IPv6 Adoption Analysis (Extended)

1  Summary Statistics

Of the 1,000,000 hosts surveyed, 357,282 (35.7%) had a AAAA record visible from at least one vantage point, and 316,779 (31.7%) were reachable over IPv6 from all five. A further 5,530 hosts (0.6%) were reachable from some vantages but not others, 10,082 (1.0%) published AAAA records but were unreachable over IPv6 from everywhere despite responding over IPv4, and 24,891 (2.5%) were unreachable over both protocols during the measurement window.

The March study asked whether hosts were reachable over IPv6 from one point in California. This study asks two harder questions: whether they are reachable from everywhere, and whether the answer depends on where you look from. For 98.4% of hosts with a AAAA record anywhere, it does not: the five vantages agree unanimously (reachable from all five, or from none). The remaining 1.5% of AAAA hosts, the partially reachable, are examined in section 5.

2  Adoption by Rank Bracket

The March study found IPv6 adoption falling from 71% in the top 100 hosts to 32% at ranks 50,001–100,000. Extending the survey to 1,000,000 hosts shows the decline continuing into the tail, but not monotonically.

Reachability in the top 100,000 of this list is 39.5%, falling to 26.8% at ranks 250,001–500,000. The deepest bracket then rises to 32.8%. We have not yet isolated the cause of this uptick; a plausible contributor is mass-hosted and parked domains, which concentrate deep in the tail and inherit IPv6 support wholesale from a small number of hosting providers, but confirming that requires an address-clustering analysis we have not done. We flag it as an open question rather than a finding.

The gap between announcing IPv6 and delivering it widens with depth. Per vantage, hosts in the top 100,000 of this list show a 1.4–2.0 point gap between AAAA presence and reachability; over the full million that gap grows to 2.2–2.7 points, uniformly across vantages. Highly ranked operators who publish AAAA records almost always serve working IPv6; deeper in the tail, stale and broken AAAA records become more common.

3  Adoption by Top-Level Domain

Splitting the million hosts by their final DNS label separates operational cultures in a way rank brackets cannot. Among the twenty most common TLDs, .io is the outlier: 79.1% of its 37,837 hosts are reachable over IPv6 from all five vantages, two and a half times the population average, consistent with its developer-heavy population hosted on CDN-fronted platforms that enable IPv6 by default. The newer Google-operated registries follow the same pattern (.dev 50.3%, .app 38.6%), and .nl leads the country codes at 42.6%. At the other end sit .jp (13.9%) and .ru (10.0%).

Where the two bars diverge widely, published AAAA records are not being honoured, and the per-vantage data says why. For .ru, 488 of its 1,947 AAAA-publishing hosts (25%) were unreachable over IPv6 from every vantage while remaining alive on IPv4, nine times the population-wide broken-v6 rate of 2.8%; .in shows the same pattern at 14%. Vantage-dependent partial reachability plays almost no role in these gaps (at most 56 hosts per TLD); the rest is hosts that were down on both protocols during the window, a form of list rot concentrated in .fr and .in.

4  Vantage Agreement

The headline result of adding four vantage points is how little the headlines move. Aggregate AAAA presence spans 34.3–34.7% across the five vantages and aggregate reachability spans 32.0–32.1%, a spread of a few tenths of a point. Notably, the tunnelled California vantage agrees with the four native ones to within 0.1–0.2 points on aggregate reachability: for measuring adoption rates, one reasonable vantage point is enough.

Agreement on aggregates does not mean agreement on hosts. Each pair of vantages disagrees on the reachability of 0.4–1.1% of the hosts they both saw with AAAA records, with the tunnelled vantage and Singapore disagreeing most.

Each vantage probed every host twice, and pass-to-pass "flips" (reachable on one pass, unreachable on the other) affected 0.1–0.4% of hosts per vantage. Single-pass, single-vantage reachability numbers therefore carry an inherent noise floor of a few tenths of a percent, which pairwise inter-vantage disagreement only slightly exceeds.

Splitting that disagreement by what is being observed is revealing. For reachability, the any:all ratio sits between 1.01 and 1.03 in every rank bracket: connectivity consensus is equally strong at the top and in the tail. For AAAA presence it climbs with depth, from 1.000 in the top 100 to 1.077 at ranks 250,001–500,000: what becomes vantage-dependent deeper in the list is not whether IPv6 works but whether DNS shows it to you at all, consistent with partially deployed geo-DNS and less reliable authoritative DNS among tail hosts.

5  Where Vantages Disagree

5,530 hosts (1.5% of hosts with AAAA records anywhere) were reachable over IPv6 from some vantages but not others. These partial hosts are the clearest payoff of a multi-vantage design: from any single vantage point they are indistinguishable from working or broken hosts, and only the disagreement reveals path- or policy-dependent behaviour.

The instructive example is www.ncbi.nlm.nih.gov (rank 67 in the March list), which was reachable from four vantages but not from New York, on both IPv6 and IPv4. That pattern identifies network-level blocking of the New York data centre's address space, not an IPv6 fault. Probing IPv4 alongside IPv6 at every vantage is what makes this disambiguation possible: a v6-only probe would have miscategorised a policy decision as an IPv6 reliability problem. Other partial hosts show classic regional patterns: www.foxnews.com and www.foxbusiness.com answered IPv6 probes only from Singapore, and mp.weixin.qq.com was reachable from everywhere except New York.

Partial reachability is not a tail phenomenon: in every rank bracket it affects between 1.3% and 2.8% of the hosts that publish AAAA records, with no strong rank trend. Even the top 100 contains a partially reachable host. Wherever IPv6 is deployed, a small and roughly constant fraction of it works from some networks and not others.

6  Geo-DNS Divergence

Reachability is only one axis of vantage disagreement; the answers DNS gives are another. Of the 339,041 hosts where all five vantages observed AAAA records, 112,248 (33.1%) returned a different set of IPv6 addresses to at least one vantage. This is overwhelmingly deliberate: CDNs and anycast operators steer clients to nearby edges via geographically differentiated DNS answers, and the most-linked hosts on the web (Facebook, Google properties, Wikipedia) are almost all in this category.

7  Top 100,000 Hosts

The table below covers the top 100,000 hosts of the July 2026 list with their consensus IPv6 status across the five vantages, and the number of distinct AAAA record sets observed for hosts where every vantage saw AAAA records (a value above 1 indicates geo-differentiated DNS). Use the search and filter controls to explore the data; at most 500 rows are shown at a time.

8  Change Since March

The California tunnel vantage re-probed the exact host list from the March study, giving a controlled four-month delta from the same coign of vantage. Of the 100,000 hosts, 2,054 gained AAAA records and 1,004 lost them (net +1,050, taking AAAA presence from 38,076 hosts, 38.1%, to 39,126, 39.1%); 2,147 gained IPv6 reachability and 994 lost it (net +1,153, taking reachability from 36,864 hosts, 36.9%, to 38,017, 38.0%). IPv6 adoption in this population is growing at roughly a percentage point per four months, and the net figures hide churn: combined gains and losses run nearly three times the net change in both metrics. Notable arrivals include www.ft.com and www.bleepingcomputer.com; departures include www.engadget.com and www.php.net.

9  Rank Stability

A finding about the ranking rather than about IPv6: only 44% of the hosts in the March top 100,000 remain in the top 100,000 of the apr-may-jun Web Graph release, computed with the same harmonic-centrality methodology. Of the 56,024 hosts that left, 22,415 fell deeper into the top million (median new rank 755,260) and 33,609 dropped out of the million entirely. Harmonic-centrality rank is much less stable across graph snapshots than the stability of the top of the list suggests, and longitudinal studies keyed to "the top N hosts" are largely comparing different hosts a few months apart. Our October follow-up will therefore re-probe this campaign's frozen 1M list as well as the then-current top million.

The IPv6 findings themselves replicate across both populations: the 43,976 hosts present in the top 100,000 of both runs produce consensus categories that agree within a fraction of a percent (for example 291 versus 284 partially reachable hosts, and 7,215 versus 7,224 geo-DNS divergent).

10  Temporal Skew

The five vantages did not observe each host simultaneously, and honest interpretation of inter-vantage disagreement has to account for that. Observation windows for the 1M run ranged from 2.44 hours (California tunnel) to 30.4 hours (New York); all quantities are taken from the second probe pass at each vantage. Per host, the maximum gap between any two vantages' observations had median 13.3 h, p90 24.76 h, p99 27.68 h, and maximum 27.96 h.

Two checks argue that skew is not driving the disagreement findings. First, partially reachable hosts have a lower median inter-vantage gap (11.33 h) than the full population (13.3 h): if skew were manufacturing partials, they would skew wider, not narrower. Second, the 100k run used synchronised launches across all five vantages, and on the 43,976 hosts common to both runs every consensus category agrees within a fraction of a percent (section 9). Skew widens the error bars on any single host's story; it does not move the aggregates.

11  Data Downloads

The complete merged dataset is published alongside this report. Vantage machine identifiers have been replaced with the public labels used throughout this page; no addresses or measurement values were altered. See the dataset documentation for the record schema and caveats.

• merged-1m.jsonl.gz (42.9 MB)
  Run B (top 1M), one merged record per host, 1,000,000 records
• merged-100k.jsonl.gz (4.8 MB)
  Run A (top 100k, frozen March list), 100,000 records
• summary-1m.json (10.3 KB)
  Aggregate statistics, top 1M run
• summary-100k.json (11.7 KB)
  Aggregate statistics, top 100k run
• hosts-1m.tsv.gz (10.2 MB)
  The frozen top 1M probe list
• overlap-check.txt (1.0 KB)
  Run A vs run B comparison at rank ≤ 100,000
• merge-report-1m.txt (2.0 KB)
  Per-vantage headline numbers and observation windows
• transient-flip-reports.txt (1.1 KB)
  Pass-to-pass flip statistics per vantage

The md5 of the (uncompressed) frozen host list hosts-1m.tsv is f913bab8793d18e61c253188b8d4d9af. The October 2026 follow-up will re-probe this exact list; the hash makes the "same frozen list" claim externally verifiable.

12  Methodology

Vantage points

Four vantages are cloud instances with native IPv6 in DigitalOcean data centres; the fifth is the March study's California measurement host, whose IPv6 connectivity is a Hurricane Electric 6in4 tunnel. The tunnel vantage doubles as a continuity control against March and as a check that tunnelled and native IPv6 see the same web: on aggregate reachability they agree to within 0.1–0.2 points. DNS at the DigitalOcean vantages used the provider's regional resolvers; the tunnel vantage used Google Public DNS (see the caveat in section 6).

Probing

Measurement at each vantage follows the March methodology: a DNS AAAA lookup per host, followed by an HTTPS curl -6 --head probe with a 10-second timeout for hosts with AAAA records, any HTTP response counting as reachable. This campaign adds an IPv4 control probe (curl -4) per host per vantage, which is what allows unreachable hosts to be split into "broken IPv6" (IPv4 alive) and "dead" (both protocols down), and network-level blocking to be distinguished from IPv6 faults (section 5).

Each vantage ran two complete passes, and hosts whose reachability flipped between passes (0.1–0.4% per vantage) take their pass-2 value, damping transient network noise. One exception is footnoted: at the Amsterdam vantage, the 100k run's first pass was invalidated by a host-level DNS misconfiguration (since fixed and verified), so its 100k-run figures come from a single pass and flip-filtering could not be applied there. All five vantages completed both passes in the 1M run.

The 100k run used synchronised launches across vantages against the frozen March host list. The 1M run's host list was extracted from the cc-main-2026-apr-may-jun Web Graph release by the same streaming method as March, and frozen for reuse in October.

Limitations

The limitations of the March study carry over: harmonic centrality measures link-graph importance rather than traffic, and the measurement does not distinguish native IPv6 from CDN-mediated IPv6. New to this study, temporal skew between vantages (section 10) means per-host inter-vantage comparisons are not simultaneous, and the geo-DNS divergence count excludes hosts visible to only some vantages by construction (section 6). Five vantages on three continents is a longer baseline than one, but it is not a global view: none of the vantages sits in Africa, South America, or mainland China, and reachability from those networks may differ.