Method — DNS Census

Where the numbers come from

Every day we download the complete zone file for each top-level domain that grants us access — the generic TLDs through ICANN’s Centralized Zone Data Service, and .com and .net under Verisign’s zone file agreement. A zone file lists every name delegated in that zone, so the census counts all of them; it is not a sample or a survey. On October 23, 2026 that was 259,746,591 names in 1,078 zones.

Births, deaths and moves

Comparing two consecutive snapshots gives three lists: names present today and not yesterday (born), names present yesterday and not today (died), and names present on both days whose nameserver set changed (moved). These are delegation changes, not registrar events. A newly registered name usually appears the next day, but a name can also appear because it was restored or released from a hold, and disappear because it was suspended. Only zones present on both days are compared, so a zone joining the census does not show up as millions of births.

What is missing

The census starts on July 25, 2026; the zone files are served only for the current day and never backfilled, so nothing earlier exists. Days with no data: 2026-07-30, 2026-08-26, 2026-10-11 (no usable snapshot: the capture repeated the previous day or none was published; shown as gaps, never as zero). Days on which some zones were excluded from the comparison: 2026-07-25, 2026-07-27, 2026-07-28, 2026-07-29, 2026-08-04, 2026-08-07, 2026-09-10, 2026-09-12. Country-code zones such as .uk or .de do not publish their zone files and are not counted.

Name trends

Each newborn label is split into its most probable sequence of English words using a word-frequency list (the open-source wordninja list, plus a short list of newer words such as “crypto” and “agentic”). Words shorter than three letters (except “ai” and a few others), rare dictionary fragments, filler words and a list of explicit terms are ignored, as are internationalised names. Brand-new coinages that are not in the dictionary cannot be found this way.

A single registrant can register tens of thousands of names in a day, which would swamp every trend. So any one nameserver set that carries 10% or more of a day’s births is treated as a bulk cohort (shared nameservers suggest, but do not prove, one registrant), and trends are measured only on the other births — on every day, so yesterday and the baseline describe the same population. For each word, the expected count is its rate over the previous 28 complete days times yesterday’s births. A word is listed when it reached at least 20 names and a smoothed ratio, (count + 5) ÷ (expected + 5), of at least 2; words are ranked by (count − expected) ÷ √(expected + 4), a smoothed Poisson-style score of how surprising the jump is. The +5 and +4 keep rare words from producing huge ratios on tiny counts.

A word moves to “likely bulk registrations” when its jump exists only because of such a concentrated nameserver set, when 60% or more of its names share one nameserver set, when 30% or more are the same name bought in many zones, or when 60% or more share one other word (one naming pattern, often built around a brand). These are heuristics about who is registering, not facts about any owner.

What we publish, and what we don’t

We publish counts and statistics derived from the zone files, and a few example names per featured word (at most 6 each, 40 words a day). We never redistribute the zone files or any bulk list of names, never link to the example names, and never use them for marketing. The machine-readable summary is /data/census.json.

Who runs this