English Español Deutsch Français Italiano Português (Brasil) Русский 中文 日本語

#SEOisAEO: How Does RankBrain Affect Query Results Today



Jason: Hi and welcome. We are going to talk about how does RankBrain affect query results today with Omi Sido, David Harry, and Doc Sheldon. Thanks for coming along guys.RankBrain is three years old this month. A lot of good stuff was written about it in 2015, then quite a lot of inaccurate stuff in 2016 and 2017. So we're going to try and put all of that to bed.

Definition of Rankbrain

"RankBrain is a continuation of the advanced query parsing and interpretation Google's been doing since Hummingbird; getting better at identifying intent and serving up content that satisfies that intent. When you intelligently target topics, Google will automatically help you rank for phrases searchers use to access that information." (Rand Fishkin)

Bill Slawski grabbed this from the original patent.


That was 3 years ago… How has RankBrain evolved since then?

David: Originally they told us it was for dealing with unknown queries. An event happens today, it could be a hockey score, it could be a world event or whatever and that is going to get a query that didn't exist yesterday. Google says that's 15% of all queries.RankBrain was there to try and figure out these queries by looking at word context and saying, "Well, we've never seen this query before, but these words are semantically related, so we think it means this." And thus garner an understanding of the intent and be able to bring good results back. We don’t know if they've applied it to larger index beyond that, nobody at Google said. You gotta assume that in 2015 it was doing baby steps and from there it's probably moved into a larger index, but we have no confirmation.

Jason: That was a brilliant answer.

Doc: The biggest misconception about RankBrain that I see repeated over and over is that it's a ranking algorithm. That's because someone at Google actually used those words early on, and that one statement is what's caused so much confusion is. Important to remember that it is only indirectly involved with ranking. It never looks at the content on your page. It looks at the query, clarifies it and then passes the query on to other algorithms. David and I have spent a lot of time over the last couple of years looking at RankBrain. What is really interesting is that RankBrain is playing with ML technology that we're going to see throughout many, many Google functions. If we're not already.

Jason: It's getting clearer with every comment. Now Omi can you make a little bit clearer still.

Omi: Understand the fact that RankBrain is only working on the front end, reinterpreting the query. Then, as Doc said, the fact that, because it's reinterpreting the query and getting a handle on the real intent of the query, it is can indirectly affect rankings.

Jason: Okay, super. So, we have a clearer vision of RankBrain. It rewrites the query to better express the intent, and that rewriting can indirectly affect the ranking. It was initially limited to just never-seen-before queries (about 15%) and has almost certainly expanded to affect a much larger percentage of queries.

David: But it has no actual scoring mechanism that we know of.

Omi: Remember, Google is a business, the searchers are their customers. Like every business, they want to satisfy their customers, make some money, and bullet-proof their commercial future. Whether they call it RankBrain OmiBrain, ElephantBrain or whatever, doesn't really matter to me. They are using ML to improve and future-proof their business.

Jason: Yeah! And we're asking this commercial business to kindly send us their customers to become our customers.

Back to Rankbrain. It reinterprets the query and that changes the relationship between the question and the answer, which in turn must change the results in some way. But how does it reinterpret the question? Word Vectors?

What is RankBrain's relationship with word vectors? Just good friends, lover, or married?

Omi: I really love Word2Vec. And the idea behind Word2Vec is very simple - The meaning of a word can be guessed by the company it keeps, like people! "Jason, tell me who your friends are and I'll tell you who you are." A specific example: you look at my pictures online, you see a lot of people like Jerry White, Nick Wilsdon… so you can guess that I'm an SEO. Going back to the mechanics of the Word2Vec, it literally takes the text as an input and produces vectors in multidimensional space as an output. Words that are similar in context will have vector coordinates that are close to each other. Word2Vec has two parts: one part that tries to predict neighboring words from a given word (skip-gram), and then another part that tries to predict a word from neighboring words. If you look at Tensor Flow, it looks like Google prefers the skip-gram method. Beyond Google, Word2Vec is used by Spotify to suggest music, and Airbnb to suggest listings. You can use the Word2Vec algorithm to extract words or sentiment from text. If other companies are heavily into using the same idea to make suggestions, why wouldn't do the same? In the beginning, RankBrain was only responsible for queries that they've never seen before… it must be more than that now.

David: Right, and that's what I keep thinking. It's logic that they started here, but they've got to have taken it and put it in the wild elsewhere. They're just not talking about it publicly. And if you look at even further than that, the Google Brain project was the grandfather of it. They're using GoogleBrain in robotics, in healthcare, translations and other things like that. There's gotta be pieces of that running in the wild all over the place.

Jason: So in fact, RankBrain's relationship with word vectors is: it was married and had lots of children, wonderful. Now on to Doc. I'm a big fan of the Knowledge Graph, as you might have noticed.

What is RankBrain's relationship with the Knowledge Graph? Just good friends, lovers, or married?

Jason: I'd love them to be married, but I think it's not the case.

Doc: Yeah, I'm inclined to agree. As Omi is saying, we're going to see this technology Word2Vec in a lot of areas and it's going to be widespread. We're going to see it in our homes if we're building a smart home, we're going to see it everywhere eventually. But I think the Knowledge Graph is probably a bit more subtle in its use of vectors. They started that early enough but I don't think they're really moving ahead at a breakneck pace with the Knowledge Graph. Where I think that we're going to see some relationship to it is as they start to refine further their Word2Vec usage - it's a fantastic technology, especially so at the scale they operate. So, yes, at some point Google are gonna start using Word2Vec to looking at content as well. So we will see the ability to interpret documents as well as queries, and that will help put together a Knowledge Graph that is a lot more meaningful, a lot more comprehensive. So they're not married at the hip, but I think it is going to happen. Everything we're going to get out of Google is going to have some Word2Vec or perhaps paragraph to vect, document to vect. We're going to see a lot of vectorization in its analysis because it's really the only way that they presently have that is capable of handling the scale at which they operate.

Jason: Brilliant. Okay. So they're just good friends, but they're moving towards some type of closer relationship.

David: Yeah, the Knowledge Graph itself is, to a large extent, a corpus of entities. Where Word2Vec is looking at relationships it will eventually use entities. The Word2Vec patent states that they take the entity “New York” and treat it as a single vector even though it's two words. In the context of using entities and multi-word entities as a single vector, you get can see a point where the Knowledge Graph melds into what's being done with Word2Vec.

Omi: Doc mentioned the content. Here’s an example - it is just a small needle in a big stack, but with one of my neighbors is a younger data geek and he uses Word2Vec to extract sentiment from some of my competitors. I use that sentiment, plus the keywords and upgrade it as better content and that translates pretty well. And I'm only a small user. I don't even have a Ph.D. in data science and all this stuff. Imagine what the people who work for Google are doing.

Jason: That comes back to what we talked about last week, which was using machine learning and those technologies against the machine itself.

Omi: I strongly recommend people actually explore Word2Vec. It's so easy, Jupiter is free, and you've got the data out there on the internet, you can play with it, see if it works for you.

Jason: The other day I made up a RankBrain recipe so it's going to be called...

Jason's RankBrain recipe


That is what I'm telling my clients to do. Obviously, it's not directly RankBrain, but I thought it was nice to say it. Omi, tell me does my RankBrain recipe flip, flop or fly?

Omi: The way you asked me, I have to say it's brilliant. Although I generally agree, those are good pieces of advice… although I'm not sure I understand number seven to be honest with you.

Jason: Oh fluffy prose it's all the poetic stuff that isn’t clear.

Omi: Ah. OK. Problem is, that this recipe is limited to content, article creation. In my mind, RankBrain is much bigger than that. My first point in any recipe is “know your audience”. When you know your audience you'll be able to see what questions they're asking online. And also, very important, related topics they are talking about. Next, you should know your competitors' audiences. Very often I learn more from my competitors' audience than my own audience. Then, it is very important to know what's bad about your product, and not just focus on what is good. That is in point 1 of your recipe but isn’t clear. But to really be good at RankBrain, to really win, you should look at the SERPs - see who is ranking for the keyword or set of keywords that you wanna rank for. Then figure out why they're ranking in position one, two and three. On a wider scale, to really win RankBrain game, you should be an authority in every topic your audience is talking about.

Jason: Okay, super. My recipe is good advice, but it doesn't apply to RankBrain. I rather skimped on the most important thing which is understanding your audience.

Omi: Everything starts there. RankBrain is all about relevance.

Jason: Brilliant, that's perfect. Thank you very much, so it kind of flopped, I saw this the other day on an article somewhere. "I think that train has left the station- we use machine learning in so many places, it doesn't make sense to try and single out RankBrain and to guess individual factors involved. Ranking is complex. Sorry for not having a simple answer, the question is irrelevant."Unfortunately for him, David then answered that comment with: "To be honest, John Mu is kind of right.... tune in next week (I.e this week) and I'll be happy to explain things a bit more." So I'm going to take you up on your word there and ask you to explain more.

It doesn't make sense to try and single out RankBrain and to guess individual factors involved. Ranking is complex

David: The vital point is "... we're using machine learning all over the place”. To get that larger context, look at the Google Brain project, which spawned the Google AI team, who developed TensorFlow and loads of other things. And the work started back in 2007, although the official Google Brain team started in 2011. They pushed out these Word2Vec patents from Greg Corrado and Jeffery Dean. Both of these guys’ names are on the patent and both of these guys work for the Google Brain team. And the Google Brain project is there to develop AI algorithms, then encourages all the other Googlers to take these algorithms and go mess with them - Google Local, crawling, link analysis… or whatever. So yeah, John's completely right in that ML is probably pervasive across many other areas beyond RankBrain.I think people are getting lost is that they focus on RankBrain and lose sight of the larger context of the Google Brain team and what they're doing with this stuff.

Jason: In short, they announced RankBrain, we all got terribly obsessed by it, and over the last three years they've rolled ML out to other areas of the search algo, but they just haven't named it.

David: I know what you mean - why wouldn’t there be a LocalBrain or the CrawlBrain, LinkBrain, or whatever… Google brain is literally saying to ALL the teams "Hey take these algorithms, go play with them, come back to us, and tell us if you found anything neat." The problem with the SEO world is we hear the word Rank and everyone starts running around losing their minds. The Google Brain project has been around since 2011; anyone ever mentioned it before 2015? Nope. It wasn't until Greg Corrado, unfortunately, put the word "rank" on it that we all went, "What? Hmm? Somebody said rank! What's going on?" That's what happened right?

Jason: I really like that idea.That it's something that became this big thing because it happened to have “rank” in the name… Worse, that is a bit of a misnomer. Plus, since then, the “Brain” (Machine Learning) has been expanded to loads of other areas and we should stop getting obsessed by RankBrain and just treat it as another part of the algorithm… which, as Omi said, at its most basic is about understanding your audience and being pertinent.

David: Yeah, right. It started with query rewriting. But then Hummingbird was a lot about query classification as well. That was in  2013 and everyone talked about that for a couple of years. But no one really talks about it now. That's always been the way with Google, especially when Matt Cutts use to be around. He'd say something this week, everyone loses their mind. Then he'd say something next week and everyone would forget about what he said last week. Google just needs to say something else, then everyone forgets about what they said before.

Jason: Oooh! I'll change the name of my recipe to "The Hummingbird Recipe" since it is probably more applicable to Hummingbird. Brilliant stuff. That was really clear, thank you.

Create great content. I don't like when people say that because it doesn't really mean anything

Omi: I hate when people say "great content" without giving you the context. I can write great content, but if it's not relevant to my audience, so what. Who cares?

Jason: Yeah, you've put your finger on it. That phrase annoys me too because it doesn't have any context. People imply that we create great content, Google will figure it all out with word vectors, the Knowledge Graph, and a bit of Machine Learning magic. Users will find you, love your stuff, buy your stuff and then we'll all live happily ever after.

Doc: As Omi is saying, you might have the best piece of content one could hope to find, but it is useless if it does not satisfy one’s needs. Great content is a very smoky term, it really doesn't define well. Provide value. That is the key. Your content has to be attractive, it has to draw people in, it has to hold their interest, it has to convince them it's offering value, whatever your conversion may be. Importantly, you have to recognize the fact that there's a lot of different users out there. We should have been providing good content for users versus search engine from day one. RankBrain doesn't have zip to do with this.  But we're kinda stubborn in this industry and we don't always let things sink in really quickly.

"What RankBrain was in 2016 is not where it will end up. It will move into other areas of search"

Jason: David said that in 2016, I think we all agree now, it has moved into other areas of search. RankBrain was the first machine learning that Google admitted to… and apparently also the last! We looked at where it started, where it is now… I'm interested to know where you think it's heading. David?

David: It doesn’t matter if they use the term RankBrain or another name we are talking about AI and machine learning. And that is going to be everywhere - in everything they're doing. RankBrain is just one instance that everyone has become hooked on. But the implantation is crawling through all sorts of query spaces and query types. Even just with query rewriting, there are potential needs everywhere - It could be in translations, understanding content… and those are just simple examples! Before with Hummingbird where auto, car, automobile and other synonyms of a word can have very similar meanings. Easy. You can add context if you're RankBrain and say, "Well, what if I mix this up?" So they mix it up with ML and then they go back and test these results to see if they have better user satisfaction. What were the click-through rates? Did it satisfy the user intent? That is big. But the beauty of AI / machine learning is it can train itself, which takes a big chunk of the leg work out. So I can't see why they aren't going to use that elsewhere.

Jason: Yeah! You also talked about translating. I do SEO in France and in French. From the moment Google uses Word2Vec to translate, which is what Cindy Krum was suggesting, being English speaking on the French market makes it much easier to do SEO in France and French. Is that a reasonable comment, Doc, you think?

Doc: Oh, absolutely. In fact, there's a lot of technologies in that sphere. Skype was talking about one three years ago, where you could speak into it in English and it would output in German, or Russian, or French on the fly. Just think about that universal translator from Star Trek. Using word vectors to interpret language seems inevitable. The number of new documents that spawn daily was already astronomical, and it's increasing exponentially on a daily basis. So in order to be able to hope to look at the majority of those documents and determine whether they're worth even considering placing in a SERP for anything, they're gonna have to do something faster and more productive and the best way to do that is to teach the computer how to understand the document.

Last question. How does RankBrain, present and future, affect our search marketing strategy?

Jason: I think having understood what RankBrain is now, I'll open that up a bit more and say how does RankBrain relate to machine learning and how does its implementation affect your marketing efforts?

Omi: RankBrain literally opened my eyes. The moment it came out it literally opened my mind. Before RankBrain I never ever thought about machine learning - it was something alien to me. RankBrain opened my eyes and I think it made us all a bit more human. I know it's a machine learning algorithm, but we stopped and thought, "Oh my god, we have to start thinking about our customers, we have to stop thinking about the search engines.”Search engines are serving our customers. As I said in the beginning, it's a business for everybody. It's business for search engines. It's business for our customer. It's business for us as SEOs. The future is customer-centric. And the idea of RankBrain is actually customer-centric. Once I understood that I started going back to the SERPs, looking at my competitors, sentiment analysis, and lots of other things. All those things came from the idea of RankBrain, believe it or not.

Jason: I agree with you, the irony that all this machine learning has made everything more human again. And that is brilliant.

David: Yeah - that’s all about future proofing. Know history and you keep an eye on technology. We do that and nothing ever surprises us. Do that and you're already positioning your clients, your work and everything you're doing to be Google-proof! Meaning that every time Barry Schwartz writes about the latest algorithm, change and says, "Did anyone see anything?" I say, "Nothing but good things." Because I knew it was coming. I have a time machine. Next, look at Google voice. It is slowly but surely going to become the norm. My kids are teens, and they don't want to type in anything, they just wanna talk to their phone. You know, "Google go find me this. Go do that." Yeah, Google Assistant can potentially do it, but that needs more processing power. To solve that problem, they're building these corpus's and these vectors and these neural networks and once it becomes everywhere, you want that phone to learn from you. When you're talking to the Google Assistant it's gonna get to the point of anticipating what you want - where it thinks for you When it comes to SEO you need to be thinking ahead to how to position yourself and your clients. You wanna be ahead of these things - that's what a lot of this machine learning, AI and vectors are going towards ultimately. Staying one step ahead so when it happens, you are ready for it.

Jason: John Mueller was talking about people missing the RankBrain train. You're saying get to the station before the train even gets there.

David: Right. Because you're gonna miss it if you're not there. I copyrighted that whole train thing last week.

Doc, how has Machine Learning affected your marketing efforts?

Doc: Something we should have been doing for a long time was focusing our content on context. It's easy enough for a reader when content shifts directions - as a human, that’s easy enough to catch and follow. That context switch is not as easy for a machine. Stick on topic - that makes it much easier for the search engine to interpret. Avoid the slang and idiosyncrasies comments. Humor is still difficult. For example, If you're talking about a recipe for making a Thanksgiving dinner, and you suddenly talk about softball games, just because your family happens to always play softball on Thanksgiving, that's not helping the search engine understand. For a human, it may make the article more readable. It's a fine balance. There are things you can do that makes it a more entertaining or informative piece for human users. But you have to recognize the fact that you may also be making it more confusing for the search engine.

Jason: Yeah, great. If we go back to my recipe. You could put that anecdotal story in an aside. Which means that Google would then skip around it and not treat it as a central part of the content?

Doc: Yeah, you can.

Jason: Great! I have a much clearer idea of what RankBrain is and what RankBrain isn't. I think I can certainly put this to bed.

David: Yeah, RankBrain has already left the station so now we're looking ahead to the next stop on the train. And remember, even the Word2Vec patent from Google with Greg Corrado and Jeffry Dean was originally filed in 2013 but granted in 2015 - that is several years out. So we should be looking into the future; Patents are a big clue - before they get confirmation, they've obviously played with it, tested it, messed around with it. So here we are six years later, just catching up.

Jason: That says it all. Thank you to you three for coming along and sharing. That was absolutely brilliant. I think, for me, RankBrain is much clearer.  I'm very pleased about that and we put RankBrain to bed. As we saw earlier on, John Mu said, "We use machine learning in so many places." This conversation about ML will continue next week, looking beyond RankBrain Please do join us next Tuesday for Episode 12 of #SEOisAEO: What Does Machine Learning Do in AE Today and What Might it Do Tomorrow?" We've got Patrick Stox, Jan William Bobbink, and Dawn Anderson.


Check out other webinars from this series