资源 · 39

在现场数据中诊断核心网页指标

在修复之前,将 LCP、INP 和 CLS 与实际元素和交互关联起来。

已更新 · 15 min

虚构的无人网络测量台,配有黄铜支架上的手机、抽象曲线、光学仪器和酒红色盒子 由人工智能生成的编辑插图

本指南有助于实现的目标

  • 选择可比片段
  • 定位元素或操作
  • 复现
  • 验证变化

快速检查

  • 信号来自现场还是实验室?
  • 哪些设备和页面构成测量值?
  • 哪个元素是 LCP?
  • 哪个操作导致了 INP?
  • 哪些添加项改变了布局?

分步指南

  1. 1

    定义目标人群

    在比较之前,将页面、模板、设备、时间段和旅程分开。 记录流量或内容的覆盖范围和变化。

    交付成果:测量范围和限制。

  2. 2

    找出主要症状

    将 LCP、INP 和 CLS 视为不同的体验。 检查它们的分布情况,而不是使用全站平均值。

    交付成果:确定受影响的优先级部分。

  3. 3

    将其与实际内容关联

    对于 LCP,找到元素以及网络或渲染延迟。 对于 CLS,找到移动的元素。 对于 INP,定位交互和主线程工作。

    交付成果:带有跟踪信息的元素和操作。

  4. 4

    在实验室中重现问题

    构建一个接近受影响情况的场景。 使用开发者工具测试假设,而不是将实验室输出视为现场证据。

    交付成果:重现问题和可能的原因。

  5. 5

    修复原因

    根据证据确定资源优先级、预留维度或减少阻塞工作。 同时测试功能和可访问性回归。

    交付成果:针对性更改和本地比较。

  6. 6

    发布后跟进

    比较相同的页面、设备和合适的时间段。 记录样本限制和总体变化。

    交付成果:现场跟进和后续决策。

管理指标

指标衡量内容首要行动
LCP实际旅程中的主要元素显示检查负载和渲染
INP对实际交互的响应减少阻塞工作
CLS布局稳定性预留空间
覆盖面具有可解释样本的片段扩展测量

常见错误

  • 从单个 Lighthouse 分数推断现场体验
  • 混合移动端和桌面端
  • 在不定位元素的情况下优化平均值
  • 即使流量发生变化,也应归功于修复

常见问题解答

为什么实验室和现场测试结果不同?

网络、设备、缓存、内容和实际行为可能与测试场景有所不同。

良好的 LCP 是否能保证良好的 INP?

不能。加载速度和交互响应速度衡量的是不同的阶段。

何时确认修复?

在对等效段进行充分比较后,考虑流量和其他变化。

参考资料:数据、用途与局限

文档核查日期 .

此语言版本的详细参考说明以英语提供。

参考资料:数据、用途与局限
来源与目的已核实的数据或指标使用方法需要保留的局限
Google · Core Web Vitals

Distinguish loading, responsiveness and visual stability.

Good thresholds are LCP ≤ 2.5 s, INP ≤ 200 ms and CLS ≤ 0.1 at the 75th percentile, segmented by mobile and desktop. Separate field measurement from laboratory tests. Record period, URL, device, population and sample size before comparing changes.

深入了解方法 →

These thresholds are not measured results for this site. A good lab score alone does not establish visitor experience.

Core Web Vitals:阈值和实测数据

此语言版本的详细参考说明以英语提供。

A useful Core Web Vitals diagnosis starts with a measurement whose scope is known. Identify page or origin, device, period and source before discussing a fix. A page may load quickly in a local run yet remain difficult for part of its audience. This dossier helps interpret values without confusing field results, lab traces and search promises.

Google sources consulted on 23 October 2026. Thresholds follow PageSpeed Insights; examples are fictional CIE data, never measurements of this site. The proposed checks form an acceptance method, not a ranking guarantee.

Google · About PageSpeed Insights · Living documentation · consulted 2026-10-23

Three metrics with explicit thresholds and units

Three metrics with explicit thresholds and units
MetricGoodNeeds improvementPoorReading and unit
LCP≤ 2500 ms> 2500 ms · ≤ 4000 ms> 4000 msLoading: appearance of the largest eligible content element. Tie the signal to the observed element before blaming every image.

Google PSI

INP≤ 200 ms> 200 ms · ≤ 500 ms> 500 msInteraction responsiveness. An action trace helps separate delay before handling, execution and presentation; a loading score does not replace that investigation.

Google PSI

CLS≤ 0.1> 0.1 · ≤ 0.25> 0.25Visual stability: a unitless score. Find what shifts elements, then reserve space or fix insertion; do not convert this score to milliseconds.

Google PSI

Eight reading points to retain with the measurement

Eight reading points to retain with the measurement
PointWhat to retainDecision consequence
p75The 75th-percentile value for each metric.This percentile is neither the mean nor the slowest visit. Do not average two population p75 values to invent their combined percentile.
DeviceThe PHONE, DESKTOP or TABLET dimension, or its absence.Omitting a CrUX API dimension aggregates devices. Compare the same device before and after a fix.
URL or originThe identifier actually returned by the source.PSI can show origin data when a URL lacks data; do not attribute that result solely to the opened page.
PeriodThe collection dates for the inspected result.The CrUX API aggregates the preceding 28 days and updates daily. Adjacent dates do not represent independent cohorts.
Missing INPThe absence of a measurement and its associated message.PSI permits its exception when LCP and CLS are good. This does not establish fast interactions.
CoverageCrUX eligibility and population limits.A popularity threshold exists but its exact value is not published. Do not invent a minimum visit count to make a report available.
HistogramBucket fractions and the returned p75, separately.CrUX derives p75 from available data, not just the three displayed buckets; do not reconstruct it exactly from those buckets.
Measurement contractSeparate entered values, source data and diagnosis.This local reader classifies supplied values. It runs neither Lighthouse nor CrUX and observes no visitor.

Chrome for Developers · CrUX API · Living documentation · consulted 2026-10-23

Chrome for Developers · CrUX methodology · Living documentation · consulted 2026-10-23

本地p75数值阅读器

Enter p75 values for the same device, period and scope. LCP and INP use milliseconds; CLS is unitless. A blank field means unavailable data, never zero. Dot or comma decimals are accepted, up to four decimal places, without thousands separators. Input caps of 600000 ms and CLS 1000 belong to this tool; they are not official metric limits. Fine fractions in the examples test the reader boundaries; they do not reproduce CrUX report output precision.

Local reading using PageSpeed Insights thresholds and its INP exception. No site test, remote call, stored history, percentile calculation or seconds/milliseconds conversion. The reader cannot verify whether values are comparable: record their provenance before drawing conclusions.

Eight fictional examples, accessible without JavaScript

V01 · Three values exactly on good boundaries
LCP
2500 ms
INP
200 ms
CLS
0.1

Assessment passed for the entered values under the PSI rule.

Equality remains good: 2,500 ms, 200 ms and 0.1. No preliminary rounding moves the comparison.

V02 · A fraction above the LCP threshold
LCP
2500.0001 ms
INP
200 ms
CLS
0.1

Assessment not passed for the entered values under the PSI rule.

LCP needs improvement. Rounding its display to 2,500 does not justify reclassifying it as good.

V03 · Upper boundaries of the middle range
LCP
4000 ms
INP
500 ms
CLS
0.25

Assessment not passed for the entered values under the PSI rule.

All three values need improvement. Poor starts strictly above these boundaries.

V04 · Three upper-boundary exceedances
LCP
4000.0001 ms
INP
500.0001 ms
CLS
0.2501

Assessment not passed for the entered values under the PSI rule.

All three metrics are poor. Investigate loading, interactions and layout shifts separately.

V05 · INP unavailable in PageSpeed Insights
LCP
2000 ms
INP
Data unavailable
CLS
0.05

Assessment passed for the entered values under the PSI rule.

INP unavailable: PSI exception applied. Interaction responsiveness remains unknown.

PSI permits a passed assessment when LCP and CLS are good despite missing INP. This proves no interaction responsiveness: the exception remains visible.

V06 · LCP absent, two other metrics present
LCP
Data unavailable
INP
150 ms
CLS
0.05

Assessment unavailable: LCP or CLS missing.

This scope cannot be assessed under the PSI rule. Missing data is neither zero nor a poor result.

V07 · CLS slightly above 0.1
LCP
2000 ms
INP
150 ms
CLS
0.1001

Assessment not passed for the entered values under the PSI rule.

LCP and INP remain good, but CLS needs improvement. Averaging the three numbers makes no sense: their units and experiences differ.

V08 · Unsupported thousands separator
LCP
2 500 ms
INP
150 ms
CLS
0.05

Correct invalid inputs before any assessment.

Input is rejected. Enter 2500 for 2,500 ms; comma or dot is reserved for a decimal separator. No performance assessment is inferred from invalid input.

Choose evidence suited to the question

Choose evidence suited to the question
SourceUseful questionReading limit
CrUX APIWhat p75 is returned for this identifier, device and period?Aggregated data with conditional coverage; missing reports are not proof of failure.
PageSpeed Insights · fieldWhat does available experience describe for this URL or origin?Check the scope actually shown, missing metrics and dates.
Lighthouse · labWhich hypothesis can I reproduce in a controlled scenario?One run and its score do not prove the field outcome for all users.
Service-owned RUMWhich component or journey explains the issue in our instrumentation?Document method, population and collection differences before comparing with CrUX.

Two 28-day windows overlap substantially

Fictional calendar example: 1–28 September and 2–29 September share 27 days. This does not mean 27/28 of visits are identical: traffic varies. A recent fix initially occupies only part of the window; record its date and follow the same segment rather than promise an immediate switch.

Six acceptance cases tied to real journeys

Six acceptance cases tied to real journeys
CIE caseObservation to collectJustified action
P-IMAGE · Main contentLCP element for the affected template, request order and reserved dimensions.Prioritize the right media and check rendering; do not prioritize every image.
P-INTERACTION · Opening a recordKeyboard or touch action, handling trace and visible feedback.Reduce the work causing delay without removing state announcements or control behavior.
P-STABILITY · Late inserted blockContent shift and control position before and after insertion.Reserve space or revise insertion; verify understandable focus and touch targets.
P-SCOPE · Sparse URL dataRequested URL and origin actually displayed by PSI.Present the result with its exact scope and add a template-specific observation.
P-RELEASE · Recent fixChange date, collection dates and compared population.Separate immediate local validation from field follow-up; record other traffic or content changes.
P-REGRESSION · Removed third-party scriptPerformance measurement and business journey after removal, including consent and forms.Retain the gain only if expected functions, permissions and accessibility remain verified.

Core Web Vitals document part of the experience; they certify neither accessibility, security nor editorial quality. Improvement guarantees no search ranking. For useful decisions, connect each number to an element or action and retain functional journey checks.

web.dev · Web Vitals · Living documentation · consulted 2026-10-23

Google Search Central · Core Web Vitals · Search guidance · consulted 2026-10-23

Provenance, date and credits

Google documentation consulted 2026-10-23. CIE: original explanations, fictional examples and local threshold reader.

Google · About PageSpeed Insights · HTML SHA-256: 3ca78c96f9df1778aed6aaed634df06b534489176ce8b7824f870661f15e4c26

web.dev · Web Vitals · HTML SHA-256: 024989d5bce351f2a95be14d3a03cba110e5440d055ad5087425a72890ee754a

Chrome for Developers · CrUX methodology · HTML SHA-256: 825f49d646eb05c476a6ec2fbd3a9e12e094e081286489c8ab8def43b9627157

Chrome for Developers · CrUX API · HTML SHA-256: 814c2f1ec4e9cd14028ab30f196f1973ed3722af50fe98f2668e6be5f7338641

Thresholds and reading rules adapted from the cited Google documentation under Creative Commons Attribution 4.0 unless otherwise noted. Original CIE explanations, fictional examples and local reader; no Google endorsement is implied. CC BY 4.0

Google Developers · Site Policies · Documentation licensing and attribution

A/B分配:数据与SRM

此语言版本的详细参考说明以英语提供。

Compare populations before comparing metrics

A measurement change can alter who appears in the analysis. Check counts at each stage before attributing a difference to the interface.

  • Name the unit: user, device, session and event are not interchangeable.
  • Compare telemetry losses and exclusions in each group.
  • Keep scope, window and configuration changes with the result.

Explore the allocation check and its limitations →

Microsoft Research · Diagnosing Sample Ratio Mismatch · Published 2020-09-14 · consulted 2026-10-23

NIST/SEMATECH · Binomial distribution · e-Handbook §1.3.6.6.18 · consulted 2026-10-23

官方参考文献

参考文献支持该方法。 请根据您的实际情况调整检查;它们并非认证。 原始参考文献标题和源文档可能使用其他语言。

参考资料查阅日期 .