In a scoop by Bloomberg Business during an interview with Greg Corrado, Senior Research Scientist with Google, it was revealed that Google has been using an Artificial Intelligence known as “RankBrain” to improve search results for the last few months.
The system known as RankBrain has been helping to interpret some of the many millions of user searches carried out each day to improve the relevancy of some of the more complicated \ unusual queries.
What follows is a comprehensive guide to the new AI-based algorithm, including what it is, how influential it is in the search results, and whether you need to adapt your SEO strategy in light of the new information.
What is RankBrain?
It is the name Google has given to its new artificial intelligence system that uses machine learning to improve the search results for unusual \ difficult words or phrases. By machine learning, it is meant that the AI teaches itself rather than following human-created directions or programming.
More simplistically, if it sees a word or phrase it doesn’t know, it can guess as to what words or phrases have a similar meaning and provide search results accordingly. As a consequence, it has significantly improved Google’s ability to return accurate results for never-before-seen search queries.
According to Bloomberg:
For the past few months, a “very large fraction” of the millions of queries a second that people type into the company’s search engine have been interpreted by an artificial intelligence system, nicknamed RankBrain, said Greg Corrado, a senior research scientist with the company, outlining for the first time the emerging role of AI in search.
Is it Learning on the Fly?
No. According to Gary Illyes and then expanded on by the Bloomberg reporter who broke the story, it is not adapting on the fly. Instead, it is periodically retrained and then rolled out into the Core Algorithm.
@glenngabe it's not. The team was working on it for months and its effects are expectable, not assumable.— Gary 鯨理／경리 Illyes (@methode) October 26, 2015
@methode @glenngabe yeah. It's periodically re-trained, but it's not learning on-the-fly.— Jack Clark (@jackclarkSF) October 26, 2015
The Sempost asked Google about the frequency of the updates, and they stated that it would be updated as needed (i.e., no definitive \ set time scales):
We’ll keep experimenting with and testing new models, and we’ll make updates as we come up with models that do a better job than the existing one. That could be about refreshing the data or developing new neural net architectures.
In essence, it will be treated a little like how they roll out Penguin or Panda.
How does it Work?
According to Bloomberg, it uses AI to embed the written language into mathematical entities called “vectors” that can then be understood by computers. This allows the AI to make guesses at which words might have a similar meaning.
Jack Clark followed him up in a tweet:
@ollieglass @ejlbell it's converting words and phrases into vectors. Closely related to Hinton's work on thought vectors.— Jack Clark (@jackclarkSF) October 26, 2015
Jack Clark is referring to Professor Hinton, who was hired by Google a couple of years ago to help develop intelligent operating systems. According to a fascinating interview with Professor Hinton in The Guardian earlier this year when discussing machine learning, he said that:
Google is working on a new type of algorithm designed to encode thoughts as sequences of numbers - something he described as “thought vectors.” [..] Although the work is at an early stage, he said there is a plausible path from the current software to a more sophisticated version that would have something approaching humanlike capacity for reasoning and logic. “Basically, they’ll have common sense.”
The Sempost noted that Jack Clark posted additional comments on the Hacker News that it may be related to Word2vec and Sequence to Sequence learning and also linking to the paper Sequence to Sequence Learning with Neural Networks written by three Google employees. It makes a fascinating read, and although quite technical, you can see that they were surprised with the results of their tests:
We were surprised by the extent of the improvement obtained by reversing the words in the source sentences.
We were also surprised by the ability of the LSTM to translate very long sentences correctly. We were initially convinced that the LSTM would fail on long sentences due to its limited memory, and other researchers reported poor performance on long sentences with a model similar to ours. And yet, LSTMs trained on the reversed data set had little difficulty translating long sentences.
Google told Sempost:
It’s related to word2vec in that it uses ’embeddings’ - looking at phrases in high-dimensional space to learn how they’re related to one another.”
So while we can’t be 100 percent certain exactly how RankBrain works, the information does give some fascinating insight as to where Google is heading with machine learning and the search results. Of course, for more information about what Google is up to more generally regarding AI and Machine learning, their public research area on the topic is well worth a look.
How Influential is it in the Search Results?
According to Bloomberg, RankBrain is now the third most significant signal in the Google Algorithm, although we are not sure there is sufficient information for this claim at present. This is significant, considering there are hundreds of different ranking factors.
According to the Bloomberg article, it helps **Google deal with the 15% of queries they receive daily **that have never been seen before. One example of such an ambiguous question is, “What’s the title of the consumer at the highest level of the food chain?”. You can see the top results indicated the correct answer (“Predators”):
Because of RankBrain, Google was able to connect the long query to a much shorter and more relevant one, increasing the accuracy and relevance of the result. We suspect as it continues to develop, it will become increasingly more adept at doing so.
From the excitement and enthusiasm shown by Google for this, it is clear that it has a significant impact. Greg Corrado described it as “as having gone better than we would have expected. "
Experiments show that RankBrain beat Engineers who were tested against it
Some Google Engineers were shown some web pages and then asked which they thought would rank on top. While the engineers guessed 70% of the time correctly, RankBrain picked 80% of the time correctly. Corrado said that turning off the new Ranking Signal " would be as damaging to users as forgetting to serve half the pages on Wikipedia.”
Do You Need to Revise Your SEO Strategy?
This very question was put to John Mueller, webmaster trends analyst at Google, who clearly stated that nothing particular was required to optimize your website for the new AI algorithm.
@StelinSEO You don't have to do anything specific for RankBrain.— johnmu is not a bard yet 🖇️🖇️ (@JohnMu) October 28, 2015
However, if we were to consider how it all works, it is clear that having high-quality content is going to help. Of course, this has always been the case, and John Mueller’s response makes sense; there is nothing particular you need to do.