Agent Rank: the Google patent behind the idea of Author Rank in SEO

Agent Rank is a reputation score assigned to the people and organizations who sign content, used to rank the content they sign. Patent US 7,565,358 B2, "Agent rank," describes it. Google filed it on August 8, 2005 (application 11/200,406), the application was published as US 2007/0033168 A1 on February 8, 2007, and the patent was granted on July 21, 2009.

What does patent US 7,565,358 describe?

Patent US 7,565,358 describes a system that links content to the agents responsible for it through digital signatures, then scores those agents from the content they sign. Its 2 inventors are David Minogue and Paul A. Tucker. The patent holds 14 claims and 3 drawing sheets.

The abstract defines the method: receive content items, receive "digital signatures each made by one of multiple agents," and assign "a score to a first agent of the multiple agents," based "upon the content items associated with the first agent by the digital signatures." The summary adds that "the score is used in ordering results of a search."

Which problem does Agent Rank solve?

Agent Rank solves the difficulty of knowing who is responsible for a piece of content. Link-based scores measure the authority of a page or a site, but "it is difficult to correlate content on an information source with the agents responsible for creating or publishing that content." The patent gives 3 reasons:

1. An author writes on several sites: "an individual author may contribute content to multiple information sources."
2. A page mixes several agents: content "may originate from multiple agents," such as "advertisements which appear alongside news articles."
3. Ownership changes: "ownership and control of information sources may change over time."

What is an agent?

An agent is "any individual or entity that either provides content pieces, edits existing content pieces, or reviews existing content pieces on a web page." The owner of a page is "the agent that has ultimate control over the web page," including the content pieces provided by other agents. Other agents generally have limited control: a commenter can place a comment but "not edit or delete the comments of other agents." The patent's sample page (FIG. 1) holds main body text, 2 comments, an image and an ad, each of which can come from a different agent.

An agent signs content in a role. The patent lists 4 roles: "author, publisher, editor, or reviewer." The signature "provides evidence that a particular agent has asserted its role," because "the agent has exclusive access to the private key used to sign the content piece." The signature can carry metadata such as "creation date, review score, or recommendation."

How do digital signatures work in Agent Rank?

Digital signatures work at the level of a content piece, not only of a page. Any suitable signature protocol can be used, and the patent gives the W3C XML Digital Signatures standard as an example. It describes 6 properties:

1. Partial signing: an agent can sign a document "while excluding any ads which are being served alongside the document."
2. Any granularity: "Signatures can be applied to anything from an individual hyperlink to an entire document," and to text, images, audio or video.
3. Integrity: the signature lets anyone verify "that the content that is signed has not been materially altered."
4. Portable or fixed reputation: a syndicated columnist can sign a column once and "have the signature follow the document wherever it is published," or tie the signature to the URL of a site.
5. Several agents per page: "a message board or web log can allow each post to be signed by its respective author."
6. Unsigned content: it is attributed to "a synthetic agency identified by the host, site, or URL on which the content piece appears."

The signature can sit in 3 places: appended to the content piece, behind a link (a URI) from the content piece, or in a central file or directory that links to the content it covers. The agent can also sign a digest or hash of the content instead of the content itself. A single agent that controls several websites can sign the content on each of them to show it is responsible for all of it.

The agent's real identity is not required: "it is not necessary to know the actual identity of the agent." Each agent uses "a private key unique to the agent," and "no personal information about the agent is necessary." If the private key is compromised, the agent can revoke it, and the signatures made with it "will not be accepted as valid."

How is an agent's reputation computed?

An agent's reputation is computed like PageRank, on a graph of agents instead of a graph of pages. "An agent should have a higher reputational score, regardless of the content signed by the agent, if the content signed by the agent is frequently referenced by other agents or content." References are not equal: "a reference by another agent with a high reputational score is of greater significance."

The patent builds a linked database where each node is an agent with all its content, and each edge is a link from a content piece of one agent to a content piece of another. The rank r(A) of an agent sums the ranks of the agents that link to it, each divided by the number of forward links in the content that agent signs. The formula also uses N, the total number of agents in the database, and a constant α between 0 and 1, interpreted as "the probability that a user will jump randomly to any content piece instead of following a link." An alternative starts from "a seed group of trusted agents" whose endorsements build reputation, the same idea as seed-based PageRank. In both cases, an agent's reputation "ultimately depends on the quality of the content which they sign."

The ranking can be query-independent and "calculated offline, prior to accepting a user query," or computed within a specific query by limiting the corpus "to the query-relevant content."

Can an agent have several scores?

Yes: an agent can have a score per role and per topic. "An agent might have a first score for content that the agent has written, and a second score for content that the agent has reviewed." The patent's example: an agent behind an entertainment magazine "could have a high reputation score for content related to celebrity news, but a low reputation score for content related to professional medical advice." Queries can be classified into topics, "e.g., sports or medical specialties," with a different agent rank for each.

How does Agent Rank change search results?

Agent Rank changes results by promoting content signed by reputable agents. Content "authored and signed by that agent will be promoted relative to unsigned content or content from less reputable agents in search results." A reputable reviewer raises the rank of the content it reviews.

In one implementation, the patent builds in a protection: "reputational scores are relatively difficult to increase and relatively easy to decrease, creating a disincentive for an agent to place its reputation at risk by endorsing content inappropriately."

How does an agent sign its content?

An agent signs its content with a content authoring tool that manages its keys (FIG. 2). The tool receives the content pieces, then the agent's login credentials, checks whether the agent already has a public/private key pair, generates one if not (for example through an authentication service), and signs each content piece with the private key. The signature can carry "a timestamp indicating the time and date that the digital signature was created, keywords relating to the content piece, or the URL of the website where the content piece is located."

An editor or a reviewer uses the same process: the signature then includes metadata "indicating that the agent edited or reviewed" the content, and can list "edits performed or review score." Anyone can check a signature with the agent's public key. In another implementation, "trusted signing authorities" record when content was signed, which "would allow priority to be determined if two agents attempted to sign similar content." The content need not be on the web: the patent mentions "e-mail, CD-ROMs, or DVDs."

What do the claims protect?

The claims protect a reputation score computed from signed content and from the reputation of the agents who reference it. Claim 1 receives content items, each tied by a digital signature to an agent and to an "assertion" that agent makes about the item, validates every signature, then computes the score of a first agent as a function of 2 things:

1. The reputation of the agents who reference it: for each signed item of the first agent referenced by another agent's item, "a reputation score of the another of the identified agents."
2. The role of a second agent: the reputation of a second agent whose item references the first agent's content, a score "based on a second assertion made by the second agent that specifies a role of the second agent."

Claim 1 also requires that at least one signed item be "a first subset of a web page," and that the score is "not a function of the second subset of the web page" that is not tied to the agent. Dependent claims cover signatures appended to the content or stored in a separate repository (claims 5 to 8), a reputation "portable across electronic storage locations" (claim 9), the ordering of resources by the score (claim 10) and unsigned content credited to the owner of its location (claim 11). Claims 12 to 14 repeat the method as instructions on a machine-readable memory device.

Did Google use Agent Rank?

No, not officially: Google has not confirmed that it used this patent. Google announced support for authorship markup in June 2011: the rel="author" markup linked articles to an author profile. In August 2014, Google stopped showing authorship in its results and stopped tracking that markup. The SEO industry calls the idea behind both "Author Rank." That program relied on markup, not on the cryptographic signatures and declared roles the patent describes, so the 2 must not be confused.

The idea that trusted people vouch for content runs through the trust patent, and the author reputation patent scores authors from the reviews of their content.

What does Agent Rank change for your SEO?

Agent Rank means that a named author with a consistent body of cited work can carry reputation from one piece of content to the next. 5 consequences follow:

1. Name your authors. A reputation can only attach to an identifiable agent: anonymous content falls back to the site owner.
2. Keep one identity across sites. The patent lets the same agent identity attach "to content at multiple locations." Use the same author name, profile and structured data everywhere you publish.
3. Write in your field. Reputation can be computed per topic: an author strong in one subject can have a low score in an unrelated one.
4. Separate your content from what you do not control. Ads, comments and guest contributions belong to other agents: mark who wrote what.
5. Protect your reputation. A score that is hard to raise and easy to lower means that endorsing poor content puts more at risk than it can bring.

The patent describes what Google's system can do. It does not confirm that Google scores authors this way today.