TrustRank
TrustRank is a semi-automatic link analysis algorithm that starts from a small set of human-vetted seed sites, propagates trust along links, and flags pages that trust barely reaches as likely spam. It was proposed in 2004 by researchers from Stanford University and Yahoo! as a technique for combating web spam.
- TrustRank is a semi-automatic algorithm that begins with human-verified, trustworthy seed sites, spreads trust along links, and classifies pages that trust seldom reaches as spam.
- It was introduced in 2004 by Zoltan Gyöngyi and Hector Garcia-Molina of Stanford University and Jan Pedersen of Yahoo! in the paper Combating Web Spam with TrustRank.
- Its core premise is the principle of approximate isolation: good pages tend to link to other good pages and rarely point to spam.
- Reportedly, fewer than 200 high-quality seeds are enough to filter spam effectively across a large portion of the web.
Overview
TrustRank is a link-based trust propagation algorithm designed to single out web spam that tries to claim undeservedly high rankings in search results. Because manually inspecting the entire web is prohibitively expensive, the method starts from a small set of seed pages that humans have judged trustworthy and then spreads that trust automatically along the web's link structure. The further a page sits from the seeds, the less trust it receives, and pages that trust barely reaches are treated as likely spam.
What distinguishes this technique is its semi-automatic design, combining a human-driven step that selects the seeds with an algorithmic step that propagates trust. By pairing human judgment with machine scalability, it aims to filter spam across a wide swath of the web with only a modest amount of manual effort.
How It Works
Oracle Function and Seed Selection
TrustRank treats the human expert who decides whether a page is spam or high quality as an oracle function. Since such review is costly, the algorithm picks seed candidates so that as few pages as possible need to be evaluated. Candidate selection relies on inverse PageRank, which flips the web graph's inlinks and outlinks (running PageRank on the transposed matrix) to surface pages with strong propagation reach that can efficiently spread trust to many other pages. Human experts then evaluate these candidates to assemble the final trusted seed set.
Trust Propagation and Attenuation
Once the trusted seeds are fixed, the trust scores assigned to them propagate to neighboring pages along the links. This proceeds iteratively, much like PageRank: at each step a node's trust score is split among the neighbors it points to and reduced by a fixed factor (attenuation). As a result, the greater the link distance from a seed, the lower the trust that arrives, and pages that trust hardly reaches end up with low trust scores.
Approximate Isolation Assumption
What makes this propagation meaningful is the empirical observation that good pages link to other good pages and rarely point to spam, known as the approximate isolation of the good set. Thanks to this assumption, trust originating from reliable seeds flows mainly among high-quality pages and is expected to seldom reach spam regions.
History and SEO Misconceptions
TrustRank was presented in August 2004 at the 30th VLDB (Very Large Data Bases) international conference, a joint research effort between Stanford University and Yahoo!. The paper showed that a high-quality seed set of fewer than 200 pages is enough to effectively separate spam across a substantial part of the web. The idea grows out of the concern that Google's PageRank is vulnerable to link manipulation, and it differs by explicitly anchoring the source of trust in the seeds.
In the SEO industry, beliefs such as "Google uses TrustRank as is" or "every site has a published TrustRank score" circulate widely, but these reflect an inaccurate understanding. Sources like Wikipedia note that ideas from this family of algorithms have been incorporated into the Yahoo! and Google search engines, yet Google has never officially confirmed that it uses the paper's TrustRank in its production rankings. The "Trust Rank" trademark that Google separately filed, as well as the proprietary metrics of some tools, should also be kept distinct from the TrustRank of the original paper. The practical takeaway lies not in any score itself but in the principle-level insight that link distance from trusted sources and good neighbor relationships can serve as signals for identifying spam.