Download websites that use HTTP/2

About HTTP/2

Nobody buys HTTP/2. A site owner picks a host or a content delivery network, and the protocol arrives with it, switched on by a default the owner never saw. Reading the marker as a property of the site is therefore a mistake worth avoiding from the first query. Presence is a property of whatever machine terminates the connection, which is why the field works better as a lens on infrastructure than as a description of the people who own the domains.

As of September 2026 we negotiate HTTP/2 with 71,447,951 of the sites we crawl, and the Miscellaneous classification is the honest place to file it. The protocol carries many requests over a single connection, compresses headers and moves framing to a binary format, so its practical effect is fewer round trips on pages that pull a long list of separate assets. In browsers it is used over TLS in practice, which means a site that speaks it has already solved certificates and redirects.

The join is where the column earns its place. On its own, a list of domains that answered with the newer protocol is a flat fact. Crossed with the server and delivery network fields, the same list becomes a map of which providers have modernised their edge, which have not, and how far behind their customers sit as a result. Hosting companies use exactly that crossing to compare their own footprint with the competitors they are trying to displace. Vendors selling performance tooling use it to size the population that has not yet moved and to price the pitch.

United States accounts for 30,595,791 of the domains where we see the protocol, and the suffix weight sits on .com. The pattern follows edge networks rather than borders. Large providers headquartered in one market serve customers everywhere from the same configuration, so a country column built from serving infrastructure will always lean toward the markets where that infrastructure is registered. The geography is best read as a statement about where connections terminate, then corrected with the domain suffix and the language of the page.

A single lookup against one domain answers a question anyone can answer for free in a browser. The download answers the questions that need a population behind them. How adoption differs between a country code suffix and a generic one. Whether a given hoster migrated its customer base wholesale or left it in place. Which segments of a market still terminate on configurations nobody has touched in years. Questions like those require every domain and its markers in one file, joinable against a customer list or a prospect list and comparable against a copy pulled a quarter later.

Two limits deserve stating. Our record describes the connection we made at crawl time from our own network, so a site fronted by a modern edge can still run an origin that speaks nothing of the sort behind it, and the field says nothing about that origin. Detection also depends on the crawl succeeding at all, so blocks, timeouts and filtering by country remove domains from the count rather than marking them unsupported. Absence in the file means we did not observe support, not that support is missing, and any analysis that turns on the difference should carry that caveat forward rather than round it away.