors. Use funnel analysis to locate early step loss and the broader product analytics guide to place retention beside activation, engagement, adoption, and business outcomes.
留存应与相邻诊断方法连接。使用漏斗分析定位早期步骤损失，并通过更广泛的产品分析指南把留存与激活、参与、采用和业务结果结合起来。
User retention FAQ用户留存常见问题
What is user retention?
什么是用户留存？
User retention is a product's ability to bring an eligible group of users back to a meaningful value-producing action over a defined period. A defensible measure states the cohort, return event, interval, and counting rule.
用户留存是产品让一组符合条件的用户在明确时间内回来完成有意义价值动作的能力。可靠指标必须说明同期群、回访事件、时间区间与计数规则。
How do you calculate user retention?
如何计算用户留存？
For a fixed cohort, divide original cohort members who complete the defined return event in the chosen interval by the eligible original cohort size, then multiply by 100. Label whether the interval is exact, bracketed, or on-or-after.
对于固定同期群，用所选区间内完成既定回访事件的原始成员数除以符合条件的原始同期群人数，再乘以 100。必须标明区间是精确时间、区间还是“该时间点及之后”。
What is a good user retention rate?
什么样的用户留存率算好？
There is no universal good rate. Compare cohorts at the same age under the same definition, product cadence, lifecycle stage, and segment, then judge movement against your own mature baseline and business constraints.
不存在通用的良好留存率。应在相同定义、产品节奏、生命周期阶段和细分下比较同龄同期群，再根据自身成熟基线与业务约束判断变化。
How can a product improve user retention?
产品如何改善用户留存？
Define meaningful repeat value, locate the lifecycle stage where loss concentrates, combine behavioral and qualitative evidence, prioritize a specific friction hypothesis, test one change with guardrails, and compare mature cohorts before scaling.
先定义有意义的重复价值，定位损失集中的生命周期阶段，结合行为与定性证据，优先处理一个具体摩擦假设，用护栏指标测试一项改变，并在扩大范围前比较成熟同期群。
What is the difference between user retention and engagement?
用户留存与用户参与有什么区别？
Retention asks whether eligible users return over time; engagement describes the depth, frequency, or breadth of activity while they use the product. A user can return but engage shallowly, or engage intensely once and never return.
留存关注符合条件的用户是否随时间回访；参与度描述用户使用产品时活动的深度、频率或广度。用户可能回访但参与很浅，也可能一次参与很深却再也不回来。
Sources and next steps来源与下一步
This guide uses official product-analytics documentation for measurement mechanics: Google Analytics cohort exploration for inclusion and return criteria, and Adobe Customer Journey Analytics retention analysis for start events, return events, duration settings, retention, and churn. Apply these concepts to your own documented event model rather than copying vendor defaults blindly.
本指南使用官方产品分析文档说明衡量机制：Google Analytics 同期群探索用于说明纳入条件与回访条件，Adobe Customer Journey Analytics 留存分析用于说明起始事件、回访事件、时长设置、留存和流失。应把这些概念应用于自己的书面事件模型，而不是盲目复制供应商默认值。