Skip to content
CWV Checker

Why your score is different on every run

Lighthouse is a single simulated run, not an average. What the variance means, how big a swing is real, and how to measure so the number can be trusted.

The same page tested five times scores 72, 58, 81, 69 and 74 — single runs scatter, so you read the median, 72.median72run 158run 281run 369run 474run 5same page, five runs — none of them wrong
This diagram is free to reuse with credit — all twelve, as SVG or PNG.

The short version

Lighthouse is a single simulated page load, not an average. One run, on one machine, at one moment, with the network and CPU artificially throttled to a mid-range phone profile. Run it again and you get a second sample from the same distribution — and samples differ. The score was never a constant; treating it like one is what makes it feel like it’s lying.

Where the variance comes from

  • CPU contention. The machine running the test is doing other things — especially your own laptop with thirty tabs. Lighthouse throttles relative CPU speed, so a busy host machine changes the result. Datacenter runs (like this tool’s, via PageSpeed Insights) are calmer but not perfectly so.
  • Simulated networking. Throttling models a 4G connection mathematically. Real request timing still varies — DNS answers faster or slower, a TLS handshake lands differently — and the model amplifies some of those differences.
  • Your page genuinely differs per load. A/B tests serve different variants. Ad slots auction different creatives with different weights. Personalisation, cookie banners on first view, a carousel starting on a different slide — the runs differ because the page differs.
  • Cold versus warm caches. A CDN edge that just evicted your page, a server-side cache that expired, a database that had to warm up: the first request after quiet hours is honestly slower than the tenth.

How much swing is real

A useful rule: a 5-point swing is noise; a 25-point swing is a finding. In between, suspect the page changed between runs — an A/B variant, a heavier ad, a cold cache — and look at the metric that moved rather than the composite score. The score compresses six measurements into one number; two runs can share a score for different reasons.

The method that gives you a trustworthy number

  1. Run the test five times, a minute apart.
  2. Read the median, not the best and not the mean — the median ignores one unlucky outlier in either direction.
  3. Compare medians before/after a change, never single runs.
  4. When the site has enough traffic, prefer field data over any lab number: it is a 28-day aggregate of real visits and has no run-to-run variance at all.

This is also why this tool leads with the specific fix and its estimated saving rather than the score: fixes are stable between runs even when the composite number wobbles.

When variance itself is the symptom

If your runs are wildly spread — 40 one minute, 75 the next — that spread is information. The usual culprits are third-party scripts that differ per load (ad auctions above all) and a server cache with a short lifetime, so half the runs hit a warm cache and half don’t. Fixing the variance source usually improves the median too.

Common questions

Which score is the real one?
None of them individually. Each run is one honest sample from a distribution. The real answer is the shape of several runs — which is why the method is median-of-five, not best-of-one, and why field data beats all lab runs when you have it.
PageSpeed Insights gave me a different score than Lighthouse in my browser. Why?
Different machine, different throttling, different extensions. PSI runs in Google's datacenter on calibrated hardware; your DevTools run competes with your open tabs and browser extensions, which inject scripts into the page being measured. PSI's number is the more comparable one.
My competitor's score jumps around too. Can I compare us?
Compare medians, not single runs, and better still compare field data — the Chrome UX Report numbers on both results. Field data is a 28-day aggregate of real visits and doesn't have run-to-run noise at all.
Should I chase the last few points?
Past 90, the remaining points are usually noise-sized and cost more than they pay. The score exists to point at problems; once the specific fixes are done, the milliseconds — and the conversion rate — are what matter, not the rounding.
Does Google rank me on my Lighthouse score?
No. Ranking uses field data — real Chrome visits — not anyone's lab run. A noisy lab score changes nothing about how Google sees your site; it only changes how you feel on a Tuesday.

Check a page against this

Free, no signup, no ads. We keep nothing except the shareable result, for 30 days.

Enter any public URL. You’ll get the one change that recovers the most load time, what it’s worth, and the measurements underneath.

Other guides