You Ask, I Answer: Understanding BERT’s Role in SEO?

In today's episode, Tara asks, with the confirmation of Bert's role in search ranking as Google Deep Rank, uh could you explain more what this means for SEO? So it's a good question. This is comes from uh revelation on search engine roundtable that uh the BERT model is being used as part of what Google's called deep rank, uh they're part of their ranking algorithm. And to understand this, uh, we have to understand what the BERT model does and what it's good at. Uh BERT's a piece of software, right?

It's no different in some ways than say, oh, Microsoft Word. And so it's less about it, the technology being used, and more about how it's being used. Uh you know, you could write uh a book with Microsoft Word, or you could write a shopping list. Uh there's so many different applications of it. Same is true for the AI model.

So in order to understand this from an SEO perspective, we need to understand a few things. What it's good at, uh what those capabilities have offer for you know search and ranking, and then uh what we should be doing about it. So Bert's great at really about what four or five big things. One classification of text, being able to say, you know, this text is good, this text is bad. Um it's good at named entity recognition, being able to say, like, okay, this is a you know a person, this is a place, this is a thing, understanding context.

Um it is good at question answering, being able to uh give a reasonable answer based on data it's extracted from text. Um it's good at extracting pieces of text uh out from a larger passage and understanding the importance of any given sentence or paragraph. And all that comes from this its ability to look at a a block of text as a whole, not just look at, say, the next word in the sentence. That's what makes it such a different model from uh other older natural language processing models. So, based on those capabilities, what what's the impact on search?

Well, first, you know, classification of text, is this good or bad is pretty straightforward. As Google gets uh more and more training data uh from its search engine search quality rating guidelines, as uh Google gets more user data from people how people use Google, um, BERT gets better and better at classifying text. This is a high-ranking page, this is a medium ranking page, this is a low ranking page, and so on and so forth. When you think about concepts like uh expertise, authority, and trust, which is a core part of uh the search quality rating guidelines, those are things that are not uh explicit in text, uh, but you human users rate pages, sample pages, uh, and then a deep learning model like BERT goes through and says, okay, all the high-ranking pages, what do they all have in common? How do they use prepositions?

How do they use named entities? Uh do named entities show up uh and what's their relationship to the text? And all this happens in the software that's not influenced by humans that we know of. Um I should caution this is speculation because nobody that I know of has an inside line to Google to ask them exactly how they're training the model, and if they did, they wouldn't tell us. Um but this is based on the understanding of the model.

So the search quality rating guidelines essentially is providing a training data set for Google to say, okay, what do pages that are rated favorably have in common. What are pages that don't? Um, and relying on things like those named entity recognition capabilities and snippets of text uh to understand the context of a page. All of that is fed into BERT, and then live pages on the web are fed into it, and it classifies them, right? It says that this is what this it should be ranked as.

Now again, BERT is a really big computationally intensive model, and we know it is not the core part of Google search for algorithm. It is part of it, just like core web vitals are part of it, just like you know, the the knowledge graph is part of it, and we don't know the percentages that's but we do know that it's in there at least somewhat. So, with all this technical information, what are we supposed to do about it? Well, a few things. One, everyone should be reading the search quality rating rating guidelines and understanding the training data set that's being fed to BERT, because that's critical for understanding how Google wants its engine to be trained.

These human raters are going in and rating pages and saying, you know, this is this looks good, this looks bad, you know, thumbs up, thumbs down, high, medium, low in in their rating guidelines that gives us the information we need to know for understanding what the the models likely to see on a page. Now, is it for sure? No. Part of deep learning is that deep learning can engineer features that are not there or not explicitly declared. That's one of the reasons why it's so valuable.

So we need to understand what the quality rating guidelines are and then use them as guidance to help improve our our content. Um we are using Google, we should be paying close attention to things like, for example, uh, when it outlines a snippet in their new uh passage uh results. Why? What made that snippet so important, that piece of text so important on that page? Can we get a sense of what makes that credible enough to display it as part of the search result?

You've probably seen on Google too. When you click through on a result, um, at least in Chrome, it will highlight the the passage of text that is most relevant to your query. Uh again, this is something that based on Bert's ability to uh extract text and understand uh, you know, individual passages and sentences within it to be able to say, yes, this is important. So when we think about how we write our content, we have a create our content, be thinking at about is there something that from a language perspective that you would say this is like the most important point. How would you write um your content to be determined this is the important passage that belongs here that if uh Google highlighted it, it would answer the user's question.

And then uh third, when it comes to interpreting how these models work and how our content works, we need to be thinking about the the features of the text and other things we publish. Um things around named entity recognition. If you're writing, for example, a post about uh I don't know, um coffee, right? What are the major expected entities that you would expect to see named entities in a in a piece about coffee, right? If you're writing about the history of coffee, you know, Lloyd's of London should be in there because it's an integral part of the history of coffee.

Um and so that demands domain expertise. That demands subject matter expertise. And that again is is why the BERT algorithm is so powerful, because you can look at the stuff that would be considered credible and see how many named entities are in there, how many uh certain types of texts and prepositions and all these things are used. Our goal as marketers is to create content that looks and reads and sounds like the stuff that would be rated the highest quality in the search quality rating guidelines. So original, in-depth, well researched, lots of citations, kind of like you know, writing a really good paper, writing a really good article.

The good news is that if you actually are good at what you do as a as a brand or as a company, and you have subject matter expertise and you can put that subject matter expertise on display, the algorithm will naturally tend to favor you as long as you check off a few of the tactical things, you know, like the the use of proper named entities and things, because it's going to be comparing your content versus what's in the quality rating guidelines, which is the training data set. So as long as you continue to pay attention to that and keep your content up to date with its recommendations and you write for the humans that will be uh reading your stuff, you will do well. You will do reasonably well with Google's algorithm. On the other hand, if you create low quality stuff that's thin, that uh doesn't have any citation, that doesn't look like authoritative credible content, you won't do well. It's just the nature of the way that the model is going to examine your text.

And there's other things that go into that too. Um we can save that for another time about you know, sort of the technical page layouts and things, but at the very least, um, you want high quality content as governed by the what's in the training data set, so that models like BERT, when they look at your page and they look at your content, will be able to compare it side by side with the things they know are valuable and go, yep, this is valuable too. Right? So i it's a lot of uh mimicry in some ways. It's a good question.

We could spend a long time on this. Uh but that's a good start. So if you got follow-up questions, leave them in the comments box below. Subscribe to the YouTube channel and the newsletter. I'll talk to you soon.

Take care. Want help solving your company's data analytics and digital marketing problems? Visit TrustInsights.ai today and let us know how we can help you.