Discuss your task

Search

Start typing to search articles, cases, and services.

↑↓ navigate Esc close

Why Your Site Is Slow Despite a Good PageSpeed Score

The score in PageSpeed Insights is the result of a single synthetic run on an emulated device, not a measurement of what your visitors experience. A site can be slow while PageSpeed looks good because the speed real people get is shown by a separate block of field data from CrUX. Those field metrics are the ones Google takes into account for ranking, and they are the ones you need to fix - not the score.

PageSpeed Insights shows two independent sets of numbers on one screen. The field block comes from the Chrome User Experience Report: according to Google documentation, the service re-aggregates the data every day, but each time over the previous 28 days. That is a rolling window, not a snapshot taken at the moment you run the check.

The second limitation matters more than the first. CrUX only includes Chrome users who have opted in to usage statistic reporting, sync their browsing history and have not set a sync passphrase - the CrUX methodology lists these conditions explicitly. Chrome on iOS, Android WebView and other Chromium-based browsers are not in the sample at all.

In Exceltic.dev site audits this is the most common source of argument between client and contractor: both sides look at the same report and read different numbers. Below - what the score actually measures, where field data comes from, five reasons for the gap, and what to check instead of the overall score.

The scenario is almost always the same. The contractor sends a screenshot with a green ninety-five and considers the job delivered. The founder opens the same site on their phone while out and about, waits four seconds and writes back that nothing has changed. The argument goes nowhere, because nobody names the key fact: these are two different measurements, and working out who is right is pointless - both are.

What the PageSpeed Insights score measures

The 0-100 score is a weighted sum of lab metrics from a single Lighthouse run, and it has no direct bearing on your visitors.

Lab data: the result of a synthetic test of the page under fixed conditions - one device, one network connection, one moment in time, no real user.

The run happens on Google infrastructure with an emulated mobile device and throttled network and CPU. The specific device model and the throttling values depend on the Lighthouse version, so the score of the same page can shift after a tool update without a single change in the code.

The score includes several lab metrics with different weights: FCP, Speed Index, LCP, Total Blocking Time and CLS. The weights also change from version to version. Which leads to an uncomfortable conclusion: the score is not a metric but an index assembled from metrics. You can raise the index without improving a single real user scenario.

And the main point. The lab score is not a ranking factor. Google relies on field Core Web Vitals, and those are different numbers on the very same page.

Where field data comes from and why it lags

Field data in PageSpeed Insights and in the Google Search Console report comes from one source - CrUX. That source has four limitations that explain almost every discrepancy with the lab.

Field data: aggregated speed figures collected from real visits to your page over a past period, not in a test environment.

First - the window. The aggregate is recalculated daily, but always over the previous 28 days. A deploy that fixed LCP yesterday is diluted in a month-long sample and will only show up in full after four weeks.

Second - the source. Chrome only, and only from users who meet the collection conditions. If a noticeable share of your audience is on iPhone, those visits are not in CrUX in any form.

Third - the data sufficiency threshold. Google does not publish the exact number of visits required to make it into the dataset, but states plainly that pages and origins below the threshold are not included in CrUX. If there was not enough data for a specific URL, PageSpeed Insights substitutes data for the whole domain. You read a site-wide average while thinking you are looking at one page.

Fourth - the percentile. The “good” or “poor” rating is assigned at the 75th percentile of visits, not at the average. A quarter of users can be seeing the site noticeably slower than the summary badge suggests.

Five reasons a site that is fast in tests is slow for real people

The gap is almost never a measurement error. It is five specific differences between a test run and a real session.

Devices and networks are worse than the emulated ones. Lighthouse throttles the CPU by a fixed multiplier. A real budget Android with background apps and a hot processor behaves worse than any emulation, and a mobile network on the move produces a spread of latencies that a stable simulated connection does not have.

Geography and distance to the server. The test runs from Google infrastructure, which may be ten milliseconds away from your hosting. For a user in another country with no CDN it is two hundred and fifty. On a heavy page that is several extra seconds.

First-visit and repeat-visit cache. The lab run always uses a clean profile: no stored data, no cache, no logged-in session. A real audience is mixed. If you have a lot of returning visitors, the field will look better than the lab; if the traffic is paid and almost entirely first-time - worse.

Third-party scripts that are absent from the test. A chat widget, an A/B testing tool and ad tags often fire only after consent to data collection, or under certain session parameters. In a clean run nobody clicks the consent banner, so part of the third-party load simply never executes.

Interaction that the simulation does not reproduce. Lighthouse does not measure INP: nobody clicks during a lab run. Instead of responsiveness it counts Total Blocking Time - an indirect indicator that can look acceptable while real INP is poor.

Which metrics to watch instead of the overall score

What you need to watch is three field metrics and their thresholds, not the final score.

According to the web.dev documentation, the Core Web Vitals thresholds are as follows. LCP: up to 2.5 seconds is “good”, above 4 seconds is “poor”. INP: up to 200 milliseconds is “good”, above 500 is “poor”. CLS: up to 0.1 is “good”, above 0.25 is “poor”. All three are assessed at the 75th percentile and separately for mobile and desktop.

INP replaced FID in Core Web Vitals on 12 March 2024. Any guide that still mentions FID is outdated in full, not in part.

The practical point is that these three metrics lead to different tasks and different people doing them. LCP is almost always the weight and loading priority of the main image; there is a step-by-step walkthrough in the piece on LCP optimization with the AVIF and WebP formats. INP is JavaScript sitting on the interface response path, and it has a separate breakdown of what causes poor INP and how to fix it. CLS is layout instability; if what bothers you is the spread of results from check to check, the causes of those jumps are covered in the article on site quality scores that jump around.

The overall score does not separate these tasks, it averages them. That is why it does not work as a brief for a contractor.

How to check your own site against field data

The check takes about half an hour and runs in four steps, in increasing order of reliability.

1. The Core Web Vitals report in Google Search Console

Start here. The report shows field data for every URL on the site at once, grouped into clusters of similar pages, separately for mobile and desktop. It is the only place where you can see what share of pages is problematic, rather than one page you picked by hand.

2. The field data block in PageSpeed Insights

Check a specific page and make sure to look at whether the block is labelled as data for this URL or as data for the domain. In the second case there was not enough data for the page, and what you are looking at is a site-wide average.

3. Splitting by device type

Compare mobile and desktop separately. The typical picture is an acceptable desktop and failing mobile, and the overall score smears that out. The CrUX API splits data by the values PHONE, DESKTOP and TABLET.

4. Collecting metrics from real users

The only way to see the whole audience, including the Safari and iPhone users who are not in CrUX, is to collect the metrics yourself on your own site. Google’s web-vitals library reports the same three metrics right in the visitor’s browser, and from there you can store them wherever you like.

Minimal metric collection on the page
<script type="module">
  import {onLCP, onINP, onCLS} from 'https://unpkg.com/web-vitals?module';
  const send = (m) => navigator.sendBeacon('/vitals', JSON.stringify(m));
  onLCP(send);
  onINP(send);
  onCLS(send);
</script>

What to do when field data is bad and lab data is good

The rule is simple. Fix what the field shows, and leave alone what is already green in the lab.

The order of work is this. First split the audience by device and take the worst segment - almost always mobile. Then work out which of the three metrics falls outside its threshold, and work only on that one. After that narrow down to page type: a catalogue, a product page and the homepage break for different reasons.

In a typical audit of a content site with 20-40 thousand visits a month we see the same gap: a lab score of 95-98 alongside field LCP on mobile of 3.4-4.1 seconds at the 75th percentile, with the share of “good” visits at around half. Three things usually account for most of it: no CDN with an overseas audience, a heavy main image with no loading priority, and third-party scripts that never run during the test.

After the fixes field LCP moves into the 2.1-2.6 second range, but that does not show up in Search Console straight away. Because of the 28-day rolling window, the final picture is worth assessing four to six weeks after rollout, not the next day. If there is nobody inside the team to dig into the gap and push the metrics to their thresholds, this gets handled as a separate piece of work - a Core Web Vitals audit and site speed-up with a measurable result in field data.

Who this hits hardest

The gap hits hardest at companies whose audience is geographically far from the server and whose traffic is mostly mobile and paid. That is the typical situation for a business serving the US and Europe while keeping the site on hosting that was once chosen on price. The second vulnerable case is sites on website builders and WordPress with a stack of plugins, where third-party load is heavy and almost all of it is tied to a real session. If you pay for ad traffic, the cost is calculated directly: a slow landing page raises cost per lead at an unchanged bid.

Frequently asked questions

Why does PageSpeed show 95 while the site is slow for users?

Because these are two different measurements on one screen. The score of 95 is the result of a single synthetic Lighthouse run on an emulated device with set network and CPU throttling. What real people experience is reflected by a separate block of field data from CrUX, collected from visits over the previous 28 days. The gap comes from real devices and networks, distance to the server, the share of first-time visits with no cache, and third-party scripts that do not fire during a clean test run. For ranking Google uses field Core Web Vitals, not the lab score.

How long after the fixes will field data update?

Field data is recalculated daily, but each time it is aggregated over the previous 28 days. The day after rollout you will see a sample in which 27 of the 28 days are old. A noticeable shift appears in about two weeks, the full picture in four. In practice it makes sense to assess the result after four to six weeks, especially if traffic is uneven across days of the week. The lab score, meanwhile, changes immediately, which is exactly why it is handy for confirming that a fix landed, but not its effect on your audience.

My page has no field data at all, what does that mean?

It means CrUX did not gather enough visits for that URL. Google does not publish the exact threshold, but states that pages and origins that do not pass it are not included in the dataset. In that case PageSpeed Insights substitutes data for the whole domain, and the block is labelled as origin data rather than page data. You cannot read it as a characteristic of a specific page: the homepage and a deep catalogue item behave differently. For low-traffic pages there is only one working option - collect metrics from real users yourself.

Is there any point in looking at the lab score at all?

Yes, but as a debugging tool, not as a target. A lab run is reproducible: it gives instant feedback on a specific change and a list of diagnostics pointing at problem resources. That is convenient while you are changing code. It cannot be the target for two reasons: it does not affect ranking, and it does not measure INP, because nobody interacts with the page during the test. A contractor brief is written in field thresholds - LCP, INP and CLS at the 75th percentile, separately for mobile.

Key takeaways

  • The PageSpeed score is an index from a single lab run, not the speed your site has for real people.
  • Ranking relies on field Core Web Vitals from CrUX, and those are different numbers on the same page.
  • CrUX sees only part of the audience: Chrome on supported platforms, with usage statistic reporting enabled, over a rolling 28 days.
  • The brief is written in metrics and thresholds - LCP, INP, CLS at the 75th percentile, separately for mobile - not in score points.

If you have a green score in PageSpeed Insights and speed complaints from real users, send your site address and a rough idea of where your audience is located to Exceltic.dev. We will work out exactly where lab and field data diverge, and name the specific causes before any work starts.

More articles

All
Get in touch your way
WhatsApp Telegram

Before you go: a free estimate

Describe your task in a few words and we propose a solution, a stack and a quote with timelines.

Three areas with one team: software, web development and CRM implementation. Over 100 projects, and the project is run directly by an engineer, in Russian and English.