How we measure post-quantum adoption
A measurement you cannot check is worth less than no measurement. This is the panel, the probe, and the limits of what the number means.
The measurement
Once a day, a worker opens two or three read-only TLS connections to each host in a fixed public panel. The first sends nothing at all. For each host it records whether the server selected the hybrid group X25519MLKEM768, the negotiated TLS version, the leaf certificate’s signature algorithm and expiry, and whether the server volunteered a session ticket without being asked for anything. The headline figure is the share of reachable hosts that negotiated the hybrid group on that run.
A host that volunteers no ticket gets a second connection carrying a single HEAD / request, because many servers withhold NewSessionTicket until a request arrives and would otherwise be recorded as offering no resumption at all. Measured on 28 August 2026 over a fourteen-host panel run twice: silent connections drew a ticket from 8 of 14 hosts, one HEAD drew one from all 14. A final connection offers that ticket back, to the same address and the same server name that issued it, and records whether the server resumes. A ticket is never offered to a host that did not issue it: that would be testing someone else’s access control with credentials minted by another configuration, and we have no standing to do it. No credentials are ever sent, connection refusals are honored, and no exploitation is attempted. This is the same class of measurement public scanners such as SSL Labs and Censys perform.
The panel
The panel is fixed and public. It lives with the probing policy in the open-source pqc-observatory repository, where anyone can propose a host or check the one we used. A fixed panel is what makes the series comparable over time: a panel that grows as post-quantum adoption grows would report a trend that is partly an artefact of its own sampling.
To have a host removed, open a pull request or write to [email protected]. It drops out within one measurement cycle.
What the number is not
It is a share of hosts, not of traffic. Traffic-weighted figures are dominated by a small number of very large properties and run much higher; they answer “how much of what people load is protected”, while this answers “how widely has the capability actually spread”. Both are useful. Quoting one as the other is the most common error made with post-quantum adoption statistics, including on our own site before we fixed it.
A host counted as hybrid has negotiated one hybrid key exchange on one endpoint. It implies nothing about the origin behind a CDN, other ports and protocols, data at rest, or the certificate chain, whose signatures remain classical almost everywhere. Key exchange is the surface that protects against harvest-now-decrypt-later, which is why it moved first and why it is the thing worth counting.
Citing it
Cite the Observatory and the date of the run. Every figure is stamped with the day it was measured and the number of hosts probed, so a claim can be pinned to a measurement rather than left standing indefinitely. If you think a figure is wrong, the panel and the policy are public: we would rather be corrected in the open than quietly approximate.
Questions about the data
What exactly does the headline percentage measure?
The share of reachable hosts in the panel whose TLS handshake selected the hybrid group X25519MLKEM768 on that day's run. It is a share of the panel, not of internet traffic, and the two are very different numbers. A traffic-weighted figure is dominated by a handful of very large sites; a host-weighted figure like ours says how widely the capability has spread.
How often is it measured, and can I cite a specific day?
Once a day. Each run is stamped with its date and the number of hosts probed, so a figure can be cited against the day it was measured rather than as a standing claim. If you are citing us, cite the date.
Why is your number different from Cloudflare's?
Because it measures a different thing. Cloudflare reports the share of requests reaching its own network, weighted by traffic and dominated by browsers that already ship post-quantum key exchange. We probe a fixed panel of hosts once each, unweighted, which measures server-side deployment instead of client-side capability. Neither is wrong; they answer different questions, and the server-side number is the lower and slower of the two.
Does a host counted as hybrid mean the whole site is quantum-safe?
No, and this is the most common misreading. It means one endpoint negotiated a hybrid key exchange on one connection. It says nothing about the origin behind a CDN, about other ports and protocols, about data at rest, or about the signature algorithms in the certificate chain, which remain classical almost everywhere.
Is the probing intrusive?
No. Two or three read-only TLS connections per host per run, at most one HEAD request per connection and nothing else, no credentials, refusals honored, and no exploitation attempted. A session ticket is only ever offered back to the host that issued it, never across hosts. It is the same class of measurement public scanners such as SSL Labs and Censys perform. Hosts can be removed on request within one measurement cycle.
Can I use the data?
Yes. Cite the Observatory and the measurement date. The panel and the probing policy are open source, so a disputed figure can be checked rather than argued about.
See the current figures on the Observatory, or check one host yourself with the post-quantum TLS check.