RankBrain is a new subsystem of algorithms that Google uses in the delivery of search results or SERPs. RankBrain uses machine learning and artificial intelligence to answer queries that have never been posed. The system embeds spoken language in the form of searches in mathematical patterns which can be processed by the search engine. RankBrain supplements the different algorithms for the categorization of these new searches. It also uses semantic search, which was implemented with the Hummingbird update, and associates unfamiliar words semantically to already known linguistic entities. The system is designed to learn this way in the future.
Google receives millions of search queries on a daily basis. Approximately 15% of all incoming Google searches are entirely new. This includes new concepts, phrases and questions. In order to evaluate these often very complex language patterns correctly and deliver relevant results, Google, according to its own statements, uses more than 200 different signals, a complex infrastructure and various updates that are designed to improve search results.
After years of research, Google implemented the RankBrain system which was rolled out globally in early 2015 after a short trial period. RankBrain was first tested and compared to real users, who correctly assigned 70% of new searches while RankBrain reached 80%. Because the system can learn almost autonomously, it quickly became an integral part of Google search. In addition to links and semantic signals as probably the most important ranking factors, RankBrain has already become one of the most important signals.
Five developers who were driving forward the integration of machine learning and artificial intelligence in the context of research were responsible for it. With the acquisition of DeepMind in 2014 and the appointment of Ray Kurzweil in 2012, the company laid the foundation for their own research in the field of artificial intelligence years ago. The recently integrated RankBrain system supplements the main product, Google search, a framework that is meant to support in particular, conversational search, slang terms, voice search and also set up Google in the area of AI for the future.
Just as with other updates or modifications to the core algorithm, Google does not provide a lot of explanations on how the exact mechanisms of RankBrain works. Only one press release including an interview with Greg Corrado, Senior Research Scientist. According to this, RankBrain uses artificial intelligence for embedding language into mathematical entities, which are called vectors. If Rank Brain receives an unknown word or search phrase in the form of a query, it guesses the meaning of these characters by collecting important similarities to words that already exist in the database; historical searches are fed into the system. If these associations are correct, they can be integrated into the live system. The delivery of the search results is then changed based on the data so that relevant websites and applications can be displayed to the user, in particular for searches never made before, with long-tail terms or which contain ambiguous words.
Rank Brain draws on the Hummingbird algorithm and semantic search to at least understand on a rudimentary level the meanings of certain statement. As part of Hummingbird, individual searches are treated as entities and not strings (things not strings), and annotated by the Knowledge Graph to understand, for example, the type of query (transactional, navigational, informational searches). But the system has not yet been able to handle complex queries with many terms in different contexts.
Therefore, RankBrain is meant to specifically handle queries that consist of combinations of several terms, long-tail keywords, questions, and conversational searches, for example. Search engines have always had problems with complex language in man-to-machine communication because meanings could not be modeled. RankBrain is apparently based on a conversational model embedded in a “sequence-to-sequence framework.”
Accordingly, the model uses a context of previous searches (sequence) to guess the next sequence. If the system is correct, the sequences which had been guessed, will then be also transferred into the context. By implementing machine learning, the system is meant to improve the processing of complex queries in the future. Currently, these searches are still being processed by engineers who teach the system how it can deal with unfamiliar terms. It can be assumed that in the long term it will do this task independently.
As of today, RankBrain is considered the third most important signal in the sorting of search results. Greg Corrado was surprised that the system worked better than expected. Although there was no date for the roll out of RankBrain, but the system most likely already supports the search query interpretation for months now. RankBrain should not be construed as an update or algorithm, but rather as a process that supports the search engine in understanding search queries.
RankBrain does not necessarily process all searches, as is the case with the main system. Only searches that have never been done before or complex queries with multiple words or phrases go through this system. Synonymy, ambiguity, meaning (intension) and significance (extension), are linguistic problems Google is trying to solve by teaching the system how to interpret terms and relate them to each other.
The bigger the database, which Google engineers continuously increase with searches that have already been made, the more accurate the predictions. From artificial intelligence or machine learning in the strictest sense, RankBrain is still currently some development steps away. Nevertheless, the system will make spoken language more of a topic for search engine optimization. This also refers to aspects such as content marketing, semantic markup, structured data, and general useful website content.