Bulk Registration Report: Methodology
The Bulk Registration Report covers one day of domain registrations and identifies groups of names that a single buyer evidently registered together. This page explains how those groups are found, how they are checked, how the brief is written, and how to read the results. All times are UTC and all dates are ISO 8601 (2026-10-20).
Data and coverage
The report day is one UTC day, 00:00 to 24:00 UTC. Registrations come from DataPulse’s proprietary datasets, which include most domain name registrations on the Internet. Registration data can arrive late. When the day’s volume is under half the usual volume for that weekday, the brief says so first and gives the share received.
The report analyzes the 50 registrars with the most registrations that day (at least 8 each), plus any registrar whose volume is unusually high against its own 30-day baseline (a robust z-score of 3 or more). Every registration at those registrars is analyzed. Registrars outside that selection are not.
Finding clusters
A cluster is a group of at least five names that one buyer evidently registered together, at one registrar or several. The signals include:
- Numbered series: the same name with a changing number (
h5zl22,h5zl24). - Shared name templates: names built from a common pattern, such as one core with a few different prefixes.
- Word-prefix runs: dictionary-like names sharing a prefix, counted only at one registrar and only when most were registered within an hour, because unrelated buyers share common prefixes.
- Timing: registrations made close together, most within the same hour.
Patterns are written in a short notation: N is one digit, [a-z] one letter,
[x|y] one of the prefixes shown, [??] one of several prefixes, and a trailing *
any ending. h5zlNN means h5zl followed by two digits.
Which clusters are examined
A day typically has thousands of clusters. The 25 with the strongest naming evidence, weighted by size, are examined in depth: names containing a common phishing keyword are ranked higher, and word-prefix runs lower. Only examined clusters can be rated moderate or strong; the others are reported as found but not examined. A cluster that was not examined is not a cluster that was cleared.
Sampling
For each examined cluster we check a random sample of members rather than every name:
- Live DNS: a simple random sample sized to estimate shares of the whole cluster within ±10 percentage points at 95% confidence (Cochran’s formula with the finite-population correction).
- Web pages: a random subsample of the DNS-sampled members that have an address, sized for ±20 percentage points.
Each draw is seeded from the date and the cluster’s pattern, so a report can be reproduced. Shares are reported as “k of n” with a 95% Wilson score interval for the whole cluster, for example “24 of 24 names checked (95% CI 86–100%)”. Percentages reported for a cluster are estimates from these samples.
Shared infrastructure
A nameserver set, address set, CNAME target or mail server set counts as shared only when a majority of the cluster uses it at 95% confidence: the lower bound of its Wilson interval is at least 50%, and a sampled name without that record counts against it.
Some infrastructure never counts, because large numbers of unrelated customers use it:
- pooled CDN addresses
- registrar defaults: default nameservers, URL forwarding and default pages
- parking services
- DNS providers and mail providers whose servers every customer gets
- hosting-platform defaults
- any value that three or more clusters share that day
Defaults come from one list that DataPulse maintains. Each entry names the operator, the evidence that ties the value to it, and witness domains, and the list is re-checked against live DNS every week. A value is added only with evidence of who operates it, never on prevalence alone.
Cloudflare gives each account a pair of nameservers. A cluster whose names all sit on one pair is consistent with a single Cloudflare account, but Cloudflare reuses pairs across accounts, so this is evidence, not proof.
Evidence ratings
A cluster earns a point for each of:
- names: a numbered series advancing by one constant step, a tight burst (80% or more registered within one hour at the main registrar), or a uniform template (90% or more share one label length and suffix). Together these count for at most one point: they show names were bought together, not what they are for, and most such batches belong to domain investors;
- shared nameservers;
- a shared address, CNAME target or mail server set;
- page content: a sampled page with a content finding.
Three or more points is strong evidence and two points is moderate. Zero or one point is weak. A naming pattern alone is therefore weak evidence: a moderate cluster also needs shared infrastructure or page content, and a strong one needs two of shared nameservers, shared address or mail infrastructure, and page content.
Page signals
- Sampled pages are fetched and classified as live, parked or failed. A registrar’s default page counts as parked.
- Content findings are possible impersonation (a page presenting a brand the name does not spell), an off-site redirect, and adult content.
- A page presenting the brand that the domain name itself spells is not counted as impersonation.
- Page signals are automated leads, not verdicts, and are always reported as “possible … (unverified)”. Verify them before acting on them.
- A name whose page may contain illegal content involving minors is removed from every output (brief, email, data files) and reviewed for reporting to the appropriate authorities.
Suspected operators
Two clusters appear to share an operator when they share the operator’s own infrastructure (the same shared nameservers, address set, CNAME target or mail servers, never a default from the list above) or the same distinctive ending of their name stem. Linked clusters are shown as one suspected operator, labelled anonymously (“Operator 1”): the evidence suggests a common party but never identifies one, and the brief says “appear”, never “proven”.
Recurring operators
Each day’s suspected operators are recorded with their fingerprints: their
own shared infrastructure values and their name patterns with the counters
removed (356dxNNNN becomes 356dx#; a pattern with fewer than four fixed
characters is not used, because unrelated buyers produce it). Today’s
operators are matched against the previous 30 days, and a shared fingerprint
counts as a sighting. The brief reports it as “appears to have been active on 11
of the previous 27 days on record”, where days on record are the days a report
was produced with a cluster analysis. Matching is evidence of a recurring
operator, not proof: Cloudflare reuses nameserver pairs and two parties can pick
the same stem.
The flow chart
The chart follows each cluster’s names to the registrar that sold them and to the TLD manager: the organisation the IANA root zone database lists for the TLD, which may contract another company to run the registry. Band width is the number of names, every column holds the same names, and each cluster keeps its colour from end to end. Registrars and TLD managers beyond the six largest are grouped as “Other (n)”. A suspected-operator column appears only when an operator links two or more clusters.
How the brief is written and checked
Every number in the brief is computed by the analysis above. A language model writes the brief from those computed figures only, with no other source. A second, independent model then checks every claim in the draft against the same data, and the draft is revised until no claim is unsupported or overstated, up to three times. A brief that does not pass is not sent by email. Briefs are short: at most five clusters, labelled Cluster A to E, are described, and the brief says how many more reached moderate evidence.
Data files
Each day’s STIX 2.1 bundle (for SIEM and SOAR platforms) and RPZ zone (for DNS resolvers) list every name in every cluster with moderate or strong evidence. Moderate evidence can include parked names and unrelated names that happen to fit a pattern: review those before blocking.
Conventions
- Names, hostnames and addresses are defanged (
h5zl22[.]vip) so they never become links. - Registration windows are the main registrar’s: when other registrars sold some of a cluster’s names, the brief says how many the window covers.
- All times are UTC and all dates are ISO 8601.
Reading the report
Registrars and TLD managers are where names were bought, not who bought them. A cluster at a registrar says nothing about the registrar’s own involvement, and a registration volume is given only as context.