Operational article · published
Use Field and Lab Data Together for Performance Diagnosis
Decide whether a reported problem is broad, route-specific, device-specific, or only reproducible under a test profile. Includes a worked example and a release checklist.
Reviewed 2026-11-18 · Chrome and Google guidance, fictional exampleThe task and the failure mode
Built for: Developers, designers, analytics owners, and site leaders improving real-user speed and responsiveness without removing useful proof or functionality. This guide is for the person who must decide whether a reported problem is broad, route-specific, device-specific, or only reproducible under a test profile and leave a decision trail that implementation, editorial, analytics, or operations can review.
A green Lighthouse performance score does not mean that mobile Core Web Vitals pass. If the lab run looks fast but mobile field LCP is slow, first confirm that the field result is for the same page and device class. A field percentile describes a distribution of real Chrome experiences over time. A lab run describes one browser under controlled conditions. Use the field result to decide which experience needs attention, then use a trace to test one plausible cause.
Decision brief
Save the PageSpeed Insights URL or origin, device class, metric, and collection window. An origin fallback leaves this page’s field result unknown. Use first-party real-user monitoring when it has enough comparable observations.
For repeated lab runs, pin the requested and final URL, release, region, browser, phone profile, CPU, network, resource cache, and session state. Report the median and range so one fast run cannot represent the set.
Vary one condition the reference run may miss. Compare the same LCP element and its timing. For responsiveness, reproduce a named interaction. Total Blocking Time can identify main-thread work to investigate; it is not Interaction to Next Paint.
Questions to answer before changing the system
- 01Does the field panel actually name this page, or does it report the origin? Is it the mobile result?
- 02Did the field window include traffic served before the last release? What does the matching window cover?
- 03Which real visits might differ from the lab profile in device speed, network, region, cache, navigation, or session state?
- 04Does the same element become LCP in the runs being compared? Which phase accounts for most of the difference?
- 05What one change would reproduce the problem under controlled conditions, and what result would make that explanation less plausible?
Workflow
- 01Record the mobile URL-level LCP p75 and its collection dates if available. PageSpeed Insights reports the previous 28 days of eligible Chrome experiences. Compare with the origin separately if useful. A new or lightly visited page can lack enough eligible samples. Without a URL result, record unknown for this page.
- 02Record what visitors need to see or do. Repeat a cold-cache navigation on the same public URL and release with one fixed profile. Record the response, redirects, LCP element, and first-visit or returning state.
- 03Break the trace into server response, resource discovery delay, resource load duration, and element render delay. If a different element becomes largest, explain the change before comparing phases.
- 04Choose one variable supported by the trace or visit context. In the example, repeat the same cold runs with only the initial consent-choice state changed. Keep the page, region, cache settings, device, network, and code constant. Do not trigger production tracking to reproduce a timing condition.
- 05Compare the element and timestamps across both conditions. A repeatable delay can justify an isolated fix in that flow. If the delay does not repeat, leave the cause open and check another condition separately. Do not infer what share of real visits used a particular condition from the CrUX percentile.
- 06Test the candidate’s content, primary action, and privacy choices in both conditions. Exclude unrelated release work from the comparison and record the deployment time.
- 07After release, separate early lab verification and any release-tagged first-party measurements from the rolling Chrome field result. Compare matched URL and mobile windows once the reported later window falls entirely after release. If the URL has no reliable matching field result, leave field impact unverified.
Evidence to retain
- The requested page, final URL, and exact page or origin entity shown by the field report.
- The device class, collection dates, retrieval date, and the reported percentile and metric. Keep the origin value in a separate row.
- The exact lab build, runtime profile, region, cache and session setup, and raw values from each recorded run.
- The LCP element, its request start and finish, server response time, and render time in the compared traces.
- The one changed condition, the settings held constant, and a trace or run that would contradict the proposed explanation.
- The candidate recheck in both conditions, including the primary action and the consent boundary where it applies.
- The deploy time, the later comparison window, and any missing or mixed field coverage that limits the conclusion.
Worked example: a fast lab run and slow mobile field LCP
- Field result
- For a fictional October 1–28, 2026 collection window, mobile PageSpeed Insights URL-level LCP p75 for https://example.com/services/drain-cleaning is 4.4 seconds. The mobile origin result for https://example.com is 2.2 seconds over the same window. LCP above 4 seconds is in the poor range. The origin result does not replace the page result.
- Controlled reference
- The same URL gives 2.1, 2.2, and 2.4 seconds in three illustrative cold-cache runs. Each uses the same emulated phone, throttling, region, browser, and release. That fixture starts with the consent refusal already recorded. The lab median is 2.2 seconds for that condition.
- One trace variable
- Keep the resource cache cold and change only the initial consent state to no stored choice. The three illustrative runs now show 3.9, 4.1, and 4.2 seconds. The same hero image is the LCP element, and its request starts later. Server response and image transfer are similar. The trace suggests delayed image discovery in that condition.
- Decision
- Test making the public hero discoverable in the initial response without waiting for the consent interface to initialize. Keep consent and tracker behavior intact. Re-run both conditions, confirm the quote action and consent controls still work, and verify that the image request starts earlier.
- After release
- Record the deployment time. Use matched lab checks for immediate feedback. Compare mobile URL-level CrUX again when a full reported 28-day window falls after the release and URL data remains available. If scope falls back to the origin or the window still mixes releases, keep the field outcome open.
Field-to-lab diagnosis record release checklist
- Field and lab results are labeled with their different scopes and time windows.
- An origin fallback is recorded as origin evidence and the exact page result stays unknown.
- The lab sample records each value, the median, and the range.
- A named LCP element and phase explain the next test.
- The comparison changes one condition while retaining the same build and profile.
- A candidate fix is checked in the baseline and slower states, including the visitor action.
- Any later field comparison uses the same URL and device class and documents whether the windows include prerelease visits.
- Consent behavior stays intact in both states.
- TBT is not substituted for field INP.
- Missing URL-level field data stays unknown.
What to measure—and what it does not prove
- Report the lab LCP median and range per condition with the number of runs. Record the phase and element that moved.
- Track the exact URL’s mobile LCP p75 in a separate field table. Name the source, window, and scope every time.
- Record whether the page content and primary action work in the tested states and whether the candidate preserves consent.
- Keep field improvement pending until comparable postrelease evidence exists. Rankings and sales need separate evidence.
Boundaries and caveats
The example is fictional. Its numbers illustrate a decision, with no measurements from VOID or a client site.
The 75th percentile is the value at or below which about 75% of eligible real-user observations fall. It does not describe the typical lab run or identify which users, visit conditions, or code changes created the slower results.
A first-party performance sample has its own coverage and collection rules. Record those before comparing it with CrUX.
When looking at responsiveness, a load trace can reveal a long task. Field INP measures interaction latency on eligible page visits, so a good TBT result alone cannot settle an INP concern.
Primary sources
- web.dev: Why lab and field data can be differentweb.dev
- Google: About PageSpeed Insightsdevelopers.google.com
- Chrome: CrUX methodology and eligibilitydeveloper.chrome.com
- web.dev: Web Vitals and percentile thresholdsweb.dev
- web.dev: Optimize Largest Contentful Paintweb.dev
- web.dev: Total Blocking Timeweb.dev
- web.dev: Interaction to Next Paintweb.dev
- Chrome: Lighthouse performance scoringdeveloper.chrome.com
Start with one bounded case
Pick one page where field and lab results disagree. Save the report scope, record a reference profile, and compare one controlled trace condition. Carry the proposed mechanism and the date of your next field check into the release review.