Accuracy & methodology

Every number on this page is computed live from the same database that serves every lookup — nothing here is a marketing snapshot. The methodology behind it is on the data page.

Clustering benchmark

A hand-audited set of 91 labeled pairs of real registry records — pairs that must belong to the same organisation, and adversarial pairs that must never be merged (name collisions like Virgin Media vs Virgin Atlantic, national operating companies of shared brands, carriers vs their customers). The clustering output is scored against every label on every request; each label was added after a real investigation, and any algorithm change that regresses one blocks release.

1.00
Precision — clustering benchmark (n=91)
1.00
Recall — clustering benchmark (n=91)

This scores clustering decisions only — did two netblocks that must (or must not) belong to the same organisation end up in the same cluster. It is not a measure of end-to-end footprint completeness or of attribution precision across the whole dataset; see "Global coverage" below for the address-weighted honesty check on that.

Global coverage

Address-weighted share of all delegated IPv4 space, per attribution confidence — the full per-registry table (including IPv6 and the honest "not indexed" column) is on the data page. This is the honest denominator for the whole dataset: read it alongside the corroboration table below, which only covers seven cloud/CDN providers.

3.69B
Delegated IPv4 addresses
39.7%
Registrant-grade attribution
59.1%
Carrier-inferred
0.0%
Not indexed

Ingestion recall — self-published ranges (not precision, not global accuracy)

These blocks are ingested from each provider's own published list, then checked that our clustering still attributes them to a cluster carrying that provider's name — so this measures recall against a source we already hold (did an algorithm change silently misroute space we started with), not precision, and it is self-corroborating rather than independent ground truth. It says nothing about false positives, and covers only these seven providers — not the millions of registry and BGP-inferred netblocks that make up the rest of the dataset (see "Global coverage" above for that). A drop from 100% is a regression alarm; 100% itself is partly guaranteed by how the check is built, not proof of overall accuracy. Re-checked every pipeline run.

ProviderBlocksStill correctly attributed
microsoft62,581100.0%
aws11,598100.0%
google1,245100.0%
oracle1,109100.0%
github345100.0%
cloudflare22100.0%
fastly21100.0%

What we deliberately don't claim

No public dataset can attribute literally every IP address — unregistered, unrouted, and carrier-NATed space is invisible to everyone. Where our evidence is an inference (a block attributed to its announcing carrier) rather than a registry record, it is labeled as such on every page and counted separately above. Ranges related to an organisation without merge-grade evidence appear in a separate "possibly related" section on its page, never in its own footprint.