SquishPixel Blog · Web performance · Core Web Vitals · Tools

Lessons from running 43,738 sites through the PageSpeed Insights API

According to DEV Community, a team used the PageSpeed Insights API to run Lighthouse, Google's automated page-testing tool, against 43,738 UK trade business websites. Each site was tested on mobile and desktop settings. Collection ran in two passes in August and September 2026, and 31,030 sites returned a usable result on at least one setting.

The authors say their collector was too lean. It kept only four category scores and six metric values. It did not store the Lighthouse version, the page finally audited after redirects, or the emulated device. It also discarded the loadingExperience and originLoadingExperience fields, which hold Chrome UX Report data from real visitors. The report is therefore lab data only. They also note the response does not include throttling settings, so they relied on Google's published description of the tests.

They also describe how to read errors and handle messy input. Timeouts, connection errors and HTTP 500s say little about a site, because they reflect the request. Google's own 400-class responses, such as FAILED_DOCUMENT_REQUEST, do show it could not fetch or parse a page. Stray spaces in five addresses caused repeat audits, and the authors advise normalising URLs first.

On results, half of the 30,399 sites with a mobile load time took 6.7 seconds or more to reach largest contentful paint, the point when the main content appears. 72.5% took over 4 seconds. The authors stress that each figure rests on one run per setting, and that lab data cannot be compared with field data.

Why it matters

If you test your own site or a client's with PageSpeed Insights, a single score is a rough reading. The article quotes Lighthouse's guidance that the median of five runs is twice as stable as one run. A site near a pass or fail threshold can move either way from noise alone.

Lab and field numbers also answer different questions. Lab tests simulate a device and network, while field data reflects real visitors. If you record results over time, keeping the full response, the Lighthouse version and any field data makes later comparisons far more trustworthy.

Source: DEV Community. This brief summarises their reporting; read the original for the full detail.