Machine Learning for SEO

How is machine learning changing SEO? And how do you need to adapt your SEO workflows to take full advantage of machine learning in SEO?

Joining David Bain for episode 28 of Old Guard vs New Blood is Lazarina Stoy from Intrepid Digital, Susan Connelly from Milestone and Jessica Peck from Local SEO Guide.

Watch On-demand

Listen to the Podcast

Transcript

David Bain

Old Guard vs New Blood, Episode 28, Machine Learning for SEO. Hi, I’m David Bain, your temporary host while Dixon demonstrates that you can’t be in two places at the same time. Welcome to Episode 28 of Old Guard vs New Blood, the show brought to you by majestic.com, mapping the web to help you, the SEO, dominate your market. We’re available on Apple podcast, Spotify and all the biggest podcast platforms, so if you want to watch the next episode live, just sign up at majestic.com/webinars. Already got some people watching us live. Simon Cox is saying, hello folk. Hello, Simon, great to have you on board. If you’re watching us live as well, try and mutter some words in the comments and we’ll try and incorporate whatever you’re saying, even if it’s just an applause. Hi Danielle in the chat there, and we’re going to have a great discussion.

David Bain

Today we’re going to be discussing how Google is using machine learning to improve its search results. So that impacts SEO and what you can be doing with machine learning to maximize the output of your current SEO activities. Joining me today are three great panelists. So in traditional Dixon tradition, let’s find out who they are by seeing if they can remember their panelist numbers and saying, Panelist Number One, what’s your name and where do you come from?

Lazarina Stoy

My name is Lazarina Stoy. I’m an SEO and Data Science Manager at Intrepid Digital and I come from Sofia, Bulgaria.

David Bain

Superb stuff. Lazarina, thanks so much for joining us. Panelist Number Two-

Lazarina Stoy

Thanks for having me.

David Bain

Thank you.

Jessica Peck

I am- I’m Jess Peck-

David Bain

Go for it, Jess.

Jessica Peck

Yeah, I’m Jess Peck. I work doing machine learning operations at Local SEO Guide. I’m originally from the UK but I live in Boston now. Yep.

David Bain

Super. Well, welcome from Boston. Welcome from wherever in the world you happen to be watching this. And Panelist Number Three, what’s your name and where do you come from?

Susan Connelly

Hi, I’m Susan Connelly. I’m a Senior Product Technical Manager at Milestone, which is a technology company located in Silicon Valley. And I am located right now in Scottsdale, Arizona, and have actually come from the East Coast outside of Philadelphia, Pennsylvania.

David Bain

Well, there we go. We’ve got three wonderful panelists today talking about machine learning for SEO. So it’s probably good if we start off with focusing on Google and getting a feel for how Google is using machine learning at the moment to improve its search results. So let’s go back to Lazarina. Lazarina, how would you describe Google at the moment in terms of what they’re doing just now to use machine learning?

Lazarina Stoy

I’d say getting better. I would describe them as getting consistently better with how they’re using machine learning. They are doing a lot of stuff with video, recognizing key moments, using neural networks to match particular topics and subtopics to the query of the person, using natural language understanding where it is getting better and better. If I have to pinpoint something that I’m not as confident in, that would be search engine result pages in different languages. I would be very interested to know in the future in a few couple of years, how that improves and would that ever reach the extent that English search engine result pages have reached.

David Bain

So why is-

Lazarina Stoy

So yeah, getting better.

David Bain

Well, why is that, would you say? Well, why is Google not necessarily as effective at other languages apart from English? Is it because there’s not as much data there for it to learn from?

Lazarina Stoy

Yeah. So fundamental problem in machine learning, your model is only as good as your data. So I’m not as in the weeds with particular research that is being done. I know that there’s quite a lot of research being done in multiple languages. So there is development in recent years with regards to that, but historically there is a lot more data in English and a lot more data that is in a format that is usable for machine learning models in English. So I think that is the problem. And also the fact that English is a language that is spoken widely and at the end of the day, that is the focus of Google, at least from what I’m understanding, but I might be completely wrong. So please, other panelists, feel free to correct me if I’m wrong, but yeah, that’s my hypothesis.

David Bain

Well, Simon Cox saying, Google’s not that good at English either. It gives me American English results… Tut. Well, I think we’ve got a bit of a mix of American English, British English, and other English going on here today as well. I come under the other English category. Jess, what’s your interpretation of what Google, how Google SERP maybe has changed over the last few years because of machine learning and perhaps is currently changing, evolving at the moment because of machine learning?

Jessica Peck

So I think it comes down to, we know what Google wants to do, or at least what they say they want to do, which is serve awesome content. And I think the way they’re trying to use machine learning is to create SERPs that are useful and good without having to rely on the kind of methods that they’ve traditionally used to figure out if content is good or not. And I think when that comes to SERPs, the using more and more natural language processing to collate and bring together data to surface things like quick answers or comparison, or try and figure out which bits of different sites are the best or the best to display.

And I think that’s just, Google doesn’t like to be embarrassed, and sometimes the sites that they surface are embarrassing to them, they have bad information or are not good results or are garbage or just click bait, right. And so I think that Google is trying to steer SERPs in a way where they can kind of control the flow of information a bit better and don’t end up being embarrassed in one way or another. They don’t want to end up with New York Times headlines saying, “Google surfaced a terrible result” and here’s the result of that.

David Bain

They can blame the machines instead.

Jessica Peck

Exactly. It’s completely out of human hands. They can kind of be like, they also, SEOs manipulate searches, manipulate the search results, it’s what we do. And if they can just be like, make awesome content and we’ll surface it, because we have machine learning that does that, that takes a lot of the liability out of the hands of the SEO, Google. And doesn’t necessarily mean the results are better. It just means fewer humans have manipulated them or at least they can say that, right?

David Bain

And that’s a hot topic at the moment, certainly, to actually try to ensure that the content that is being read is, well, first of all, I mean not necessarily unbiased but different arguments or if there is an alternative argument or view in place in that article can also be easily discovered as well. Perhaps another topic by itself. Susan, is, does machine learning mean that Google is delivering a better SERP experience nowadays?

Susan Connelly

Well, I think for sure, it’s going to be delivering a better experience as we talk about the changes in the search landscape and how all of this machine learning being used by Google’s kind of being translating into different areas of search. What Google’s doing is kind of understanding a lot better, not only just what types of searches to surface by intent and really looking at if you’re searching for a red lipstick, for example, are you looking at it because you want to learn information about the content specifically and learn tips and tricks or are you looking for it because you want to purchase that. And I think that what Google’s doing is really looking at the searches before you get to this, to Google as well, as well as within the search rankings to say, okay, can we understand a little bit more about this user? Can we combine different types of data sets to understand a little bit more about them so that we could deliver them the right experience and the right search results for a particular time?

So I think that everything from images to video, understanding more about the contents on the pages, all of those things in context, not just what’s on the page, but the context of those pages is going to allow searches to become a lot more personalized into the future.

David Bain

Right. Okay. Well, you certainly mentioned many things there as well, starting with intent and talking about content and other areas as well. So I guess machine learning then gives Google the opportunity to probably incorporate more signals into their algorithms to make a more nuanced decision as to what content is right for each individual. I’d like to watch, anyone watching, ask anyone watching live, are you actively using machine learning in a way that you deal with your own SEO activities at the moment? If so, what are you using machine learning for just now? So share that in the chat, if you can. But let’s move the conversation onto what SEOs can be actively doing at the moment if they’re not actively using machine learning at the moment. Susan, you started off with intent. You also talked about content and other things there as well. How would you recommend that SEOs get started with machine learning?

Susan Connelly

Well, I think one of the biggest things is having a really good base in analytics. So you should be not only just looking at analytics, but understanding what metrics make changes. So if you’re making a change to a piece of content, what specifically have you changed and then kind of recording and storing all of that analytics data. So for the most part, I would say that the very first thing that you should be doing is learning as much as you can about analytics data, and then as much as you can just about the very basics of how do you take that data and make meaningful stories out of it. And some of that is very going into the Google Analytics data forums, kind of seeing what other people are doing as far as how to take in and adjust data, how to pivot data, store it and retrieval, which is all a lot just to get started with.

I myself, I’ve been taking some classes over at Stanford University. So I definitely recommend, they have some free classes that are online available via YouTube, and they can get you started in just the basic processes to start to gain an understanding of just how it’s being used. Because once you understand how it’s being used as a technology, then I think it starts to open your mind, oh, I can do this with SEO or I can change meta titles easily, or I could use this to change a piece of content. So it starts to open those minds when just having the initial just basic understanding of how it kind of works together.

David Bain

So Luke saying that he uses machine learning for natural language classification to improve content tagging and entity matching. Lazarina, Susan mentioned there to dive into your analytics initially. What areas within an analytics tool like Google Analytics, would you recommend that an SEO focuses on in order to see what areas they can actually get started with machine learning?

Lazarina Stoy

That’s super interesting what Susan said, because I actually have a totally different approach when I’m using machine learning. But to answer your question, David, I would say perhaps if you are, depending on the size of the client that you have, if you’re working for a very big enterprise organization that has a lot of historic data, the easiest, most beginner-friendly product that you can do is doing forecasting, and that’s something that is also very valuable for the organization. And then if you want to take into consideration all the things that are happening in the SEO industry and typically why most people don’t want to do forecasts or things like algorithm updates, external events, you can start digging into neural networks with your forecasting in order to actually incorporate all of these different effects and the different, the effect on the data that they have for your particular site. I would say that would be kind of where I would start if you have to limit me just to Google Analytics data and if I’m a beginner.

But my approach and maybe the way that I personally have go and started with machine learning and SEO was just, first of all, understanding what machine learning can do. And that means seeing different projects that people are doing, maybe reading a little bit more about what different machine learning tasks are out there and then trying to find the opportunity to implement that, and even just a small project, something that you’re working on. And to me, very contradictory to a lot of people that do machine learning in SEO, I would say that you don’t have to start necessarily with the coding. You don’t have to start with understanding how everything works. You just have to kind of start testing things and breaking things and just trying to get some impact out of it because once you do that, and once you, for instance, classify a bunch of keywords using GPT-3 or whatever else it may be, that is, you can do in literally less than 10 minutes.

If when you start doing that and you start finding those little opportunities, then you actually start to develop that passion for it. And once you have that in you, you’re constantly going to be looking for ways to do more complex things, and you’re going to start branching out into Python SEO, you’re going to start branching out into more complex automations, building your own systems, connecting Apps Script, connecting Google Sheets, connecting Google Tables, which is something I discovered like a month ago and I didn’t know existed. Maybe some people in the chat are going to be like, I cannot believe you didn’t knew this, but yeah, just get your hands dirty is my kind of tip for people that are just starting out. It doesn’t have to be a super complex project. It can just be something super simple, but if you can deliver the impact through it, nobody’s ever going to ask show me your script, or how did you learn this? They’re just going to be interested in hearing the story behind it.

David Bain

So there are many things that you mentioned there that I would like to dive into, but it’s not just a discussion with Lazarina. And also Lazarina, you mentioned that you’ve got a slightly different perspective or a way of doing things compared with Susan. If you don’t necessarily agree with another panelist, jump in, shout and say you disagree-

Lazarina Stoy

Oh no, no-

David Bain

Feel free to do that.

Lazarina Stoy

I completely think Susan’s approach is amazing. I just feel like my personality needs instant, quick results and Susan’s approach is so theoretically founded and super long term and definitely a lot more sustainable, but I just had a different perspective. That doesn’t mean I disagree. Absolutely.

David Bain

Okay. Well just before we delve into Jess’s perspective on maybe how to get started with machine learning, I just want to delve slightly deeper into one thing that you said Lazarina, and you said dig into neural networks with your forecasting. What does that mean in practice?

Lazarina Stoy

So anytime that you’re building a model from scratch, which is not as often, at least, I don’t do it that often. Maybe Jess does it a lot more often than I do. But if you are building a neural network or using neural networks in your machine learning model, you can provide different parameters based on the data set that you have. So for instance, you can essentially say at what extent, or what weight do you want to give to certain things that are happening? And this is not a project that I have personally done. I have done neural networks back in the day in university. But if you are interested in doing a little bit more advanced modeling, I think that that is the way to go, because this is something that Google is using as well.

But yeah, if I were to be limited with Google Analytics data, I would definitely look into, if you have the dataset that is big enough to do neural networks, which is mind you, that is a million pages and literally no less than 10 years of the site being up and running and things like that, high traffic, I would definitely look into this as kind of the machine learning modeling that you can do.

David Bain

Now you also mentioned the importance of using machine learning to classify keywords. So hopefully we can park that thought and actually get back to it in a little bit. But let’s go to Jess. Jess, what are your thoughts about how to get started with machine learning for SEOs?

Jessica Peck

I’m kind of, I feel like I’ve got kind of an approach between Lazarina and Susan where I’m like, you’ve just got to kind of get started, but it is good to have a theoretical approach. It’s good to have, I’ve some software engineering kind of not background necessarily, but to get some of that experience just so you understand how the data you are piping in is going to affect the results and how to clean that data and that kind of thing. I think the best way to get started with machine learning, especially in SEO, is to find, look at places where there are going to be patterns that machines can see that humans can’t and try and find those patterns because machine learning is all about that kind of pattern recognition and Google is using it to do that pattern recognition. So there is a perspective that you can get from machines that you can’t get from people.

And that can be keyword clustering. That can be, I mean, it could even be just simple data analysis and pattern matching with content, just being like I’m going to TF-IDF, these, this text and see what the results are. I’m going to pull all of the entities from the Google top 10 results for a keyword I’m looking at and compare them and see if my page is missing a bunch of entities that all of the competitors who are beating me have. It’s, I think you can get started with code, or you can get started just by using the libraries that exist and things like Streamlit. There’s a lot of flexibility out there.

David Bain

Jess, can any size of website harness machine learning, or is it something that’s only appropriate for sites of a million pages plus, or can a site with just 100 pages on actually use machine learning? Is it worthwhile using it for that size of site?

Jessica Peck

So it depends. I think it’s all cost benefit analysis. If you have 100 pages, but they’re 100 pages that you would really like to be ranking pretty high, and you’re in a competitive niche, you should definitely try using machine learning to kind of see what patterns you are missing. I think if you have a hobby site and you don’t want to use machine learning on it, you don’t have to, right. It’s not a must have necessarily, but it can give you a competitive edge, and if you are trying to seek out as many competitive edges as possible, machine learning is definitely one of them that can really have cool dividends.

David Bain

So another phrase that crops up is natural language processing, and I think a lot of SEOs will be aware of that phrase, but not necessarily know exactly what it means and how it can really benefit what they do. I’m just looking at faces here as who to ask this question to. Who wants to talk about natural language processing to begin with?

Jessica Peck

I can volunteer for that.

David Bain

Okay.

Jessica Peck

I’m on the edge of my seat.

David Bain

Go for it, Jess. Go for it Jess. Okay, so what’s your summary of what natural language processing is and the benefits that it offers SEOs?

Jessica Peck

So historically computers are really bad at understanding language, right? You look back, beginning search engines, Google even, basically Exact Matching. Does it have all of these words on this page, then I’m pretty sure it’s good. And we’ve kind of moved from that into natural language processing, which is a kind of grouping of the methods and algorithms to try and understand how human beings naturally speak and naturally type and then creating nodes and paths to understand that and apply it to let computers comply it, to let computers apply it to understand language. So we talk about BERT a lot recently. That was a big thing recently, and that’s a way of letting computers understand sentences in ways that go forward and back, it’s letting computers understand context and slang and subtlety and the way words interact with each other, not just I see the sentence, I don’t know what it means, but it also applies over here, and that kind of easy pattern matching.

David Bain

Okay. Okay. So Susan, I mean, it’s about understanding the relationships, I guess, with different texts published in different places. How can an SEO practically make use of that to improve the rankings?

Susan Connelly

Well, one of the things that we talk a lot about is using the knowledge graph. So I think that Google does have a database API access that allows you to take a look at what is existing in that knowledge base. In addition, it’s good to see what other data points, are there other specific data points? So say if you have a travel site and you might want to focus on, let’s just do hotels because Google has the most natural language processing information storage on that particular niche. But as we talk, look about having some understanding of those specific hotels’ information and looking to see what other data statistics are out there, and then if I’m writing a, example would be if I’m writing a travel guide or a travel blog, what are specific data metrics that people are searching for that’s already existing out there?

So if you’re using, featuring a map, you can get nearby walking distance and you can get nearby landmarks and attractions. You can get a lot of information as far as people see different information such as FAQ queries and things that people are asking for and what those topics of understandings are. So as an SEO, I think that’s a really great project is to take a look at a page and say what is some existing content that would enhance this page and what is existing that’s out there that we can kind of draw in and then it’s kind of filling in the blanks to say, all right, now we have a good idea of some of that automated, the automated aspect of it. Now if we can fill it in with an actual person going through that content and we’ll get a lot better of that combination of what works from the machine standpoint, but also from a user having that final review of going through that content.

David Bain

Okay, great. So I guess we’re using machine learning to analyze what other similar websites rank out there, have a look at their pages, see how many words, see the format of the pages that also exist in relation to those queries and then feed you back some advice in terms of what you need to do differently to give yourself a better opportunity to rank for those keywords as well, which I guess brings us up into the content that’s written on those pages. Can we trust the machines to actually write the content as well? Lazarina, what are your thoughts on that? Are machines good enough to write content now?

Lazarina Stoy

You ask me this every time I come on this podcast. So my response has not changed since the past two times. So I think no, but there is this big, this bot is getting bigger and bigger and I will tell you why, in my opinion. I can, what Susan said, 100%. If you can use publicly available APIs to get data and combine this with first party data that you have on the website, you can build literal hundreds of pages that are programmatically generated. And this is a super scalable, automated content system and it doesn’t involve any sort of writing. I do apologize for some yelling, by the way. So, and it doesn’t involve any sort of writing by an AI model like DPT-3, GPT-J or any sort of equivalent.

From another point of view, if you have the parameters that you want the model to be written with, you can provide very strict restrictions on what kind of content you want. And even if you can break it down by paragraph, if you can break it down by different sections, what each section has to cover, then that also allows you to build very, very scalable page templates or content briefs that then you can provide just for an editor to review, which again provides a lot of opportunity, not only for small sites, but also for sites, like Susan mentioned in the travel niche, that you can build neighborhood guides or even real estate and things like that. So lots of opportunity there.

I wouldn’t just put it on autopilot and forget about it, although I’m sure that there are a lot of SEOs that are doing this at the moment, and they’re just waiting patiently to reveal their case studies of how they made a million dollars using GPT-3. I am sure there are, but I just think that at some point, Google is going to be smart enough to recognize that, or at least I’m hoping, and once that happens, you don’t want to be on the wrong side. It’s kind of like 10 years ago when it was black hat link building and at some point, the algorithm comes in and you’re just like, oh my God, I just lost my livelihood. And it’s like-

David Bain

10 years ago plus.

Lazarina Stoy

Just, you should have known this was coming. Yeah, you should have known. So, yes.

David Bain

Lazarina, I’m going to keep on asking you that question as well. So we’ll see how your opinion changes over the years.

Susan Connelly

Yep.

David Bain

Susan, do you-

Susan Connelly

I know that-

David Bain

… agree? Disagree?

Susan Connelly

Oh, I was just going to say one of the big things I think that a lot of SEOs are talking about right now is the fact that John Mueller came out and said basically that they are not in the fan point of looking at content from AI perspective. I think that I agree with what Lazarina said. So as far as it being a long way off, so I think we are a long way off for that content to be completely customizable as far as that. But I think that when talking to other SEOs and other people that are in the industry, like she says, that window is getting smaller and smaller and the content is getting smarter and smarter.

And so while I think that it’s, Google has come down and said this, the idea is that you won’t be able to distinguish the difference, right? So that’s where people are kind of waiting to get to is where you can create that content and it has a story and the pieces that are collecting in an automated way, but that there is no difference in it, and then as you said, you definitely wouldn’t want to set it and then let be.

David Bain

I mean, rather than the window getting smaller and smaller, is it not maybe the shade of gray in SEO getting lighter and lighter, and perhaps you can get away with using machines to write certain elements within pages that aren’t massively high trafficked? And perhaps you could do things like having the initial couple of paragraphs or section of a page written by a human, but if you decide that it needs additional content, then perhaps automating the creation of content for lower down on the page? Jess, what are your thoughts on that?

Jessica Peck

I have three conflicting thoughts on this. My thought one is please don’t use machine learning to make content. It sucks. I hate, I know people are already doing it and I hate when I search for something and the first, this programmatic result. I search for a lot of code related stuff for my day job, and there are sites that just scrape Stack Overflow and then spin it using sometimes machine learning and you can tell, and it isn’t good content and it’s just the same result 10 times but in different ways, and I want a human to have looked at it. So that’s my personal, please don’t.

But then point two is people are already spinning content. People are already writing kind of crappy SEO content and getting it to rank, right? We can admit that here. And we’ve all done a Google search and been like, this is terrible. Why am I looking at this? So I don’t know if a well trained AI content spinner is all that different from I paid someone 20 cents for 100 words so they just kind of wrote stuff because they’re underpaid and overworked, and I’m just trying to make as much content as possible.

I think another aspect of this though is that people have been doing automated content and ranking, and it’s actually pretty okay in niches where it makes sense. So stock prices and stock sites do a lot of content spinning and do a lot of just looking at the stock and being like this stock went up, this is why we think the stock went up. You should buy it, not buy it, and they can do that without human involved at all. And it is actually useful content because it is observing something automatic saying this is what we see with it, putting it in a human readable format and then people have alerts on their phones for those kinds of articles and a human wasn’t there at all.

And I actually have a fourth aspect of this, which is also all of this technology can look really impressive, but have a lot of cognitive biases baked in, like we were talking about language. A lot of these were developed in Mountain View or in California by a bunch of white dudes who speak English, and nothing wrong with being a white dude who speaks English. But it means that if you’re building something, you may not see some of the content come in or some of the data that you are building your machine around have biases.

David Bain

But that’s scary. I mean, machine learning also by its nature, looks at what’s happened in the past, what currently exists and attempts to make sense of that and perhaps create something based upon that. How do you remove those cognitive biases to try to ensure that there’s as true a representation of what’s happening or what is the truth within the content that you’re producing?

Jessica Peck

I mean, at the moment it seems like everyone’s way of doing that is I’m just going to let my algorithm go wild and let people use it. And then someone on Twitter will be like, hey, I used GPT-3 and got a lot of racist stuff. And then they go, oh wait, that wasn’t supposed to happen. We didn’t want that to happen, and it’s like… So they go back and fix it once the problems have been pointed out to them. So you really need to be QAing, you need to clean your data, if you are building stuff, right? You need to be, have a data pipeline that understands that it is building on things that already exist, and then understand the flaws of things that already exist.

If you are using algorithms that someone else made, which probably you are, because the best ones are the ones that someone else made, you still need to be aware of these biases. If you are a site, if you run a site that talks about historical events or talks about diversity and inclusion, do not just spin stuff up with an algorithm someone else made, because you could get in real big trouble real quick, right? This is, if you’re-

David Bain

Great points.

Jessica Peck

Yeah.

David Bain

Talking about QAing, Simon Cox said, how about AI writing content that is then edited by humans? So is that a reasonable solution there, Susan? I see you nodding your head there.

Susan Connelly

Yeah. I think that, as I mentioned, if you’re going to use some of these sources and data sources to create content, then the last step is to have a queue and have a team of editors that are going to go through that content. As Lazarina has mentioned, you definitely don’t want to launch it without having some editors go through all of that content and look and observe and read through all of that content, for sure. And I’ve had sites, I’ve worked with clients and sites who have launched content and where, while it wasn’t necessarily the automation of the content and the data from that content, but even if you’re creating that content, you want to make sure that it’s actually rendering as it’s supposed to. So you want to check to make sure it’s formatted correctly and it’s grammatical and it reads like a person does. And I really think that saves a lot of time for a lot of clients. So they’re able to create the overarching of that content and then just have an editorial team really go through and provide and make sure that it does resonate with the consumer.

David Bain

Yeah. The quality of these types of machines are improving all the time. In the last year or so, I’ve been using automated transcripts services. And even now compared with a year ago, the quality of what’s being produced is absolutely incredible. It’s just getting better all the time. One other word that we’ve discussed today has been intent. Lazarina, you talked about classifying keywords. Can you also classify keywords for intent using machine learning?

Lazarina Stoy

Yeah, absolutely, you can. But again, depends on how good the output is going to be and how much data you can provide for the examples of what you’re going to give. So if you are building your own custom model, it requires a lot of data in order to learn what actually is informational intent, what is transactional intent. There are a few different models that you can use such as Naive Bayes. That’s one of the most super rudimentary, easy-to-implement type of models that’s rule based. So rule based means essentially the same thing that you can do in Excel or Data Studio, where you say if it contains this and this words, or a similar type of word, like these, then it is informational. So for instance, if it’s what, why, how, or whatever, then it’s that. But more advanced models would like to look into the nuances of things, incorporate synonyms, incorporate other things that are more important into how we actually classify intent and thinks as well, like what is the niche, what is the context and things, which I have not actually seen done that well.

I have actually experimented with using GPT-3 to classify keywords based off intent and it generally works really, really well. You can also classify keywords using GPT-3, if you provide sufficient examples based off of whether they are part of the industry or not, like if they are related to your kind of subject category or your site, you can do all sorts of things. But again, the limitation, as Jess said, that this is a pre-trained model, so the customizations that you might want to achieve with it are very limited. So-

David Bain

Susan-

Lazarina Stoy

… take it as you wish.

David Bain

Susan, are we going to get to a stage fairly soon where machine learning is all we need to classify keywords for intent, or will we always, at least for the next few years, require human input on that?

Susan Connelly

One of the most interesting things that I’ve seen is that there’s a lot of companies that are doing supervised learning for intent. So they’re taking it a step further. So an example would be, if you are worried a little bit about that modeling as far as from the consumer standpoint, you can bring different audiences in and different age groups, different demographics, and you can do some supervised learning and start to have people search for particular products in the, it’s, I’ve been in digital marketing for almost 20 years and one of the things that we did a lot was bringing people in and doing similar tactics as far as saying if the site navigation structure works well or how does a user in a different age group or different audience type get to the site and kind of navigate through your website.

So I think that as far as intent goes, I think that that’s kind of the perspective that Google is also tagging, though they haven’t gotten to the level of, I’m not sure how much level they are looking at demographics and things like that, but that kind of opens it up into really being able to get more specific for the audience types that we’re, that are kind of searching for that information. So that’s where I, people always complain about it, but I like it when I get the requests on social media sites or Facebook sites or things like that, sometimes when it works well, that is kind of understanding what you’re looking for before you’re looking for it kind of modeling that is coming soon and is being used in the industry right now.

David Bain

Let’s try and cover one more question. Would you recommend machine learning for identifying images in order to generate alt-text for them? I see a slight nod from Lazarina. I’m not sure what I see from Jess. Lazarina, because of your slight nod there, what’s your thoughts on that particular question?

Lazarina Stoy

Well, I think it’s a very good first step when you have a lot of images to tag. Again, same approach applies here. You cannot just run a model, and by model, I mean Microsoft Vision API is one API that you can use for generating captions for images. I know Google has one, Amazon has one, but at this point there should be someone that actually goes through and provides a sufficient context because most of the time, the alt-text that you’re going to get from a model like that would be something like man in a picture, which is obviously not enough. It’s not enough. Sometimes it might be okay, at least in order to identify what is a graphical element, which we know they don’t need alt-text, but what is actually an image that is visual, that has a lot of things going on, and in some cases with every image recognition API, you might also come to the point where a cat is recognized as something totally different and you need to have someone to actually go through that and see if it makes sense.

From what I know, especially with projects like this, it’s always good to have a starting point because the whole exercise seems less daunting. Even if it means you have to edit it out and enhance it, it’s always better to enhance it than to give someone a spreadsheet with 2 million images and say, go tag that. Nobody likes to do this and nobody wants to do this task. So I think it’s good to have kind of a starting point, using machine learning but then from that you need the human input.

David Bain

Superb stuff. Now we’re getting lots of comments and questions in at the moment. You’ve got to get your questions in earlier. Don’t know if we can cover them all just now, but it’s great to see so much interaction there as well. I just want to finish up by asking everyone what’s one thing that SEOs need to be using machine learning for right now? Perhaps it’s something that they’re doing manually at the moment or using some other tool to do just now, but it’s a fairly auto, it’s a fairly manual thing that they’re doing just now. So Jess, should we go to you? What’s one thing that SEOs need to be using machine learning for right now and it’s a much more efficient thing to be doing?

Jessica Peck

Keyword classification and cleaning up your keywords and the data that you use. You can use, you can split it up using different like N-gram methods and just K-means it and get clusters without, and then you can just explore it a lot more quickly. You can also just use some general data science and clean out stuff you don’t care about, and it will make your life so much easier once you have everything tagged and clustered for you and you don’t have to scroll through 10 million rows of data to try and figure out what you want. That’s my-

David Bain

Great stuff, Jess. Thank you. And Chris is just saying the chat didn’t catch Lazarina’s MS AI suggestion. Do you want to repeat that, Lazarina?

Lazarina Stoy

Yeah. Microsoft has a Vision, Microsoft Vision API is the suggestion for Microsoft, but I know Google has one, Amazon has one. There are plenty of captioning APIs if you are looking to go down that route. But yeah-

David Bain

Lovely.

Lazarina Stoy

It works generally well.

David Bain

Okay. And Lazarina, I might actually just throw you by asking Noor’s question instead of actually what’s one thing that SEOs need to be doing just now, because I’m sure you’d able to handle that one as well. And that’s how to learn machine learning for SEO, what’s one good resource that you could point people to?

Lazarina Stoy

Oh, one good resource. That is hard to think from the spot, but-

David Bain

Okay. 510.

Lazarina Stoy

Okay. Okay. Okay, let’s go with a few. Definitely start looking at code as early as possible. So if you have a particular model or task that you want to do, you have to break down the task in the terms that machine learning developers use. So if it’s for instance, generating content, then it is what kind of algorithm is it? What are the big players in the market? How you can get started. From then I would say maybe read a paper or two. They’re not that daunting, I promise. Look at the code. And then if there are any sort of tools developed with that, if you have any no code alternatives, go down that road, if you are a beginner, of course. But if not, and you want to go in the deep straight away, then Stack Overflow is going to become your best friend and just search whatever you want to do plus the keyword tutorial and just wrap your head around four hours of troubleshooting and error handling, and then repeat the process maybe a few different times and you’re going to get it for sure. If not-

David Bain

Okay.

Lazarina Stoy

… then, yeah.

David Bain

Thank you. And Susan, I’ll give you a choice of either two questions to answer. One is either, if you can share another resource that Lazarina hasn’t mentioned that an SEO can find out more about machine learning for SEO or on the other hand, what’s one thing that an SEO needs to automate now, because what they’re doing manually just now is just not cutting it.

Susan Connelly

I think that as far as learning more, there’s some great courses that are on Coursera. I know that they’re free, which is always great to start learning, because I know that when I started learning, I didn’t want to make that commitment until I could figure out if I could start to understand some of these processes. So getting a lot of those free courses available. In addition, I think that starting with the APIs and starting, even if you’re starting with Google Sheets, starting to look at what APIs are out there and just play with bringing in the data and kind of taking a look and seeing what’s available. And then as I mentioned, just having some kind of idea from an analytics perspective. One of the things that SEOs don’t do that often, which I’m always surprised about is as you’re executing meta titles or meta descriptions, as you’re creating contents and changes to just start storing those changes. So if you have like a 25% difference or you’re having a 10% difference or 5% ranking difference.

I think that can help a lot as far as understanding should a meta title be structured with the keyword on the front? Should it be, have it in the middle? Do we see it more of a user based? So definitely SEOs should be testing and just recording as much as possible so that they can have a really good understanding of what works in Google’s eyes or what doesn’t work and start to really model that as far as taking that, because that kind of, that’s the starting place, as everybody has mentioned. Even if you’re doing something as meta titles and description, playing with that and playing with the recipe and seeing what changes and how that is being affected. Are more users clicking on that or not, and getting that performance right. I think it gets into, okay, now if I’m going to create a model that’s much, much larger, it’s really the same recipe. It’s just on a larger scale.

David Bain

Always be testing. That’s another episode in itself I think. Great advice and thank you so much all for coming on. I think we should just finish off but just going around our panel again and reminding people who you are and where they can find you and feel free to share a website, social profile, whatever you want. So Lazarina, please remind, if you were the listener, who you are and where you’re from.

Lazarina Stoy

I’m Panelist One. So I’m Lazarina Stoy. You can find me over at Twitter at @lazarinastoy. My website is also titled lazarinastoy.com. So yeah. Feel free to send over any other questions that we didn’t answer.

David Bain

Super. And Jess, please.

Jessica Peck

Yeah, I’m Jess Peck. You can find me on Twitter at @jessthebp. I tweet way too much. But also you can find the company I work for at localseoguide.com and we actually have links to some forecasting code that you can plug your info into and do some forecasting with on the site. So Lazarina talked about forecasting earlier. If you want to give it a whirl, we have some stuff for that.

David Bain

Wonderful. And last but not least, Susan.

Susan Connelly

Perfect. I am available, I’m Susan Connelly. I am available at @bizdevelopwiz, actually, which is B-I-Z-D-E-V-E-L-O-P-W-I-Z on Twitter, as well as you can find more information about our company at milestoneinternet.com.

David Bain

Lovely stuff. Thank you, Susan. I’ve been your temporary host, David Bain. You can always find me producing podcasts and video shows for B2B Brands over at castingcred.com.

Join us for the next session on Wednesday the 8th of June. That’s when it is, on 8th of June at 500 PM at BST, that’s 1200 PM Eastern Daylight Time. We will be discussing strategic alliances. Dixon Jones should be back in the host chair for that one, and joining him will be Jean Lachapelle and Melody Sinclair-Brooks from Agency Analytics. Sign up to be part of the live audience for that one over at majestic.com/webinars. But thanks so much for the panelists, thanks so much for you and for being part of today’s show. Bye bye for now.

Aditional Resources

Want to know where to get started? Then take a look at Lazarina’s ‘Beginners Guide to Machine Learning for SEO‘.

As mentioned during the webinar, there are also limitations of AI and how intelligent machines can be tricked.

Previous Webinars

Follow our Twitter account @Majestic to hear about more upcoming webinars!

Or if you want to catch up with all of our webinars, you can find them on our Digital Marketing Webinars page.

Comments

  • Christopher Romero

    Looking forward to this one as it’s a great topic and am using some Machine Learning in my process.

    April 28, 2022 at 5:01 pm

Comments are closed.