• by Laurie Sullivan  @lauriesullivan, March 10, 2016

Google released details of its artificial intelligence (AI) technology, RankBrain, in October -- but the practical implications of the algorithm on search were not realized until this year, according to a study. The original intention was to decipher and put into context about 15% of the queries Google had not seen in the past.

The study by Stone Temple Consulting set out to determine whether the machine-learning algorithm used to process search queries really understands the meanings.

As recent as June and July 2015, Google's search engine had trouble processing many queries, and to prove it Stone Temple Consulting CEO Eric Enge found more than 100 examples in about one million search queries.

The study analyzes how much Google’s search results have improved after implementing RankBrain as the third most important ranking factor. By using machine learning, RankBrain is able to better understand user queries. Since the initial findings, Google improved the answer to queries about 54% of the time. Added maps about 2.2% of the time, and improved answer box results about 43.8% of the time.

Some queries are complicated. In July 2015, the search query "Why are PDFs so weak?" -- where the user clearly asks a question about security capabilities of the file format -- returned results with the top article, a PDF file, titled "Why the Iraq resistance to the coalition invasion was so weak."

In January 2016 that same query returned "What security scheme is used by PRD password encryption" as the title of the No. 1 article on the page.

When asked for another example, Enge mentioned the use of the word "coach." Sometimes the word "coach" might refer to a coach on a sports team. Other times, the Coach brand that makes handbags.

"Stop words" also can confuse search engines. These words -- what is, who is, where is, without, not, and convert -- are typically ignored when processing the meaning because it's easier for the technology to understand the context. It becomes a problem, however, when the user's search query contains the title of TV show like "The Office," because stripping out the word "The" changes the user's intent.

Machine learning allows Google to make observations about strings of words and their relationship with each other.

The study shows that the real impact of RankBrain has been an increase in overall search quality, and the ability to use the core search algorithm for other projects.

Enge said as Google's technology provides a better understanding of language and word relationships, RankBrain will do a better job of matching user queries with Web pages, making marketers less dependent on having to add all the query keywords on the Web page.