TrustRank

The critical manual process of identifying a small, pristine set of highly reputable pages that serve as the origin points for trust propagation. Unlike PageRank's democratic link voting, TrustRank relies on an absolute human judgment to bootstrap the algorithm. A seed set typically includes university domains (.edu), government resources (.gov), and established scientific journals. The quality of this initial set is paramount; if a spam page is accidentally included, the trust score becomes contaminated. The selection process often involves reverse link analysis, identifying pages with no outbound links to spam and a high ratio of authoritative inbound links.

The mechanism by which trust flows from the seed set through the citation graph. A page's trust score is a function of the trust of its inbound links, divided by their out-degree. Crucially, trust decays with each hop away from the seed set. A page linked directly from a seed page receives high trust, while a page four links away receives significantly less. This is modeled using a damping factor, similar to PageRank, but the random surfer is biased to jump only to trusted nodes. This creates a trust radius where pages too distant from the core are automatically flagged as suspicious.

A differential analysis technique that quantifies the impact of spam on a page's ranking. It works by calculating the difference between a page's PageRank (which includes spam links) and its TrustRank (which filters them). A high PageRank coupled with a disproportionately low TrustRank indicates a high spam mass. This metric allows search engines to demote pages that have artificially inflated their importance through link farms, comment spam, and paid link networks without needing to identify every individual spam link.

The inverse of TrustRank, this technique starts with a seed set of known spam pages and propagates distrust backward through the link graph. While TrustRank identifies good pages, Anti-TrustRank actively hunts for bad actors. Pages that link to known spam sources are flagged as suspicious. This is particularly effective at identifying link farms and hijacked sites that have been modified to link to spam. The combination of TrustRank and Anti-TrustRank creates a robust classifier for web quality.

An evolution of the original algorithm that segments the seed set by subject matter to create topic-specific trust scores. A page can be highly trusted for 'medical information' but have zero authority for 'financial advice'. This prevents the conflation of general popularity with domain expertise. The algorithm uses ontology-based classification to group seed pages and then propagates trust within those topical silos. This directly addresses the E-A-T (Expertise, Authoritativeness, Trustworthiness) framework used by modern search quality raters.

The primary adversarial target of TrustRank. Link farms are densely interconnected networks of sites created solely to inflate link counts. TrustRank neutralizes them because they exist in a zero-trust neighborhood—they are too many hops away from the seed set to receive any meaningful trust score. Detection algorithms look for dense bipartite cores, where a set of 'hubs' link to a set of 'authorities' with no organic connectivity to the rest of the web. Modern systems combine TrustRank with temporal link velocity analysis to identify farms that appear overnight.