Start with the customer path, not the home-page score
List the landing pages that attract or qualify prospects, the proof they inspect and the enquiry action they need. Test representative service, article, case-study and form pages on mobile and desktop instead of treating one home-page run as the whole site.
Record the page, device, network profile, consent state, test location, tool version and date. Without a reproducible setup, score changes can reflect test variance rather than a real improvement.
Separate field evidence from lab diagnosis
Field data reflects eligible real visits across different devices, networks, locations and behaviours. Lab data uses controlled conditions and is valuable for reproducing a problem before and after a change. The two can legitimately disagree because they measure different populations.
Use field data to understand whether real users have a problem and lab traces to identify likely causes. On a low-traffic page, acknowledge when public field data is unavailable rather than substituting one synthetic result as real-user evidence.
Read the three Core Web Vitals correctly
Largest Contentful Paint describes loading performance, Interaction to Next Paint describes responsiveness to user input and Cumulative Layout Shift describes visual stability. Each metric points to a different class of failure and should not be collapsed into one unexplained score.
Google’s current guidance identifies good targets of 2.5 seconds or less for LCP, under 200 milliseconds for INP and 0.1 or less for CLS. Treat the thresholds as experience diagnostics, not guarantees of rankings or conversions.
Diagnose slow largest content
Identify the actual LCP element on each template. Common causes include slow server response, a late-discovered hero image, oversized media, render-blocking styles, font delays and important content that is inserted only after scripts run.
Serve appropriately sized modern images, preload only genuinely critical assets, reduce unnecessary request chains, cache stable files and keep the primary message available in the initial response. Do not improve the metric by hiding useful proof.
Diagnose interaction delay
Reproduce the slow interaction: navigation opening, filter changes, form input, consent controls or another real action. Inspect long main-thread tasks, heavy third-party code, large client-side work and handlers that do more than the user action requires.
Break up expensive tasks, reduce shipped code, defer non-critical work and provide immediate visual feedback. A page that loads quickly but freezes when a customer tries to enquire still has a serious performance problem.
Stop layout movement at the source
Reserve dimensions or aspect ratios for images, embeds and dynamic modules. Keep consent banners, notices and fonts from pushing the primary action after the user has begun reading or tapping.
Test logged-out, consented, returning and error states. Layout can appear stable in a clean lab run but move when a banner, validation message, late font or personalised component arrives.
Control third-party and consent-dependent code
Inventory analytics, advertising, chat, maps, video, review widgets and session recording. Record the owner, purpose, load condition, data boundary and performance cost for each provider.
Do not load optional tracking before valid consent. After consent, delay non-critical providers until the page and primary interaction are usable. Removing an unowned or unused script can improve speed, privacy and operational reliability together.
Review persuasion and performance together
A fast page can still fail when the service is vague, proof is weak, the next action is hidden or the form asks for too much. A slower page can sometimes convert despite poor performance, but that does not make the friction harmless.
Measure whether qualified visitors reach proof, begin the form and create a stored enquiry. Keep CTA clicks and form starts separate from leads so an interface change cannot appear successful by inflating shallow interactions.
Set a performance budget around templates
Define limits for critical images, fonts, CSS, JavaScript and third-party requests on the priority templates. Add checks to release review so a future redesign, campaign or widget does not silently erase the improvement.
A budget is a decision tool, not a universal byte target. Allow an asset when its customer value justifies its cost, then optimise it and document the exception.
Release and verify the real outcome
Capture before evidence, change one clear cause where practical, run repeatable lab tests, verify the live page and watch field data as its reporting window refreshes. Retain a rollback path for changes that break layout, tracking or conversion.
Report implemented, lab-improved, field-improved and conversion-improved as separate states. A higher Lighthouse score is useful evidence of a controlled test; it is not proof that revenue or rankings increased.