research · web graph

The web graph

A public, web-scale hyperlink graph, rebuilt quarterly. It is the one view of a publisher's linking behaviour that no seller can edit, and it costs nothing to check against.

methodology version 2026-09-05 · all methods

The source

The domain-level graph of the summer 2026 release: 2.45 billion edges between 120 million registrable domains. We keep the raw release and extract the sub-graph touching our inventory. Subdomains are folded to their registrable domain, so a subdomain carries no graph reading of its own; platform and free-hosting domains are excluded from every aggregate.

Outbound, inbound and saturation

For each publisher: the number of distinct outside domains it links to and the number linking in. For a site that sells placements, the first is roughly how many customers it has had. A publisher linking out to 500 or more domains with at least five times as many out as in is saturated and fires the saturated_outbound quality test. Of 189,774 sellable publishers with graph data, 11,208 are saturated; the median saturated publisher links out to 1,053 domains against 10 for the rest.

Leader links

For each of thirty topics we take the most competitive US-English keywords, the top-ranking domains for them (the leaders, public by definition and the only domains we name), and ask which inventory domains link to them. Publications link to what ranks; farms link to whoever pays. A publisher’s leader links per topic are reported on every offer and feed the relevance and graph components of the score.

Probable buyers, in aggregate only

A domain linked from three or more of our sellable publishers, with a modest outbound count and no web-scale PageRank of its own, is probably a buyer of placements. We use that only as aggregate demand (how many such domains a topic has, what they buy) and never name one. The signal is structural and can be wrong about any single domain, which is a second reason not to.

Limits

A sample
The crawl covers a fraction of the web each release. Absence from the graph is not evidence of anything.
Undated within a release
Edges are from a three-month crawl window; a link added or removed last week is not here. Fresh readings come from the backlink index and our own checks, not the graph.
Domain level
One page selling links and a thousand that do not look identical here. The graph flags a publisher's pattern; the page audit after delivery describes the page.
Refresh
Re-extracted when a new release lands (quarterly). Cross-release deltas (out-degree jumps, PageRank collapse) are planned as expired-domain evidence, not yet published.

4 published figures on this page, measured 2026-09-10. Dotted numbers carry their sample size on hover. Methods: web graph.

Every figure on this page is a field in the API. Query it on a free key, connect the MCP server to your own agent, or ask Papi in a chat.