What a Lighthouse 100 doesn't tell you

Local Lighthouse runs kept scoring a print shop site I rebuilt between 80 and 94, and flagged 115 KB of unminified JavaScript. The whole site ships one script. It is 13 KB before compression.

The other 100 KB belonged to an ad blocker. Browser extensions inject scripts into every page you visit, and a Lighthouse run inside your own browser measures them as part of the site. Lighthouse knows this about itself: when it notices extension code eating time, it prints a warning and suggests incognito. PageSpeed Insights runs the same audit on a clean machine: 100 in all four categories — performance, accessibility, best practices, SEO.

Measure in a clean environment

The rule I took from this: numbers for publication come from pagespeed.web.dev, not from a local browser. Your machine brings extensions and background load into the measurement; Google's test rig brings nothing of yours.

If you own a website, this is a two-minute check. Open pagespeed.web.dev, type your address, wait for the report, and read the four scores — Mobile tab first, because that is how most customers arrive. No account, nothing to install.

What the score is made of

The performance score is a weighted sum of five metrics: 30% Total Blocking Time, 25% Largest Contentful Paint, 25% Cumulative Layout Shift, 10% First Contentful Paint, 10% Speed Index. The weights are public, in Lighthouse's scoring calculator.

Read the weights and the strategy falls out. A site with almost no JavaScript has a blocking time near zero, so 30% of the score comes along by itself. Most of what remains rides on LCP — how fast the largest element on the screen appears.

Recommendations are not levers

Lighthouse attaches suggestions to every report, and they read like a to-do list. Five of them went to a test bench on the same site.

Dropping text-rendering: optimizeLegibility saved 0 ms. Removing a backdrop-filter saved 0 ms. All five came back the same way.

That is not a verdict on the suggestions — on a script-heavy site some of them may matter. It is a verdict on treating the list as a work order without measuring first.

One lever moved the metric: fonts. They load with font-display: block — a deliberate choice: the text waits for the real font instead of starting in a substitute and changing shape mid-read — which puts every font kilobyte on the critical path. Cutting the font files took 0.8 seconds off LCP on a slow-network profile; no recommendation from the list came close.

That choice has a cost: on a very slow connection, the text stays invisible for up to three seconds, then a fallback with matched letter widths steps in — if the real font arrives after that, the swap moves nothing. A different site, a different symptom, the same ending — the page that jumped on every reload also came down to lighter fonts.

One page, three numbers

The same home page, measured three ways:

EnvironmentLCP
Chrome DevTools, Fast 4G, CPU slowed 4× (September 2026)1.0 s
PageSpeed Insights emulation1.9–2.3 s
Real desktop0.5 s

All three are true, and none of them means anything alone. A speed figure without its measurement profile is decoration; when you quote one, name the profile.

What a 100 doesn't see

Everything above is the score measuring badly. There is a harder limit: what it does not measure at all.

With four categories at 100, the site still carried a real accessibility bug. The lightbox markup had aria-hidden set once and never removed — an open, fully visible lightbox did not exist for a screen reader.

Lighthouse cannot catch this. The accessibility audit scans the page as it loads, lightbox closed, and in that state the markup is correct. A manual review found the bug; no tool did.

The fix was to stop managing that state by hand. The native <dialog> element gets the top layer, the focus trap and the Esc key from the browser; the script only remembers which thumbnail to hand focus back to:

var open = function (i, trigger) {
  lastTrigger = trigger || null; show(i);
  if (isDialog) box.showModal(); else box.classList.add('open');
  document.body.style.overflow = 'hidden';
};
if (isDialog) box.addEventListener('close', cleanup); /* also fires on Esc */

One close listener covers the close button and the native Esc, and cleanup() returns focus to lastTrigger. Browsers without <dialog> fall back to the class-based version.

Lighthouse has bugs of its own, too: issue #17080 makes the image-delivery insight flag correctly sized images when device emulation hides the host's display scale — still open as of November 2026.

So the score is hygiene, not truth. A 100 says the measurable problems are gone; it does not say a person can use the page.

Lighthouse even admits it: the report ends with a list of manual checks it cannot run for you — keyboard order among them. So put the mouse away and Tab through the page. Open the menu, the gallery, the form with the keyboard alone, then once more with a screen reader on. What breaks there never shows up in the number.

← All notes

mihai@papush.ro