Download websites that use Google PageSpeed

About Google PageSpeed

The name covers two different things, and the difference matters before any number does. Google publishes a hosted analysis tool that scores a page on request, and it also published a server module that rewrites HTML, images and stylesheets as they leave the origin. A hosted analysis tool leaves no trace on the sites it measures. A server module leaves plenty, in response headers and in the rewritten markup it emits, and the module is what our detection actually finds.

So the entry named Google PageSpeed in our Caching data should be read as the server-side optimiser, not as evidence that somebody once ran the scoring tool against the site. As of September 2026 the module fingerprint appears on 164,180 sites, 47,325 of them in United States, with .com recurring through the rest. The figure describes a population of origin servers whose administrators installed an Apache or Nginx module and configured it to rewrite output automatically, which is a narrower and more interesting group than the population of people who care about performance scores.

Version data is the most revealing field on this technology. We have seen 210 distinct builds, and the one most commonly present is 1.13.35.2, an unusually wide spread for a component that operators install once and then forget. A server module has no update prompt and no user-facing surface, so it stays at whatever version the package manager delivered on the day the machine was provisioned. Reading the version distribution is therefore a way of reading server ages across the hosting market, and a heavy tail of old builds points at providers running long-lived base images rather than at thousands of independent decisions.

Presence tells you something specific about the operator. Somebody with root access on the origin made a deliberate performance decision, which separates the site from the much larger group whose optimisation consists of a caching plugin or a switch flipped at a CDN. Hosting companies also enabled the module across entire fleets, so a cluster of detections on shared infrastructure usually reflects one provider's default configuration, and the dataset only makes that visible when the hosting field and the technology field are joined across the whole list.

The project dates from 2010 and was distributed as Freemium software, and development has since gone quiet, which makes the population interesting for a reason that has nothing to do with performance. A list of servers still running a retired optimisation module is a list of machines nobody has rebuilt in a long time, and that is a useful filter for infrastructure vendors, for security researchers and for anyone selling migration work. The caveats are ordinary but real. Headers can be suppressed in configuration, rewritten markup is not always distinguishable from the output of other optimisers, and a proxy in front of the origin can strip the evidence entirely, so the count is a floor and the version field is an approximation of build age rather than a certificate of it.