Main menu


Google is using a machine learning technology called Rankbrain to help deliver search results. Here's what we know about Rankbrain so far.

Just yesterday, news broke that Google uses a machine-learning artificial intelligence system called "Rankbrain" to help categorize search results. Wondering how he works and fits into Google's entire system? Here's what we know about Rankbrain.

The following concealed information has three sources. First, the Bloomberg story revealed about Rankbrain yesterday. Second, additional information provided by Google directly to Search Engine Land. Third, our knowledge and best assumptions about what Google doesn't answer. We'll make it clear where these sources are used, separate from general background information when necessary.

What is Rankbrain?

Rankbrain is the name of Google's machine-learning artificial intelligence system used to help process search results, as reported by Bloomberg and confirmed to us by Google.

What is machine learning?

Machine Semester is when a machine can learn by itself, without a human to teach it or follow complex programs.

What is artificial intelligence?

Artificial intelligence means that a computer can also become as smart as a human being, at least in the sense that it can acquire knowledge and make new connections from being taught and building on what it already knows.

Of course, real AI exists only in science fiction. In fact, AI generally refers to computer systems designed to learn and make connections.

How is AI different from computer learning? From the word Rankbrain, it seems to be synonymous to us. You may have heard them used interchangeably, or you may have heard machine learning used to describe artificial intelligence that can be used for labor.

So Rankbrain is Google's new way of ranking search results?

no. Rankbrain is just one part of Google's overall search algorithm, a computer program used to rank billions of known pages and find the results that best match a query.

What is the name of this Google search algorithm?

He's called "The Hummingbird," as we've covered in the past. For the past few years, the overall algorithm has not had an official name. But in mid-2013, the algorithm was completely overhauled and he was given the name – Hummingbird.

So Rankbrain is part of Google's Hummingbird Algorithm?

This is our understanding. Hummingbird is the whole search algorithm, just like a car has a whole engine. The engine is composed of many parts, such as: oil filter, fuel pump, cooling system and so on. Likewise, Hummingbird contains many parts, and Rankbrain is the newest part.

We learned that Rankbrain is part of the whole Hummingbird algorithm, because the Bloomberg article already stated that Rankbrain is not responsible for handling all searches, because only the whole algorithm can handle it.

Hummingbird also includes other algorithm names that are familiar in the SEO world, such as: Panda, Penguin, Payday to fight spam, Pigeon to improve local search, Top Heavy to downgrade pages with too many ads, and search engine friendly to to improve mobile-friendly pages, and Pirate to combat copyright infringement.

I think Google's algorithm should be called "PageRank"

PageRank is part of the whole Hummingbird algorithm, and it has a special way of giving weight to links from other pages.

PageRank is special because in 1998, when the search engine first started, PageRank was the first time Google gave a name to some of its algorithms.

What are these "signals" that Google uses to rank?

Signals are used by Google to help determine page rankings. For example, he reads the text in the web page, so the text is a signal. If some words are in bold, that might be another sign. These calculations are usually part of PageRank and are used to give pages a PageRank score, which is then used as a signal. If a page is marked as mobile friendly, that's another sign of being registered.

All signals are processed and analyzed by various parts of the Hummingbird algorithm to determine which pages should ultimately be displayed in various searches.

How many signals are there?

Google talks about over 200 primary ranking signals, and possibly over 10,000 variables or secondary signals. He usually says there are "hundreds" of factors, as in yesterday's Bloomberg article.

If you want a more intuitive guide to ranking signals, check out our " Periodic Table of SEO Success Factors " article which

How many signals are there?

we think is a really good guide to help your pages get listed on search engines like Google Get a ranking.

And is Rankbrain the third most important signal?

That's right. Surprisingly, the new system has become what Google says is the third most important factor in ranking pages. According to the Bloomberg article:

Corrado said: Rankbrain is one of the "hundreds" of ranking signals in the algorithm that determines how it appears and ranks on Google's search results pages. He said: "It has already been launched in the past few months, and Rankbrain has become a third ranking factor and contributes to search.

What are the first and second most important signals?

Google doesn't tell us what the first and second most important signals are. We have already asked. And asked twice...

Google doesn't tell us the first two signals, which is really annoying and confusing. Google wants to use PR to break through the development of machine learning.

But if you really want to assess the breakthrough and the impact of Rankbrain after that, it helps to understand other important factors that Google is using now. That's why Google should explain.

By the way, my personal guess is that links will still be the most important signal, and Google will aggregate those links.

 This is also a very old system. 

As for the second most important signal, I'm guessing it will be "text", which includes the text on the page and how Google understands the text people type into the search box in the Rankbrain analysis.

What exactly can Rankbrain do?

From Google's letter, I speculate that Rankbrain is primarily used to parse people's submissions for searches that may not contain precise text.

Can't find a page without a precise query Google?

No, it's been a long time since Google could find pages without exact words. For example, many, many years ago, if you typed in something like "a shoe," Google might not be able to find a page about "a pair of shoes," because technically those were two different things. word. But "stemming" can make Google a little smarter, in order to understand that a pair of shoes is a deformation of a shoe, just as "running" is a deformation of "running".

Google also has the wisdom of synonyms, so if you search for "sneakers," Google may know you're looking for "jogging shoes." To understand the tech company "apple" vs the fruit "apple", even get conceptual wisdom.

What is a knowledge graph?

The Knowledge Graph, developed in 2012, is Google's way of making connections between words smarter. More importantly, he learned to search for "things, not strings," as Google once described it.

A string means a search for a string of letters, like a page with the spelling "Obama". Conversely, things mean that Google understands that when someone searches for "Obama," they may be referring to real-life people connected to other people, places, and things by US President Barack Obama.

A knowledge graph is a database of things in the world and the connections between those things. That's why when you search for "when was Obama's wife born" instead of his name, you get answers like Michelle Obama:

What is a knowledge graph?

How does Rankbrain help optimize searches?

The methods Google uses to redefine search generally go back to the people who work somewhere, as are the stemming lists that have been created or the synonym lists or databases that make connections between things. Of course, some are automated. But most of them depend on manpower.

The problem is that Google processes 3 billion searches a day . In 2007, Google said that 20 to 25 percent of queries were seen before. In 2013, that number dropped to 15 percent , which again was used as a Bloomberg article and Google confirmed it again. But 15 percent of those 3 billion that haven't been searched yet is still a huge number -- 450 million a day.

Those can also be more complex, multi-group word queries, and can also do "long tail word" queries. Rankbrain is designed to help those queries and translations more efficiently, behind the method is to find the best pages for searchers.
As Google puts it, he can understand compound queries that seem unrelated on the surface, and their similarity to each other. And this study can learn more about compound queries, and whether they are related to other topics. Most importantly, Google tells us that it can connect query phrases with search results that searchers will love.

Google doesn't provide examples of query phrases or details on how Rankbrain guesses the best page. The latter may be because if he can translate ambiguous queries into explicit ones, he can bring better results.

How about an example?

While Google doesn't give examples of query phrases, and Bloomberg articles have a single example for searches, Rankbrain may help. Here's:

Who is the highest consumer in the food chain?
To a layman like me, "consumer" sounds like someone who buys something. However, he is also a scientific term for a species that consumes food. There are multiple levels of consumers in the food chain. What about the top consumer? He calls it "the carnivore."

Enter that word into Google, and Google provides a good answer, although the query itself is still very strange:

How about an example?

Now think about the similarity of results for a query like "top of the food chain", as follows:

How about an example?

Imagine that Rankbrain takes that original Long and complex searches link to the shorter one, which is probably more common practice. He was able to understand that they were very similar.

 The result is that by getting answers from the more common queries to improve the less common ones, Google can make everything it knows work. 

I want to stress that I didn't know Google joined the two queries. I only know that Google gave the first example. This is a brief description of how Rankbrain may be used to link uncommon and common queries to improve search results.

Can Bing do this with RankNet?

Back in 2005, Microsoft started using their own machine learning system, called RankNet, which is part of today's Bing search engine. In fact, the lead researcher and creator of RankNet was only recently honored. But over the years, Microsoft has rarely talked about RankNet.

You can bet this will probably change. Interestingly, when I typed the same example on Bing as the Google Rankbrain above, Bing gave me good search results, one of which was the same as Google's.

Can Bing do this with RankNet

A search does not mean that Bing's RankNet is as good as Google's Rankbrain, and vice versa. Unfortunately, it is difficult to come up with a list to compare.

Are there more examples?

Google does give us new paradigms: "How many spoons are in the cup?" Google says that Rankbrain will provide different search results for Australia or the US, because each country is measured differently, despite the similar names.

I tried searching on and google Australia to test. I found not much difference.

 Even without Rankbrain, the results are often different simply because the "outdated" method of serving pages is to serve known Australian sites to those searchers using Australian Google.

Does Rankbrain really help?

Although the above two examples are not convincing evidence that Rankbrain is strong, I really believe that he may be having a huge impact, as Google claims. 

Google is fairly conservative with its ranking algorithm. He's always doing little tests. But when he has enough confidence, a big change will unfold.

To the point where it is generally believed to be the third most important signal, integrating Rankbrain is a huge change. I think he did it because he really helped Google.

When did Rankbrain start?

Google told us that Rankbrain will roll out gradually in early 2015, and it has been rolling out globally for several months now.

What search terms are affected?

Google told Bloomberg that a "majority" of queries go to Rankbrain. We asked Google for an exact number, but we still got a large portion of the answer.

Will Rankbrian keep learning?

  • Google tells us that all learning at Rankbrain is offline. 
  • He is given batches of historical search records and learns from them to make predictions.
  • Those predictions are tested, and if the predictions are verified to be correct, the latest version of Rankbrian will be used. And the cycle of offline learning and testing will continue to repeat.

Can Rankbrain do more than query optimization?

  • Usually how a query is optimized, through stemming, synonyms, or now Rankbrain is not considered a ranking factor or signal.

  • Signals are usually related to content, such as words on a page, links to a page, whether the page is on a secure server, etc. It may also be related to the user, such as the user's location or their search and browsing history.

  • So when Google talks about Rankbrain being the third most important signal, is it really being used as a ranking signal? Yes. Google reassured us that there is a component, Rankbrain, that directly affects page rankings in some way.

  • Exactly how? Is there some type of "Rankbrain score" that can be used to assess quality? Maybe, but it seems more like that, Rankbrain somehow helps Google sort pages better based on content. Rankbrain may be better at summarizing web pages than Google's current system.

  • Or not at all. Google doesn't say anything other than about the ranking component.

How can I learn more about Rankbrain?

Google tells us that anyone who wants to know more about the word "vector" - the word and phrase are mathematically linked - should check out our blog post about the system (which also goes into the article). (not named Ranlbrain) how to learn the concept of cities by scanning the news:

Can Bing do this with RankNet

There is a long search paper based on this . You can even use Google's word2vec tool to build your own machine learning program. In addition, Google has an entire section on his AI and machine learning papers, as does Microsoft .
Original source: click here


table of contents title