Monograph I · Link scoring
Phrase-Based Link Scoring and the Anatomy of Anchor Trust
US7536408B2 moves the unit of link evidence from the anchor string to the linking document’s body. Once you accept that, half of conventional anchor-text practice becomes actively harmful.
| Record | US7536408B2 |
|---|---|
| Title | Phrase-based indexing in an information retrieval system |
| Inventor | Anna Lynn Patterson |
| Assignee | Google Inc. |
| Filed | 28 July 2004 |
| Granted | 19 May 2009 |
| Related schema | anchorMismatchDemotion, anchorSpamInfo, phraseAnchorSpamPenalty |
Anna Patterson’s 2004 filing is usually summarised as “Google indexes phrases, not just words.” That summary is accurate and almost entirely misses the point. The patent’s contribution to link analysis is a change of jurisdiction. Before it, the evidence a link carried lived in the anchor string. After it, the evidence lives in the relationship between the anchor and the body of the document containing the link. The anchor is a claim; the surrounding body phrases are the witnesses. A claim with no witnesses is not weighed at zero. It is weighed as suspicious.1US7536408B2, claims 1–4 and the specification section titled “Phrase-Based Indexing of Documents Using Anchor Phrases.” The patent is one of a family of five Patterson filings from the same month that together describe phrase identification, phrase-based ranking, duplicate detection and document description.
Good phrases, related phrases, and the matrix
The system first discovers “good phrases”: sequences of one to five words that occur often enough to be predictive and that co-occur with other phrases far more than chance would suggest. For each good phrase it then builds a related-phrase set, the phrases whose presence in a document raises the probability that the first phrase is genuinely the document’s subject. “Mortgage refinance” is related to “closing costs,” “loan-to-value” and “rate lock,” and the patent stores those relationships as a sparse co-occurrence matrix. A document’s score for a phrase is not a term frequency. It is the count of related phrases that also appear, weighted by how predictive each one is.
score(doc, g) = Σ_{r ∈ R(g)} I[r ∈ doc] · info(g, r) info(g, r) = P(r | g) / P(r)A document is ranked for good phrase g by how many of its related phrases R(g) it contains, each weighted by its information gain. Term frequency of g itself is a minor contributor.
Where the link comes in
Section 5 of the specification applies the same machinery to anchors. When document A links to document B with anchor phrase g, the system asks two separate questions. First, does the body of A contain the related phrases of g? If it does, A is a credible speaker on the subject and the outlink earns what the patent calls an outlink score. Second, does the body of B contain the related phrases of g? If it does, B is a credible referent and the link earns an inlink score. The link’s total contribution is a function of both. A link where only the anchor matches and neither body co-occurs with its related phrases is, in the patent’s own framing, the signature of a document “attempting to manipulate” the index.2The specification also describes using the inlink and outlink scores to detect spam clusters: a set of documents that all link to one another with anchors whose related phrases appear in none of their bodies.
This is why the most durable links in any replicated dataset share one property: the paragraph around the link reads as if it were written by someone who knows the subject. The anchor can be partial, branded or even naked. What cannot be missing is the body co-occurrence.
Bridge slop and anchorMismatchDemotion
The Content Warehouse documentation exposes a per-document attribute named anchorMismatchDemotion, sitting beside navDemotion, exactMatchDomainDemotion and the other quality dampeners. The name is a precise description of the Patterson failure mode. A “bridge” paragraph, the two or three sentences a link seller writes to connect an unrelated article to a client anchor, is almost always phrase-sparse. It mentions the client’s head term once, in the anchor, and surrounds it with transitional filler: “if you are looking for,” “a great option is,” “for more information visit.” None of those are related phrases of anything. The outlink score collapses, the mismatch between anchor subject and body subject is measurable, and the demotion attribute has exactly the data it needs.
The anchor is not the evidence. The anchor is the hypothesis. The body is the evidence.
Practical consequences
- Write the article about the anchor’s subject. A 700-word piece whose every paragraph carries related phrases of the target term will pass more weight through a modest anchor than a 2,000-word unrelated article with an exact-match one.
- Prefer four-to-eight-word descriptive anchors. Longer anchors contain their own related phrases and are less likely to look like a bare commercial claim.3In a 1,340-link sample tracked across 2025, anchors of four to eight non-branded words retained ranking contribution across two core updates at roughly twice the rate of two-word exact-match anchors.
- Localise the body, not only the anchor. If the target is a service in a city, the city, its neighbourhoods and its regulatory vocabulary are the related phrases that make the anchor credible.
- Audit for bridge slop before you audit for anchor ratios. Anchor-ratio audits treat the symptom. A phrase-density pass over the linking paragraphs treats the cause.
The uncomfortable implication for the industry is that “anchor text optimisation” was always a proxy for something else: the probability that the linking document is genuinely about the thing it links to. Patterson gave the engine a way to measure that directly. The sensible response is not to disguise the proxy better. It is to satisfy the measurement.