← Voices of Search // A Search Engine Optimization (SEO) & Content Marketing Podcast
Voices of Search // A Search Engine Optimization (SEO) & Content Marketing Podcast
Working w/Embeddings & Building Your Own Code
00:42:53
Loading summary
Transcript38 lines
- [00:00]
A
The Voices of Search Podcast is a proud member of the I Hear Everything Podcast Network. Looking to launch or scale your podcast, I Hear Everything delivers podcast production, growth and monetization solutions that transform your words into profit. Ready to give your brand a voice? Then visit iheareverything.com welcome to the Voices of Search Podcast. A member of the I Hear Everything Podcast network, ready to expedite your company's organic growth efforts. Sit back, relax, and get ready for your daily dose of search engine optimization wisdom. Here's today's host of the Voices of Search Podcast, Tyson Stockton.
- [00:44]
B
AI Tools continue to reshape SEO One particular area of opportunity is further utilization of vector embeddings. They enable us to map queries and content into vector space, allowing us to understand how search engines interpret context and relevance. That opens up a new way to expand beyond single keyword targets into capturing entire clusters of semantic semantically related queries and have great combined impact. Although many SEOs may not have the technical background to take advantage of this shift, that's where custom tools and workflows come in. By building embedding analysis that utilize emerging tools like Cursor or Windsurfer, we can visualize semantic clusters, plot vectors on T S and E charts, and show clients exactly where the opportunities lie. This is the Voice of Search Podcast. My name's Tyson and today I'm joined by Rylan Bacorn, founder and growth advisor, Bokeh Day. Bokeh Day is a growth consulting agency in Santa Monica. Ryland's going to be walking us through how embedding analysis works and how we can use this in our workflows to further improve our impact in SEO. So with that, Rylan, welcome back to the podcast.
- [01:53]
A
Thanks a lot, Tyson. It's great to be back.
- [01:56]
B
I've had you on here a couple times. I mean, we've worked together across various companies. I feel like at this point. So it's great to see you again. How you been?
- [02:05]
A
Likewise, I've been great. Family's good, health is good. Can't complain. I'll find reasons though.
- [02:11]
B
And I mean since. And actually I forgot maybe where you were. How long ago is even last time I had you on the podcast, but this year too, you've broken out on your own. You've started Boca Day. Like, I mean, I know we have a whole other topic to get into, but just maybe a little bit on that. Like, I mean, I'm sure it's been quite the journey kind of making that transition.
- [02:33]
A
Yeah.
- [02:34]
B
How are you doing with that kind of Change?
- [02:36]
A
Yeah, it's been a lot of fun. So I get to work with a lot of different type of clients, so lots of different problems to solve, and I get to actually bring in a lot of the cool AI stuff that everybody's talking about and actually implement it. So I've been working in corporate for a long time, like 15 years, and the appetite is there. But then when it comes to actually executing, it's challenging to actually get these things on the roadmap versus companies that I've been working with. I'm able to build things for them and actually implement some ideas, some hypotheses, and help them flesh out their SEO strategy a little bit more based on what's happening today. So embeddings is one of those things that I spent a lot of time with. I built a couple of tools, not a developer. And so it's been a lot of fun actually producing code. Although I do share this stuff with my developer friends, and they scorn, but that's beside the facts.
- [03:28]
B
Well, and I think that's like, that's one of the things that I want to touch on in this conversation is the. The barrier of entry on a lot of these topics, I feel like, is lower than it's ever been. And in our kind of side conversations prior to this, I think that was something that really kind of shone through and that I'm hoping, you know, more people can take away from this conversation, is feeling more motivated, empowered, whatever. You have to actually also play around in the space and see that it's something that is more approachable than maybe it was years in the past.
- [04:04]
A
Yeah, it's. It's actually. It's the easiest it's ever been to actually get started on these things. So, I mean, I've been pushing code to GitHub more than I've ever been. Again, not a developer, but with vibe coding, which I guess is dead, by the way, it's like agent swarm coding is the way to go, according to some other more sophisticated people. But you can just talk to an agent. It'll build code for you, it'll ship code, it'll push code, and it'll test the code. And of course you're gonna run into things where it's like, it's perfect. It does exactly what we had talked about, and it's gonna. And then you'll test it, and then it's like, okay, cool, this isn't working. You know, it's throwing all these errors. That's actually where the fun has been for me. That's the learning. And you really start to get into, like, what actually works, what doesn't work, but the barrier of entry. You're right. Like, it's never been lower. It's probably only going to get lower. But people who don't code, like, I don't know how to set it, I don't know how to write Python in some editor, but I can look at it and get a sense of what's going on. And I can also copy and paste the error into the chatbot and say, help me fix this error, give me some diagnoses, give me some proposals, we'll figure out the best. And so where it's actually been a lot of fun for me is trying to wrangle these agents and these, these chats to actually do what I'm looking for. That's where the prompt engineering comes in. And if you haven't had anybody on prompt engineering, I definitely recommend that. But getting into the space and just playing with it, that's really it. You just have to install stuff and do free trials. I'll talk a little bit about that. Nice.
- [05:36]
B
Well, maybe to bridge into the topic, let's kind of start with the basics. And so first, would love to hear kind of like in your words and stuff, like, what are embeddings? How does that fit within SEO? What is the relevance? And why is this something that as SEOs, we should utilize more in our daily work?
- [05:58]
A
Yeah, so embeddings, they're just a type of vector. Vector is just a string of numbers, multidimensional, which at first, when I was introduced to this concept several years ago, it's like multi dimensional space just kind of makes my head hurt. Fun fact. Like, obviously we're only seeing 2D right now, but as three dimensional beings, we can only see 2D. And if we were in 2D, we'd only see like lines, we'd only see one dimension. So just kind of get past that whole concept. But embeddings are vectors. They're expressed in multidimensional space. And these are things that basically they help machines understand words and sentences and documents. And, you know, we're not going to talk about images and videos, but it's getting into that space as well. So it's a mathematical representation of a sentence, so. Or concepts. So one of the classic examples that is always shared is the, the queen king example where you have a 3D chart and you have king and then queen and then you have man and woman. So if you take king, you subtract man, add woman, and then you'll ultimately get queen. And this is how machines understand semantically that this is what you get with royalty and gender. So this is just a way for machines to understand concepts that humans get from language, which of course is challenging. I mean, English, for instance, you know, it's very, very, very difficult for none, native English speakers to deal with most of the way that we talk with just casually. That's because it's, you know, this conglomerate of all these different languages. In any case, this is important for search engines because this gives machines relevance. So instead of saying, you know, I'm looking for a cake and you're only doing the strict exact match query for cake, it'll give you all the different types of things that you might find at a bakery. This is where it's useful for machines to help you find things that are semantically relevant because it turns all this stuff into embeddings. And then the embeddings inform this multidimensional space to help search engines, other machines give you results that are relevant. So for SEOs to bring this all back in, understanding how embeddings create these relationships with your content, the queries that users are searching for, this is what helps you build a strategy and some of the tactics that help you optimize to improve your relevance.
- [08:30]
B
And I feel like a lot of SEOs listening to this, regardless of how familiar, how much time you spent on these, you know, topics of embeddings, vectors and stuff, I feel like the writing is already right there on the wall. As we've talked about relevancy in SEO, we've talked about, hey, it's not just one topic, but how you're addressing the larger scheme of things. So I think like probably in some form or another, every SEO, regardless if they know it, probably has some familiarity in these topics. Even if the exact words or phrasings is something that they hadn't been as exposed to.
- [09:09]
A
Yeah, that's right. Search is no longer just giving you the results that are exact match. That's, that's gone. So moving into a space that's based on relevance, based on, you know, I mean, the Amazon example, you go down, you see related products. You're not going to find like the exact USB cord. You're going to find things that you can plug into it because it, you know, involves this space. But the fact of the matter is, is these things have actually been around for a long time. Word2Vec, been around since 2013. It's older than a decade now. So Google's had AI in their search results for a long, long time. It's just become popular in the lay vernacular to talk about AI and what things are doing. And now Everybody is using LLMs and chatbots. And so like that's where the fun is to try to get into these chatbots. But it's not, we're not diverging from like helping companies become more relevant for these new spaces. We're still applying some of the same tactics. We just have an additional lens of trying to understand how to represent this and how to measure this. And so that's where building your own custom tools is really, really helpful.
- [10:23]
B
And I mean, I guess the further double down on the importance or significance here maybe too, a lot of people are probably thinking like, okay, so this I can identify or use as a tool to identify content with optimization opportunities or maybe identify missing pieces of content. But I would say like the applications of this can go a lot farther beyond just like that content layer. And maybe you can kind of just touch on some of the different forms that you could actually use this information beyond just written copy.
- [10:57]
A
So there's a lot of different applications beyond content. So obviously we want to solve for SEO, but if you're building tools that help turn your web's product into vector embeddings, you can get a lot more out of it. So you can apply this to your site search, for instance. So most site search products aren't that good because obviously they're not Google. But if you were to run a search against a vector database, for instance, with your product catalog, you're going to get similar items. So you're searching for a winter coat, but you're not necessarily searching for like the layers that are applicable. If you were to search vector database, you would find other related products in that way. So that's something that's beyond content. Additionally, sticking on the e commerce here, the recommended products that you're going to find that I mentioned earlier, those are going to be things that are semantically similar. If you're using a vector database, as opposed to just a regular old database of here are the things that are in this product category, then in SEO we can do a lot of things where we can measure the distance of where your content, your particular topic is to a primary query, the related queries, the supporting queries. I even built a tool that can help you evaluate whether or not all of those related queries are valuable. Everybody's gotten a big massive keyword list from all the tools and then you start parsing through manually. It's like this isn't even relevant. It's the same word, but it's completely irrelevant. So doing things like cosine similarity, you can help eliminate a lot of that stuff. And you can move into building content or just building assets for your SEO strategy to optimize just some of the on page elements. So the applications are really endless. But for SEOs, this is really, really important because this is how search engines are looking at your content. And more importantly, I think for everybody who's wanting to figure out the LLM optimization or the other acronyms that are being thrown around now is how do you actually become relevant for the chatbots.
- [13:04]
B
And what are some of like the differences that you're seeing in that as far as like how maybe it would be seen in traditional search versus AI mode or ChatGPT, whatever other tool you want to throw into the sun there. But like, would you create a separation in how that is being utilized in like one form or another? Or would you say it's like it's still a fundamental layer that's being leveraged across both.
- [13:38]
A
This is getting into like this year's been interesting because it feels like the late 90s again. It feels like the Wild West. It seems like everything is changing, massive disruption. And there's a great debate around, like there's a great debate on what even to call this stuff right now. My opinion is that if you're doing what I call SEO, which is helping companies create experiences that are actually useful for their users, with a lens on how search engines, answer engines, are actually rendering your content, that's the stuff that is still useful and relevant. How many articles have been written about SEO being dead? Countless. We're still, this podcast still exists. There's companies doing SEO that are more successful than ever, but there's also significant struggles that people are facing in trying to figure out what exactly works. And so that's where the debate. Is. So trying to figure out where a very small portion of your users are coming from is interesting. It's fun trying to get the citations as opposed to backlinks. That's one differentiator. Trying to get fresh content in front of an answer engine that maybe isn't going to be super relevant to your organic search results. That might be something that you differentiate from, but by and large the focus still is how are you creating experiences that are useful for your users? How are you fulfilling user queries? That's the fundamentals for what I do. And that's where I think SEO has been and will continue to go Obviously there's a lot of cool things and like tricks and maybe some black hat stuff that you can explore now, like, you know, putting text on images or white text on white backgrounds. That seems to be working for some LLMs. There was a recent podcast with Sam Altman talking about how they didn't anticipate companies would actually want to show up better. They were more worried about copyright infringement. So being relevant in LLMs, this is the new shiny thing we want to be here, but it isn't dissimilar to how are you creating experiences that users want to actually come to. The only challenge now that I see is really a differentiator is the funnel is completely changing and the measurement is extraordinarily difficult. So typically with LLMs, ChatGPT. Typically. So a lot of people I talk to, they're like, yeah, people coming in from ChatGPT, they convert like crazy. So talking a little bit more about that, the reason that is, is they've gone through the entire funnel in ChatGPT. They don't need to engage with your site, they don't need to come to your site, they don't need to read your great copy because it's all being consumed by an LLM and it's being produced in their chat experience. And so when they say, cool, I want to put money down, there's a link they get to the transaction page and that's where people are interacting. So of course their conversions are going to be great. But it's still like at best, 5% for most sites. Like 2 and a half is the number that I hear regularly quoted. And as we know from some other publications, like people using ChatGPT still by and large are using Google to find things. So sorry, SEO is not going away as ever. We're adapting to the new technology. That's exactly what we're talking about.
- [17:06]
B
Absolutely. And it's in a lot of ways too. It's like, yeah, the game fundamentally hasn't completely changed. I think the theoretical approaches and like the mindset of an SEO of hypothesis test reiterate like those things are all still relevant and valid. And it's just like, yeah, we have different pieces to play with. The field's changing, where we're competing is changing, and maybe our strategies, our prioritization, but there are all these fundamental layers are still intact. And with that, I want to dive in a little bit deeper into like that tool aspect. And we touched on it in the beginning here because I think it's something that's also quite interesting with like the shifts and changes right now is being that, that, that barrier, that bar, that kind of gap that you had to bridge has been changing. I think a lot of people can now approach into that space. So how do you like approach building tools and kind of letting non developers tap into some of these areas like embeddings, but maybe they don't have that experience in actually writing code.
- [18:24]
A
So I think we talked about IDEs, integrated developer environments, right? There are many, there's more and more coming and the struggle is real for them to remain relevant. I started using these a little over a year ago and the interface is the same, but the application is improving every day. Some people I've talked to is like in AI, it's a war basically because last week something was relevant, now it's completely thrown out. But these tools, these are the things that people can get into and they can start writing code like you don't need to know how to write Python to develop something that is useful and actually runs on Python. So like I built a tool on Hugging Face, the first little cool thing that I built. I didn't even have one of these IDs. I was just writing, I was creating the Python code, copying and pasting the Python code and putting it into a text editor and uploading it to Hugging Face manually through Terminal because you know, you can, that's where a lot of this stuff happens. You know, uploading things to hugging face or GitHub. Easiest way to do it is Terminal, provided you understand that space a little bit. With an ide, you don't need to know how to operate Terminal. You can just say, yeah, do this, do this in Terminal. It'll give you the command to run it. You know, getting things set up is how you, you can actually make the magic happen. But the first thing that I built on Hugging Face was the, the keyword cosine similarity tool. And I had a need. A client had a need. They wanted to get a clean list of queries to focus on instead of this big list from some of our favorite tools that didn't actually give them a real good target.
- [20:12]
C
Time for a one minute break to hear from our sponsor, Pre Visible. So you're looking for SEO help and you got a couple of options. You could start replying to spam from agencies that claim they can get you to rank number one on Google. You can pay an hourly rate for a consultant who will inevitably nickel and dime you with hourly charges. Or you can work with a cookie cutter agency to quickly launch a strategy less project with Low success rate. None of those sound very good now, do they? Well, that's where Pre Visible's integrated consulting model comes in. Pre Visible draws From a collective 40 years of SEO and digital marketing experience to unlock your organic growth opportunities. They build custom solutions that combine strategy, technical expertise, content and reporting to effectively operationalize SEO for your business. Pre Physical's four stage approach ensures that your SEO programs thrive by starting off with a strategy first approach. Then they support you in your efforts to create quality content, help you identify technical issues, and most importantly, they'll work with your cross functional teams to integrate your SEO strategies to make sure that your SEO budget actually drives results, not just your agency's bottom line. So join brands like Yelp, eBay, Canva, Atlassian Square, all who rely on the SEO consultants at Pre Visible. For more information, go to Previsible IO. That's Pre Visible. P R E V I S I B L E I O.
- [21:41]
A
So you know, cosine similarity is basically the distance between these vector embeddings. And there's other people who can go into detail around the exact math on it, but again, you can just ask these machines to help you build these things. So the way that this tool works is I wanted a tool that you can plug in the primary keyword because all clients want to rank for this word. That's great. You need to rank for a lot of other things too in order to actually be relevant in this space. And that concept isn't new because of AI. That's a concept that has been around for over a decade. So I guess maybe it's built on AI, just everybody wasn't aware of it. And I just went back and forth with ChatGPT. I did testing, troubleshooting, and I got closer and closer to my vision. I had a field for the primary query, a field for all of the keywords that I got from a tool. And then I select a model and then I hit, you know, do calculate the cosine similarity. The resulting CSV, which you can download is just a clean list. It gets rid of most of the stuff that's completely irrelevant. I didn't write this python code. ChatGPT helped me write this code. I was able to put it up to hugging face and create the hugging face profile and create the space in hugging face. And I'm basically just standing on the shoulder of giants who taught me how to build stuff in Streamlit. I even posted this on my LinkedIn. And then some of those people actually stepped in and said, like, here's how you can make it a little bit better to have that. That tsne, which is T distributed stochastic neighbor embedding, by the way. I wrote that down. A little bit of cheating here. So knowing precisely what you're doing isn't essential. It's good. That's where the learning happens. But basically, in a NutShell, download an IDE cursor, Windsurf, Google AI Studios, you can vibe code with now, I tried it. It's okay. It makes pretty stuff. But you have to go through the process of actually testing it and you have to have a conversation with the chat. This is where you're going to get better at your prompts. This is where you're going to get better at understanding where it's going to mess up. It's going to mess up. And also one. One key tip that I'll share is context pollution is a very serious problem with chatbots. So what this basically means is you go down this rabbit hole and, you know, things get really mucky, and then you think like, okay, let's zoom out and start from the beginning. The problem is it's going to cite all of that stuff that got you to that black box in the first place. So in cursor, you can scroll up, go to where the problem began, and reset that. So that's really, really helpful. You can flush all that junk out, and then you can start anew. So that's one of the pro tips that I'll give to the audience. But ultimately working through this, focusing on what your vision is, really understanding what your problem is, and then just doing the free trials, that's. That's all it takes.
- [24:34]
B
No, I want to dive into, like, both of those aspects because one's interesting, too. It's like how you're, you know, defining the problem that you're wanting then to create this solution for. And I think, you know, there's an entire probably approach to that aspect, but from someone that's like, just wanting to play in, and they're like, all right, I'm in. I'm bought in. It's time to just jump in, play around with this. And I'm going to give a caveat to this, too, just kind of hedge bets. Granted, this is a rapidly evolving space. So if you're listening to this, you know, three months, two months even from now, maybe have a little asterisk on this as the tools can change. But as of today, what recommendations would you have for, like, a tech stack for someone to get into creating some of these, like, tools and solutions on Their own.
- [25:30]
A
I don't know if I would call it a tech stack, but one way to get into this space as an SEO, presumably the audience is largely SEO folks. Start with your Google search console data export from the Performance tab and then like I had mentioned, like, download one of these ides and open it up. Create a sandbox folder. You don't want one of these agents working in your, your main machine that can, that can be fun for you. But download this data and create a project in one of these tools and say some really simple prompt like, I want to explore vector embeddings. Like, that's, that's seriously the starting point. You're going to need an API key for something like, you know, Gemini has a free tier that you can get into, which is pretty generous, and Google has the funding for it. OpenAI, you have to pay a little bit of money and figuring out how to set up APIs if you've never done it before is a little frustrating. It's something you can ask ChatGPT about. And getting this data, bringing the data together in your cursor, Windsurf, whatever, that's where it starts. So just have this conversation. I want to do these things. Give me some suggestions. It'll likely just go like, it goes off the rails. It'll say, cool, it looks like you want to build all these things. Sounds great, I'm going to build all these things. So it'll probably set up a virtual environment for Python. It'll create a requirements. Txt, it'll create a hidden env file for you to plug into your APIs and so you can work through this process. You're going to run into the errors that I had mentioned. But if you're walking through and you're running this and you're saying, let's run it, let's test this and let's, let's explore this space. That's the easiest way to get into vector embeddings. It'll take a while to actually get something that's ultimately useful, I think, for you. But this is the track that everybody should be on so that they can understand what's going on in the space. But like a more academic approach, I think would be like getting started with embeddings on hugging face. Hugging face. Just start a hugging face account. Start there. Obviously, OpenAI is a, as an embedding guide that's pretty useful. And then again, on hugging face, they have these spaces with all these demos and so you have all these resources at your disposal. Obviously you can Google things or find videos on YouTube to try to find those things. Those are always useful. And there's a lot of great people on LinkedIn that you can follow that I can do a call out for like Dan Petrovic or Metahan yesiliert. I'm probably not pronouncing that right, but they, so they regularly riff on like some of the tools that they're building and it's amazing. Like just find something that approaches the problem that you're trying to solve. For me, I wanted to demonstrate how content looked or how content could be better if we were to optimize it with that vector embedding lens. And so that was the framing. So I built, I built some tools that are publicly available if anybody wants to try them on GitHub for clients. And this is basically the process that I go through to help build tools for clients. So you can build them yourself. And there's a lot of off the shelf tools that can help you with SEO, but they're not going to be tailor made to your specific use case. And you're the one that really knows this and the people who are developing the workflows with AI right now. So what's the Jensen quote? He said AI is not taking your job. It's the people who know how to use AI. Those are the ones that are taking your job. So this is how you can get into this space and start learning about it and you can actually be more useful for your organization or just developing things for your clients that have that AI aspect on it that's going to give you the leg up.
- [29:15]
B
Okay, so building from that, I feel like a lot of people too are going to be questioning there's the OpenAI, Gemini or you know, what version of these models have come out. And obviously these are rapidly evolving. So when you're going to create your own tools, how are you thinking about what model might be best for that specific use case?
- [29:41]
A
So picking a model is an important decision. There are, I don't know, a thousand, maybe 100,000 models and they keep releasing one all the time. So like one of the ones that I always use, it's in, in the tool that I built on Hugging Face is the all mini LM L6 V2 or whatever. It's like a base model that I use for everything. It's great for prototyping. Super small, the output's going to be okay. So this is the model that helps you build things and figure out the proof of concept. But then like, so say for instance, you have an international client, multilingual client. You want to use something more like labsa, like language agnostic basically. So this is multi parity tool. This gets into the, you know, amia. It does an okay job on the Asian languages. I haven't had those use cases yet. But there's obviously significantly larger models that you can use. OpenAI has many models. They have the text embedding three small for you know, other small, fast, cheap things. But then they also have the large model that's better for higher fidelity, a lot more fine grained semantic matching. So things like query to content scoring or internal linking similarity, those are things that are really important. Gemini has, there's a whole host of models that you can get with Gemini. The gem embeddings is one that I haven't played enough of with. I know that's been around for a month now. So like basically prehistory at this point. There's other models too that are good at like analyzing SERPs, the keyword coverage against your on page optimization, like multi qa, NP NET based, something like there's a lot of versions there. You know, these are, these are the models that I typically go to. Each model is going to have a slightly different application. And the other thing that's cool is this is a little bit going beyond what we're talking about here is that you can build on other models. Like that's the area that I'm exploring now is you can start off with like a pretty small model that's doing what you need. It's cost efficient with your tokens, the output is looking good and you fine tuned it to you know, this is exactly what I'm looking for. And then you can bring in your own data and build a model on top of that. We're not going to talk about that here but the possibilities are really endless. It's the sentence transformer models that I spend most of my time in and these are all available on Hugging Face. So thinking about the model that you're going to use, it sounds daunting to people who are just getting started in this space. Really it just comes down to like what's your application? And we didn't even talk about like audio, video imaging. There's all those models that are useful there too. Like didn't touch on that at all. Like Google's Nano Banana for instance is great for the selfies that everybody's been producing lately. So model is an important part of the concept. But don't get intimidated. Just start with something small and you know the All Mini LM based model from Sentence Transformers that's going to get you good enough to build a prototype that makes sense.
- [32:45]
B
And it's basically don't get too hung up on any one aspect because the models will keep evolving the tools, everything else. So it's more of have your process, have your approach and then don't be too married to one tool, one language, one whatever.
- [33:04]
A
Yeah, you don't need to be married to any specific thing. Like, like the people who are doing this that are more sophisticated than I am. Like the problems that they run in into is like the Python version that they're using Python is notorious for, you know, it's compatible at like version 13, but then version 13.3 doesn't work for whatever it is you're building. Those are the problems that they're running into. So know that the deeper you get into this, the more good problems that you'll run into. And don't be intimidated about it. Just like don't be intimidated by the code. It shouldn't be a barrier to entry because these IDEs can basically do everything for you. And that's where the learning begins. Is starting to have this relationship with this software to actually produce things that are based on. I have a need, I want to automate this stuff, I want to build something for this specific use case. So yeah, don't get intimidated by all the models. They're just going to keep rolling in. That's what this disruptive phase is all about.
- [34:02]
B
Yeah, it's all part of the game at this point. Now from that though, let's hit on some real life examples, maybe give some inspiration to the listeners here. What can you tell us about kind of like your own journey in creating these and maybe some real life examples that you've developed and ran into?
- [34:22]
A
Yep. So I touched on this a little bit in the earlier part of our conversation, but I had an international client who wanted to optimize specific GEOs. They weren't in English. And so this was a perfect way for me to test like, okay, how do I help them understand that if we optimize their content foundation is SEO that they'll get, they'll become more relevant for what people are searching for. So most people, most companies, most executives want to rank for a keyword that's well and good. I totally get the vanity, but the reality is there's this myriad of queries that people are typing in that are relevant for your space. So I developed that keyword cosine similarity tool on hugging face as a way into this route. So I developed that, building a way to clean up this list of keywords that you can get, which can be endless, and turning them into things that are really semantically relevant to the space. Then we'll get content from competitors and content from the client site and then bring all these queries together. And what I wanted to build was a visual representation. So this gets into the T SNE products that you can pretty easily build on Plotly. Like I look at the, the source of these, these pages that I've developed for various clients over the years. Like, I don't, I don't know exactly what's going on here. I know it's built on Plotly, but I wanted to do an analysis on the title, the description, the headers, some of the list items, some of the paragraph components that were actually part of the actual article content. Do this for the client, do this for the competitor, and do this for the query landscape. This required building an embedding analysis tool. So I pulled in that multilingual model and I worked on basically crunching all of this data. So I scraped that data, I built some code to actually go through and turn this all into vector embeddings and then plot them into the three dimensional space. That was fun. Just like you can drag it around and it moves around, it looks cool. All these dots in 3D space, which is a simpler representation of that multidimensional space. Where this becomes useful is by turning this into the mean embedding value. So you have the queries, you have the client, you have the competitor turn this into a mean value. And the objective, the hypothesis that I had is if we update the client's content and move that mean value closer to the query, the query mean value, then we're going to be more semantically relevant for the queries, not just the query, but the queries that people are searching for. So what I did is I built this, it looks cool. I had some examples to show the team, they could play around it and move that 3D plotly chart. So that was always fun for them. And I also was able to give the code to the engineers because they wanted to do this internally and scale it themselves. It showed the semantic relevance of the content. And so ultimately we proposed some changes. So this is the SEO part, we proposed changes to the content on the site. So some things, some sections we would update based on the query or the supporting queries, things like the headers, improvements on the title. But I think the difference here is that we got rid of some of the content on the page that yes, design wise, made a lot of sense for the user, but actually it took away from the semantic meaning of the page overall. So we demoted those into like H5s as opposed to H2s. And so what this did was after we did the updates, rescraped the page and rebuilt that plotly T sne we saw that there's a difference in the cosine of the cosine value of the old version and the new version. And so by getting closer, my hypothesis is that you'll actually earn more users, qualified users ideally. We launched this and I was pretty surprised, but like the clicks were up 10%. No, sorry, 10x, not 10%, 10x for this segment. So it wasn't a perfect a B test because it was segmented by, you know, a couple of pages in the larger group of pages. But I have ran this test with other clients and have seen similar results. But, but 10x on clicks, pretty great. That's the biggest number. It was like 8x on impressions, 7x on click through rate and then 50% on the average ranking, which I hate the metric average ranking, it's so annoying. And then also after we got rid of all the results past 100 now like, oh, like where's my ranking, my average ranking going? I don't want to talk about that. So I was able to demonstrate a significant increase in users coming to the site for the primary query. The queries that were coming from the tool that were combed through by doing the cosine similarity analysis. And I was able to demonstrate this in a cool visual that was able to get the buy in from the client. And so this is something that I've been repeating with clients and having really good success with. I haven't had any bad cases where it hasn't worked. That's because we're improving the relevance of your site. Of course you want to do this in a way that is actually useful for the user. You don't want to just do keyword stuffing. That's not helpful. But this was something that I was surprised and basically led into. It's true. You can, I'm. I'm not a developer, but I can make tools, I can give tools to engineers and they can replicate this internally. And that was, that was a lot of fun. I love sharing those types of results for clients.
- [40:07]
B
Interesting. And maybe just to dig in a few areas of that. So from that description it sounds like the visualization aspect was strictly used, more of like a stepping stone for buy in and Kind of like stakeholder alignment to then actually conduct the work. I assume you could probably get the same outputs just without the visual of knowing, like, okay, was the cosine like did we improve that? So that was like use that to basically like facilitate the conversation, get the buy in to the action that then you're working towards and then the actions that you put in place. Whereas it's strictly content actions. Or did you also have like internal linking plays that played into it?
- [40:53]
A
Like, yeah, this is expanded into a lot of different areas. So internal linking. This model isn't the best for internal linking, the multilingual model, because you want to think, you want to find things that are semantically relevant in that specific language. But this is expanded into a lot of different areas of what is the right content that should appear on the page. Choosing the right content to link to is absolutely an application that this has turned into. So if you have all the content that is important to your site or content ideas that you're thinking about building, you can start filling in these gaps between the semantic space so you have content that fulfills these needs. Let's get this really tight on fulfilling that need. But there's this whole gap that your competitor is filling and the visual is just, here's what it looks like, it just looks cool. So this is a way to get buy in with everybody on the team. But yeah, you can absolutely just crunch the number number of like okay, the cosine similarity value is 0.7 now and it went up to 0.8. And now we actually know that this is becoming more semantically relevant to the queries that we're targeting. But ultimately you can apply this to many different facets of SEO. It's not just on page optimization. It does get into internal linking, it does get into informing your content calendar or the content that you're actually wanting to produce. It informs what content you actually want to trim as well. That's a concept that's really important. You don't want to over optimize and just have everything on one page. You want to have different parts of your site do different things. You also want to fulfill different parts of the funnel. So you can get into the intent people who are looking for information about your product versus I want to buy your product nearby. And getting into the hyperlocal component of it, it applies through the entire funnel as well.
- [42:44]
B
I really love the visual aspect too. I know, it's like so often it's like we'll have conviction and like, hey, this is, you know, we can back it with evidence and for that. But then sometimes it is needing to win the room over to really bring something to life. So I love that you combined something that was like, both technical in the SEO application, but also addressing the real challenges that we have as practitioners and actually bringing to life some of these activities too.
- [43:17]
A
Yeah, I've heard a story from a colleague of ours that said, he said, I did this great big presentation. It was like 40 slides. I got to slide 17, and all the executives in the room stood up and clapped. I ended the presentation there.
- [43:34]
B
So 17 just never went through the rest of the slides. And you're just like, that's all I had planned.
- [43:39]
A
I mean, if you've, if you've gotten your point across, then, you know, let's not go down the rabbit hole. If you get the buy in. Enough said.
- [43:48]
B
Yeah, you're only, you're only going to work against yourself at that point. So with that, that's going to wrap up this episode of the Voice of Search podcast. Thanks again to Ryland Bacorn from vocaday for joining us. If you'd like to connect with Ryland and you can find a link to his LinkedIn profile in the show notes, be sure to go on over and check out bokeaday.com to check out his work, some of the stuff that he's been talking about here on the show, and all in all, just get in touch with him. If you haven't subscribed yet and you want a daily stream of SEO and content marketing insights in your product feedback, hit that subscribe button on your podcast app or on YouTube and we'll be back in your feed soon. So with that, that's all for today. Look forward to seeing you in the following episode.