Lighthouse score changes every time? Why it happens and what to trust

The same page scores 79, then 92. Here is why Lighthouse scores move between runs, how many runs to take, and which number Google uses.

Published Updated

Lighthouse performance ring at 92 in green, with a yellow pill reading 79 → 92
In short

It is maddening when the same page scores 79 on one run and 92 on the next. Lighthouse measures a single page load, and the conditions around that load change every time: the network, the test machine, the browser and your own page code. The fix is to stop trusting one run. Take the median of 5 runs, and judge real speed by field data.

Why does the score change when nothing changed?

The score changes because the test conditions change, even when your code doesn't. The Lighthouse team says it plainly: "Lighthouse performance scores will change due to inherent variability in web and network technologies" (Lighthouse variability docs).

The same doc lists the sources it rates as high impact:

  • The network. Packet loss, traffic prioritization and last-mile congestion differ from one minute to the next.
  • The test machine. "The hardware on which the webpage is loading can greatly impact performance."
  • Other work on that machine. Concurrent runs "can skew performance results due to resource contention." So can browser extensions and anti-virus software.
  • Your own page. A/B tests, rotating ads and anything else that loads different code on each visit make the page itself different on every run.

Lighthouse throttles the connection to smooth out network noise. Its default simulated throttling replays network activity on its own model, which helps a lot. No throttling mode fixes the last point: if your page loads different code each time, the score will differ too.

Why a small change moves the score so much

The performance score is "a weighted average of the metric scores" (Chrome performance scoring). Three metrics carry 80% of the weight, so a wobble in one of them shows up fast.

MetricWeight
Total Blocking Time (TBT)30%
Largest Contentful Paint (LCP)25%
Cumulative Layout Shift (CLS)25%
First Contentful Paint (FCP)10%
Speed Index10%

Each raw value, such as an LCP of 3.1 seconds, is first turned into a metric score from 0 to 100 on a log-normal curve. A log-normal curve is steep in the middle and flat at the ends, so the same number of milliseconds is worth more points in some places than in others.

A worked example: 79 becomes 92

Here are two example runs of the same page, as metric scores. Only TBT moved, because the main thread happened to be less busy on the second run.

Metric (weight)Run 1Run 2Points gained
FCP (10%)90900
Speed Index (10%)90900
LCP (25%)90900
TBT (30%)509312.9
CLS (25%)94940

Run 1 adds up to 9 + 9 + 22.5 + 15 + 23.5 = 79. Run 2 adds up to 91.9, which shows as 92. One metric, one noisy run, 13 points, and the page jumps from the orange band into the green one.

The curve also explains why the top is so hard to reach. Google notes that "taking a score from 99 to 100 needs about the same amount of metric improvement that would take a 90 to 94."

How to get a number you can trust

Run the test several times and use the median, not the best or the latest run. The Lighthouse team's own figure: "The median Lighthouse score of 5 runs is twice as stable as 1 run."

  1. Run it 5 times. Use the same tool each time, either PageSpeed Insights or the Lighthouse panel in Chrome, never a mix.
  2. Write down the median. Sort the 5 scores and take the middle one. Ignore the best and worst runs.
  3. Keep the device type fixed. PageSpeed Insights mobile simulates "a mid-tier device (Moto G4)" on a mobile network, while desktop is "an emulated-desktop with a wired connection" (PageSpeed Insights docs). Their scores are not comparable.
  4. Run tests one at a time. Parallel runs on the same machine compete for CPU.
  5. For automated tests, use dedicated hardware. The Lighthouse docs recommend at least 2 dedicated cores and 2GB of RAM, and warn that "running on inconsistent hardware will lead to inconsistent results!"

If your median still swings by more than a few points from day to day, look at your own page. Rotating ads, A/B tests and chat widgets are the usual suspects.

Which number does Google use?

For the Core Web Vitals assessment, Google uses field data, not your Lighthouse score. Field data is what real Chrome users experienced; lab data is one simulated load on one device.

Lab data (Lighthouse score)Field data (Core Web Vitals)
Comes fromOne simulated page loadReal Chrome users, via the Chrome UX Report
Time spanA single runThe previous 28 days
Pass ruleScore 90 to 100 is shown greenThe 75th percentile of all three metrics is Good
Best forDebugging what to fixJudging what users feel

PageSpeed Insights shows both field data and the Lighthouse lab score in one report. Google's own docs describe lab data as "useful for debugging issues" and field data as "useful for capturing true, real-world user experience." When the two disagree, trust the field data for how your site performs, and use the lab run to find out why.

Small sites often have no page-level or even origin-level field data, because not enough Chrome users visit them. In that case the median lab score is the best signal you have, so the 5-run habit matters even more.

When a score drop is real

A drop is real when the median moves, not when one run does. Use this table to decide what to do.

What you seeWhat it likely meansWhat to do
One low run, median unchangedTest noiseNothing
Median drops right after a deployA real regressionCompare the five metrics before and after
Mobile drops, desktop holdsHeavier JavaScript hurts the slower simulated deviceCheck TBT and the scripts added
Lab looks fine, field data is poorReal users have slower devices or networksTrust the field data and test on mobile

For a real regression, the metric that moved tells you where to look. A jump in LCP often comes from render-blocking resources or a hero image that is lazy loaded when it should load at once. Speed is one line in the 2026 SEO checklist, so check the rest while you are there.

Tracking scores over time instead of one run

A history of scores tells you more than any single test. UpCrawler runs Google's Lighthouse on every page of your site and keeps the score history, so one bad run reads as a blip in a trend rather than a crisis. Paid plans test desktop and mobile; the Free plan tests desktop.

Each opportunity lists the affected pages, with fix guides for Next.js, WordPress and Shopify. See how it works on the Lighthouse feature page. Like every Lighthouse tool, it gives lab data: it does not measure field Core Web Vitals, and TBT stands in for INP.

Want to see the scores for your own site? Run a free audit and check your homepage in about a minute.

Frequently asked questions

Why is my PageSpeed Insights score different from Lighthouse in Chrome?

They run on different hardware and different networks. The Lighthouse variability docs name both as high-impact sources of change. Compare medians from the same tool only, and never mix the two.

Do I need a Lighthouse score of 100?

No. Scores from 90 to 100 are already in the green "Good" band. Going from 99 to 100 takes about as much metric improvement as going from 90 to 94, so effort past 90 rarely pays off for visitors.

Why is my mobile score so much lower than desktop?

PageSpeed Insights tests mobile on a simulated mid-tier phone over a mobile network, and desktop on a wired connection. The same JavaScript takes longer to run on the slower device, which often shows up in TBT. A gap between the two is normal.

Found this useful?

Share it with whoever runs your site’s SEO.

Ready when you are!Crawly, the UpCrawler mascot

Catch the next drop on day one.

UpCrawler watches your uptime, crawls every page and tracks Search Console, so you see the turn the day it happens.

14 days freeNo cardNothing to install