A 0 to 100 score for how deeply a domain is woven into the trusted layer of the web: the sources AI answer engines already cite. This page is the reference for how the score works and what it does and does not claim.
Agent Citability scores a domain’s citability: whether the parts of the web that answer engines treat as trustworthy have chosen to reference it. It measures a capacity, not an event. A high score means a domain carries the kind of provenance those engines select for, not that any engine has cited it or will.
The citation record explains why provenance is the thing worth scoring. Across 22,295 AI answers in one study of business-software prompts, the ten most-cited domains accounted for only 11 to 13% of citations on each engine, with around 88% of citations outside the top ten. A Washington University measurement study of 55,393 trending queries found that nearly 30% of the domains cited in a Google AI Overview don’t appear in the first-page results shown beside it. Citations are dispersed and engine-specific. What a domain can hold is the provenance those engines draw from.
The computation is seeded trust propagation: personalized PageRank over our domain-level web-graph index, following the TrustRank method published by Gyöngyi, Garcia-Molina and Pedersen in 2004. The current release covers more than 100 million domains and billions of links between them. Trust flows outward from a curated, tiered seed list of over 1,000 sources, and final scores are calibrated to a percentile within each release.
Links are weighted by who published them. The seed list is tiered; a few example sources per tier, with no weights attached.
Encyclopedic, governmental, academic and wire sources. References here are rare, hard to earn, and move a score the most.
Newsrooms and reference publishers with editorial control over every outbound link, plus university institutional pages.
Aggregators and moderated communities: placement is open, but curation and review are real.
Heavily cited by answer engines, and fully open placement. A link anyone can create in about a minute carries almost no weight here. That is by design, not an oversight.
We weight links by whether the citing domain chose to publish them, not by how often agents read that domain.
The seed list is a calibration anchor, not a whitelist. Domains outside it are still scored; they inherit conservative defaults by classification rule rather than being zeroed. The domains above are examples: the full list and its weights are not published.
Recency. A link’s weight decays with the age of the page carrying it, measured from when that page was last updated rather than when it was first published. Answer engines respond to maintained content, so a well-kept page from 2009 can outweigh an abandoned one from last year.
Reachability. Whether a source’s content can actually reach a given engine is a live input, not a fixed attribute. A source that blocks an engine’s crawlers cannot put content in front of that engine, whatever its quality, and posture like that changes over time. We track it per release rather than assuming it.
Scores are recomputed for each rebuild of the graph index and shipped as versioned releases, each with release notes naming what moved and why. The current release is May–July 2026 index · preview.
Scores fall. A domain drops when trusted sources stop referencing it, when a linking domain is re-tiered, or when recalibration shifts the percentile under it. A falling score is usually the metric working: trust was withdrawn somewhere, and the number registered it. Scores are comparable within a release; we do not promise the scale means exactly the same thing across releases, and the release notes exist so you can see what changed.
Agent Citability does not predict that any engine will cite a domain. Citation data is dispersed and engine-specific; the score measures the capacity, not the event.
The score says nothing about whether a domain’s content is correct, or whether an answer that cites it will be.
It measures provenance: where a domain’s links come from. Provenance is an input. Traffic, rankings and citations are outcomes, and none of them are promised here.
The core hypothesis is untested in public data. Published studies measure which domains get cited; we have not found one that measures whether a link from a trusted domain predicts that the linked domain gets cited. Until that test runs, the seed set and its tiering are a well-grounded prior, not a validated model.
Citation behaviour is also noisy and unstable. Repeated identical prompts return different citation sets, engines reshuffle their preferred sources within months, and most published measurement is US-only while much of our inventory is Australian. Strong correlation claims against a surface that noisy deserve suspicion, including from us.
One current-release limit is worth stating plainly. Many major news publishers restrict automated crawling or hold their pages behind paywalls, so their sites emit almost no outbound links into this release of the graph. In practice, the preview’s trusted-source signal comes mostly from reference, government and academic domains: news weight exists in the model but has little to act on. More news seeds would not fix that. Broader crawl coverage will, and the release notes will say when it moves.
The validation work: a fixed prompt panel held constant across engines, run in the US and Australia, with each prompt repeated per cycle so we know our own noise floor before claiming signal; matched domain pairs differing mainly in inbound-link profile, tracked over time, which tests the hypothesis directly; and tiering re-derived from that panel on a schedule rather than from vendor studies. If seeded trust turns out not to predict citation outcomes, that is a finding we will publish.
The method is published: seeded trust propagation, the tier philosophy, example seed domains, and the release process, all on this page, with a changelog to follow from the first stable release. That is the part you can hold us to.
The exact tier weights are private, and we would rather say that plainly than imply an openness we don’t have. A public, stable scoring function is a spec sheet for the people who sell inflated authority scores: documented experiments paid between $15 and $85 per domain to land Domain Rating between 44 and 59 on domains that were days old.
What that means in practice: you can audit what we claim to measure and argue with the tiering, but you cannot recompute our number. You can check the ingredients and the machinery; you cannot reproduce the output.
The strongest published argument against any link-based metric, and why we think this one is worth testing anyway, is covered in the announcement post rather than repeated here.
Agent Citability is a v0 preview computed from the current release of our web-graph index. Scores appear on marketplace listings, in the free domain report, and in the free checker, which will look up any domain we hold a score for. Treat preview scores as directional: tiering and calibration will move before the first stable release, and the release notes will say how.
