SEO links that bring clicks: how Google scores quality from links and the traffic they send

A link-traffic quality score is a measure of a page's quality based on how many pages link to it and how many people actually click those links. Patent US 9,558,233 B1, "Determining a quality measure for a resource," describes how Google computes it and spreads it across the web. Google filed it on December 31, 2012, from a provisional application of November 30, 2012, and the patent was granted on January 31, 2017.

What does patent US 9,558,233 describe?

Patent US 9,558,233 describes a "pull-push" process that scores seed pages from links and clicks, then propagates quality to the pages that link to them and to the pages those sources link to. Its 6 inventors are Hyung-Jin Kim, an inventor of the click re-ranking patent, Paul Haahr, Kien Ng, Chung Tin Kwok, Moustafa A. Hammad and Sushrut Karanjkar. The patent was assigned to Google Inc. and holds 24 claims, 3 of them independent (a method, a system and a computer storage medium), and 10 drawing sheets.

The abstract sets out 3 scores: "a seed score for each seed resource," based on "a number of resources that include a link to the seed resource and a number of selections of the links"; "a source score for each source resource," based on "the seed score for each seed resource linked"; and "a resource score for each source-referenced resource," based on "the source score for each source resource that includes a link."

Which problem does the patent solve?

The patent solves the gap between links and real attention. A ranking process may rank a page "that has a large number of other web pages that link to" it higher. But "a resource may be linked to by a large number of other resources, while receiving little traffic from the links." The patent names the cause: "an entity may attempt to game the ranking process by including a link to the resource on another web page." To prevent this skew, the system evaluates the "mismatch" between the number of linking pages and the traffic those pages generate.

The rule follows: "If a resource is linked to by a number of resources that is disproportionate with respect to the traffic received by use of those links, that resource may be demoted in the ranking process."

How does the pull-push process work?

The pull-push process works in 3 steps:

1. Score the seeds: seed resources are pages "for which sufficient link and traffic data is available." A page with less than a threshold number of linking pages or less than a threshold number of selections may be treated as a target rather than a seed. In one implementation, the seed score is "the ratio between the number of selections of links to the particular resource and the number of source resources that link to the particular resource": the clicks per linking page. A seed with a disproportionately large number of linking pages relative to its clicks gets a low seed score. Instead of or in addition to link clicks, the patent mentions other traffic measures, "such as a total number of views for the resource, or a number of selections of search results for the resource."
2. Pull to the sources: seed scores are "pulled" up to the pages that link to the seeds. Each source page gets a source score from the seeds it links to (a sum, an average, a weighted average or another combination). A source that links to many low-scoring seeds gets a low source score. Each source is then classified as "a 'qualified source' or an 'unqualified source'" by comparing its score with a threshold; the patent suggests deriving that threshold, for example with a sigmoid function, from the link-to-click ratios of known qualified and unqualified sources.
3. Push to the targets: source scores are "pushed" to every page the sources link to, seeds included. Each of these pages gets a resource score from the sources that link to it, including target pages with too little data for a seed score of their own. For a seed, the resource score can also combine the source-based score with the seed's own score.

Links from unqualified sources "can be discounted in a ranking process that ranks resources based on the number of links." The process lets Google "distribute link-based quality measures to resources for which sufficient data may not be available."

How does the patent judge the quality of a click?

The patent judges a click by what the visitor does after it, so that low-quality clicks are discounted. Each selection of a link to a seed can get a selection quality score:

• Dwell time: a click followed by a long dwell on the seed page (above a threshold) scores higher than a short one, because "automatically generated link selections are often of a short duration." The claims sort clicks into "a short click," "a medium click" and "a long click."
• Onward navigation: a click is worth more when the visitor then clicks a link on the seed page, which is "indicative that the selections were made by a real user rather than an automatically generated selection."

The selection quality scores are combined into a seed selection quality score, which can be weighted by how recent the clicks are. The system can also keep only the clicks above a quality threshold and compute the seed score from that adjusted count. This is not a side option: in all 3 independent claims, the seed score is determined "in accordance with a seed selection quality score."

How are the scores used in ranking?

The scores are used to demote, promote or discount, alongside the relevance of each page to the query:

• Demotion and promotion: a page with a resource score below a first threshold can be demoted, a page at or above a second threshold can be promoted, and pages between the two thresholds may keep their position.
• Discounted link count: the system can subtract the number of unqualified sources linking to a page from its link count, so that "the ranking does not give a resource credit for links included on unqualified" sources.
• Combined rank score: the resource score can serve as an overall quality measure, combined with a relevance score to reorder the results.

How do queries change the weight of links and clicks?

Queries change the weight when they tend to surface spam. The patent defines a spam-related result as one for a page with a disproportionately large number of links relative to the selections of those links, for example a link-to-selection ratio above a threshold. The system computes a score for each query from its n-grams: each source page's score is assigned to the n-grams of the anchor text of its links, and the scores of a query's n-grams are combined. A low query score indicates a query likely to surface pages linked to by unqualified sources. For queries with "a high propensity for surfacing spam-related web pages, the system can put a higher weight on selection counts for the search results and a lower weight on resource link counts."

The patent gives 3 examples:

• Video sharing sites: "publishers of many video sharing web sites attempt to manipulate rankings by creating links to the sites, resulting in a disproportionately large number of links compared to the number of selections, while national news web sites typically do not attempt such manipulation."
• "Payday loan": this n-gram may be common in the anchor text of links to payday loan sites that get many links and few clicks, so it may have a low n-gram score.
• "Atlanta ER": the query may get a low score only because "ER" appears in links to video sharing sites hosting episodes of the TV show. To correct this, the system can combine the query score with the score of the whole query as one n-gram, for example by averaging the two.

How does the system resist manipulation over time?

The system resists manipulation with smoothing and alerts on sudden changes. The patent's scenario: a seed with a high score receives new links that bring no traffic, so its seed score "may suddenly drop below its previous value if only current data is used." Scores can therefore be "the weighted average of a particular number of previous scores and the score determined using current data," each weighted by the time since it was computed, or use "a smoothing algorithm" to avoid sudden changes. "If a score changes by at least a threshold amount, this score change may trigger the use of previous scores in the score determination, or may trigger human review." The patent's example: "a publisher's attempted manipulation may cause a spike in the score that triggers the use of historical data."

Scores can be computed at different levels: "domains, web sites, URLs, or partial URLs." At the site level, the link counts and click counts of the site's pages are aggregated.

How does this patent relate to other link patents?

This patent combines 2 families of signals: links and clicks. Google's link weighting patent predicts which links get clicked; this one measures the clicks links actually receive. The trusted seeds patent propagates trust from seed pages through link distance; this one propagates quality from seeds chosen because enough link and traffic data is available for them.

What does link-traffic quality change for your SEO?

Link-traffic quality means that a link is worth what it brings in real visits, and a page whose links bring no one is suspect. 6 consequences follow:

1. Earn links that people click. In one implementation, the seed score is clicks per linking page: links in places that send real visitors are the ones that build quality.
2. Send visitors who stay. Short clicks can be discounted, while long visits and onward clicks on your page count as quality clicks.
3. Distrust links nobody follows. A page with many links and little traffic from them can be demoted, and links from unqualified sources can be subtracted from its link count.
4. Get cited by pages that link to quality. A source page earns its score from the seeds it links to: links from pages that themselves cite pages with real traffic carry more weight.
5. Be careful in spam-prone niches. For queries that attract manipulation, such as payday loans, the system can weigh selection counts more and link counts less.
6. Grow steadily. A sudden change in scores can trigger the use of historical data or a human review.

The patent describes what Google's system can do. It does not confirm that Google combines links and clicks this way today.