In March this year we measured IPv6 support across the 100,000 most-linked web hosts, from one machine in Santa Clara, California. The obvious objection, which we raised ourselves in the limitations section, is that the Internet doesn't look the same from everywhere. This month we ran the whole thing again from five places at once (Amsterdam, New York, San Francisco, Singapore, and the original Santa Clara machine with its Hurricane Electric tunnelled IPv6), and then went ten times deeper down the list while we were at it.
Of the top million hosts in the cc-main-2026-apr-may-jun Web Graph, 35.7% publish a AAAA record somewhere, and 31.7% actually answer over IPv6 from all five vantage points. Adoption still falls as you go down the ranking, 71% in the top 100 to about a third overall, but not smoothly. Strangely, the deepest half-million hosts do
better than the quarter-million above them. We suspect mass hosting is involved. We haven't proven it, so for now we're just flagging it as a kind of curiosity.
The TLD split is where the chart gets fun. Nearly four in five .io hosts serve working IPv6, which says more about who registers .io domains and where they host than about anything else. .dev domains are also fairly far out in front of the rest. Nerds at play. .jp and .ru are under 14%. And .ru publishes a lot of IPv6 that doesn't actually deliver. A quarter of .ru AAAA-bearing hosts were actually dead over IPv6 from everywhere while answering fine on IPv4. That's nine times the global rate for broken IPv6.
So, did the four extra vantage points change the answer? For the headline numbers, the answer is no basically. That’s somewhat embarrassing, by the way. Every vantage measured adoption within a few tenths of a point of every other one, the HE tunnel included. If you just want an IPv6 adoption rate, it seems like one sensible vantage point will do.
The per-host picture is a different matter, though. 5,530 hosts answered some vantages but not others, and from any single viewpoint you'd probably just file them under “working” or “broken” and be wrong either way. Our favourite: www.ncbi.nlm.nih.gov answered four vantages and stonewalled New York on IPv6 and IPv4. That's probably not an IPv6 problem, we think that's a firewall that doesn't like the data centre. Without the other vantage points and the IPv4 control probe we'd have miscounted it, and the same goes for every single-vantage-point study that hits a host like this.
There's a pattern to where there’s disagreement too. The vantage points agree about whether IPv6 works pretty much the same all the way down the list. What they increasingly disagree about in the tail is whether DNS admits it exists, as geo-DNS setups and wobbly nameservers make the answer depend on who's asking. Among hosts where all five vantages did see AAAA records, a third of them handed different addresses to different vantage points.
Since the Santa Clara machine re-probed the exact same host list we used in March, we also get a clean four-month delta: reachability in that population went from 36.9% to 38.0%. Under that net number there's apparently a churn: 2,147 hosts gained IPv6 and 994 lost it, nearly three times the net movement. And one finding we weren't actually looking for was that only 44% of the March top 100k is still in the July top 100k. Harmonic centrality rank churns quickly between graph snapshots, which means longitudinal "top N" studies are mostly comparing different hosts. Our follow-up planned for October will re-probe this campaign's frozen million-host list for exactly that reason, alongside whatever the top million looks like then.
Want to know how your own host did? Have a gander:
Interestingly www.fcc.gov publishes AAAA records to all five of our vantages, each one a different Akamai edge set, so its geo-DNS is working beautifully. It then refused every probe we sent it, over IPv6 and IPv4 alike, from all five vantages, on both passes. Its sibling hosts apps.fcc.gov and licensing.fcc.gov, on the same Akamai platform, answered over IPv6 without complaint. That pattern is consistent with the main site blocking traffic from data-centre address space wholesale. To measurement infrastructure, the US telecoms regulator is indistinguishable from a dead host. This is exactly why we probe IPv4 alongside IPv6 everywhere; without the v4 signal we'd have filed the FCC under "broken IPv6" ... and we'd have been wrong.
The full dataset is published alongside the report, one merged record per host with per-vantage-point detail, and the frozen host list is checksummed so you can hold us to the October re-run. The full report has the methodology and the caveats (temporal skew between vantages is the big one), plus interactive charts and a map that shows which routes work for any host you click. The March study is still up for comparison.
.svg)
