This is The Search Session, and I’m Gianluca Fiorelli. Today, my guest is Lazarina Stoy, marketing consultant and founder of Women in Marketing in Bulgaria. We talk about the rising importance of entity search and semantics, how personalization and query fan-out are reshaping intent, and why machine learning is more reliable and predictable than powerful—yet often unstable—generative AI.
The discussion centers on:
The rising importance of entity search: why entity search is now central as AI pushes SEOs to understand ranking patterns, semantic structure, and entity relationship mapping.
The limitations of generative AI: why unresolved prompt injection issues limit agentic systems and drive enterprises toward task-specific ML models.
Responsible use of LLMs in SEO: why tools like ChatGPT, Claude, and NotebookLM fit content transformation and workflows, not for tasks that require critical thinking, research, data forecasting, or deep analysis.
Strategic brand positioning in the AI search era: using competitor comparison, persona mapping, and entity–attribute analysis to decide where to appear, who to compare against, and who to convince.
Strategic brand validation: consistent third-party mentions help brands surface in Google’s Knowledge Graph and AI systems—making digital PR essential for surfacing them.
Query fan-out and personalization: how AI-generated follow-ups reshape intent, shifting SEO from keywords to context-aware strategies.
Machine learning education for SEOs: MLforSEO builds confidence through ML theory, real-world use cases, and practical automation tools.
Explaining knowledge graphs to stakeholders: framing them as a way to help AI and search engines better understand, connect, and surface your brand and its key information across traditional and AI-powered search.
We explore how these ideas work in practice and where they can help. Let’s dive in.
Video Chapters
Transcript
Gianluca Fiorelli: Hi, and welcome back to The Search Session. I’m Gianluca Fiorelli, and today we have a fantastic new guest. She's a marketing consultant, trainer, and speaker. She's the founder of ML for SEO, or more precisely, machine learning for SEO.
In fact, that’s her own definition: ML for SEO is a platform to learn practical AI machine learning as a marketer, which I think is a great, great idea. She's also the community lead at Women in Marketing. She lives in Bulgaria, and her name is Lazarina Stoy. How are you?
Lazarina Stoy: Hello, hello! Thank you. It’s great to be here; thank you for having me. I’m doing great!
Gianluca Fiorelli: It’s great to have you! I think we’re going to have a really, really nice conversation, especially because we share so many interests related to linguistics, entity search, and all these things that, all of a sudden, have become super important for SEOs across the world.
Lazarina Stoy: Yes, super trendy!
Gianluca Fiorelli: Yes! We’re lucky to be on the trendy side of things. So, speaking of SEO, how’s SEO been treating you lately, during all of 2025 and the beginning of 2026?
Lazarina Stoy: I think it's a very exciting time to be an SEO, to be honest. So in 2025 and 2026, I've been building the MLforSEO platform, and I've also been working on building a community as well.
Obviously, these are business entities first and foremost, and it's been really fun to invest in SEO as a channel from the get-go, especially in times like these. I think it's super exciting.
So, I think SEO in 2025, and especially in 2026, is going to be the channel for small business owners. And it's becoming even more important for bigger enterprise brands too, especially now, with AI search becoming part of SEO, at least in most departments. I think it’s being considered as such. So yes—exciting, super dynamic, but very exciting.
Gianluca Fiorelli: Yes, super dynamic for sure. It's almost—not daily, but almost weekly—that new features are added to Gemini, to AI Mode, and now the Web Guide is being tested live in the main search, and then the new models of ChatGPT, the battle between all the competitors in the AI search sphere—it's really a frenzy moment, let’s say.
The Resurgence of Entity Search and Semantic SEO
Gianluca Fiorelli: So, before we were off the record, we were joking about how much all the things related to entity search—let’s say, which we might consider as the base of everything we’re optimizing for lately, plus many other things obviously—have all of a sudden become so, so, so fashionable to talk about.
And even if, sincerely, it’s been around for more than 10 years already—starting with the advent of the Knowledge Graph, and, if you want, even with the advent of Schema—entity search is something that should have been put on every SEO’s radar.
What do you think about this sudden importance being given to entity search and all the derivatives? And why do you think SEOs finally decided that Schema isn’t just for rich results, for instance, and that entity search isn’t just some theoretical, foggy definition of things?
Lazarina Stoy: Yes, I’ll start with the first question: why is it suddenly super popular? I think in the past year, SEOs have really been pushed to understand how a new technology works. I’m talking about AI search systems—and when I use that term, I’m thinking of ChatGPT, Perplexity, AI Mode, AI Overviews, and anything with LLMs in it.
SEOs have been pushed to really understand how LLM works. I’m not saying that everyone does understand it, but we’ve definitely been at the forefront of answering questions within an organization. Regardless of whether you're working agency-side, in-house, freelance, or whatever, you’re the person people come to with questions like, “How does this work? How do we appear? What happens?”
So, I think the first wave of discussions around this stuff—like AI search systems in general—didn’t really look into how these systems work. And of course, there was a ton of discussion, most of which wasn’t really correct, about how to appear, what to do, and so on.
And then, once people really understood that you have to look into the patents, understand how the systems work, and understand how they rank, I think three things emerged as really important.
The way to reduce, you know, hallucinations and improve the model’s performance—from the point of view of Google, Perplexity, ChatGPT, and so on—was done through adding the live index (by Google) and also by adding entity relationship mapping, adding Knowledge Graph access, and adding access to things like quick data for validation. And that was like the first thing they did in order to improve the models.
And so, when SEOs started to look into how these things work, one thing came up again and again: entities, semantic structure.
And of course, one thing that’s not really being talked about by a lot of people—Mike King is one of the few who talks about this very frequently and credit to him for that—is how these systems utilize context as well, like personalization.
I think it’s very interesting, because when you look at traditional search systems, they’ve been using personalization aspects—components—to rank results for ages now, at least five to ten years. But no one’s really talked about this.
And, you know, the people who do understand the patterns and understand semantic search, they know that there is a very high component of what you see—both in traditional results and in AI search systems—that is very, very personal to you: what you’ve clicked on before, your user history, where you're located, what languages you speak, blah, blah, blah. All of these personalization aspects.
And so yes, that’s why I think semantic SEO is currently the trend, you know? And I think the next one is going to be: how is personalization actually affecting everything—how we track, what we rank for, how we appear, and so on. But I think we’re not there yet. People are just catching up to entity SEO and semantics.
Gianluca Fiorelli: Well, I remember—many, many years ago—I was still a Moz associate, and I think it’s still there, buried in the Moz blog—a post I wrote about SEO in the Personalization Age, which was about personalization in classic search.
And yes, it’s not just about the classic example of mobile search, IP detection, device detection, and search history. I was also talking about a very old Google patent, which was about search entities—not really the classic entities, but about search entities. And if my memory serves me right, I think it was Bill Slawski who first started talking about them.
Lazarina Stoy: Yes, yes.
Gianluca Fiorelli: And it was really talking about this idea of the web document as an entity, the query as an entity, and so on.
And yes, I think people didn’t really think about personalization—first of all, because they were somehow conditioned by classic keyword tracking, which basically gives you tracking results as if from a, let’s call it, “neutral” condition.
And then, sincerely, because it’s true that there is personalization in classic search but the differences sometimes are not so big as we can see now with AI search. Surely, when it comes to AI search, the personalization is evident. I mean, for what I’m asking, the answer that I’m going to receive is surely going to be different from the one you’re going to receive for the same question.
And that’s probably why people are starting to understand that prompt tracking is not really a good idea—not just because of the cost, but also because… what are you even tracking, really?
Lazarina Stoy: Is traditional tracking also a good idea? I don’t know.
Gianluca Fiorelli: I think that traditional tracking, maybe not now, also because, at least the most advanced tools for tracking—including our host, Advanced Web Ranking—give us the pixel position. Because one thing is being in position one, a classic position one, after—I don’t know—a block of ads, an AI Overview, and maybe an image block. So, substantially, you’re way below the fold.
Another thing is that, at times, I still happen to find the classic 10 blue links—search results with no AI Overview, just the 10 blue links. So that’s a big difference.
But maybe because the differences between a personalized SERP and a neutral SERP were not so big—or not so evident, at least. That was more just a kind of visualization of, let’s say, a visibility score of a system, for instance.
So that’s maybe why that kind of tracking—not really for ranking itself, but more like a screenshot of where I’m visible, and especially in what kind of SERP feature I’m visible—that is maybe the most important function of rank trackers right now.
But for prompts? I don’t know, I mean...
Lazarina Stoy: Yes, and also synthetic queries as well, because right now we have a ton of AI search platforms generating query funnels, synthetic queries… So the concept of search volume… search volume by whom? Is it by your real users? Is it by agents?
Gianluca Fiorelli: Maybe the smartest tool providers could say, “Oh, generating volume is not really important, because in the future your website might only be visited by bots.” But we’re still not there in the agentic search as we think it should be.
Machine Learning vs. Generative AI: Choosing the Right Tool
Lazarina Stoy: They also recently came out to say that agentic search and the idea of AI agents and AI browsers are flawed—because there’s no way to stop, at least at the present time, the AI prompt injection issues. So it basically has a fundamental operational issue that prevents it from being scaled.
Like, when Shopping started—or Atlas, I think it was, by ChatGPT—where you had this like looking at your browser and stuff, there were a ton of reports and they said that, in order to generate a response, the chat within Atlas was looking at—for instance—shared Google Drives, where someone had given permission to one file, and then the system would go into the shared drive of another person to look at files and generate a response, and all of these different things.
So, yes, I’m a little bit hesitant to say that this will be the future.
Gianluca Fiorelli: Me too. Sincerely, I tested them, but I haven’t really transitioned to any of these agentic browsers as my default browser.
Lazarina Stoy: Yes.
Gianluca Fiorelli: I also read something just a few days ago, and I think it was from the guys from Gemini themselves, who were basically implying that the prompt injection issue will probably never be totally solved.
Lazarina Stoy: Yes, it was like the team at OpenAI. They said that at the moment, they can’t solve it. So of course, they’re still raising billions and saying, “Yes, this is the future, everyone,” but yes, we have to look at it through a critical lens.
I think there’s a ton of stuff in machine learning and AI… I mean, AI, of course, I’m saying that in quotations because everything is AI right now. But the biggest organizations, like enterprise companies, are adopting AI, but not generative AI, because it’s still so new and so uncertain, and it’s very difficult to scale with control and with predictability—like, predictable outcomes every time. And what they’re doing instead is pausing AI—generative AI—investments, and pivoting to classic machine learning as well.
So, for those who don’t know the difference: machine learning sits as a subset of AI, and generative AI is also a subset of AI—but it includes a ton of other things, right?
In classical machine learning, the main difference is that you train a model, and that model is trained on optimal data points. It has one singular task—or maybe two tasks, whatever—but it’s very narrow. Its main aim is to provide a prediction based on whatever the task is.
So, let’s say, for instance, you are trying to classify whether a page is a news article or an opinion piece—that’s a very simple example. That would be one ML model.
And if we look at LLMs, for example, they’re kind of like, you know, creative machines. So when we look at the data that we have—we have image-based data, we have numeric data, we have video data, we have text data, of course, tons of it.
And so, what we want to do is try to build very narrow models that do the job that we do and then orchestrate them to work in tandem to automate small parts of our role so we can be more efficient when we do things.
Let’s say, for instance, a content audit or a technical audit. But we’re still in charge, and we’re, you know, fully in control of the output, as opposed to just saying, “ChatGPT, audit this website for me.” Then what it does, it’s not what it's supposed to be doing, because that’s just not what it has been trained to do.
Gianluca Fiorelli: So, in this context, can’t we consider that something which has all of a sudden popped up in these last two weeks, which is Context Graph, is going in this very direction?
And considering this concept of the Context Graph—which is something that maybe, and I talked with Andrea Volpini about this—I was asking him, “Isn’t this Context Graph what I was substantially calling a proprietary knowledge graph?”
So, a knowledge graph based on all your enterprise documentation, closed into your enterprise communication. And it was coming to my mind that maybe Google is the smartest guy in the room for having created NotebookLM, where you can basically put all your documentation.
Lazarina Stoy: Yes.
Gianluca Fiorelli: Even with a Pro subscription like mine, I think you can put in, like, 300 documents, which is immense for the kind of things I do, let’s say, with a client.
And I think if you combine that with an LLM for certain kinds of things—like, for instance, sheet analysis or something like this, in order to find patterns, etcetera... but using everything inside the notebook—what do you think? Am I totally wrong, or am I in the right direction?
Lazarina Stoy: No, I think you're in a great direction. So, I think it's important for me to say this, because I just want to make it clear so people don’t misunderstand me, I am 100% a proponent of LLMs. Like chatbots, ChatGPT, Gemini, NotebookLM, I love them. I use them on a daily basis.
And I think what you’re doing is, of course, a great use case—because you're utilizing your own data in order to, you know, transform it into another format, right? You're taking text, visuals, and so on, you're transforming it into text, but it's based on what you have created.
My pet peeve with the use of LLMs is when we’re trying to replace thinking or research or planning, which are steps that we should own. And I also think that there are some use cases where people are trying to—let's say, for instance—do entity analysis with LLMs, which they can do, but they’re just not the best at this. And you also don’t get the full suite of data that you would get if you use an entity extraction API.
Or let’s say, for instance, you’re doing content outlining with an LLM. If that’s the only thing you do—without you researching the topic or pulling in some user research or researching entities and attributes—you’re just not doing the most that you can do. And everyone can replicate it.
So, what I would recommend using LLMs for are exactly those kinds of transformations. So you’re taking a PDF document, and you want to transform it into a social media post or a blog post; that’s fine. You still have to do editing, of course, as we all do, but you’re taking content from one format and transforming it into another.
Now, we are seeing LLMs transforming, let’s say, images to video or text to image, and so on. These types of content transformations work fantastically. What you shouldn’t use an LLM for would be text classification or forecasting.
I’ve seen this, actually; trying to forecast data or doing data analysis with an LLM is not really very reliable, especially at scale. I’m not saying here mostly with what you chat with Claude or ChatGPT; it can work, let’s say, in an isolated example. But if you're running the API and trying to analyze your Search Console data, that’s just not a good use case, because that’s not what it’s built to do.
"Vibe Coding" and Essential Skills for the 2026 SEO
Lazarina Stoy: Also, coding. I think especially if small business owners and, in general, beginner or intermediate SEOs learn how to code with Claude, ChatGPT, and Gemini, that will be the key differentiator between them waiting for something to happen—like approval, or proof-of-concept building, or whatever it is. You can now actually resolve most of the issues on a small to medium-sized website, on WordPress or on Wix or whatever, with Claude, let’s say, or ChatGPT.
So, you need to learn how to prompt correctly—and what people call vibe coding. I think that is the skill to acquire in 2026 as an SEO, as someone who works with technical instruments, that is websites.
So even if you’re a content strategist, if part of your role is to strategize posts for social media, for YouTube Shorts, for whatever else, you have to know what tools you can use to transform posts from one format to another. And also what tools you can use to build automations, because no one is going to pay the salary of someone, an employee, who doesn’t know automation in 2026.
At least the basics, you know, how do you use Claude in your day-to-day work? How do you use ChatGPT?
And building a system, building SOPs, like documenting things—those are things that Claude, ChatGPT, and Gemini can be insanely valuable for. And building internal knowledge systems as well. I can go on and on about this.
Gianluca Fiorelli: Yes, everything is interesting. So, returning to entity analysis, entity extraction, and so on, never forget there's something called Named Entity Recognition (NER ) there, which is very, very important to do as the basis of everything, even before doing an embedding analysis, for instance, in that case.
So you have a real mapping—for instance, doing a competitive NER entity recognition between your content and the content of your competitors in order to understand what you both are covering, what you are not covering, and what you are covering that they are not, etc.
And from there, starting to say, “Okay, let’s go with embedding analysis and cosine similarity” to see the very specific details of why your competitors are more visible than you, or what you have to defend your visibility. Which is something, for me, very important—and many people forget it: how to defend your already existing visibility and not throw it away trying to be more visible for other things.
Buyer Personas in the AI Era
Gianluca Fiorelli: And I want to go back to the personalization thing. I think that personalization is, somehow, a strange consequence of this AI search frenzy that made SEO rediscover something that already existed—something very classic in marketing—which is the buyer persona.
Because maybe with buyer personas—while we cannot create one piece of content for each individual that might be searching for something through an LLM—we can create cores of personas that we want to target. So we can specifically try to answer them, to these personas, and create “personalized” content or chunks for these different personas to be visible to all of them, hopefully.
So, do you think that this is maybe not only the correct way to think in terms of SEO for AI search but also a good way to finally make SEO, as a discipline, and talk the same language as the CMO, the marketing leader of the company, the clients, etc., etc., etc.?
Lazarina Stoy: Yes, absolutely, I love this question. It’ll be kind of a long response, but I’ll try to keep it under five minutes.
So, the first thing is, whenever we talk about personas, I think it has been a discussion, at least in my career, for a very long time. But I know that not every SEO has been doing this. And now, yes, I think they should be focusing on this more and more, especially if they want to be relevant in AI search systems, but also in traditional ones as well.
So, one concept I want to credit Beatrice Gamba for really nicely developing in one of the courses we have on the ML for SEO Academy for AI search and entity SEO in the AI search era is the concept of co-citation.
So, actually building credibility by co-occurring with other brands that you can compare with. Before you can compare with any brand, like a competitor, based on a product that you offer, based on a service, or based on your brand attributes—we are already surfacing what I want to talk about—you have your entities, like your brand identities, the people in your organization, your product, your services, and then you have the attributes of those entities.
So you have, for instance, for a product: the price, how it’s delivered, what color it is, and so on, depending on what product it is. So you would have other competitor brands that you will appear in conversations with when people decide whether to choose a product or service from your brand versus one from a competitor.
So, what Beatrice talks about in this course is essentially strategically placing yourself in the conversation—whether through social media posts, through forum posts, through blog posts, whatever kind of content you're producing—strategically placing yourself in proximity to other brands and pivoting the discussion based on the attributes of your entities that make you look most relevant to your audience.
So that doesn’t mean saying, “I’m the best brand,” regardless, you know, for everything. If you are, that’s great, but most of the time, you will be the best brand for a certain persona, right?
So this is where we kind of circle around to the persona discussion. I think it’s very important nowadays not only to do this kind of mapping: Who are your competitors? What kind of entities and attributes do they serve? What kind of entities and attributes do you serve?
And I’ll give an example. For instance, MLforSEO as an academy: we do courses. We're not the only academy to do this. Let’s say, for comparison, Traffic Think Tank is another academy that does this. What differentiates us is the topics that we cover; we only focus on machine learning, marketing, and SEO. And also, what differentiates us is the price. So these are some attributes that we cover as a brand.
So in the context of, you know, the exercise that I’m explaining, it would make sense for us to write and to say, “Yes, we know that this competitor exists. We know that they do a similar thing to us, but we only focus on these topics, whereas they focus on a full suite, and they have more experience than us, and they have more authors than us.”
And it’s not a problem, especially in the AI search era, to say this. To say that, for instance, you are a newer brand or, you know, kind of talk about your flaws is what I’m trying to say. Because your flaws might be your competitive advantages for the right audience.
So, it’s important for you to do the user persona mapping exercise, to do the entity-attribute-variable exercise, and to also do your competitor analysis, because then you have the three pieces of data that you need in order to actually execute the strategy: Where do I need to appear in the conversation? Who do I need to compare myself against? And who do I need to convince?
Basically, those are the three questions that you need to answer as a brand in every discussion that you do online, whether that’s on social media, on YouTube, in webinars, in blog posts, or whatever—you name it.
And it can go as niche as you want, and that will be dependent on how niche your user personas are. And I think that will be a very interesting exercise going forward. I think we’ll see a lot more interesting, creative ways to actually map out—based on prompts, based on user queries—to actually extract data for personas.
But again, this is something that only Google has at the moment, and of course, ChatGPT, and so on.
Gianluca Fiorelli: Yes, but it’s something that can eventually be retrieved. And I was thinking about this is a kind of persona that actually, Rand Fishkin built an entire new tool around, which is SparkToro, and that is the user persona.
So, the user persona is not really just the one that you target as a customer or potential customer, but the one that you target because they can influence your potential customers.
Digital PR and Fact-Checking in the Knowledge Graph
Gianluca Fiorelli: So this is the kind of information that you need in order to understand: Where do I have to be visible? With whom do I have to talk? With whom would it be great to create a co-marketing occasion? Where should I have a mention, and where should I not? Etc., etc., etc. And eventually, even creating digital PR campaign and stunts.
Because we know that one of the many legs for being visible, especially when it comes to LLMs, is the differentiation between training data and fresh data. We also want to be visible.
And LLMs do not use fresh data, or we are one of the trusted websites already included in the training data, so that we can eventually be mentioned. But if we want to be mentioned, we must actually be mentioned on these kinds of websites—social media, forum websites, etc.
Lazarina Stoy: And I think—also, one small insert here as well—and this actually goes for both how Google adds facts to its Knowledge Graph and how LLMs decide what is a trustworthy piece of information that they can include in their response.
And that is actually where third-party mentions become important, because they need statistical amounts of websites or sources of data to confirm something as fact before they present it.
And that goes for how Google discovers new information to add to the Knowledge Graph. They do tuples and inverse tuples. So they kind of ask a question, and if there are enough responses on the web to confirm this as a fact, they would assess it and add it to the Knowledge Graph.
And also, LLMs would take the information from your website. Let’s say, for instance, X person is the founder of this company. And then they would actually search other websites to see: Is this information something we can confirm?
I’m not going to go into the discussion of how this can be hacked, because, of course, we know it can be. You know, there’s plenty of experiments in, like, black hat SEO. I’m not going to go into that. It is a flawed system, but it’s the best system that they’ve come up with for the time being. But that’s essentially what you need. You need the third-party mentions. You need the mentions on social media.
You need to have authoritative sources essentially saying—and re-emphasizing—the same thing. And “the same thing,” in this context, should be the most important facts related to your brand, your products, your competitive advantages, and all of that stuff.
So yes, you do need digital PR—but strategic, not just like, “X company has been mentioned on CNN.” So what, you know? I think that is where SEO and digital PR can be a lot more intertwined in the future, to actually make those mentions really strategic.
Breaking Intent: Understanding "Query Fan-Out"
Gianluca Fiorelli: Lazarina, earlier, we were talking about the query fan-out, and you wrote up a really great article for IPullRank about it. I think it was focused on AI Overviews, right? The AI Overview query fan-out. And you were saying that, substantially, query fan-out is breaking the intent. What do you really mean by that, breaking intent? Because we’ve all built our SEO strategies over the last few years around search intent, user intent, etc. What do you mean exactly?
Lazarina Stoy: The article actually covered all of the systems. Not just Google does query fan-out, but also ChatGPT and Perplexity—all of them. So, the article covered all of them and how they work. Some have been more transparent about their algorithms, others more secretive, but they all do fan-out in some way.
The very interesting thing is the way query fan-out works: whenever you enter a query—let’s say, for instance, “best running shoes for marathons” the fan-out can actually take you further away from the core intent.
For example, they might do a comparison query, or they might do something like “marathon training tips,” which is, of course, really different from your original query.
So, the way that we should approach search intent in the future, in my opinion at least, should be different depending on whether the query is a user-generated query versus a synthetic query. Now, making the distinction between the two is going to be a lot more difficult, and how we distinguish between synthetic queries and user-generated queries. But synthetic queries are meant to expand the user's horizons because that’s essentially the aim of systems like Chatbot, AI Mode, ChatGPT—to introduce the user to a new topic that’s related to their interests, so they can keep engaging with the chat.
Gianluca Fiorelli: Yes, the difference is the Web Guide. Web Guide is not meant to do that. It's vertical. The query fan-out there is very, very vertical on the topic. It’s going like this: the main topic—which is the topic intercepted or implied by the query—and then segmented into subtopics, and eventually, somewhat similar to the classic web search, because you can still find mixed intent.
So, for instance, let’s say the query is about a vacation in Canada. You can see kind of an informational cluster topics like guides, things to do, things to see, but also tours, best resorts, best hotels for this, etc.
So it can be mixed, but it’s very different. And maybe that’s where we can find the difference, because we know that Web Guide is still Gemini-based and query fan-out, but treated differently.
And I would love to see something like the QForia, which we already have for AI Overviews and AI Mode, to also add something for the Web Guide. That would be great, because it could give us the complete picture.
Lazarina Stoy: Yes, I don’t know. I think it’s interesting how the different mechanics of how queries are expanded will affect how we do keyword research, or at least content planning in the future.
For those who don’t know, in the patent for AI Mode, there are eight different types of queries that query fan-out generates. When I say types, I mean—for instance—they would do entity-based query fan-outs, they might do rephrasing, and they might do expansions. So again, each of these different types would kind of have its own—let’s say—category of how far it can advance your query away from the original intent and the original query that you had.
It’s also really important to bring up again the point of contextualization and personalization of the query fan-outs. So let’s say, for instance—and this is kind of a sneak peek of an article I’m working on for iPullRank again—but imagine you have someone who has been searching for “best shoes for marathon runners,” and then they have a chat. And then, two weeks later, they search for “Why does my foot hurt?”
The AI chat—Gemini, ChatGPT, or whatever—will actually pull the other relevant chat from its memory. And when they build the fan-outs and generate a response, they’ll incorporate the information that this user has been training for a marathon.
And they might—even if they have information, let’s say via ChatGPT’s agent buying certain shoes for them—they might contextualize this information even further in the response to say, “Your feet might be hurting because you bought these specific shoes and have been training based on this plan.”
I’m just giving a very generic example here—it’s not super detailed—but it does include that kind of personalized context in the response.
So, the response that this user is going to get is going to be very personalized, based on the plan that ChatGPT previously recommended, the shoes that they bought, and how long they’ve been training for—maybe the system even has information that the user has flat feet because they discussed it somewhere, in some chat ages ago.
And it’s super, super interesting to think about how this actually changes the nature of search. Because if you can imagine that Google, for instance, will have all of the knowledge about you as a user—they already do have a ton of this, of course—but also information that is, you know, implicit, like things they don’t fully know or that haven’t been explicitly stated in the query.
Right now, we’re seeing more of that implicit information being integrated into chat searching, like agent behavior online when they’re searching, and into the responses they give.
So users are not—I think—very in tune with this, not aware of it. But that’s why I think most people are now searching on ChatGPT and changing their search behavior, because it feels so much more personal. Like, “Oh my god, they really know what it is.” Yes, they know, because they have the data on you.
Gianluca Fiorelli: They should have been trained on everything that’s social media, because social media actually knows you in the same way.
Lazarina Stoy: Yes, yes. But of course, they’re sharing this data, and they’re collaborating.
Gianluca Fiorelli: I think there’s another reason why people are adopting it, apart from the fact that it is, literally, more natural to approach. To have a conversation instead of typing. Even if we’ve been trained to do it for 20 years—typing something like “hotel Madrid” instead of “hotel in Madrid”—we tend to write in a sort of strange language.
But I recently conducted a semiotic analysis of interfaces, and they’ve been designed this way. They always talk in the first person. They never talk in third person or in an authorial tone. The default mode—even if there are differences between models—is always slightly friendly, at least, not very approachable.
There are a few things that differentiate them. For instance, the fact that Perplexity puts the sources before the answer—that, somehow, is reassuring. So you instantly think: “Okay, it’s using sources, so it must be true.”
The fact that others are inlining the sources—at least mentioning one by name—helps too. They’re really using all the signs in order to be credible. And that’s why people trust and believe them, even if—substantially—they are hallucinating things.
Gemini is even worse in this sense. It’s good, but even from the use of the spark icon, like magic—when you see an answer—it feels like something that it found. But actually, this answer is an interpretation of something found—like, for instance, on a Reddit thread—and rewritten just because it's a probabilistic sequence of words.
But this is sometimes very hard to make people understand—normal people who are not familiar with all the mechanics of these systems or how AI search works.
Machine Learning Education for SEOs
Gianluca Fiorelli: And let’s talk about your mentoring, tutor, or teacher role in MLforSEO. What kind of students do you have?
Lazarina Stoy: Yes, so we recently crossed 100 students in our academy, which I’m very excited about.
We only have three courses at the moment but we will have a ton more. We’re working in the background on courses for Google Cloud, machine learning for marketers, and then Gemini for marketers as well. And yes, a ton more, so stay tuned!
What I do is essentially plan the program—what I think is most relevant for SEOs, and also for marketers as well. Because the way that we define MLforSEO is that SEO is a function, like a search function, that exists on different platforms. So, we’ll eventually have courses for YouTube, for TikTok, for Amazon—how to actually optimize the search function for those platforms using machine learning.
And essentially, at the moment, I have two courses that I created on the academy. And then, Beatrice Gamba from WordLift recently launched a course as well.
And the people that we have as students vary a lot—from whether they’re enterprise organizations, in-house SEOs, or freelancers that are looking to understand machine learning and how they can streamline their operations.
We also work a lot with agencies. And we also do training for agencies that have, let’s say for instance, bought team access for our courses, for their team. And then we can take some of the concepts from the courses and apply them to specific scenarios that the agency team has, or that they’ve been looking to resolve with the help of automation. And essentially putting the concepts that we teach in practice for the specific use cases that the agencies have in their internal work and pipeline. So, it’s a different mix.
One thing, a characteristic that defines all of them, is that they’re all hungry to learn more about machine learning and AI, not through the lens of ChatGPT or generative AI, but actually to understand machine learning as a discipline.
To understand, for instance, what the difference is between clustering and classification? What is the definition of generative AI in the context of machine learning and in the concept of AI? What are the limitations of LLMs? And things like that.
So we do go into a lot of the theoretical aspects—let’s say, concepts like entity analysis, named entity recognition, EAV models, and all of that stuff, and knowledge graphs, ontologies. We do a lot of theory in our courses and in the materials that we’ll launch as part of the paid academy that we’ll launch this year.
When we are creating resources, we want to start with the theory so people understand the concepts and terminology. They feel confident in explaining this to stakeholders, because I think that’s the first step when you’re working with machine learning. You need to be confident in the results that you’re presenting and how you got to these results.
And then, of course, we have a ton of practice and a ton of resources that we share, like coding scripts, notebooks, and so on, that actually help people to get the first step in and get to the dopamine and the results from implementing machine learning. And once they do that, they’re hooked. They want to automate everything and build their tools and all of that. So it’s very fun to see.
We also have a community on Slack, which is free to join for anyone listening—which is, again, another place to share projects and tools and what you’re working on, and news, and so on.
I’m sorry for taking so long.
Gianluca Fiorelli: No, no, no, it’s fine. It’s a wonderful project, and not everybody’s been born for learning all these things. One is me. Maybe because I have so much—I don’t want to say “so much knowledge,” but I know these things from the classic linguistic background of my studies and so on. And I have a sort of idiosyncrasy with learning new languages, which is substantially the base if you want to do machine learning. You have to, substantially, learn a language, which is Python, and so on.
Lazarina Stoy: So, just a caveat, the barrier for entry now is the lowest it has ever been. Especially, as I mentioned, they can do the coding for you, so you don’t even need to learn Python.
Gianluca Fiorelli: Yes, I know! I’ll give it a try in this sense. Sometimes I do my experiments. Some of them I’m quite proud of, because I can use them also for my work. But I want to…
Lazarina Stoy: Yes, that’s the best thing to do, I think.
Gianluca Fiorelli: Yes, yes. Break things, as we always did.
Lazarina Stoy: Yes, exactly. Exactly.
Explaining Knowledge Graphs to Stakeholders
Gianluca Fiorelli: And one thing, I want to talk, even if briefly, about the knowledge graph. Because usually when we think of a knowledge graph, people always think of the classic Google Knowledge Graph—so, Knowledge Panel—which is just like a very rich result. It’s just the tip of the iceberg of a knowledge graph.
But a knowledge graph is something more. It’s a really structured vision of everything related, for instance, to a brand, or to a person, or to a topic.
Lazarina Stoy: Yes.
Gianluca Fiorelli: So, thinking of agent search—as, for instance, we were talking about before—the knowledge graph is maybe, as a concept, the best thing to use in order to give all the information you want to give to a machine in a structured way. And Schema is just part of it.
Lazarina Stoy: Yes, absolutely.
Gianluca Fiorelli: But the question—the one million dollar question—is how can you explain this to a stakeholder who doesn’t know anything about the knowledge graph? About even, maybe, ontology? If you cite Umberto Eco, let’s say, they think of The Name of the Rose.
Lazarina Stoy: They think in money terms, yes. It’s a very interesting question; you’re kind of putting me on the spot here. I don’t know why, but I thought you were going to ask me, “How do you explain this to students?”
Gianluca Fiorelli: No, well, somehow the stakeholders are “students.”
Lazarina Stoy: Well, yes. So, I think the easiest way to think about it is that you have to tell the story that this is how the machine learns about you. Yes, machines—bots and so on—are getting better at understanding unstructured text, but they’re not fully there yet. They still don’t capture the nuance, and they don’t capture the relationships.
So structured data and entity relationship mapping—done through a knowledge graph—is actually how you help communicate better with the bots visiting your website. And that applies to traditional search and also to AI search. If you want to increase your visibility, you have to increase the machine’s understanding, in the way they receive the world.
One big pet peeve of mine is that, in the recent discussions around AI search optimization, I’ve seen people claim that structured data is no longer important—especially the more nuanced, niche structured data types. I think this is super false. This couldn't be further from the truth, because you need to actually pinpoint the relationships between the entities. And the way that you do that is the very detailed schema.
I think the misconception comes from the fact that LLMs don't use or parse structured data during the training process. But they do that in the live system. So, AI Mode uses structured data. It reads it, it comprehends it. Traditional search—Google Search and so on—of course, uses that. So, you do need that.
And building your own knowledge graph is, I think, super, super valuable. Again, I'm mentioning her two or three times in this episode—but shout out to Beatrice Gamba and the entire WordLift team, by the way. They are pros at semantic SEO.
And she has these wonderful resources about entity relationship mapping and building your own knowledge graph. And I think it's very interesting for people who are absolute beginners to really grasp this concept.
Step one is to identify the entities that you are known for, that are important, that you want to rank for, and that define you as a brand in the online space. And then to map them. Like, how do they relate to competitors? How do they relate to one another? There’s a person, so who is this person? Are they a founder? Are they an author? What kind of content do they produce? Why is this content important? Have they created a product? How is this product priced? Blah, blah, blah.
Gianluca Fiorelli: Yes, totally. And I think that the best example is how Google usually says, "Okay, for this article, these are the types that we really need, and these are optional and good to have."
And if you sum them, they’re not all. But there is one type where Google says, “I don't tell you anything, but the more things you can tell us, the better.” Which is Organization.
Lazarina Stoy: Yes.
Gianluca Fiorelli: And why Organization? Because it’s the center. Because you put the Organization with an @id in every type, in every schema, and it's your passport.
Lazarina Stoy: Also, for all of them. Let’s say you are organizing events, you have courses, you have products, you have people, especially for people. You want to capitalize on their personal brands that exist, let’s say on LinkedIn or whatever. You want to capitalize on that. You want to actually showcase your organization to the machines.
And you know, it’s kind of like — don’t optimize for bots, right? But you’re not. You’re actually opening a visibility layer so that your brand can be surfaced to users better, right? So the end goal is to be seen. Be seen by the users.
Gianluca Fiorelli: I call it “translating to machines.”
Lazarina Stoy: Exactly. Yes, yes, yes. Absolutely. Very important point. All of this text, all of this information that you put in the structured data, should exist on your page. It shouldn’t be like cloaking. Like, you have super massive structured data, and then your page doesn’t say anything. It shouldn’t be like that. But once you have it on the page, just translate it to structured data.
This is, of course, so important. And to your point as well, I think a lot of people miss out on the linking part. As you said, with @id, you can link the organization to all of the types. A lot of people miss that. They create structured data in silos.
Gianluca Fiorelli: Yes.
Lazarina Stoy: Because, you know, that’s how the agency model works. "This quarter we’re doing this type of project; we’re doing structured data next month.” Blah, blah, blah.
Gianluca Fiorelli: I think that Google is also guilty of this because of rich results. When it presents in the structural data section of the dev subdomain, it says, “For having the product rich result, use this.” But it’s not telling you that you also have to say that this product is part of a webpage, which is part of a website, and to use the @id to connect all these things.
Lazarina Stoy: Yes.
Gianluca Fiorelli: Google is not telling this.
Lazarina Stoy: They should. They absolutely should.
Gianluca Fiorelli: Yes.
Lazarina Stoy: Google is not saying a lot of things, as we know. But of course, you should.
I think something that can move a beginner to an advanced level is understanding that the linking part is one of the key components to actually making it click, you know? So yes, thanks for bringing that up. I think it's an important thing to hone in on.
Empowering the Community: Women in Marketing Bulgaria
Gianluca Fiorelli: So, let's talk about you so people can know Lazarina Stoy a little more, not only about the machine learning side of things and SEO. So I wanted to ask you, what pushed you to create and be the community leader of Women in Marketing in Bulgaria?
Lazarina Stoy: Ooh, yes. So, I love Areej AbuAli, the founder of Women in Tech SEO. I’ve been part of that community since literally the first steps I took in SEO. And I always loved the welcoming vibes and how nurturing it is for women supporting women. It’s very much an environment of giving first. And then, of course, it's a community, so you receive as well, but you have to give more than you receive. I think that's a very core principle.
Then I started going to BrightonSEO, and I always thought that this was a fantastic environment for learning and for networking. The people are so friendly, welcoming, and so on. I always developed my career abroad, and even though for the past five years, I’ve been living in Bulgaria, before that, I used to live in the UK.
And the first conference that I attended in Sofia, in Bulgaria, and in my home country in general, the first event that I attended as a professional, I really had a lot of feedback from women, and also, I’d say, a little bit more negative interactions. Something that I would typically associate with how I’ve heard people talk about the SEO industry 15 or 20 years ago in the West.
I kind of saw that kind of negative interaction in my hometown. So I really wanted to do a pulse check and see how women are actually connecting in Bulgaria. How are they finding peers to talk with about different topics? How are they learning from one another? Are there women-only communities?
And of course, there are some. But I found that there was a really big lack of a marketing-focused community for women. And once we launched—initially, it was a slower experience—but one thing I wanted to implement from the get-go was different types of events: networking dinners, mini conferences, webinars, blah, blah, blah, all the different things.
And I saw that a lot of us were actually very similar in the way that we would work, in our ambitions, and very similar in the challenges we’ve faced as well, regardless of whether you had a career focused in Bulgaria or a career focused abroad.
And yes, it’s been such a great journey. It's been kind of slow, to be honest, because of me. I am the problem! I had a bigger growth plan, but in the first year, when we launched, we launched, and then three months later, I found out I was expecting my first child. And then this year, of course, I gave birth. So we only came back, and we only had six months each of the years, basically 2024 and 2025. So we’ve essentially been running for two years, but we’ve only been active for one of those two.
So 2026 will be the first year where we have a full-blown event program, with multiple events happening every month. And we recently crossed 300 members in our Slack community—which doesn’t seem like a lot—but almost everyone from those 300 people has been to an in-person event and has actually seen how we do things.
And the feedback that we’ve received is really positive. I’m very, very happy because this has been one of the most rewarding initiatives, and I have met absolutely incredible people as a result of it.
And I think, again, just to emphasize this: different countries are different in the way they approach in-person events, the way that they approach networking, and the way they approach, you know, equality and equity between genders. And I think it’s important to kind of push the boundary when it comes to, you know, what is the standard that we should be focusing on. So I’m very happy to be part of this initiative in Bulgaria.
Gianluca Fiorelli: Cool. Congratulations, because I think that these kinds of initiatives are very welcome and very needed—besides the gender discourse or things like this—for the growth of the entire community.
Lazarina Stoy: Yes.
Gianluca Fiorelli: They are really needed and welcome. So, I mean, it's one hour and six minutes we've been talking.
Lazarina Stoy: We can go. Everyone listening, how long should we go?
Gianluca Fiorelli: Well, we can ask an LLM to summarize our video.
Lazarina Stoy: Yes, TL;DR.
Gianluca Fiorelli: So let’s stop it here, but let’s promise ourselves to meet up again. Maybe, why not, we can ask Beatrice Gamba to join us and have a good, very nerdy conversation about ontologies, entities, and semantics.
Lazarina Stoy: Yes! And in that conversation, I will take a sideline, and I will be observing the knowledge from you two.
Gianluca Fiorelli: No! But I think that you are needed, because you could be the one who can ask us, “Let’s do this. Everything you’re saying, let’s put it into practice.” Because of your knowledge in everything related to machine learning.
Lazarina Stoy: Yes. It’s been really great. And yes, let’s absolutely invite Beatrice. I didn’t know that’s how it’s pronounced in Italian. I love it!
Gianluca Fiorelli: Yes, Beatrice. It’s a very nice name. We have our national poet, Dante Alighieri, and Beatrice was the name of his loved one.
Lazarina Stoy: Oh, love it!
Gianluca Fiorelli: So it’s a very famous name in Italian.
Lazarina Stoy: Love it. That’s amazing. Thanks for sharing this little bit of knowledge. Awesome.
Gianluca Fiorelli: Thank you. And thank you to all of you. And remember to ring the bell to be notified when a new episode of The Search Session is coming out, and obviously subscribe to the channel. Thank you and bye-bye.
Podcast Host
Gianluca Fiorelli
With almost 20 years of experience in web marketing, Gianluca Fiorelli is a Strategic and International SEO Consultant who helps businesses improve their visibility and performance on organic search. Gianluca collaborated with clients from various industries and regions, such as Glassdoor, Idealista, Rastreator.com, Outsystems, Chess.com, SIXT Ride, Vegetables by Bayer, Visit California, Gamepix, James Edition and many others.
A very active member of the SEO community, Gianluca daily shares his insights and best practices on SEO, content, Search marketing strategy and the evolution of Search on social media channels such as X, Bluesky and LinkedIn and through the blog on his website: IloveSEO.net.
stay in the loop