# Mobile Website Performance: Why a 96 Score Didn’t End Our Audit

A green PageSpeed score is a useful result. The next run can still tell a different story.

On 2 November 2026, Pakistan time, a mobile performance review of [StackEndBox’s homepage](https://www.stackendbox.com/?utm_source=hashnode&utm_medium=article&utm_campaign=mobile_performance_review) produced a score of 96. A second live PageSpeed Insights run scored 81. Both tested the same deployed release.

Keeping both reports gives a website owner a better basis for a decision than keeping the more flattering screenshot.

## What the two mobile tests actually showed

| Metric | First run | Repeat run |
| --- | --- | --- |
| Performance score | 96 | 81 |
| First Contentful Paint | 0.9 seconds | 2.3 seconds |
| Largest Contentful Paint | 2.7 seconds | 4.4 seconds |
| Total Blocking Time | 60 milliseconds | 80 milliseconds |
| Cumulative Layout Shift | 0.008 | 0.008 |

These were Lighthouse lab results from PageSpeed Insights under its mobile test conditions. The first report also scored 100 for desktop performance. Accessibility, best practices and SEO scored 100 in both mobile reports.

You can inspect the [first report](https://pagespeed.web.dev/analysis/https-www-stackendbox-com/cl491u9qyq?form_factor=mobile) and the [repeat report](https://pagespeed.web.dev/analysis/https-www-stackendbox-com/pmibc4mycc?form_factor=mobile).

The reports had no Chrome User Experience Report field data. They therefore could not establish that real visitors were passing Core Web Vitals. Google’s explanation of [lab and field data](https://web.dev/articles/lab-and-field-data-differences) is useful when a supplier presents one test as proof of everyone’s experience.

## Make the first image the right size

The deployed update added a 640-pixel hero image between the existing 480- and 768-pixel variants. The image source set, preload and client-side image helper used the same size list.

The assembled 640-pixel image was 35,290 bytes. Its 768-pixel counterpart was 46,188 bytes, a 23.6% reduction. The alternative hero image dropped from 44,840 to 34,608 bytes at the new size.

A browser test at a 412-pixel viewport width and 1.75 device-pixel ratio selected and decoded the 640-pixel asset. That matters more than merely generating a smaller file and leaving the page to request the old one.

Those byte savings are directly measurable. They do not explain the entire difference between the two PageSpeed scores. Two live runs are too little evidence to isolate every cause of timing variation.

## An image can finish downloading before it becomes visible

The slower run’s diagnostic breakdown showed a short image download and a longer delay before the browser displayed it. That directs the investigation toward rendering, styles and script startup as well as image compression.

Our checks also covered a normal mobile screen, an unusually tall mobile viewport, desktop and JavaScript-disabled rendering. The heading, description and calls to action needed to remain visible without horizontal overflow.

This is an important part of a mobile website speed audit: a faster request is only useful if the page still shows the right content and people can use its navigation.

## A small improvement can be the wrong release

A separate cloud-browser experiment delayed the main page module until the hero image and fonts were ready. Two alternating comparisons saved 172 and 84 milliseconds of Largest Contentful Paint, averaging 128 milliseconds.

One sample also increased layout shift slightly. Delaying the module introduced an interactivity trade-off that needed more evidence than a modest timing gain.

We left that experiment out of production. A performance project should be able to explain which changes were rejected, as well as which ones shipped.

## Verify the deployed page, not just the build

A release check exposed another useful distinction. The browser response included a Cloudflare analytics beacon that the server-side verifier did not receive. Hashing those two responses produced different values.

The fix was to compare the actual responses and recognize the exact reviewed previous server response. The new release still had to pass its HTML and asset checks. Disabling the check would have hidden the cause.

For a website owner, the practical question is simple: has the team verified the page that visitors now receive?

## Follow-up: smaller fonts with full character fallback

The next production release reduced the two preloaded font files from 89,956 to 80,212 bytes, a 10.8% reduction. It uses a narrower first-screen character set while keeping the complete original fonts available for characters outside that set. New filenames prevent an old immutable cache entry from being mistaken for the new file.

Build checks verified glyph advance widths, vertical metrics, variable-font axes and font metadata. Eight candidate browser scenarios passed, including an explicit extended-Latin test that loaded the unchanged full font. The release then passed staging and production checks.

That is evidence of a smaller payload and compatible rendering. It is not evidence that the entire LCP problem has disappeared. A font change should be judged alongside fresh mobile measurements, without treating every difference between two runs as an effect of that change.

A post-release [main-URL report](https://pagespeed.web.dev/analysis/https-www-stackendbox-com/1hvu8bo14k?form_factor=mobile) scored 72 with a 7.6-second LCP. A separate [query-parameter diagnostic](https://pagespeed.web.dev/analysis/https-www-stackendbox-com/1oxt00lsv0?form_factor=mobile) scored 81 with a 4.5-second LCP. The second URL included a performance-check parameter to obtain a fresh capture after a repeated main-URL request reused its earlier capture. Different test and cache conditions prevent a clean causal comparison. Neither result supports calling the mobile speed issue resolved.

## What to ask for in a performance handover

Before signing off, ask for:

- Dated mobile reports, including a repeat result and the tested URL.
- The exact images, styles or scripts changed, with measurable evidence.
- Checks that menus, calls to action and the enquiry journey still work.
- A clear distinction between lab measurements and real-user field data.
- Remaining limitations and a way to recover the previous release.

The initial results varied, and the later measurements mean the work cannot honestly be described as consistently finished. That is useful information for deciding the next step.

A website performance review earns its value by making the next decision clearer, not by selecting the greenest screenshot.

---

Muhammad Abdullah is co-founder of StackEndBox. For teams planning changes to an existing site, our [website redesign and development service](https://www.stackendbox.com/services/website-redesign-development?utm_source=hashnode&utm_medium=article&utm_campaign=mobile_performance_review) starts with the current site, the important customer journeys and a bounded scope.

Cover photo: Myriam Jessier, via Unsplash.
