Measure representative pages and interactions, identify the actual bottleneck, compare a controlled change and monitor real-user Core Web Vitals separately.
Select important page types and tasks, including navigation, search, forms or checkout as appropriate. Record the build version, device, browser, network, cache state and test location. Review available field data and its coverage, then choose local test conditions that answer the observed problem. Define the performance budget you intend to compare.
Run repeatable lab captures and preserve their reports, network waterfalls and traces. Inspect available field LCP, INP and CLS for the relevant device segment and period. A navigation-only Lighthouse run cannot measure field INP; exercise actual interactions and inspect their timings. Keep missing field data distinct from a passing result.
Identify the actual LCP element and its request/render path, the work around slow interactions and the sources of layout shifts. Inspect transferred versus decoded resource sizes, dependencies and cache behavior. Use the header parser for one pasted response block when examining Content-Encoding or Cache-Control; it does not measure transfer time or effective caching by itself.
Choose a change supported by the trace: for example, improve discovery of the LCP resource, remove unused work, split a long task or reserve space for content. Avoid applying compression quality, preload or cache rules universally. Verify that images remain acceptable and navigation, forms and other critical interactions still work on the production artifact.
Repeat the same pages and interactions with the same declared conditions and enough runs to understand variation. Compare metric values and distributions rather than only a single composite score. Use percentage arithmetic for a defined before/after quantity with matching units; a percentage improvement in one lab run is not a field-performance or retention result.
Save the original issue, trace evidence, change, checks and remaining uncertainty. After the normal release process, review real-user measurements over their stated reporting period and device segments. Monitor regressions in the affected flow and keep a rollback path. Download the audit record with direct references to the captured reports.
Loading your checklist…
The current Core Web Vitals are LCP, INP and CLS. Evaluate field performance at the 75th percentile by device segment; lab testing is useful for diagnosis but does not replace field evidence.
Keep these task-specific records with the tested version and review date.
| Record | Include | Verify |
|---|---|---|
| LCP | Good: ≤2.5 seconds | 75th-percentile loading experience |
| INP | Good: ≤200 milliseconds | Real interactions; navigation-only lab runs do not measure it |
| CLS | Good: ≤0.1 | Visual stability across the measured experience |
Keep these task-specific records with the tested version and review date.
| Record | Include | Verify |
|---|---|---|
| Conditions | Build, device, network, cache and task | Comparable between runs |
| Evidence | Trace, waterfall, field coverage and dates | Supports the identified cause |
| Outcome | Before/after values, variation and functional checks | Lab results and field outcomes reported separately |
It may be text, an image or another supported element; identify it before optimizing a presumed hero image.
A fast first paint can still lead to a slow or broken task after the user clicks.