Local experts and local SEO: how Google identifies the reviewers whose opinions rank local businesses

A local expert is a reviewer who has reviewed many businesses of one category in one area, and whose reviews Google uses to rank local results for that category and area. Patent US 9,792,330 B1, "Identifying local experts for local search," describes how they are identified. Google filed it on April 30, 2013, and the patent was granted on October 17, 2017.

What does patent US 9,792,330 describe?

Patent US 9,792,330 describes a local search system that finds expert reviewers for each area and business category, then ranks local results with their reviews. Its 2 inventors are John Alastair Hawkins and Cristina Stancu-Mara, both based in London. The patent holds 19 claims and 5 drawing sheets.

The abstract sets out the process: obtain reviews "of a reviewed category of businesses in a reviewed geographic area," identify "at least some of the reviewers as being experts" once reviews from "more than a threshold number of reviewers" have been obtained, then rank "local search results responsive to the query based on the reviews of the experts."

Which problem does the patent solve?

The patent solves the difficulty of finding real experts among reviewers at web scale. Some users are knowledgeable about "a particular genre of restaurants, a particular city or neighborhood, a type of retailer, or a combination of these categories (e.g., the New York Italian restaurants)." Such experts are hard to identify at web scale, because "many users with relatively little expertise will hold forth on various topics," "some users will self identify as experts in order to propagate spam for commercial benefit," and "the number of local-search topics on which one may be an expert is very large."

The patent also names the core flaw of consumer reviews: many "are not trustworthy, either because the reviewer lacks expertise, or the reviewer has different preferences from the user submitting a local search query." Expert identification answers the first cause, personalized experts the second.

The patent's example: one user may have submitted "reviews about 35 different Italian restaurants in New York City, while most users may have submitted a single review."

How does Google identify local experts?

Google identifies local experts in 5 steps:

1. Collect recent reviews: in some cases, the system selects reviews "of less than a threshold old age, for example two years," to work "on relatively up-to-date data."
2. Filter spam and short reviews: "Submitted reviews are first filtered to remove spam." A spam detector relies on "previous correlations between keywords, users, businesses, and the occurrence of spam"; a quality score can flag known low-quality n-grams or "mal-formed sentences"; and reviews below a threshold length are discarded, "as experts tend to write longer, more informative" reviews.
3. Categorize reviews by geographic area and business category, "in hierarchical categories with varying levels of geographic and business-type specificity," such as "California bike shops."
4. Compute expertise scores: for each user and each area-category pair, the expertise score is "the number of reviews by that user (or weighted sum)" in that pair.
5. Designate the experts: users are "ranked according to the number of reviews they have submitted," and a number, "e.g., the top 10," become the experts for the pair. Experts are only identified on a pair once "more than a threshold number of users (e.g., three)" have reviewed it, so that "users with relatively little expertise who happened to have reviewed obscure combinations" are not designated as experts.

Reviews can be weighted "based on the time since the review occurred, the length of the review, or indications from other users regarding the quality of the review, such as up vote or down votes indicating that the review was helpful."

The work is usually done in advance, by batch processes run "e.g., on a daily or weekly basis," and the experts are stored in an expert index keyed by area and business type, so they can be retrieved quickly when a query arrives. Some embodiments identify experts at query time instead.

How do specificity levels work?

Specificity levels let one review count toward several areas and categories at once. A single review of a clothing boutique in Chicago counts toward expertise scores on "clothing boutiques, Chicago clothing boutiques, Illinois clothing boutiques, retail stores, Chicago retail stores, Illinois retail stores, Chicago, and Illinois." The patent's other examples: "five reviews of sporting-goods stores in Miami and five reviews of pet-supply stores in Tampa count as ten reviews on Florida retail businesses," and reviews of "twenty specialty chocolate stores in a dozen different US states" count toward an expertise score "on chocolate stores in the United States."

When no experts exist for a precise pair, the system widens the request. Its example: no experts for "Senegalese restaurants in Marfa, Tex." The system then looks for "experts in restaurants in general in Marfa, Tex., or experts on Senegalese restaurants in the state of Texas," decreasing specificity by area, by category or by alternating between the 2, "until experts are identified." The trigger is that "less than a sufficient number of reviews were received to identify experts for that combination."

This widening is at the heart of the claims: independent claims 1, 15 and 16 identify experts "based at least on respective numbers of reviews each user submitted, including iteratively modifying the category of business to identify at least a threshold number of experts." Claim 2 adds widening the area "to a hierarchically higher geographic area."

How do experts' reviews rank local results?

Experts' reviews rank local results through an expert review score for each business they reviewed. When a query arrives, the system identifies its area and category, retrieves the matching experts, and computes the score:

1. From ratings: "an average numerical rating provided by the expert during the review," such as stars.
2. From text: the text of expert reviews can be scored by a human reviewer or by natural language processing, "based on use of terms like 'awful' or 'terrific' indicating reviewer sentiment," and the scores are averaged.
3. Weighted by expertise: the score can be "a weighted combination (e.g. a weighted average) of the reviews by the experts based on the expertise score of those experts, such that the highest-rated expert has a greater influence."

Results are then ranked or re-ranked so that "local businesses highly regarded by experts tend to rank higher than they otherwise would." The patent's example: "a bicycle shop that is very highly regarded by experts may rank higher than a slightly closer and otherwise more relevant bicycle shop." In one embodiment, the ranking of a result combines "its distance to the query location, its relevance, and the weighted average rating of the experts."

A result counts as local within a threshold distance, "e.g., within 50 miles or one-hour's driving distance," or distance can weight results, "such that closer businesses tend to rank above further" ones.

When the user selects a result, the experts' reviews of that business are ranked by expertise score (or personalized expertise score) and shown "in ranked order," for instance the top ten. Ranking reviews from the most prolific experts "is expected to highlight relatively high-quality, high-relevance reviews."

What are personalized experts?

Personalized experts are experts whose tastes match the user's. The system compares the user's reviews with those of each candidate expert on 2 criteria:

1. Shared businesses: "the number of businesses reviewed both by the user and the respective candidate expert."
2. Agreement: an agreement score, for example "an average difference between ratings of businesses by the user and the respective candidate expert."

The patent's example: if "the ninth ranked expert," by non-personalized expertise score, has reviewed more businesses in common with the user than any other expert, and their ratings match "in each instance," that expert "may be re-ranked (or re-scored) as the top personalized expert for that user." Candidates ranking above a threshold, "such as the top five," become personalized experts, and their personalized expertise score can weight their reviews in that user's results. The patent explains why: it accounts for "the tendency of users to have differing tastes."

This step runs at query time, to avoid storing experts for every combination of business type, area and user in advance, or can be preprocessed for topics the user searches "with greater than a threshold frequency."

How does this patent relate to location prominence?

This patent refines the review signal of local ranking. The location prominence patent counts documents with reviews of a business; this one asks who wrote the reviews, and gives experts of the category and area more weight. The patent does not mention Google's Local Guides program, but that program, which encourages users to contribute reviews and rewards active contributors, fits the same logic of valuing prolific local reviewers.

What does the local experts patent change for your local SEO?

The local experts patent means that reviews from people who know your category and your area can weigh more than a pile of one-off reviews. 6 consequences follow:

1. Attract reviews from experienced local reviewers. Users with many reviews in your category and area are the experts whose ratings count most.
2. Encourage detailed reviews. Reviews below a threshold length are discarded, review length and helpful votes can weight a review, and expert text can be scored for sentiment.
3. Earn praise, not just stars. Words like "terrific" or "awful" can be scored: the content of a review can matter as well as its rating.
4. Keep reviews coming. Reviews older than a threshold age (two years in the patent's example) can be left out, and the time since a review can weight it.
5. Never buy or fake reviews. Reviews are filtered for spam first, and self-proclaimed experts posting for commercial benefit are the patent's explicit concern.
6. Be strong in your exact niche. Expertise is computed per category and area, from the most specific level to broader ones: a business that excels in its niche benefits from niche experts, and can outrank a slightly closer competitor they rate lower.

Our local SEO service tracks who reviews a business: the number of reviewers with a history in its category and area, the length of their reviews and their age.

The patent describes what Google's system can do. It does not confirm how Google weighs reviewers in local rankings today.