PageSpeed Insights MCP questions.

The ones people ask first

Why does it take minutes when the website is instant?

Because Google only re-tests a page about once a minute and hands out the stored answer to everyone who asks in between. Five requests in a row get you one measurement five times over. This waits for five real ones, which takes roughly two and a half minutes and is the entire point.

Can I just use pagespeed.web.dev instead?

For a quick look, yes, and it's free. What you can't do there is tell a real change from the instrument moving, because you only ever see one run. If you're about to make a decision on the strength of a number, you want to know how much that number wanders.

Do I need a Google API key?

For the lab scores, no. It works without one on a shared pool that is often exhausted, so you'll meet a 429 sooner or later. For real-user data, yes. Getting one is free and takes about a minute, and the install page walks through it including the one step nearly everybody gets wrong.

Does it cost anything to run?

No. Both Google APIs it uses are free, with a quota of 25,000 calls a day that a person checking their own sites will never approach. The tool itself is open source and free via an Apache 2.0 licence.

Using it

What can I actually ask it?

Whatever you'd ask a person. How fast is my home page, has it got slower for real visitors since spring, is the blog worse than the shop. It runs the measurements and answers, rather than handing you a dashboard to interpret.

Will it tell me my site got slower?

It will tell you what it measured and how much that measurement wanders, and leave the conclusion to you. If a score moved by less than the spread, nothing happened, and a tool that called that a regression would be inventing news.

Can I use it in CI?

Yes, and the exit codes are built for it. A failing build should mean the page is broken, not that Google's quota ran out, so those are different codes. Only one of them means the site.

Does it work on Windows?

Yes, and that's tested on every change rather than assumed. There's no virtual environment, no pip step and nothing that has to be compiled, which is usually what breaks a Windows install in a way nobody can debug.

What it won't do

Why won't it just give me one quick number?

There's no flag for it and there isn't going to be. A single run is an anecdote, and the whole reason this exists is that everything else reports one and calls it a measurement.

Why are the lab score and the real-user score so different?

Because they answer different questions. Lighthouse is a simulation on deliberately slow hardware, so it's repeatable and works on a page nobody has visited yet. Real-user data is what actually happened to people. A big news site can score 34 out of 100 in the lab while its readers get the page in just over a second, and both of those are true.

Why does it say there's no real-user data for my site?

Because Google won't publish it until a site has enough traffic to stay anonymous. That's most sites, and it isn't a fault in yours or in the tool. It says so plainly rather than leaving a gap, because an empty space next to a good lab score reads like proof it isn't.

Does it check my site's SEO or accessibility too?

It reports the Lighthouse scores for both, alongside performance and best practices, with the same spread treatment. It won't tell you how to fix them. There are better tools for that and this one is about measuring honestly, not advising.