Scores a document’s quality by reading the quality of the documents that link to it and the affiliation among them, moving link-based ranking beyond raw inbound counts into authority-aware territory.
Patent Overview
- Inventor
- Amit Singhal
- Assignee
- Google LLC
- Filed
- 2006-07-31
- Granted
- 2010-08-24
- Application Number
- US 11/461,259
The Challenge
Counting Links Is Not Counting Quality
Early ranking systems treated every inbound link as roughly equal evidence of document quality. Real authority does not work that way. A link from a respected, on-topic source is worth more than a link from an unrelated page or a link farm. The ranking system needs to evaluate not just how many links point to a document, but how good those linking documents are, and whether they form an affiliation pattern that should be discounted.
- Equal-Weight Counting Is Manipulable — When every inbound link counts the same, the cheapest way to rank is to acquire many low-quality links. Link farms exploit this directly.
- Linker Quality Is The Real Signal — A link from a high-authority source carries far more endorsement than a link from an unknown page. The system needs a recursive evaluation where linker quality flows into linked-document quality.
- Affiliated Links Should Not Reinforce — When a cluster of related sites all link to each other, the apparent endorsement is artificial. The ranking has to detect affiliation and downweight intra-cluster links.
- Need A Stable Recursive Score — If linker quality depends on its own linkers’ quality, the computation is recursive. The system needs an iterative or eigenvalue-style method that produces a stable score per document.
- Spam-Resistant By Design — Manipulation attempts (paid links, reciprocal link networks, mass linking from low-quality sources) should produce smaller ranking gains than they would under naive counting. Spam resistance is a primary design goal.
Innovation
Quality Flows Through Links, Affiliation Discounts Them
A document’s ranking score is computed from the ranking scores of the documents that link to it, weighted by affiliation. Affiliated linkers (same site, same owner, mutually interlinked cluster) contribute less. Independent, high-quality linkers contribute more. The score iterates until it converges, producing a quality estimate that resists single-vector manipulation.
- Identify The Document’s Linkers — For each document in the corpus, enumerate the set of documents that link to it. Inbound links are the input to the quality computation.
- Look Up Linker Quality Scores — For each linker, retrieve its current ranking score from the previous iteration. The score reflects how much endorsement that linker can contribute.
- Detect Affiliation Among Linkers — Identify groups of linkers that share affiliation: same host, same owner, or membership in a mutually interlinked cluster. Affiliation can be inferred from WHOIS data, IP proximity, link patterns, or content similarity.
- Discount Affiliated Contributions — Affiliated linkers contribute less to the target document’s score. The discount can be a fraction per affiliated link, or a cluster-level cap on how much the whole cluster can contribute.
- Compute The Score — Sum the discounted linker contributions to produce the candidate document’s new ranking score. The summation is iterated across all documents until scores stabilize.
- Iterate To Convergence — Repeat the score computation across the entire corpus until per-document scores no longer change materially between iterations. The converged scores represent the quality estimate.
- Apply To Ranking — The converged quality scores feed into the live ranking pipeline. Documents with high quality scores rank better, all else equal.
Affiliation-Aware Authority Flow
The patent’s contribution beyond classic PageRank is the explicit recognition that not all inbound links are independent endorsements. Affiliated linkers form a single signal, not a chorus. Discounting affiliation is what prevents internal cross-linking and reciprocal networks from inflating quality scores.
Independent Endorsements Count More
When the documents linking to you are independent of each other and of you, their links are stronger evidence of quality than the same number of links from an affiliated cluster.
- Linker Quality Recursion — A linker’s contribution is proportional to its own quality score. High-quality linkers contribute more; the recursion converges to a stable per-document score.
- Affiliation Discount — Affiliated linkers (same site, same owner, mutually interlinked) contribute less. The discount eliminates the manipulation incentive of creating link networks.
- Iterative Convergence — Scores are recomputed across the corpus until they stabilize. The converged values represent quality as a system-wide equilibrium rather than a local measurement.
Authority flows, but affiliated links are not new authority.
<\/section>Technical Foundation
The Authority Recursion
The ranking value of a document depends on the ranking values of its linkers and the affiliation among them. The system computes this recursively across the whole corpus.
- Inbound Link Set — The set of documents that link to a given document. Computed from the crawl graph and updated as the corpus changes.
- Linker Quality Score — Each linker has its own ranking score from the previous iteration. The score is what flows through the link to the linked document.
- Affiliation Indicator — A boolean or graded indicator of whether two linkers are affiliated. Sources include same-host, same-owner, cluster membership, and content-similarity signals.
- Discount Function — How much affiliated linkers contribute relative to independent ones. Can be a per-link fraction or a cluster-level cap on total contribution.
Quality Metrics
- Document Quality Score — Iterated until convergence. High Q(D) means the document has strong, independent, high-quality linkers.
Q(D) = sum( discount(L, D) * Q(L) for L in linkers(D) ) - Affiliation Discount — Independent linkers contribute full strength; affiliated linkers contribute partial strength. The exact discount is tuned per affiliation type.
discount(L, D) = 1 if independent, 0 < d < 1 if affiliated
Key Insight: Link-based ranking only works if the system can distinguish independent endorsements from manufactured ones. The affiliation discount is what makes the score resistant to the cheapest manipulation strategies. Without it, link counting would remain trivially gameable; with it, manipulation requires acquiring genuinely independent high-quality links, which is what the algorithm rewards as a side effect.
<\/section>The Process
The Score Computation
Offline computation across the crawl graph produces per-document quality scores. The computation runs on a schedule because the corpus and links change continuously.
- Build The Link Graph — From the crawl, construct a directed graph where nodes are documents and edges are links. Each edge has attributes including anchor text and position.
- Compute Affiliation Signals — For each pair of documents, compute affiliation indicators. Sources include host match, ownership data, cluster membership in interlinked groups, and content fingerprinting.
- Initialize Quality Scores — Each document starts with a uniform or seeded quality score. The initialization does not affect the converged result but does affect convergence speed.
- Iterate — For each iteration, recompute every document’s quality score from its linkers’ scores with affiliation discounts applied. Stop when scores converge.
- Publish To Ranking — Write the converged scores to the ranking system. Subsequent retrieval uses the scores as one input to relevance ranking.
Quality Control
Quality Control
Resisting Link Manipulation
Several controls keep the quality score honest in the face of adversarial link patterns.
- Strict Affiliation Detection — Affiliation is detected aggressively. Same-host, same-owner, and obvious cluster patterns all trigger the discount. False negatives cost the system spam resistance.
- Cluster Contribution Cap — A cluster of affiliated linkers cannot contribute more than a configured maximum total, regardless of how many linkers are in the cluster. The cap prevents large networks from accumulating false endorsement.
- Iteration Stability — Scores must converge. Non-convergence is treated as a signal that the link graph contains pathological structure (cycle of mutual high-confidence links, for example) and is flagged for investigation.
- Quality Score Floor — Documents below a minimum quality score do not contribute to other documents’ scores at all. This prevents low-quality pages from being used as sources of artificial endorsement.
What This Means for SEO
What This Means for SEO
Link quality is one of the longest-discussed topics in SEO and this patent encodes the foundational mechanics. The implications shape both link acquisition strategy and internal linking practice.
- One High-Quality Link Beats Many Weak Ones — A link from an independent, authoritative, on-topic source contributes more than dozens of links from low-quality or affiliated sources. Acquisition should optimize for source quality, not link count.
- Affiliated Networks Are A Trap — Sites that interlink heavily within a cluster (same owner, same network, mutually reinforcing) get discounted. Cross-site link rings, private blog networks, and reciprocal link schemes are explicitly the target of the affiliation discount.
- Internal Links Don’t Compound The Same Way — Links within your own site are by definition affiliated. They distribute authority across your pages but do not create new authority. External independent links are the only source of new quality signal.
- Owner-Level Affiliation Is Detected — Multiple sites under the same owner (WHOIS, payment info, hosting fingerprints) are treated as affiliated even if they appear distinct on the surface. Cross-property linking between owned sites does not stack the way independent linking does.
- Anchor Quality Inherits Linker Quality — Anchor text contributes to relevance but its strength is bounded by the linker’s quality. A spammy site using your target keyword as anchor text contributes little. A respected site using a generic anchor contributes a lot.
- Topical Affiliation Of Linkers Matters — On-topic linkers from your subject area contribute more to your quality signal in that area. The recursion concentrates score within topical neighborhoods.
- Quality Score Has A Floor — Below a certain quality threshold, a linker contributes nothing. Comment spam and forum-signature links rarely register because the source sites are below the floor.
- Build For The Convergence, Not The Snapshot — Quality scores iterate until they converge. Sudden link-acquisition spikes register only after the next convergence cycle and are mediated by affiliation detection. Sustainable, diverse, independent link acquisition compounds over multiple cycles in ways short bursts cannot.