How Lighthouse works
Lighthouse loads a page in a controlled Chrome instance, applies simulated network and processor throttling so the test resembles a mid-range phone rather than your desktop, records what happens, then runs a fixed set of audits against the result. You can start it from the Lighthouse panel in Chrome DevTools, from a command line, or from a build pipeline. PageSpeed Insights runs the same audits on Google’s own hardware.
The output is grouped into four categories — performance, accessibility, best practices and SEO — each with a score and a list of individual audits marked passed or failed. The performance score is a weighted combination of timing metrics; the other three behave more like checklists. Below the scores sit the diagnostics, and those are the genuinely useful part: named files, unused code, oversized images, elements that shifted while the page loaded.
Why Lighthouse matters
It runs anywhere. A page behind a login, a staging site nobody else can reach, a local build on a laptop — none of these have real visitors, so none of them can be judged by anything that depends on public traffic. Lighthouse works on all of them, which makes it the tool you use before launch rather than after.
It is also repeatable enough to sit inside a release process. Running the same audits automatically on every deploy catches the regression where a new plugin quietly adds a heavy script, long before it surfaces in the real-user measurements shown by PageSpeed Insights.
Common mistakes with Lighthouse
Treating the SEO category as an SEO audit is the one I meet most often. Those checks confirm that a page has a title, a meta description, a viewport, crawlable links and legible text. They are worth passing, but they say nothing about search intent, keyword mapping, internal linking, duplication or content quality — the things a real SEO audit exists to examine. A full score there is a formatting tick, not a verdict on the page.
The accessibility score deserves the same scepticism. Automated tests catch only part of what makes a page usable, and a page can pass every one of them while being impossible to operate with a keyboard. And because Lighthouse measures your machine as much as the page, comparing a score from one laptop with a score from another tells you very little.
How to act on it
Run it in an incognito window with extensions disabled, several times over, and work from the middle result rather than the best or the worst. Browser extensions in particular can ruin a performance score and send you chasing a problem your visitors never had.
Then set the headline number aside and read the diagnostics in order of estimated saving. Change one thing, re-run, confirm the saving was real, and only then move on. Use Lighthouse as lab data: it tells you what is wrong and where, while field measurements tell you whether it mattered to anyone. When the same audits keep failing after the easy fixes, the constraint is usually the theme, the hosting or the build process rather than any single page.