AI Mode & The Query Fan-Out Technique: How Google AI Search Works | The Latest in SEO [June 2, 2025]

The latest news in SEO and AI search with today's SEO FOMO Tool Landing Read, going through the top stories from the latest edition of the SEO FOMO newsletter. And today to go through the top stories, I have invited Andrea Volpini, who's the founder CEO of Wordlift, which is one of those SEO tools that already provide us the functionalities to get prepared for The next stage of AI search.

Thanks. Thanks for the invite, Aleyda. Super excited.

Thank you for coming. And before starting going through the latest news, I want to thank Similarweb, who's our today's sponsor. Let's go through them. The first one, as we can expect, is this massive piece that was published last week by Mike King, going through how AI mode works and how SEOs can prepare for the future of search to optimize For it. What Mike did was to go through all of the patterns pretty much, and not only analyze the patterns in an easier to understand way for SEOs, describing the implication to our work, but also analyze the different areas in the way that we do things at the moment and how they will change, giving ideas for potential tools and data that we need that we still don't have. at the moment. Andrea, I would love to hear your thoughts on Mike's piece that you would like to highlight from this. Anything that you thought that was particularly important. I was so very thankful that he took not only the time and effort, but also provided actionable advice as outcomes, because this is something that unfortunately we had already coincidentally gotten from search engines in their official AI search guidance. Google and Microsoft also published this similar pieces providing best practices for AI search. And I'm here showing the one of Microsoft because this is the one that I saw that was a little bit more specific, more actionable, but the Google ones was so very high level, not necessarily saying something different than what we already do currently in SEO, right? So I really Like that Mike went to the nitty-gritty and described the real implications.

Mike did monumental work here because it kind of combined a lot of the aspects that technically drive the new AI mode experience. Now, to make it short and kind of highlight the focus, one aspect is the user context. And this is still to be practically understood. But user context means that everything that we do across the different Google properties and tools will translate into a user embedding. So it would in the end be represented by a matrix of numbers. And so if I have visited your website and Mike's website during my journey to understand AI mode, these will become part of the experience. So every time that we trigger a request, the AI mode system will take into account the embedding representing the customer journeys. And this is something that also Cindy Kram has been explaining in the past in different ways. So user context, first area where we need to start focusing on. And practically that means that we have to understand the target personas of our brand. And so one thing that, for example, we do when we build our knowledge graph with the SEO ontology, we define user personas because if we don't understand very clearly who we are addressing, it would become very difficult to get into these new systems. So a clear understanding of the user context and customer journeys, it's key in order to understand more about how to build relevance. The second aspect that Mike highlights is the importance of semantic similarity. We're not ranking pages, but we are ranking chunks by semantic similarity. That means that you can have a long content, but there is only a short snippet snippet of text that would be described in machine learning as a text chunk that would be ranked against the other chunks across the other websites. And so we are into the world of what Google used to call passage rankings. And that means that we need to start optimizing for these passages. The third aspect is on the query fan-out technique, because Mike found an interesting pattern here that provides an highlight on the different types of expansion of the query. So from one query in a fan-out process, we will obtain multiple subqueries. Now, there are different ways in which a system like a large language model can generate these subqueries. And here in this patent by Google, we can see different angles that can be explored in order to generate these subqueries. So these are, for me, the 3 key parts of Mike's piece.

That's a wonderful summary. Thank you very much. And Many will say, oh, but we are sort of already doing that in one way or another, right? But it's the level of execution, as you mentioned.

Let's get back on the practicalities of our job. Like in the past, for example, we have expanded content by analyzing the People Also Ask feature on Google. Now the People Also Ask feature is something that now it's been crawled and we've been using it for now several years. And it starts from the user query. So we have a user query and then we have the People Also Ask that tend to expand. And so we might think that this is a similar fan-out technique. Well, the reality from some of the analysis that I've done is that the output is quite different because in a People Also Ask, Google will try to keep the user within a same context. So it will expand vertically, whether in an AI mode fan-out approach You will see that Google will tend to get a broader set of topics covered by the new queries. This is also why only 25% of the results are coming from the top rankings. And so there is a crucial difference in the data that we used to have and the data that we need to build with AI mode.

This is amazing. I thank you very much for bringing that up about the differences in traditional search results, top results versus answers in AI overviews in this case, or AI mode, which is why I want to go to the next piece or already that explains that a little bit. The one from Tomek Drudzki regarding SEO still matters for AI search engines, going through an analysis that he did of 25,000 real users across AI search platforms. And he discovered that if your website ranks number 1 in Google traditional search results, blue links, you have 25% chances of being cited as a source in the AI overviews. And he goes through the explanations and the differences of this in AI overviews versus ChatGPT, Perplexity, and At some point in the explanation here, he mentioned how on one hand, this is why ranking well in traditional search results and rankings in Google remains crucial in AI search inclusion on one hand. At the end of the day, I think that we can consider this because of the grounding systems that LLMs and AI mode in this case has like the base. But then why we see sources that don't rank highly mentioned or cited here, and he mentions 2 main reasons. On one hand, what you just mentioned, personalization and caching, and the query fan-out technique. This is a big one, issuing multiple related searches across topics and data sources to develop a response in order to make sure that they provide a comprehensive answer, taking into account many different angles to provide the synthesis to the user. But yes, it's interesting on one hand to understand that, oh, it's not necessarily that it's something completely different. What we have been doing provides an amazing ground to continue building and understanding and refining, evolving what we have done. But yes, there are some major differences indeed. And that takes us to the other top piece of the week, which it doesn't surprise me that the topic is as popular because there are hundreds of conversations about this going on right now. In LinkedIn, we can see, oh, should we still call ourselves SEOs or should we call what we are going to start doing now with AI search something different. Should we be SEO? Should we be AIO? But then Ran also made a really good piece that really connected because at the end of the day, it's also still SEO what we'll do here because SEO, it can be also the acronym for search everywhere optimization too, because realistically the optimization that we will be doing is yet for another platforms. Many other SEOs have also evolved to optimize for YouTube, TikTok, and so on. And LLMs here is far more discussed. And I believe that it's seen a little bit of the evolution of search engines. And that is why, unlike those who optimize for YouTube, there is this conversation now, right? So I would love to hear your thoughts about this, Andrea. Are your team SEO or something else?

I've always been an advocate of the idea that SEOs are, you know, an important part of an organization. When it comes to AI. So I always had this vision that the SEOs are on kind of on the pioneering side of things that have to do with AI. So I'm totally okay with the term SEO. I'm totally okay with Mike's idea of relevance engineering. I think that there is a shift in the data that we have. There is a shift in the consumer behavior. I am not totally on the side of still talking about search because there is a lot of course content recommendation at place. We've seen it for many years with Google Discover and And we'll see more with AI. So I'm not sure we should think in terms of search. I believe that the skillset that we have will remain valid as we keep evolving from one area to another.

100%. I believe that potentially the only reason to do it will be to package it differently in order to make it easy to support, let's say, across organizations. So to get more or better buy-in by certain decision makers who are very impressed with all of what is happening with AI, right? And for AI, potentially they might want to even go bigger than For SEO. And unfortunately, because of the bad reputation that SEO had for a long, long time, and still today with certain people somehow, even if it has matured a lot and we know that it's very different than 10 years ago. Yes, I can think that just because of that, it will be a potential decision to make. Although let's say it's a little bit more of a marketing of how you package it rather than a real need, because it's something entirely different. To finish our today's edition, I want to go through actually your explanation of the query findout technique. I believe that as, well, Mike highlighted also in his piece explaining how beyond personalization, this was one of the other core areas or aspects of how AI answers are different or unique in the way they work versus traditional search results and explain it really, really well. I also explain it, but more in layman's terms, let's say. So very, very quickly to start tackling this. New topic that will play only a bigger role in our SEO or AI optimization day to day is the query fan-out technique, which is a technique that AI mode uses to breaking down your questions into subtopics and issuing a multitude of queries simultaneously on your behalf. So these are done in parallel and this enables search to dive deeper into the web than traditional search on Google. So rather than taking a specific query and trying to match it with content that they have in their index, they will generate a variety of queries trying to identify and discover from different contexts and areas information that will satisfy our queries in a very, very comprehensive way. And I went through here a few examples and also mentioned here a little bit of what implication, let's say, we could expect in our day-to-day, the need of deeper intent understanding on one hand, real comprehensive topic authority and anticipation of follow-up questions. Andrea took this to the next level, also going through patterns analysis and provided also not only examples, but even a tool. This tool is completely free. You can use it. Andrea just mentioned before the call how he had even refined the tool. In order for you to use it and to get an even more comprehensive result that resembles more of what we could receive using AI mode with related queries, implicit queries, comparative queries, reformulation of queries, entity expanded queries. So Andrea, can you please explain?

Today I went live with the second release because of this opportunity that I had to discuss it with you. So the core idea is, as you ask, how do we do something practical, right? So. I didn't want to start from the user intent because the user intent, of course, would require an understanding of the user context. I want to start from the existing content that we have online. And so my goal was, okay, let's look at a product page, let's look at a listing page, let's look at an article, and then let's try to understand the AI mode readiness of that piece of content. That's my idea. And so I wanted to start from the ontological core. So what is the core entity that the piece is describing? Is this a product? Is this a comparison between products? Is it, you know, an informational piece? And here in this second release, I'm using Gemini. Gemini has a very powerful functionality for extracting context from URL. So it doesn't allow you to actually crawl the page, but it would use The page as a context for generating whatever request you have. Originally, I was scrolling the page, but then I look at what Dan Petrovic was doing, and then I say, okay, we can use Gemini to read the page and use it in context, which is very similar to what I expect Google AI mode is doing. Okay, so Google is not crawling the page because he has the page in the index, and the model now has this functionality that will enable it to get the information from the page as part of the context. So I started to ask the model, what is the main entity for the piece? And then after doing that, I build the program. So why I want to build the program is because I don't want to remain within the context of prompting. Why? Because prompting is very brittle, it's fragile. But when we build the programs with these frameworks, then we can improve the prompting ability autonomously. So if I feed now more data, for example, the data that comes from SimilarWeb about, you know, the type of questions that people are actually asking on a given page, then I can improve autonomously the prompting. So I don't do the prompting. I just do a first prompt and stick it into a program, and then the program will learn based on the data that I feed. Now, when we generate the subqueries now based on one of the patterns that Michael chunking highlighted, the system reasons on the different type of templates that, you know, the different type of facets that can be used for generating subqueries. So then I generate a list, and this list again is, you know, it's quite different from the traditional People Also Ask type of subqueries. Then I get back on the page, I chunk the page, so I create a text chunk from the article or the PDP that you had. And I start comparing the subqueries with the chunks that you have on the page to let you see if you're covering each of the subqueries already with existing content. Now, an interesting aspect is that we might have some content elsewhere on the website that might cover that subquery as well. We don't need to create these ultra-long pages now because we have to cover the fan-out queries, because we might have that content already available. And so here, the next step, of course, using things like Wordlift, where you have all the embeds headings of your website, we can immediately find the right chunks that can respond to the fun out starting from an entity like a product or an article. And so that's pretty much how things are working now. It's a collab, it's open source so that everyone can play with it. I'm going to create, of course, a web application so that it's not too technical and it's not too complicated. I'm using embeddings from OpenAI. I will probably use embeddings from Gemini to keep things as consistent as possible. Here you see the results. At the end, I'm evaluating an article and I can set a threshold. I can say how similar should be each chunk to the generated queries. I can decide to generate 5, 10, 20 subqueries with this algorithm. Then I can see what is covered, how well it's covered, because a competitor might cover that way better. Right? So I can start thinking in terms of, okay, is these subqueries well covered by my piece or is there someone else out there that is covering? Because we have access to Gemini and the way in which Gemini uses the web as context, we can actually build such a tool that will allow us to see what is the best chunk for this subquery. And of course, do we have to cover every subquery? Not at all.

That is a great point that you just made because I believe that that is one of the other, let's say, worries for a lot of people. And we can see this directly in one of the other top pieces from the latest edition of SEO FOMO, which is this fantastic guide of Chima about should we stop creating informational content? A lot of people are actually wondering this in the other way around, but now I actually need to create even more content to create like more comprehensive topic clusters that actually answer these questions from many other different angles. But how and why? Because a lot of people are actually scared right now and thinking about actually stopping to do it because of AI overviews and the fact that AI overviews will be shown much more for informational queries and provide directly an answer that will highly satisfy many scenarios the user needs and won't make necessarily or translate to a direct visit like before. And of course, Chima makes a fantastic point in her piece. At the end of the day, it's about creating something that is worthy and actually helpful for your real customer and users, not creating content just for the sake of it, just because you have a high search volume or because it's one of the related queries that we see through the funnel process right now. Is this content actually solving a problem for my audience? Or is it merely adding to the noise of the SERPs? Or in this case, it's actually one of those that AI mode at some point will assess. But if it doesn't make sense for me to add or to try to contribute, or we don't have actual insights or something that connects with our products or services, well, we might not necessarily go for it. What are your thoughts about this and how do you think that is the best way to address this?

I think there is a lot less space for bad SEO, right? With AI, there are better capabilities of sifting through the good data. Now, what is the good data? It's really what I call the ontological core. Like, how strong is your topical authority in addressing your customer needs? And how strong is your understanding of your customer personas would make a difference. So you don't need to increase the coverage. You need to increase semantic relevance for the subqueries that really are part of the customer journey for your users, your clients, and that's becoming very strategic. And there's no space with such a system to create just fluffy content that would not make value in relationship to a brand. That's the way I see it.

Amazing. I think that is a fantastic insight. It's better to have an example. If you have the capacity, let's say, to create 100 pieces of content, it's better to create 100 pieces of content to Cover your core topic, your real expert topic in a very comprehensive way rather than going for 10 different topics related that where you don't necessarily actually offer something for or unique based on your product or services, right?

I mean, the idea is that when we look at a brand, we have to look at what are the core entities of this brand? What are the core entities representing the brand's clients? And then we can take it from there. It is true that we will possibly generate more content than before, but this content is also going to be dynamic. We have to start thinking in terms of, we don't need to create a webpage every time that we need to publish content because as AI progresses, we will have chatbots that will address some of these needs. And still behind the chatbots, there needs to be an SEO that connects the user needs with the core value of the brand. And so I think that's, that's the direction where we're going.

Thank you very much for your insights, your advice. The tool, go and check it out. It's completely free. And stay tuned also for the additional web version of the tool that Andrea will publish too. And check out his tool because it's one of those tools that a lot of what he just mentioned is automatically generated or is part of the functionality, right, Andrea?

Yeah. I mean, that's basically what Wordle does is really help companies, you know, understand their core entities and build an infrastructure that creates content at scale without the risk of going off topic, the risk of hallucinating, whether it's programmatic or editorial. That's what we've done for you.

Amazing. Thank you very much for that. And to wrap up our today's episode, I want to thank again Similarweb, today's sponsor, which coincidentally, I believe that is one of the only players out there that I know at least that offers prompts, information about any page of any site out there. They don't only give you the traffic information or details of the different LLMs to your site and your competitors, but can be of a lot of help to better understand how people, users already searching out there for your products or services through LLMs. Thank you very much, Andrea, for joining. Your insights have been super helpful to deconstruct what's going on, going through the main criteria, what people should pay attention, understand better, and to focus our efforts on actually what matters, right? There's so much noise going on. So I really, really appreciate you joining.

Thanks to you.

And if you want to keep up with the latest stories without waiting for the next edition of the SEO FOMO newsletter, go and take a look at the SEO FOMO News, which is pretty much an SEO news aggregator. So you can keep up with all what is going on. On Real-Time, it actually has also an AI search category in case you are particularly interested on this topic. Remember to subscribe to the channel and like the episode if you have liked it. Thank you very much, and until the next edition, bye-bye.