Read the distribution before the score
The ‘good’ thresholds are LCP at or below 2.5 seconds, INP at or below 200 milliseconds, and CLS at or below 0.1, generally evaluated at the 75th percentile. Averages can hide recurring failures on slower devices and networks.
For low-traffic URLs, start with template-level groups rather than unstable page-level numbers. Separate landing, detail, and conversion flows to see where a small improvement can have the largest user impact.
LCP is a delivery path, not one number
If the LCP element is discovered late, image compression alone will not solve it. Split the timeline into server response, resource discovery, priority, transfer, and render delay, then fix the longest phase.
Make hero images discoverable in initial HTML and provide candidates sized for their rendered area. Avoid lazy-loading critical images or constructing their URLs only after client rendering.
- If TTFB is long, inspect caching, server work, and redirects first
- For discovery delay, inspect HTML, preload, and fetch priority
- For render delay, inspect CSS, fonts, and main-thread work
Find the shared cause behind slow interactions
INP summarizes interaction responsiveness across a visit, not the speed of one click. Reproduce slow inputs through event processing and all the way to the next painted frame.
Break up long JavaScript tasks and update only the state needed immediately after input. Keep third-party scripts, large DOM mutations, and synchronous layout reads from converging in the same interaction window.
Prevent CLS with space-reservation rules
Give images and embeds dimensions or an aspect ratio, and reserve space for late banners. Font swaps can change line breaks and shift large regions, so align fallback font metrics as well.
Do not inject experiments or personalized modules above existing content. Use overlays or reserved slots, and distinguish expected movement after user input from automatic shifts.
Validate one template at a time
Record the baseline, deployment SHA, target URL group, and expected effect. Avoid bundling unrelated changes that affect the same metric during the observation window.
Use lab data to catch regressions quickly and field data to judge outcomes. When they disagree, investigate device, geography, cache, and interaction conditions instead of discarding either dataset.
Pre-release checklist
- ✓ Did you check 75th-percentile field data and sample size?
- ✓ Did you isolate the most important journeys and URL templates?
- ✓ Did you split LCP into discovery, transfer, and rendering phases?
- ✓ Did you record long tasks and rendering delay behind slow INP interactions?
- ✓ Did you reserve space for images, embeds, and conditional UI?
Frequently asked questions
Does a better Lighthouse score guarantee passing Core Web Vitals?
No. Lighthouse is a controlled lab measurement, while Core Web Vitals assessment uses a field distribution. Use Lighthouse for diagnosis and regressions, and field data for outcomes.
Is LCP always the first priority?
No. Prioritize by failure rate, user value, and tractability. Severe INP on checkout may matter more than a mildly slow LCP on an editorial page.