リソース · 39

フィールドデータでコアWebバイタルを診断する

修正前に、LCP、INP、CLSを実際の要素とインタラクションに結び付ける。

更新済み · 16 min

真ちゅうの台に置いた電話、抽象的な波形、光学機器、えんじ色のケースを備えた架空の無人サイバー測定台 AIで生成した編集用イラスト

このガイドが達成するのに役立つこと

  • 比較可能なセグメントを選択する
  • 要素またはアクションを特定する
  • 再現する
  • 変更を確認する

クイックチェック

  • 信号は現場からのものか、ラボからのものか?
  • 測定対象となるデバイスとページはどれか?
  • LCPはどの要素か?
  • INPに寄与するアクションはどれか?
  • レイアウトを変更する追加要素はどれか?

手順

  1. 1

    母集団を定義する

    比較前にページ、テンプレート、デバイス、期間、ジャーニーを分離する。 トラフィックまたはコンテンツの範囲と変更を記録する。

    成果物:測定範囲と制限。

  2. 2

    主要な症状を特定する

    LCP、INP、CLSをそれぞれ異なるエクスペリエンスとして扱います。 サイト全体の平均値ではなく、それぞれの分布を検証します。

    成果物:優先順位付けされた影響を受けるセグメント。

  3. 3

    実際のコンテンツと関連付けます。

    LCPについては、要素とネットワークまたはレンダリングの遅延を特定します。 CLSについては、要素のシフトを特定します。 INPについては、インタラクションとメインスレッドの処理を特定します。

    成果物:トレース付きの要素とアクション。

  4. 4

    ラボで再現します。

    影響を受ける状況に近いシナリオを構築します。 ラボの出力を現場の証拠として扱わずに、開発者ツールを使用して仮説を検証します。

    成果物:再現手順と考えられる原因。

  5. 5

    原因を修正します。

    証拠に基づいて、リソースの優先順位付け、ディメンションの確保、またはブロッキング作業の削減を行います。 機能およびアクセシビリティの回帰テストも実施します。

    成果物:対象を絞った変更とローカル比較。

  6. 6

    リリース後のフォローアップ

    同一ページ、デバイス、適切な期間を比較する。 サンプル制限と母集団の変化を文書化する。

    成果物:現場でのフォローアップと次の決定事項

管理指標

指標測定対象最初のアクション
LCP実際の移動における主要要素の表示負荷とレンダリングの検査
INP実際のインタラクションへの応答ブロッキング作業の削減
CLSレイアウトの安定性スペースの確保
網羅性解釈可能なサンプルを含むセグメント測定範囲の拡張

よくある間違い

  • 1つのLighthouseスコアから現場の経験を推測する
  • モバイルとデスクトップの混在
  • 要素の位置を特定せずに平均値を最適化する
  • トラフィックの変化に関わらず修正をクレジットする

よくある質問

ラボと現場の結果が異なるのはなぜか?

ネットワーク、デバイス、キャッシュ、コンテンツ、および実際の動作は、テストシナリオと異なる場合があります。

優れたLCPは優れたINPを保証するのでしょうか?

いいえ。読み込みとインタラクションの応答性は、異なる時点を測定します。

修正はいつ確認されますか?

ボリュームやその他の変化を考慮し、同等のセグメントで十分な比較を行った後。

参照資料:データ・用途・限界

資料確認日 .

この言語版の詳しい参照メモは英語で提供しています。

参照資料:データ・用途・限界
出典と目的確認済みのデータ・指標利用方法留意する限界
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.

手法を詳しく読む →

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

Core Web Vitals:しきい値と実測データ

この言語版の詳しい参照メモは英語で提供しています。

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

ローカルp75値リーダー

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割り当て:データとSRM

この言語版の詳しい参照メモは英語で提供しています。

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

公式参照文献

参照文献はメソッドを裏付けるものです。 チェックは状況に合わせて調整してください。 これらは認証ではありません。 元の参照文献のタイトルやソース文書は別の言語で書かれている場合があります。

参照資料の確認日 .