01 / The score is not the goal
I like a good score. I just do not want to work for one.
There is something rather satisfying about watching a Lighthouse score change from red to green. You make a few improvements, run the test again, and there it is. A nice big number telling you that your work has paid off.
I understand the attraction. Numbers give teams something to discuss. They can help us identify problems, check whether a change made a difference and show progress without requiring everyone to read through a technical report.
But I think it is surprisingly easy to make the number the objective instead of the experience it is supposed to represent. We start asking how to get a website from 82 to 95 rather than why someone on an older phone is struggling to open a page, complete a form or find the information they came for.
For a public service, health information website, community organisation or any product people rely on, that distinction matters. Someone might be using a slow connection. They might be under pressure. They may be trying to complete a task between other commitments. They do not care how attractive our dashboard looks when their Submit button refuses to respond.
My view is simple: use the metrics to understand the experience, not the experience to improve the metrics. That sounds obvious until somebody proudly announces a score of 100 while the actual task still takes six confusing steps.
02 / What Lighthouse measures
Think of Lighthouse as a very useful inspection, not a final verdict.
Lighthouse is an automated auditing tool built into Chrome DevTools and used by services such as PageSpeed Insights. It can investigate loading performance, accessibility-related checks, common web development practices and search fundamentals.
It is particularly valuable during development because you can run an audit, make a change and run it again under similar conditions. It points you towards the things worth investigating: large images, slow-loading resources, JavaScript doing too much work, or elements that move around while the page loads.
But a standard Lighthouse page-load run is a lab test. It works within particular simulated device and network conditions. It does not follow every visitor through every page on their actual phone. Nor does it decide whether the information is useful, the navigation makes sense or a person can finish an important task without help.
A score may also change between runs. Browser state, test settings, page content, third-party resources and the wider environment can affect the result. Run comparable tests, record the conditions and look for consistent patterns rather than treating one reading as unquestionable truth.
I also make a point of distinguishing a real audit from an illustrative demonstration. In my Website Performance Training Lab, the score display is intentionally simulated so people can learn how to interpret it. It is a training example, not a live scan of the website you enter. For real measurements, use Lighthouse in Chrome or PageSpeed Insights.
03 / Three signals that matter
Core Web Vitals, without having to speak fluent acronym.
Google's Core Web Vitals look at three things people tend to notice: whether the main content appears quickly, whether the page responds when they try to use it, and whether it stays where it is supposed to stay.
Imagine opening a service page and waiting for its main heading, hero image or principal content block to appear. LCP measures when the largest qualifying visible element renders. It is an indicator of loading experience, not a guarantee that every part of the page is ready.
You tap a filter or try to open a menu. Nothing happens. You tap again, wondering whether your thumb has stopped working. INP measures responsiveness across interactions during a visit, including the delay before the page visibly responds. A page can look loaded and still be frustrating to use.
You are about to press a button and an image finally arrives, pushing the button somewhere else. CLS captures unexpected visual movement over a page visit. It is a surprisingly effective way to measure a very ordinary kind of annoyance.
Notice what these three measures do not tell us. They cannot tell us whether the language is understandable, whether the eligibility rules are clear, whether an interaction works for a screen-reader user, or whether the service itself makes sense. Those are still design, content and accessibility questions that deserve proper attention.
04 / What good looks like
The thresholds are useful. The context is essential.
Google publishes recommended thresholds for a good real-world experience. The values are assessed at the 75th percentile of visits, typically considered separately for mobile and desktop. In plain language, we are not aiming for one perfect test on one fast laptop. We want most people to receive a good experience.
Good at or below 2.5 seconds. Needs improvement above 2.5 and up to 4 seconds. Poor above 4 seconds.
Good at or below 200 ms. Needs improvement above 200 and up to 500 ms. Poor above 500 ms.
Good at or below 0.1. Needs improvement above 0.1 and up to 0.25. Poor above 0.25. CLS is a unitless score, not a number of seconds.
A target is a helpful starting point for investigation. It does not replace observing people completing real tasks or evaluating accessibility.
These thresholds come from Google's published Web Vitals methodology. They provide a common reference point, not a magic certificate that a website is now good.
One technical detail is worth knowing: a standard automated Lighthouse page-load audit cannot directly measure INP, because INP depends on interactions over a real visit. Lighthouse uses Total Blocking Time (TBT) as a useful laboratory diagnostic for potentially slow interaction. TBT is not the same metric as INP. If you are investigating actual interaction delays, use field measurements and interact with the page in Chrome's performance tools. Google's INP guide explains how input delay, processing and presentation each contribute.
05 / Lab versus real users
One test tells us what happened in the room. The other tells us what happened outside.
Here is where conversations about web performance can become confusing. You open PageSpeed Insights, see a Lighthouse score, then find a different result in the real-user section. Which one should the team believe?
Both, provided we understand what each measures.
A test under defined device and network conditions. Useful for reproducing problems, checking changes before release and identifying technical opportunities. It cannot represent every visitor.
Experiences across real devices, connections and circumstances. Useful for prioritising what visitors are actually experiencing, but harder to reproduce and sometimes not available for low-traffic pages.
PageSpeed Insights draws its real-user section from the Chrome User Experience Report (CrUX), using an approximately 28-day rolling collection period. If a particular page has too little qualifying data, PageSpeed Insights may show data for the whole site instead, or no real-user data at all. That distinction matters before someone attributes a site-wide result to one particular form.
Suppose the lab test looks excellent but field data shows poor responsiveness. Perhaps the automated run did not exercise the search widget that people use most, or did not reflect the devices they use. That is a reason to investigate the journey, not to dismiss the people using it.
The reverse also happens. A controlled lab run may look much worse than the typical field experience because real visitors benefit from caching, different device capabilities or other conditions. Google's explanation of why lab and field results differ is well worth a read.
My preference is to use real-user evidence to identify priorities, then controlled testing to understand causes and verify proposed fixes. Neither view is especially useful without the other.
06 / Find the real cause
A red number is a clue. It is not a diagnosis.
Once a metric tells us where something is slow or unstable, the interesting work begins. That is when the team needs to ask why.
Check server response times, whether a main image is discovered early, image formats and dimensions, font loading, and resources that delay rendering. A huge image is an obvious suspect, but not the only one.
Investigate long JavaScript tasks, expensive event handlers, heavy rendering and third-party scripts. Watch what happens when someone actually uses the controls.
Check missing image dimensions, late embeds, banners, injected content and fonts that change text geometry after rendering. Reserve space before content arrives.
Google has practical, detailed references for improving LCP, diagnosing INP and reducing layout shifts. I find those more useful than memorising a list of tricks intended to please a scoring algorithm.
And be careful about removing useful features just to gain a few points. A map, interactive search, form or data visualisation might be central to the service. The goal should be to make it lighter and more dependable without stripping out the thing people came to use.
07 / Think about the service
The most important page is not necessarily the homepage.
Imagine a public information website with a beautifully optimised homepage. It loads quickly, scores well and makes everybody feel rather pleased with themselves. Then someone tries to find eligibility information and gets lost in a dense content page. Another person reaches the final step of a form and the confirmation button takes too long to respond.
Those are hypothetical examples, but the principle is familiar across digital projects: the parts of a service that matter most are not always the parts that receive the most attention in a presentation.
I would want to understand the journeys people actually need. Can they find the right page? Is the content understandable? Can they operate the form on a keyboard or mobile device? Does the experience remain stable while data and images load? Does a slow connection make an important step unnecessarily difficult?
Performance, accessibility, content design and service design are connected. You can make a page faster while leaving it confusing. You can simplify a journey and accidentally introduce heavy scripts. Good delivery means looking across those perspectives rather than handing each one to a different team and hoping the pieces meet nicely at the end.
And we should be clear about the limits of automated accessibility checks too. A Lighthouse accessibility result is helpful, but the W3C explains why testing tools cannot determine full accessibility on their own. A technical score and a human experience are related. They are not interchangeable.
08 / My practical review
A performance review that leads somewhere useful.
Here is how I would approach a website performance review with a multidisciplinary team. You do not need an enormous testing programme to make a sensible start, but you do need to be deliberate.
Pick the journeys that matter: finding information, using search, submitting an application, accessing support or completing a transaction. Ask what success looks like for the person.
Include a content-heavy page, a search or listing view, a form and a complex component where relevant. A homepage alone rarely tells the whole story.
Capture the page URL, device type, network or simulated settings, date, audit tool and relevant browser version. Run a few comparable tests before treating a change as meaningful.
Look for real-user Core Web Vitals where available, then use a controlled Lighthouse or DevTools test to investigate. Check whether a result applies to the URL or the broader site.
Use the individual audit findings and browser performance tools. Identify the resource, interaction or layout behaviour behind the problem, not just the label attached to its score.
A slow eligibility page or broken mobile form is usually a more important fix than a minor score improvement on a page nobody struggles with. Consider accessibility and content effects as part of the decision.
Verify the same journey after the change. Record what improved, what did not and whether anything regressed. Content, integrations and publishing habits can gradually undo good work.
I would also make sure somebody owns the follow-through. A list of 27 recommendations and nobody responsible for any of them is not an improvement plan. It is just a very well-formatted list of future regrets.
09 / Turn data into decisions
The report should help someone make a decision.
When presenting findings to a client or delivery team, I would resist leading with a page of coloured gauges. Instead, start with the journey and describe what the result means.
For example: “On a slower mobile connection, this important page's main content appears later than it should. The hero image and render-blocking resources are worth investigating. We recommend testing a lighter image and changing resource priority, then measuring the same journey again.”
That is a much more useful conversation than “Your LCP is red.” It connects evidence to a user problem, a plausible cause and a next action. It also leaves room for the developer to investigate rather than pretending a tool has already designed the solution.
A short report can often cover four questions: What did we test? What did we observe? Why does it matter to people? What should we do next? Add an owner and a way to measure the outcome, and you have something a team can actually act on.
We are not trying to turn performance into a competition between developers, designers or platforms. We are trying to give everyone enough shared understanding to improve the product together.
10 / Try it in the Lab
A little practice helps the numbers make sense.
I put together a Website Performance Training Lab for people who want to move from the theory into a practical exercise. It began life as an internal training workshop, and I have since adapted it as an interactive part of this website.
It walks through Lighthouse, Core Web Vitals, performance diagnostics, accessibility, security hygiene and how to explain the findings to somebody who does not spend their day inside browser developer tools. There is also a hands-on challenge to work through at your own pace.
One important detail: the Lab's example audit uses simulated results. They are there to help explain how the interface and reports work. To measure a live page, open PageSpeed Insights or use the Lighthouse panel in Chrome DevTools, and bring that evidence back to the exercise.
Run a real audit using Chrome DevTools or PageSpeed Insights.
Choose one issue affecting an important journey, fix it and check again.
For further reading, I recommend Google's Core Web Vitals overview, its PageSpeed Insights guide and the W3C's accessibility evaluation resources. They are good places to go when the numbers raise a question this article cannot answer.
In the end, the best measure is not whether we got a prettier score. It is whether the website has become clearer, faster, more dependable and easier for the people who actually need it.
Because a fast website is a good start. A useful website is the point.