A low score is not yet a useful brief
If your business website is slow on mobile, ask the developer to state three things before approving work: which observable wait will be reduced, which resource or process causes it, and how the test will be repeated afterwards. A PageSpeed or Lighthouse score can start an investigation. On its own, it does not prove that you need new hosting, no images or a complete rebuild.
Tie performance to a customer task. Waiting to read the primary service is different from waiting for a map at the bottom of the page. A heavy homepage is different from a form that freezes during submission. The goal is for the priority page to present its content and complete the agreed action under reasonable mobile conditions, without breaking useful features to improve a score.
Describe the wait you can reproduce
Record the URL, device, connection and action where the problem occurs. Check the homepage, the main service page and contact at minimum. On each one, separate four moments: useful content appearing, the layout settling, the first control responding and the task completing. This gives a supplier a reproducible problem rather than “the site takes ages”.
Keep the original report and its date. web.dev distinguishes lab data, gathered under controlled device and network settings, from field data based on real visits across varied devices, connections and locations. The two can legitimately disagree. Lab tests are useful for reproduction and debugging; field evidence helps describe the audience's actual experience when sufficient data exists.
That distinction should not become a reason to do nothing. Ask the proposal to identify whether its evidence belongs to one URL or the whole origin, mobile or desktop, and lab or field. Where only a synthetic test exists, it can provide a repeatable baseline, but the limitation should be explicit.
Worked example: a language school with video, chat and external booking
Consider a fictional language school. Its mobile homepage downloads a hero video, two font families, an external chat widget and several measurement tags. The classes page embeds a third-party booking calendar. The team says the whole site is slow, but the test reveals three separate waits: the main heading appears late, chat consumes resources before anyone uses it, and the calendar delays interaction only on the classes page.
The first proposal does not move the website to another server. It keeps a visible poster while taking video out of the initial load, creates image variants for narrower displays and reviews which fonts the opening block genuinely needs. MDN documents how responsive images let browsers select resources suited to different screen sizes and resolutions rather than making a phone download an image intended for a larger display.
The second change treats third parties according to their purpose. Chat loads when requested or after primary content; an unowned measurement tag is removed after confirming that it feeds no business process; booking remains because it supports the main action, but it is measured separately and contact is offered if the provider fails. web.dev warns that third-party scripts can affect performance and behave unpredictably, so their cost and the effect of blocking them should be checked before deciding.
The developer delivers a comparison using the same pages and lab conditions, plus a recording of the mobile journey: open classes, read times, start booking and use the alternative if the calendar does not respond. The school accepts the work when priority content appears earlier, secondary elements no longer hold up interaction, and booking and contact still work. It does not demand a perfect score or claim that the change produced more enrolments.
Match each cause to a proportionate option
When the main content appears late, the work may involve an image, a font, loading order or the server's initial response. web.dev breaks Largest Contentful Paint into delivery of the HTML, discovery of the main resource, download duration and render delay. The distinction matters: compressing an image does not fix a slow server, while changing hosts does not help an image that JavaScript discovers too late.
For image weight, compare creating suitable sizes and formats, replacing a decorative asset, or keeping it outside the initial load. For third parties, ask what each script does, when it needs to load and what happens if it is unavailable. If the server's first response dominates the wait across several pages, then caching, hosting, redirects or backend work deserve investigation.
Targeted optimisation fits when causes are local and the current foundation can support the changes. A hosting move fits when evidence points to initial delivery and the application will not recreate the delay. A rebuild is justified only when the architecture prevents the improvements from being implemented or maintained, not because an audit displayed a red indicator.
Set acceptance checks that do not reward breakage
The proposal should name the pages, conditions and protected functions. Repeat lab measurements with the same profile and several runs because one result can fluctuate. Compare field evidence later when it exists, recognising that it will not change immediately: it represents a distribution of real visits and must accumulate new experiences.
Include functional checks. Navigation must still work with touch and keyboard; the form must validate, confirm and deliver; booking or payment must retain its journey; images must not lose necessary information. Keep before-and-after evidence, but also require a change log and stated trade-offs. Removing chat can be valid when another contact route remains. Hiding the main action to speed up the homepage is not.
Avoid accepting “we will reach 100” as the sole outcome. A score depends on the tool, its conditions and the page tested. It is more defensible to agree that the main resource is discovered earlier, an identified transfer is reduced, a third party stops blocking content, or the initial response improves while the journey remains intact.
Choose the intervention with a one-page trial
Before commissioning site-wide optimisation, choose one representative page and request a short diagnostic with four columns: symptom, evidence, proposed change and acceptance check. If the change works and the same cause recurs elsewhere, it can be extended. If other templates have different causes, the quotation should separate them rather than promise one universal fix.
You may be able to resize an unsuitable image or remove an unused integration internally when you control the system, understand its dependencies and can test the complete journey. Professional help is worthwhile when reports conflict, several suppliers are involved, nobody knows which code or account loads a resource, the server is implicated, or changes could disrupt forms, booking, analytics or accessibility.
VDNX can turn the diagnostic into a design and development scope with named pages and checks. Bring the URLs, device, dated report and customer task that must remain operational. The final question is not “can you raise the score?” but “which wait will you reduce, what will you preserve, and what evidence will let me accept the change?”
Sources and documentation
- web.dev: Why lab and field data can be different ↗
- MDN: Using responsive images in HTML ↗
- web.dev: Load third-party JavaScript ↗
- web.dev: Optimize Largest Contentful Paint ↗
Content created with AI assistance. Check the sources, date and scope of each guide.