Skip to main content

Page Speed

Page Speed measures how fast a page of your site loads and shows what slows it down. The page is opened in a real browser on the published version of the site. The report has the main load metrics, a frame-by-frame video, a request waterfall and recommendations on what to fix.

The section is in the dashboard menu, in the "Services" group.

Plan

Page Speed is available on both plans. Free gives 3 runs a month, Pro gives 500 runs a month and scheduled runs. The limit is shared by the organization and resets on the 1st.

What we measure​

One run is two passes of the page in Chrome plus a separate Lighthouse check. During the passes we record a video of the load, every network request and the main thread's work. Each metric is the median of the passes.

MetricWhat it meansGood
LCPwhen the main block of the first screen appearedup to 2.5 s
TBThow long the page didn't respond to tapsup to 200 ms
CLShow much the layout jumps while loadingup to 0.1
FCPwhen anything appeared on screenup to 1.8 s
SIhow fast the visible part of the screen fills inup to 3.4 s on a phone

The thresholds are Google's. Each metric and the score formula are explained in How we measure speed.

Under which conditions​

The device sets the conditions of the whole run: the video and Lighthouse run under the same ones. That's why phone and desktop runs aren't compared.

PhoneDesktop
Internet37 / 8 Mbit/s90 / 90 Mbit/s
Latency50 ms20 ms
CPU2× slowerno slowdown

The conditions are printed in every report. PageSpeed Insights uses a slower profile, so its score is usually lower.

How to run a measurement​

  1. Open Page Speed and click Measure speed.
  2. Pick a project and a branch. You can change the address: a path (/pricing?sort=price) or a domain connected to the project.
  3. Pick a device and click Measure.

The report is ready in about two minutes; the progress is shown right in the run list. If the same page on the same deploy and device was measured in the last 10 minutes, the dashboard shows that run. To measure again, click the re-run button in the report.

What's in the report​

  • Main metrics, each with a green, yellow and red scale.
  • What to fix first — the three Lighthouse suggestions that save the most time.
  • Metrics — every metric of the passes: min, median, mean, max.
  • Load video — screen frames over time, marked with the first paint, the main block and visual readiness. The timeline can be scrubbed.
  • Waterfall — every request on a timeline, with the main thread's long tasks. You can see which file holds up the rest.
  • Resources — page weight by file type and the heaviest files.
  • Recommendations — Lighthouse checks (performance, accessibility, best practices, SEO) and Coach checks (performance, best practices, privacy). Lighthouse suggestions show the expected savings; Coach suggestions list the files they're about.

The report can be copied for an AI assistant with Copy for AI, and the full request log can be downloaded as HAR.

How to read the verdict​

A run is compared with the previous scheduled run that has the same page, device and branch but a different deploy. One-off runs don't move the baseline, so a "just in case" check won't spoil the comparison.

VerdictWhen
Got worsethe score dropped by 10 points or more or the main block (LCP) appeared 1.5 times later and at least 500 ms later
Got betterthe score rose by 10 points or more
No changethe difference is smaller, within measurement noise
Nothing to comparethere's no previous run, or it was made under different conditions

The 10-point threshold comes from repeated runs of the same pages. On one deploy the score of a simple page varied by 9 points, of a heavy one by up to 19. A difference under 10 points is mostly noise, not a change.

Run under load

If our server was busy during a run, the numbers may be worse than the real ones. Such a run is flagged and left out of the comparison. Re-running such a run doesn't use the monthly limit.

You can also compare any runs manually: select them in the run list and click Compare. The earliest one becomes the baseline.

Schedules and Telegram​

On the Pro plan a page can be measured on a schedule: every 3 or 6 hours, daily or weekly. Daily and weekly runs happen at night, when the server is free and the numbers are more accurate. Runs of one schedule are grouped in the list under its name.

To get results in Telegram, turn on Send the result to Telegram when you set up a schedule. After each run you get a message: a snapshot of the page, the score and what changed, deploys since the previous run and what to fix first, with links to the report and the comparison.

From the terminal and from an AI agent​

npx layero@latest perf check --device mobile --path /pricing --wait

The command measures the project's main branch and waits up to 4 minutes for the result. Without --wait it only starts the run, and npx layero@latest perf show prints the result. The verdict is the same as in the dashboard.

AI agents start a run with the check_performance MCP tool.

Limits​

  • Only your project's addresses can be measured: the branch address, the project address and domains connected to it.
  • A run covers one page load without signing in. Clicks, navigation and forms after the load aren't included.