You run PageSpeed Insights on a page that feels fast on your phone, and mobile comes back red. The client sees the same report. Before anyone touches the build, find out whether the number describes your page or describes the way Lighthouse guessed at it.
What happened on 5fourdigital.com.
The page had a Calendly booking embed. Calendly brings about 4MB of its own scripts. On a normal load they arrive after the page has painted, and nobody waits on them.
The default Lighthouse run reported First Contentful Paint at 8.8 seconds. Re-run with devtools throttling, which actually slows the connection, the same page measured 1.6. Same build, same mobile settings, two numbers. Only one of them was something a visitor would ever see.
Three numbers, one page.
There are three ways to get a performance number for the same page, and they answer different questions. Most arguments about a score are really two people holding different ones.
| How it was measured | What actually happens | Trust it for | Where it misleads |
|---|---|---|---|
| Simulated lab, the default — and the top half of PageSpeed Insights | Loads the page on your own connection, records every request, then calculates what a slow phone would have done | Watching one page over time, and comparing two builds of it | Anything third-party that starts before first paint gets charged to your first paint |
| Devtools throttling | Slows the connection first, then loads the page and lets it paint when it paints | Confirming a bad number before you act on it | Slower to run, and moves around more between runs |
| Field data from real visitors | Pools measurements from the browsers of people who actually came to the page | The verdict. It is what page experience is assessed on | Lags behind a fix by weeks, and a quiet page never gets one at all |
The order to use them in is the order of that last column. Watch the lab number, confirm anything alarming with devtools throttling, and judge the site on the field data when it arrives.
Why the simulated run gets it wrong.
By default, Lighthouse does not load your page on a slow connection. It loads it fast, records every request and what each one waited on, then calculates how long that would have taken on a slow phone. PageSpeed Insights works the same way. It is quick and repeatable, which is why it is the default.
The calculation is only as good as the map it builds. If a third-party embed starts requesting files before first paint on the fast run, the model treats those files as standing between the visitor and the first paint. Then it charges the full slow-network cost of every one of them to your FCP.
Devtools throttling does the slow load for real. The browser fetches over a throttled connection and the page paints when it actually paints. The embed loads behind it, where it always did.
The simulation is an estimate. On a page with a heavy embed, it is an estimate of the wrong page.
This is not a reason to ignore Lighthouse. Most of the time the simulated number and the real one agree, and a bad score is a real problem. Confirm it first, then fix it.
How to check the number before you chase it.
Run Lighthouse from the command line with --throttling-method=devtools, or from the Lighthouse panel in Chrome DevTools with simulated throttling switched off. Compare First Contentful Paint between the two runs. If they are close, the score is real and you have work to do. If they are far apart, look at what loads before first paint.
- A bad mobile score is re-run with devtools throttling before anyone starts fixing it.
- If the two runs disagree, look in the simulated run for a third party that starts before first paint. That is usually the whole gap.
- Booking widgets, chat, maps and video embeds load after the page
loadevent either way. It costs the visitor nothing and takes the question off the table. - A looping background video keeps Speed Index open for as long as it plays. Expect that page to read lower than the rest of the site, and tell the client before they find it.
- Both numbers go in the launch notes, with the throttling method written next to each.
What to put in front of a client.
Both numbers, with the method written next to each, and one sentence saying which one a visitor would recognize. That is it. The temptation is to report only the good one, and it is a trap: the client will run PageSpeed Insights themselves, get the bad one, and now the question is not about the page any more.
Reporting both also makes the embed a decision rather than an accident. A booking widget that costs you a red lab score and nothing a visitor can feel is usually worth keeping. Whoever is paying for the site should get to make that call with the real figures in front of them.
What the score is for.
A page speed score is a diagnosis, not a grade. I aim for 75 to 80 in most competitive sets and then stop arguing with it. Past that, the time is better spent on the page itself.
What matters is that the number you report is one a visitor would recognize. Hand a client a red 8.8 without checking it and you have taught them to distrust every score after it, including the ones that are right.
Confirming a bad mobile number is part of the performance pass in the 5Four Launch Checklist, on every launch we run.
Related: which Core Web Vital to fix first, Core Web Vitals in the glossary.



