Your website on a client’s phone, not on yours

Your site is fast for you because your phone already has it and you are on office wifi. A first-time visitor on mobile data sees something else. Three numbers say what.

By João Borges, founder of Firmitas

Blog

The same site, seen from outside Two phones, one in front of the other. Behind, yours: the page is complete, because you have opened it many times, the images are stored and you are on wifi. In front, the phone of a client arriving for the first time, on mobile data: the loading bar is halfway and the page has not appeared yet. Your site, seen from outside. Your phone Everything already stored. Wifi. The client’s phone First visit. Nothing stored. Mobile data. The gap is invisible from inside.
The gap between the two experiences is wide, and it is invisible from inside.

Everyone sees their own site load fast. They see it on the phone that has already opened it dozens of times, with the images stored in memory, on the office wifi.

Someone arriving for the first time has none of that. They are out in the street, on mobile data, with a signal that is decent rather than excellent, on a phone that is not top of the range. It is the first time that handset has asked a server for those photographs and those fonts.

The gap between the two experiences is wide, and it is invisible from inside. That is why almost every owner of a slow site thinks their site is fine.

Opening the site on your own phone proves nothing

For the test to mean anything, three conditions are missing:

  1. An incognito browser window, so that nothing already stored gets used.
  2. Mobile data instead of the office wifi.
  3. If you can, a phone that is not the newest one in the house.

Done that way, the test is honest. It is still a sample of one. What matters is how every visitor fares, and for that there are measurements.

The usual test and the honest test On the left, opening the site on your own phone: the same handset as always, the office wifi and everything already stored in the browser. On the right, the same test repeated under the three conditions that were missing: an incognito window, mobile data and a phone that is not the newest. Done that way it is honest, but it is still a sample of one. The usual test Opening the site on your own phone. The phone that already opened it The office wifi Everything already in the browser It measures the trip of a returning visitor, with everything ready in advance. Result: proves nothing. It is your trip, not theirs. The honest test Repeat under these three conditions. 1 An incognito window, so nothing stored gets used. 2 Mobile data, instead of the wifi. 3 A phone that is not the newest one. Done this way it is honest, but still a sample of one. Result: honest. For every visit, there are measurements.
The three conditions do not make the test complete: they make it honest.

Three published thresholds, judged at the 75th percentile

Google publishes three page experience measures and the thresholds it counts as good. They are public, they are the same for everyone, and anyone can look them up.

LCP: the time until the largest element appears. It is the moment the page stops being blank and shows the main thing: usually the big photograph at the top, or the headline. The threshold for "good" is 2.5 seconds.

INP: the time it takes to respond to a tap. It measures how long passes between someone tapping a button and the screen actually changing. The threshold is 200 milliseconds. It replaced the previous measure in 2024, and it catches the problem nobody ever complains about in writing: the page opened, but it feels dead for a second.

CLS: how much the page jumps while it loads. It is that irritation of reading a paragraph and having it run away downwards, or of tapping one thing and hitting another because an image has taken the place in the meantime. The threshold is 0.1.

Here is the part that is almost never explained. These thresholds are judged at the 75th percentile of real loads: not the average, not the best case, but the value below which three in four visits sit.

A site with a good average and a bad quarter of visits fails, and it fails for a reason: that quarter is people.

The three thresholds, and the load they are judged on At the top, the three measures Google publishes with the good threshold for each one: LCP, two point five seconds; INP, two hundred milliseconds; CLS, zero point one. Below, the real loads ordered from fastest to slowest: the value judged is the one at the 75th percentile, the value below which three in four visits sit. The slowest quarter, marked solid, is what makes a site with a good average fail. THE THREE THRESHOLDS GOOGLE PUBLISHES LCP Time until the largest element appears. 2.5 s “good” threshold INP Response time to a tap. 200 ms “good” threshold CLS How much the page jumps while it loads. 0.1 “good” threshold Each threshold is judged at the 75th percentile of real loads Not the average, not the best case. each visit, ordered from the fastest load to the slowest the value judged 75th percentile three in four visits are below this value the slowest quarter A site with a good average and a bad quarter fails, and rightly so.
What counts is not the average: it is the value below which three in four visits sit.

Field data is real visitors, the lab is a simulation

Two free tools, nothing to install.

PageSpeed Insights (pagespeed.web.dev): you type in the address and get an analysis. It comes in two blocks, and the difference between them separates the people who can read the report from the people who cannot.

At the top, when there are enough visits, you get the field data: real measurements from real users, over the past weeks. Below it comes the lab test: a simulation run at that moment, in a controlled environment. The field data is the truth; the lab is for diagnosis.

The two blocks of a speed report A drawn report, with the address analysed at the top and two blocks below it. The first block is the field data: measurements from real users over the past weeks, and it is the truth, so its readings are shown solid. The second is the lab test: a simulation run at that moment, which is there for diagnosis, and its readings are shown faded. The three measures appear in both blocks. One report, two blocks The difference between them separates who can read the report from who cannot. aminhaempresa.pt 1 Field data real users, over the past weeks LCP INP CLS 2 Lab test a simulation run right now LCP INP CLS 1 It is the truth: real people, on their own phones, over the past weeks. 2 It is for diagnosis: a controlled environment, at the moment of the test. No field data means the site has few visits. Then the lab test is an indication, not a measurement.
Field data is the truth. The lab test is there to diagnose.

Search Console, Google’s free tool for site owners, has its own report on these three measures, split between phone and desktop, and it groups pages by type of problem. This is where you find out whether the problem is the whole site or one page.

Images are the first cause, and nowhere near the second

In order of frequency, and it is almost always the same order.

Images. A photograph comes out of a camera several thousand pixels wide and a few megabytes in weight. On a phone it will be shown at around four hundred points wide. The difference is downloaded anyway, in full, over the mobile data of the person who is waiting.

The fix is dull and mechanical: resize, compress, use modern formats, serve different sizes depending on the screen, and defer the loading of anything below the fold. Never the main image at the top, which is precisely the one being measured.

The whole photograph, and the space it is shown in A large rectangle stands for the photograph as it leaves the camera, several thousand points wide and a few megabytes in weight. Inside it, a small rectangle stands for the space where it will be shown on a phone, about four hundred points wide. The difference is downloaded anyway, over the mobile data of the person waiting. Next, in order of frequency, the fonts, the third-party scripts and the cookie banner. THE FIRST CAUSE, AND NOWHERE NEAR THE SECOND The photo as it leaves the camera Several thousand points wide. A few megabytes in weight. The space where it is shown on a phone: about four hundred points. The difference is downloaded anyway, in full. Over the mobile data of the person waiting. diagram: not to scale NEXT, IN ORDER OF FREQUENCY The fonts Third-party scripts The cookie banner None is a design problem: they are accounting problems.
The difference between the file and the space it is shown in is downloaded anyway.

Fonts. Every weight and every style is a separate file. Three weights of two families are six files to fetch before the text can be drawn properly. It is also a common cause of the page jumping, when the text appears first in the system font and is then redrawn in the right one.

Third-party scripts. The chat window, the embedded map, the advertising pixel, the statistics tool, the newsletter poster. Each one is code from another company, on another server, that the browser has to contact before it can finish the job. They add up quietly, one at a time over the years, and nobody ever goes back and removes one.

The cookie banner. It appears on top of everything, late, and pushes the page down at the exact moment the visitor has started reading.

None of the four is a design problem. They are accounting problems: things added one by one with nobody adding up the total. It is also why the initial choice between a site in code of its own and a template has consequences that only show up here, two years later.

Speed is a tie-breaker, not a lever

Google’s position is published and it is clear on three points. First: page experience measures are used by the ranking systems. Second: getting good results in these reports "doesn’t guarantee that your pages will rank at the top of Google Search results". Third, and this is the decisive one: Google Search always seeks to show the most relevant content, even if the page experience is sub-par.

If your site is the only one that answers the question well, being slow does not hide it. If there are ten equally relevant sites, which is the normal situation for a local business, then the tie-break happens and speed starts to count.

Nobody can promise you a position, and we do not promise one. What can be promised is the number, because the number is a fact about your page and not a prediction about Google. A fast site is fast whether Google notices or not, and the person who stayed instead of closing the tab noticed for certain.

Half an hour tomorrow morning, with nobody hired

  1. Run your address through PageSpeed Insights and write down the three numbers from the field data block, in the phone column.
  2. If there is no field data, the site has few visits: use the lab test and treat it as an indication, not a measurement.
  3. Open the list of images on your home page and see how many weigh more than a few hundred kilobytes. There is almost always one that weighs more than the whole of the rest of the page put together.
  4. List the third-party scripts the site loads and ask yourself, one by one, when anybody last looked at what that thing measures.

We measure before we hand over and again afterwards

We build light sites on purpose, we measure before we hand over and we measure again afterwards. The person who builds the site is the person who talks to you: whoever explains a number to you is the same person who can fix it.

And we do not leave that as a promise. We measure our own home page under the same conditions we are asking you to measure yours, and the largest element appears well inside the 2.5-second threshold above. You do not have to take our word for it: run PageSpeed Insights on firmitas.pt, in the phone column, and compare it with what your own site gives you. It is a lab figure, and the distinction still holds: field data comes from visitors, not from us.

If you want someone to look at the three numbers for your site and tell you plainly what is costing you the seconds, book a 15-minute call.

Sources

Consulted 9 August 2026. The two Google sentences quoted in the body are the original English wording, on the pages linked above.

Talk to us

Want to know what we would do with your site?

Book a 15-min call