Can AI Create a Topical Map? Prompting, Data and Progressive Optimisation with Germans Frolovs

James Dooley: Can an AI tool create a topical map? I think this is one of the most common questions that I get asked with regards to topical maps. And surely you can click a button and get an AI tool to create it. So today I've got Germans Frolovs. Now, Germans, you are a legend when it comes down to trying to set up systems and processes. And I know that you use different LLMs for a lot of different things. So it's going to be an interesting one. Question is, can you, can you use AI to create a topical map? Yes or no?

Germans Frolovs: Yes. But let's dig into this then because I like this. So obviously, you are brilliant when it comes down to different prompting, different LLMs and stuff like that. So how are you using AI to help? I'm presuming it's not just a click of a button and it's creating it. It's assisting you along the way. Is that correct?

James Dooley: Yes, exactly. I wanted to expand on the but, uh, my main kind of, like, message there, it can assist you heavily. Especially in the last two, three months, AI and then the things that you can do with Claude Code and agentic systems is crazy. So, uh, definitely it can assist you to a whole bigger degree than it was before. Maybe if we'd had that discussion eight, 10 months ago, I would say no. But now there is definitely many more possibilities. There are many more possibilities with that. And, uh, assisting from an ideation perspective, but then also once it all comes down to the process, if you know how to do things manually, you essentially need to create a sub-agentic system that will exactly do those things, uh, with the same, uh, mental models and thought processes that you have. But the biggest problem is maybe, uh, explaining that to AI. Because, uh, uh, let's say how Koray taught us about how to create topical maps and so on, not always AI will understand what does that mean. Because some of the concepts are quite abstract. They're not too deterministic. There are no only if-else statements. So probably need to, uh, yeah, train your AI and improve those prompts on the different components of that system to actually get the outputs that you want. I think a big one for this, for me, um, it was... It was something I actually watched on one of your episodes. And, like, I hated... Don't take this the wrong way because I mean it in the nicest possible way. I hated the fact that you used one of the words and it was called, "Obviously, you just do this." And it was like, for you guys it was really obvious the way you set something up, but for me I was like, "That wasn't obvious to me." And what it was, was you was talking about wine and you was... The way you worded it was, um, "Yeah, so it was obvious you needed to go into the meronyms of it." And then when I actually had to search, like, "What is a meronym?" And I needed to, like... Because I didn't really understand the hyponyms, hypernym. I've obviously understood, like, acronyms and synonyms and stuff like that, but he was like, "Oh yeah, you need to find the meronyms of the wine." And I'm like, "What? That's not obvious to me." So I had to search, "What is a meronym?" And then when I went into it, I can now understand. Yes, I understand why the meronyms are important because I understand why, and I understand. But I would never have prompted AI to say, "Can you extract me all the meronyms for premium wine?" And this is where you guys, when you're using artificial intelligence in the way you're prompting it, you prompt it in the correct manner. And people like me, I couldn't use AI to create me a good-quality topical map because I wouldn't know what to prompt it correctly. Does that make sense?

Germans Frolovs: Yeah. Exactly. And, uh, uh, I think, uh, one other thing to understand is that AI knows language. Uh, it understands language. It understands lexical relations. So when you ask AI, "Give me..." Uh, it will understand hypernyms, hyponyms, you name it, frame semantics. It will understand that. You don't necessarily need to understand it yourself, but if you want to employ some, some of these tactics, and especially when it comes to linguistics and language, then the AI can, can be a big help. Uh, but definitely prompting it properly. Uh, a few, few weeks ago, uh, in our company, one of also actually, uh... the persons from the community, Kahuna, he was developing the content brief system. And, uh, the biggest challenge was for a couple of weeks just explaining our SOP to AI to understand. So make sure it understands how it should work. Because from, from what we thought, how we were taught, let's say, from Koray and from, from the course, how our recipe was written, it was very confusing for AI to understand at first.

James Dooley: Yeah. I mean, I mean, for me, just so you understand the previous part there, it wasn't having a go at you. In fact, it was trying to explain that you guys are that intelligent that you just take certain things as being like, "Oh yeah, well you just do this, this and this." And it's like, but that's not obvious to the general public. So anyone, when this... If anyone's watching this and I'm saying, "Can you use AI to create your topical map?" Yes, if you've done all the training of understanding what goes into a good-quality semantic content network and a good-quality topical map. Because me, when I was doing it, the way I was explaining it, I was getting ChatGPT and I was getting Claude, and it was agreeing with me because so many times it just agrees with you. "Yeah, yeah, yeah, yeah. You're right there. Yeah, yeah, yeah. You're right there." And then when I tackled it and said, "Well, no, I've now understood that I need the meronyms for this," which was a... When I started to understand what it was, I'm like, "Hey, I need the hyponyms and the meronyms for this. I don't feel I've covered it enough." When I said that, they went, "Oh yeah, you're right. We need to do..." And it found a whole new set of topics that I was like, I, if I didn't... If I didn't know about that, I wouldn't have found that. And there were so many things that were then interconnected that made my semantic content network so much better. But it was like, I wouldn't... I wouldn't have known about it. Does that make sense to you? Like, it's garbage in, garbage out. You need to understand, knowing what to prompt. But when you do know what to prompt, the ideation part of AI can help you in a massive way.

Germans Frolovs: Absolutely. And this part specifically, like, finding all the, let's say, hyponyms of a hypernym manually, it would be quite challenging to find. And with keyword research tools also, you will not be able to find it. But with AI, it's, it's brilliant. It understands the general, let's say, knowledge, knowledge very well. And it can help you, uh, with that. And then the easy part is just that, obviously, you just connect it back.

James Dooley: Yeah, no, it wasn't. It was just... Like, for me, I was like, it just, it hit home. It, like, it cut me deep because it was like, that wasn't obvious for me. And that shows how, like, I am at semantic. So, but, but back onto it with regards to artificial intelligence, because I'm going to love this. And I think people are going to like the idea that, for ideation and for different things of what can be being done, it can help you along the way. Where does automation help in a topical map, but also where can it fundamentally fail if you just let it rip?

Germans Frolovs: Mhm. Yeah. So it can help a lot with the same exact example that we just discussed. Uh, expansion of the hypernym and hyponym, kind of, like, relations, uh, if you need to expand on those as well. Uh, I use it heavily when it comes to ideation. Just if, if I need a quick idea about the topical map. Again, it's, it's all prompted with a bunch of different examples and, uh, definitions of what I need and how it should be done, uh, on the whole five components of the topical map. Uh, it can help ideate those things. But, of course, you as a, let's say, topical map specialist, creator, expert, whatever, you need to validate all those, all those inputs before you can progress further. Uh, so with ideation, expansion, I would say also to this day now, with, uh, the ability that you can quite easily connect different MCPs, APIs and so on, create MCPs from APIs and so on, uh, you can essentially, uh, automate many other things of the process. Let's say, uh, you can, uh, let it find out the, let's say, tokens or the main exact-match terms that upon which you want to expand on, and find all the variations. Connect it to Ahrefs, get those variations. Connect it to another tool, cluster all of them, get the clusters, uh, back to you. And then maybe reorder those clusters and so on. Find duplicates. Uh, many parts, uh, nowadays. But, uh, I think, uh, as usual, I think, uh, across different stages there must be first a lot of human input and, let's say, the whole business context input. And then throughout the whole process, throughout the whole pipeline, there needs to be there need to be stages, like a stage-gate model, where you need to stop, adjust before it can continue and continue and so on.

James Dooley: Do you know, do you know one of the main nyms that we've started to use, um, with regards to artificial intelligence for a topical map? Antonyms. So, like, we, um... And the reason why we started to do it is because of query fan-out, mainly because of the large language models. They love to show what's good about you, but also what's bad about you, and what's safe about you, but also, like, what concerns there is about you. And, like, when you started to cover the antonyms with regards to certain articles, we connected certain things together. And some, some of the antonyms get, like, insane volume of traffic, which we never really thought... We didn't even consider it previously. And that's where artificial intelligence for us, when we learned... When I started to go, "I don't know what a meronym is, so what other nym do I not know?" And I'm, like, going through all the others. I'm like, "Well, I'm going to integrate this. I'm going to integrate that." And it was, like, something else that came along. It was like, "Oh." And then because of query fan-out, mainly the LLMs and the AI Overviews, it massively helped. Um, yeah. So you're doing it on the article level or on the topical map level?

Germans Frolovs: I did it on both. So what I ended up doing was, I ended up doing, um... What was coming back was, um, my brand name with, like, alternatives and my brand name with competitors. Now, I've done some... On some sites, I've put it on my site. But on some sites, when it came back, I was like, "I don't want this on my own site, but I'm going to do an article on it." So I'm covering the article saying, let's say, like, "FatRank alternatives." And then what I would do is I would say, pretty much, "You don't want to leave the best." You're planning to do the same now into a positive sentiment that when the LLMs are doing the query fan-out, that negative sentiment type, "What are the competitors of FatRank?" or "What alternatives is it to FatRank?" It's basically saying there is none. They're the best. And, and it was feeding back to the other ends and doing great. But H1, H2, H3, Fed number one, Fed number two. And then for content briefs as well. So, like, one of the... We normally put it in the micro, but one of the last H2s, we'll do something about, um, almost like the... Not just doing pros about something, doing the cons and showing one or two of things. Like, we don't work with... We don't, um, we don't work with all businesses within the UK. We're quite selective with who we work with. So we, we touched upon that. And because we spoke about that slight negative sentiment, for some reason, I don't know if they've got some, like, balancing in sentiment analysis. But if you only did positive, positive, positive, when you... The minute you added in an antonym and did touch upon a con, it seemed to help. That's at a page level anyway, not at a topical map. But it was, um, it was interesting.

James Dooley: When, when I started to use AI for the topical maps, we started to go deeper and it was bringing up things that we would never have thought about on the ideation. Um, but let's get back to it. Artificial intelligence. You're saying feed it the data, like Ahrefs. How good is AI now for crunching the data? And have you found that when you're getting it to do certain things, you're going, "Well, actually, because of the search volume of this, this now becomes a more important page within the semantic content network because now you've got the data"? Has that come along and AI has helped you with that, with topical maps?

Germans Frolovs: For sure. Yeah. Nowadays it can do quite sophisticated data analysis, AI, as well, compared to the early days when it was just hallucinating and couldn't really count. Uh, now it's much better. And definitely, if you provide the input, maybe also, uh, uh, add other parameters like the prominence and relevance, uh, plus popularity from the data, uh, you can basically filter out all the, uh, query terms that are, yeah, uh, going to be a part of your topical map or not. So for sure, nowadays, uh, very capable of that. And then if someone does an AI-generated topical map, how can you validate whether it's useful, incomplete or completely wrong?

James Dooley: Uh, quite hard. But I would say doing some manual checks with just taking some of the topics and just checking the SERPs, checking the volumes, checking whether these pages actually deserve a page. I would say that this is probably going to be the case where AI would fail the fastest. It will probably suggest too many pages for which, uh, there's no query. Or maybe topics that are, yeah, basically not existent, or merged with two different topics plus one another. Uh, you, you can kind of, like, sometimes maybe spot these, uh, already headings that are more, like, for click-through rates, maybe optimised from the old days. Uh, those are some of the hints. But, uh, I would necessarily... I essentially need to model the topical map on my own first and then, uh, do some of the checking whether, okay, within this category, do we have enough level of topics there? Does it actually correlate with the SERP? Uh, and whether the query deserves a page. Those would be the main ones, uh, to check. And then a couple of other questions with regards to AI creating a topical map. Is there any specific large language model that you're using, or are you using different ones? Are you using, like, ChatGPT? Are you using, um, Gemini? Are you using Claude? Like, is there any one that you're preferring? Are you trying to just use all of them and seeing what differences come back?

Germans Frolovs: Yeah, mostly Claude. But for just simple ideation and so on, uh, ChatGPT still. Uh, I haven't been testing rigorously. Like, as long as one works good enough, I, I just stick to it.

James Dooley: Yeah. I didn't know whether you might use Gemini because it's Google-owned, but it might have the data and it might come back with something slightly different. And obviously, more data, the better it can be. Something I've always thought about, um, recently with regards to artificial intelligence and the different LLMs for topical maps. I think it would be brilliant for progressive optimisation. So not the initial creation of the topical map. I feel like it needs your brains and everything of what you're doing. But let's say you created me a topical map and I knew it was expertly set up. But obviously, in another episode, we've spoke about, like, the site radius score and there isn't a scoring mechanism where we might go... I won't say we go too broad, but we might open up certain pages that we shouldn't. Or there's certain pages that we should open up and we've not yet opened them up. Once you've got that data with Google Search Console, you could load that into then, "Here's my topical map. Here's all my data. Is there anything that you think we've done wrong or you think that we've missed?" backed up with the data. Because you said it can crunch the data really fast. It's not really that you've made a mistake. It's just new things have opened up. There's new... There's new services that you might be offering. There might be new keywords. There might be much higher search volume for something that we didn't realise. And I feel, for a progressive optimisation topical map, that's where AI, I think, could excel.

Germans Frolovs: Absolutely. And we talked also, just like you said as well, that it's not a one, one, one thing and done. It, it is dynamic and it is changing. Uh, the semantic distances between the query terms are changing. And I recently did a talk about content reconfiguration. And, uh, basically there it was a big, kind of, like... The main message is go to Google Search Console, see, uh, what you, what you rank for, uh, for any given page, and then actually try to reconfigure the content according to, uh, the impressions of the main queries that are, uh, basically having the same meaning of that group. And according to that, uh, one... Uh, yeah, that is the biggest, let's say, take. If you organise the page according to whole search, uh, search demand on a specific, let's say, page, you'll increase the rankings there. But also, it's going to show you where you rank poorly, which is, again, an idea for expansion of your topical map if you haven't considered that page yet. And, uh, maybe definitely, like you're saying, that what we are doing in the beginning with topical maps, with the first initial content briefs, we are just guessing. We are just making an educated guess based on only available data that we have from Semrush, from Ahrefs, for example. But, uh, once we get the Google Search Console data, once we get some validation and some historical data, maybe some initial rankings, then we know a lot more. We have much better data to work with. And then we start working with that. And then every six, nine months, 12 months, it's going to change. So we need to continuously, basically, use that data. And definitely a system like this that would both recommend what is your topical map expansion opportunities, but also how to reconfigure the website page by page depending on the, uh, impressions for that page. It's, uh, it's possible. Something that we need to make.

James Dooley: Crazy. Yeah. I mean, we, we do, uh, something pretty similar for LLM visibility. So we've got an AI visibility tool, and what it does is it comes back with certain prompts that we're missing, that we need to get in for. And it shows up, um, gaps and opportunities. And we get, um, I think it's every three days, we get an alert. And there might be, like, six opport... Six new opportunities that we've not covered. And then we then decide internally, does this go on and part of an existing page, like as a subheading? Or does this deserve its own, kind of, page of what needs to be set up? Or does it go on a third-party corroborative source or, like, on a wasteful domain? Um, and that's, kind of, something that we're working on, uh, day to day. It's, it's, it's added quite a lot of extra work, to be honest with you, for the progressive optimisation team. But the LLM visibility has been through the roof. Like, so it's... I want to get that tool.

Germans Frolovs: Yeah, it's a great model. I'll show it, um, show you next week in The Masterminders. Um, yeah, I think you'll like it. But anyway, anyone who's listening to this, we hope you like the episode about, "Can an AI tool create a topical map?" Make sure you check out all the other episodes where I'm going through with Germans, the topical map versus semantic content networks, the importance of the source context and lots of other episodes here, all related to semantic SEO, including the misconceptions of a topical map and the common errors that's happening within semantic SEO. Germans, it's been an absolute pleasure.

James Dooley: Thanks for having me.

Germans Frolovs: Likewise. Thank you very much.