Transcript for "Office Hours: Optimizing for Google vs. LLMs — Yes, They're Different!": Hello, friends. I think we're all here. Okay. Hey, Rand. So we have a little bit of some sound issues here. I think it's because well, I have external speakers, and my headphones are not working. So I'm using external speakers. It's causing an echo with everyone else's tech. But that's good because I'm not the one presenting today. So I'm just here to greet you all and say hello, and then I will turn it over to Rand pretty soon who will then introduce our good friend, Michael King. So for now, some quick housekeeping. Friends, we love seeing you in the chat. Hello, Sophie, Sarah, Naomi. I saw Marjorie. There she is. Lots more friends. People are coming into the chat. Feel free to use this chat to connect with friends, drop in some notes, maybe even drop in a couple questions in in passing, and we'll do our best to see them. But, ideally, you will put your questions in the q and a tab. So it's in the same area as the chat box. Just click over on q and a or ask your questions there. And then if anybody else asks a question that you would really like to see answered, do make sure to upload that so that we can prioritize that during q and a. And then with that, I will turn it over to Rand and go on mute because I am also getting over a sickness. I don't sound great. And you might be thinking, is this contagious through my ears? It's not. Don't worry. Oh, man. A fundamental misunderstanding of germ theory actually sounds like the most American thing you could have right now. Fair. I'm glad you're on mute. Mute, by the way. I read in the journal of, do your own research that that will prevent the sickness that Amanda has from spreading to you, so we're all in luck. Friends, welcome to the nightmare hellscape of 2026 America. I apologize for everything that's happening in my slash our country. If you're here from, another region, you might be wondering what's going on. We're not gonna be addressing that today on the webinar. Today, instead instead, we will be focusing on important marketing things so that you can, put your head down and concentrate and hopefully make enough money to flee if an authoritarian regime should knock on your door. This our our webinar today is actually one of the most well attended, well registered ones that we have had, and I believe that is be that is for two reasons. One, because of this topic, How to optimize for AI versus classic Google search results is fundamentally a huge issue for a ton of people, and their skepticism about whether there are real differences. I'm gonna tell you right now, I shared that skepticism because a lot of the people that I follow in my LinkedIn feed and my threads and my Blue Sky feed were saying, right, from classic SEO world, which I haven't been in for a while, they were saying, hey. Good optimizing for for AI is just good SEO practices. And I was like, oh, okay. Then last November, I was in Tokyo with, Mike King, who is joining us today. And I'm watching Mike present on stage, and Mike delivered an incredible talk overall, but there was this one section of his talk that hooked me and and kept me on the edge of my seat, and it was about exactly this topic. How to optimize differently to appear in AI results versus in classic Google search, you know, rankings. And I was fully convinced. Like, Mike does not BS around this stuff. He is not coming from a place of theory. He shows he showed examples. He showed the research. He showed the data. It's it's too compelling to ignore. And so I'm I'm thrilled that Mike, after the event, agreed to join us for our first office hours of 2026 and share with you exactly how to do these things and how they are different. The two questions that we always get if you're late joining the the webinar, one is how do I get the recording? You already get it. Guess what? If you're here, if you've registered, if anybody on your team registers at that GoldCast link, you get the recording. It will be sent to you afterwards. Two, if you have questions, the q and a tab, which is which is right next to the chat and messages, and there's a q and a tab. Click on that q and a tab. Leave a question for Mike. We will answer all of them, as many as we can, at the end of the webinar. And with that, Mike, thank you so much for joining us. We're thrilled to have you. Please feel free to, take it away and and show us how to do this this AI optimization thing. Cool. First of all, thanks for that introduction. That's a great vote of confidence. So for anyone that doesn't know who I am or where I'm from, Mike King, from an agency called Eyeball Rank. I was also just named search marketer of the year for by Search Engine Land for a second time, so that's cool too. Alright. So let's just jump right into this. You know, the the primary question is, is this just SEO? It's like that all you need to do for AI search. And my first disclaimer before I get into this, my data driven beliefs don't require you to agree with me. So let's just leave that where it is. So there's been a lot of great research going on in this space, whether it be from people like Dan Petrovic, me, you know, Mehtahan. I can't I don't remember his last name, but he's been doing some great work as well. And also Josh over at ProFound has been doing some really compelling research. So this is one of the things that he shared very recently. He showed that from a bunch of analyses that he did on some of the standard ranking factors in SEO, those only really explain like four to 7% of the citations that are showing up. He showed that there was like a pretty weak correlation between things like some of the linking scores that we look at and so on. And then also in that data that he looked at, he found that there's just a lot of unexplained variance in that performance as well. And the ultimate conclusion that he found is that content expertise, freshness, or things that we would bucket under relevance broadly tend to show a higher correlation with the visibility of content. So one of the things that I keep hearing people say is that AI search uses retrieval, so SEO matters more than ever. And I don't disagree with that. But the thing is that it presupposes that our community has kept up with retrieval. I think that our quote, unquote best practices, the software that we use, and this just the general understanding of how things work really contradict that idea. So on just from, like, a data perspective, you know, Zip Tie, they put out one of the first studies that I saw on this subject where they look to see, like, what's the overlap between how things appear in Google as far as, like, your rankings in standard organic, like, the 10 blue links, and then how likely are you to appear in chat, GPT, or Perplexity. And their data said that being in the top 10 in Google gave you a 25% chance of appearing in AI search. People don't, come. am I seeing this sorry to interrupt you. Am I seeing this right? Is that is that saying that there's only a 25% chance of showing in AI overviews as well? The is that the purple line? Mhmm. Yeah. Oh, damn. Yeah. So the so the overlap between Google results and Google AI overview, not high. Right. And the thing is this, what we didn't understand when this data came out was the idea of query fan out. Like, was no discussion of that. I don't think we as a community knew that that idea existed because people were doing this analysis. And even the the early analysis that I did on SGE before was AI overviews, there was no thinking about that because we didn't know about it. Right? And then Google announced it. and, of course, query fan out, Mhmm. I completely understand that topic, like, a 100%. I'm an expert in that. But but for those who don't know, I don't know what that is. I just what what is that? Sure. So query fan out is the idea that all of these different answer engines, if you want to call them that, are taking the query or the prompt that the user puts in and then extrapolating it to a series of synthetic queries. So, you know, whatever your question is, like, you know, how far is is Earth from Venus or something like that. I mean, actually, that's a bad one because that's that's probably just coming from the knowledge graph. But let's say someone is asking, like, a complex question. They are then gonna break that into a series of questions and then run those queries and then pull different components from passages from ranking pages to then generate the response. So it's not just the pages that rank for whatever you put in that's being used for AI overviews. It can be a variety of other queries that are running. And so you have no visibility into that. And that's part of why this overlap is so small because they're pulling from a variety of other queries beyond what you typed in. Okay. Okay. So this would be like if I search for, I don't know, a a men's tailor in New York. It might also search for men's alterations or men's custom suits or, you know, those kinds of things and then pull results from all of those types of searches? Right. And those types of searches vary a lot. So let's say, for instance, you're like, I'm training for, you know, the New York Marathon. What is a good training program? So they're gonna ask a lot of implicit questions related to that. So, you know, how do you train to run 26.1 miles? What is a checklist for, you know, training for a marathon? Like, it's gonna ask a variety of different types of questions. And, you know, depending on the modality that you're using, like, you're in AI mode, there'll be a lot more of those synthetic queries than if you're just in AI overviews because AI overviews have to be fast, and it tends to be, like, five to 10 queries that they'll use. And the. chat chat GBT environment, they'll use, like, three to seven queries or so on. So we had no visibility into that. And that's why when we look at data like this, it's like, okay. Well, what do I even do? You know? Because you're like, well, I'm focused on ranking for, you know, men's tailored suits, but it's 10 other queries that you need to consider in order to get that. visibility. Got it. Got it. Okay. Thank you. I I really appreciate it. I've seen the phrase query fan out a bunch. Now now I get it. Cool. So, you know, then we asked ProFounders, like, hey, like, what are you guys seeing? And so for them, they saw it was even worse. It was a 19% overlap between the Google SERP and ChatGPT. But I actually went back to Josh to see, like, what's the latest. And now that they have query fan out in their dataset, they see that's a 39%, overlap. And so what they're doing is they're able to pull the actual query fan out because in ChatGPT, in the API response, you can see like, in the actual JSON, you can see what was used for those queries. So a variety of tools are doing this now, but I know Profound was, like, one of the first that was extracting that data and they can give it to you as part of their platform. So it Google does Google's AI overview also show you the fan out? No. does not. No. So the only way you can get what they're using is if you're using the Gemini API. So we have a tool called. Qforia. We were, like, the first one to, like, give you QFO data. And what we were doing is we were effectively using the Gemini API and saying, like, okay. Based on all this research that we did into the patents as to how this works, here's the prompt. Please generate me a series of queries that you would if you were trying to, like, answer this question. So we have that, but then you can also using the Gemini API, they have a function. It's like a function call that it does where it will go to Google search. So you can use that to see what fan out queries it uses. And again, like I said before, it can be anywhere from five to 10 queries that it might use. And so the tooling that we use internally, which I'm going to make open source very soon, does that. And it also does what I said before where it'll extrapolate based on what the model will give it, but also you. can get exactly what it would use in the Google search environment. Killer. Okay. Thank you. Cool. So another issue, though, is that so there's a study that came out from SEMRush a couple months ago where they said, you know, 28.3% of the queries that are used have no search value. And that's that's the big problem here. Like, there's a big information gap in what the typical tools are providing, whether it's SEMRush, HRAS, or whoever, about rankings data because they're not gonna track rankings for queries below a certain threshold of search volume. But the queries that are used for query fan out mostly have no search volume. And the interesting thing about that is, again, it kind of gives you that sense of, like, these models are generating queries. Yes, they do use, like, source data from Google to some degree, but they're they're trying to fulfill gaps that aren't necessarily representative of user searches. And so there's a there's a lot of of keyword research that needs to go into this where we just have gaps in data, to kind of validate that these are the queries being used. So the big question or the big comeback that I see from SEOs in the community is like, oh, show me what you do for GEO or AEO or whatever we're gonna call it that you don't do for SEO. And I find that incredibly disappointing because it just reflects how our community really struggles with the distinction between strategies and tactics. Right? So think about this. When you when you think about something like outbound sales, PR, link building, even fundraising, you use the exact same tactics for all those things. You build a list. You come up with targeted messaging. You send it out, and you keep sending it out until you get the results that you want. Right? So no one is calling those the same thing. There's a different value exchange for each of those even though you do them the exact same way. And social media marketing is the same. Right? Like, all social media marketing is is channel specific content strategy. Like, you are understanding the nuances of the channel itself. You're understanding the audience in the channel. You're understanding the formats of the channel, and then you're creating things and you're putting them out. Like, that's content strategy. You're just doing it for Instagram or TikTok or x or whatever it is. But social media marketers understood that there's value in in standing up their own discipline because there is a different value exchange for this. Now for AI search, each of the channels is different. They react differently. Your strategy for how you structure your content may be different. So treating it like it's just SEO, you're you're kinda missing the forest for the trees. And that's not to that that's to say nothing about the fact that AI search has attention from the decision makers in business right now in ways that SEO hasn't in, like, twenty years. And it's an opportunity to reframe what you're doing so you get more buy in and people actually execute on it. And you can get away from some of the reputation problems that SEO has, so on and so forth. But here's a concrete example of something that is different in AI search. So not sure if if the audience is aware of the four ninety nine HTTP response code. But it was something that was introduced by NGINX, I think, like, you know, fifteen years ago or whatever, but it's also been adopted by CDNs and a lot of, like, modern, you know, clients. Right? And so because and what it means is that the client gave up on waiting for the server to respond to the request. And because all the AI search platforms aside from Google and Bing, which have indices, they're all fetching the pages in real time. This is something that we need to look for. And so here's an example. Right? This is a client that I worked on recently that they saw a spike in the four ninety nines, which again means that, you know, ChatGPT when it's coming to these pages, it's like, you're taking too long to load so I'm just gonna move on to something else. And so what they saw was a dramatic drop in their visibility in ChatGPT because of this one thing. And, yes, fundamentally, what we're talking about here is you've gotta do log file analysis, but SEO never taught you that this is something you need to look for. In fact, if you look at any blog post that talks about, you know, the different HTTP response codes, none of them mention the four ninety nine response code. And this is a fundamental thing that is keeping you out of AI search. I'm sure we'll see a flood of these blog posts after people hearing about this, though. So ChatGPT, you know, they use both, Bing and Google as their indices. I mean, there's been a lot of discussion as of late that they've stopped using Bing. But, nevertheless, the whole point is that these URLs are requested in real time. There is no index. I've seen that things get, like, short term cached if you keep, like, saying, like, hey. You ask the same question and it goes out. It'll go out, like, maybe two or three times, and then after that, it'll just use what's in the cache. But beyond that, I don't see them having, like, a long term index in the way that Google does. And so there's ranking factors that are a bit different as well. So they care a lot about your meta description. As we all know, the meta description is not a ranking factor in Google. In fact, in most cases, something like 80% of cases, it gets rewritten by Google anyway or they they pull something off the page. Semantic URLs matter a lot. We know that URLs have a very, very small correlation in Google. They have. a Wait. Just to just to be clear, so the the meta description is used by ChatGPT? mhmm. They? yes. It's it's not just, like, for getting someone to click. And I'll show you this in a couple slides where they're they look at it to determine whether or not they're gonna request the page and then, you know, review the content and use it as part of the response. There's a recency bias. They look at a lot of the UGC content and, you know, content that has, like, extractable data in comparison. Like, they bias towards that. So when when AI search gets your search result, you know, it it's basically what you see in the SERP. Right? Like, they're getting the URL. They're getting the title. They're getting the description. And if the snippet is different, they may also get that as well depending on how it's structured. And then based on that, they make a decision. Like, are we going to request this page and then use it? So your metadata is the advertisement to the LLM to determine whether or not they're gonna use your content. And so, again, some more data from Profound. They found that if you have a URL that, you know, is heavily like, the slug is heavily keyword related or high semantic similarity of the slug, it gets 11.4 more citations. And they're seeing that, you know, if it's very specific to the query that's asked, 5% more citations there. So again, we all know the URL. Like, yes, ideally, you know, you have a good structure URL that doesn't have a bunch parameters, but we all know that Google doesn't use that as a major rank ranking factor anymore, but it is significant in the LLM environment. So Bing's Copilot, you know, they use it they use their own index with ChatGPT on top of it. The system can also take actions in, you know, the the ecosystem. Perplexity is using Bing and Google, but they are citation forward. So, you know, being that they care so much about citations, they are more looking more deeply at the content and and making determinations of, like, where do we wanna send the user? It's it's less about the light review that these other systems are doing. They are more trying to, like, synthesize the information deeply to give you that answer. So it also means that you are trying to there there's gonna be, like, different strategic considerations. So as an example, one of the things that you often will do in marketing to generate leads is not have a pricing page. Like, iPOR rank does not have a pricing page because I want you to talk to us. Right? And what that means is that I'm losing opportunities to control my narrative because in this environment, they're gonna go look for an answer. So they go to Cyrus' site. They go to Clutch. They go to Designrush, and they're pulling whatever is in there to indicate what my pricing is. So, strategically, I need to make a different decision for AI search because of that. I need to say, I wanna control my narrative around my pricing, and I may want to have a pricing page. So is that just SEO? No. It's a business decision that you need to. incorporate into what it is that you're doing. Totally. Can can I ask, Mike, on that on that pricing thing, is there is there evidence that essentially if, let's say, lots of people on the Internet are saying some wrong or old thing, but the page on your site says the right most updated thing that the LLM will will bias to the answer that's on the domain that's being asked about? It depends which LLM. And I've seen, you know, varied responses from this. Right? Like, I've seen if you have your primary page that's saying what your pricing is and other people are saying something different, AI overview will say, well, iPoolRank says that their pricing is this, and these other people say that. But you'll be, like, the first part of the response. In. ChatGBC, I see it more being the consensus rather than what you say. Yeah. We we struggled with this a bunch where where, you know, like, old descriptions of SparkToro from its early days when it's based on Twitter keep showing up in AI overviews. And we're like, gosh. How do we how do we get it to talk about the SparkToro that's been around the last four years rather than the one that launched in 2018? Mhmm. And, yeah, Yeah. that's just been a struggle. is actually way more akin to reputation management than it is, like, standard SEO. Yeah. What I mean by that is it's, like, you gotta think about the content ecosystem. It's not just the page on your site. So you gotta think about, like, how do you spread that message across this matrix of queries and across a variety of formats and sites and so on so that then the consensus is your message, not like what everyone else is saying about you. I mean, controlling stuff on our own site's hard enough. Now I gotta control it. Yeah. Yeah. So here's another one of my favorite retorts. People say things like, oh, I didn't do any of that fancy GEO stuff, and I'm in ChatGPT in twenty. four hours. I see that all the time. But I've never seen someone show me that for a competitive prompt or query space. If you can, I'd love to see it. So my DMs are open. Send it to me. This is one of my favorite ones. Chunking is a scam. So chunking, which is basically, like, restructuring your content in a way that it's, you know, more bite size and you're having, like, atomic units of information, basically. It would in in practice, it looks a lot like things that you would do where you're, like, breaking up paragraphs and not having walls of text and so on. And, you know, I highlight this blog post because it's one of the blog posts that Cyrus wrote back in the day that was really impactful on me. It's called 10 super easy copywriting tips for link building. And he comes out swinging in this in this post where he's, like, showing side by side two things that you had written and showing how the thing that had a lot more structure performed better than the wall of text. It performed better on, like, linking. It performed better on time on-site. Like, a variety of metrics, it performed better on. And I remember Jamie, Steven, he had sent it to me. He was like because when I first started writing for Moz, was I writing, these big walls of text. And he was like, yo. Check this out. Use this to inform the way that you write. And anything I've written since, you can, like, tie back to the the the concepts in this blog post from Cyrus. And, you know, when we're talking about chunking, chunking is really the word that's used on the AI and information retrieval side for what these systems are doing with your content. They're breaking them down into passages and then indexing them that way, And then those chunks are what what are being fed to a language model to generate these responses. But when we're talking about chunking as to what we're doing as marketers, it's really restructuring your pages. And so people have been saying, oh, chunking is a scam. Well, here's an example. Right? This paragraph that I have here, it's targeting the keywords, machine learning, and data privacy. And just for clarity here, both search engines and large language models are built on something called the vector space model where they're effectively plotting your queries or your prompts and the documents that we're talking about, whether they're passages or full pages or whatever, in multidimensional space. And then the queries or prompts that are physically closest to the documents or, excuse me, the the documents that are physically closest to the query are considered the most relevant, and that's measured in a variety of different, distance measures, one of which is called cosine similarity. So what you're seeing here is this one paragraph, I have done cosine similarity on it between machine learning in the paragraph and then data privacy in the paragraph. And so the scores here for machine learning are point six four eight one and for data privacy point six nine four eight. Now I have simply split that one paragraph in half, and now I've measured the same thing again. So for the first paragraph, is focused on machine learning, the cosine similarity improved to point seven four seven seven, which is a 15.4% improvement. And then the same for the data privacy paragraph is now point seven six three four, a 9.78 improvement. So in those environments, just me splitting this in half has made them more relevant and more, they have more potential to perform in those systems. So when we talk about chunking, like, this is part of what you're doing. Again, you need to be thinking about humans first, and you need to be structuring your content for humans first so that it can be more performant in the ways that we had just seen. But this also supports the nuances of the system that you wanna be more visible in. So anyone that says that that is, like, not worth doing doesn't know what to talk about. And then Danny Sullivan, just the other day, he said something to the effect of Google doesn't want you to make bite sized content and that Google may make a change down the line where that no longer works. Here's the full quote. I'm not gonna read it to you, but, you know, you you can see it on the Internets. And here's the problem with what he said. It doesn't align with how these things actually work or where the research is going for all of these different systems. It doesn't it doesn't even align with how people read. I. don't Exactly. what the what the come on, man. Does it wait a minute. Wait a minute. Am I incorrect here? If we go pull up a bunch of Danny Sullivan blog posts on Google, aren't they chunked? Aren't they, like, short little paragraphs with a couple sentences on each topic and then it's, like, all broken? He doesn't write big freaking walls of text that go whatever. Whatever. So. so this all started with passage indexing. Like, this is the the innovation that allowed all of this to happen. Right? Like, Google announced this, I wanna call it, like, 2018 or so, where they went away from just understanding pages on the full page level to understanding it on the passage level. Right? And. so that. that then ultimately gave us retrieval augmenting generation which is how all these things function. But to his point that, like, oh, things may change, you know, soon and and all the chunking you did won't be valuable, again, that doesn't make sense. So with the state of the art, these are some of the key innovations that have come out of research that Google DeepMind has been building on top of, you know, the various, like, Google research organizations in the company have been working on. So there's something that Berkeley rolled out called, ring attention that really, like, is taking long sequences and then breaking them up and then rotating the sequences. And so what that means is that they're taking passages, and they are looking for where there is, you know, unified meaning in a single passage. So, again, like, you doing chunking helps this. Right? And then Google, has something I I'm sorry. I'm ahead. sorry? I don't I don't wanna stop you. But just, like, on this specifically so I saw Daisy, by the way, great to see you in the comments. Daisy is saying, like, chunking is just good content structure. And I is that is that generally right, or is is chunking like this process of I think what I'm hearing you saying is chunking is, like, take a specific topic, separate it out, write about or provide information about only that thing, and then move on to the next topic rather than trying to blend multiple ideas, concepts, information into a single paragraph or a single sentence. So where it's different from just content structure is that we can have a measurable feedback loop. And what I mean by that is like the same way I just showed you how I measured cosine similarity of what I've done there. When you're structuring your content, there's a series of different measures that we might wanna use. We have a a ton of different metrics that we use on the relevance engineering side of things. But, effectively, we can make an adjustment, see how well it scores, and then have validation or verification that this is better than what we had before. Now the difference is that, historically, like, the content editing tools out there, they're not based on semantic analysis. They're all doing effectively like TFIDF which isn't the only thing being used in these environments. So, yes, we are physically doing what we've always done where it's like use more headers, make sure that this section is only talking about this thing using data points, using semantic triples, all that sort of stuff. But now we have the ability to measure it in the same way that these systems are so that we can say, okay. That adjustment didn't actually do anything. This one did. Okay. Okay. Cool. Alright. So, yeah, we've got Ring Attention, which is, you know, another one of these innovations. And then Google is also trying to move towards infinite context where it's like you know, we talk about context windows. I think with Gemini, it's like a million tokens that they can look at. Google's trying to continue to extend that. But there's an argument that once you have, you know, a large enough context, you don't necessarily need to, like, take a passage. You can take the whole page. You can take the whole Internet and put it in there or whatever. But even then, what they're doing is they're compressing the information. And when you compress something that's messy that has a lot of information in it, that's like a wall of text, you lose more information. If instead you're compressing these smaller chunks, you retain that information. So even with this innovation, chunking your content is better. Meta has something called MemWalker where they're basically taking long form content and then breaking it into these hierarchical, memory trees, and then they can basically traverse the tree to get to the information. But, again, a chunk works better in here in that you can build this like map of semantic anchors in the content and then reference it again. There's something called recursive language models which came out of MIT where they're breaking down these longer inputs into smaller units and then recursively looking at these different chunks, which I implicitly just said chunks. So, like, that's gonna work better in this environment. This is innovation from MIT, but Google built on top of that with something called mixture of recursions or MOR. And this is looking at more individual tokens and then, like, looking at things based on their complexity. But, again, even in in this environment, having your, content chunks means that it's gonna be less complex, and they're gonna be able to, like, review the content faster, which they're gonna orient towards. So they're not gonna, like, prefer a big block of text. And then the last one, something that Google rolled out or talked about recently called nested learning where they have something called the hope architecture, which is built on the idea of of what they're calling memory and infusion where they're building long term context. Again, in this environment, having something that is understandable on that atomic level or a chunk will also do better because anything that's outside of that chunk would be considered noisy data, and they would get rid of it. So my whole point here is no matter what they do, structured content will always perform better in these environments and also for people. I don't I don't wanna derail you too much, but a lot of time when I write, Mike, I I write in a way that is intentionally trying to not sound like AI. Like, Mhmm. not just deliver information. Like, let me give voice and tone and tenor and humor and emotion Mhmm. and all those kinds of things. Let me let me play with words in ways that I know AIs don't. Am I hurting my AI visibility when I do that? It depends. And what I mean by that is, like, if you are mixing a lot of things explanation of something, like, yes, that is gonna do it. But, like, just because you write with style does not necessarily mean it's gonna be worse for those environments. You know? Because, like I mean, I'm not as good as you as a writer, but, like, I I also write in that way where it's, like, there's a lot of random asides and, Yeah. you know, there's. personality in the writing and so on and so forth. But it doesn't stop us from performing, and that that's kinda my whole point here. Like, what what what Danny said kinda presents the idea as though doing the these structural things and, you know, writing for people are mutually exclusive, and they're not. It's just a constraint that you may wanna introduce so that your content can perform better. Like, the way I've always looked at it is, though, if we're thinking about, you know, user personas that we're creating content for, the machines are just another persona. I'm still mostly speaking to, you know, the segments of my audience that are actual people, but I do have to also account for, you know, the lowest common denominator, which will be a machine that doesn't understand things as well. And to some degree, like, that's no different from accounting for accessibility for your website in that, Yeah. you know, you're still gonna make, like, the best thing that you can make, but you also wanna make sure that, you know, someone that's blind can get the information that they want. So I I just don't like. the way that it's framed as though, you know, like, you're doing this thing only for machines, like, when we know fundamentally better structure is better for everybody and everything. Yeah. Part of me part of me wonders and this might be outside the the context of what we need to get through here, but part of me wonders whether there's a case to be made for, you know, writing for saucy humans. Right? Like like, writing like Geraldine does. for people and then writing for information retrieval of all kinds, search engines and AIs, which is which is fundamentally just a different, you know, science of writing versus kind of the art of building human connection of an emotional writing. Well, my point is that you can blend the two. You know? Like, obviously, you know, Geraldine is a fantastic writer, but, like, I don't think you what I believe is that you can do something similar to what she does, but also kinda, like, drop in data points here and there or have, like, very direct statements that are extractable. So and and that's not to say, like again, like, she's a fantastic writer. It's not to say that I want to change what she writes or her style or or someone like that in their style. It's just that they're, like, little things that can be sprinkled in. And if you're doing it stylistically, it can still work for both things. Like, as an example, rapping. Right? Like, if you are a good rapper, you technically or you typically have very strong technical skills. And technical skills as they're presented in rap usually means that you're matching a lot of syllables. You're using a lot of, you know, various literary objects in your stuff. And you may say to yourself like, hey, I want to do more of those. You still have to connect with a human while doing that, but you're saying like this is your style. So I think the same thing applies to writing online as well. Yeah. Yeah. Hugely helpful. Okay. Thank you. Thank. you. So another one of my favorite things that people say is like, oh, you don't have case studies to prove this works. Of course I do, but you're telling me that you don't. So, here's an example of, you know, a client that we work with in the vehicle sales space. You know, we grew their ChatGPT visibility by 661%, their AI overview visibility by 330% largely by doing things that I've described here. You know, one of which was well, I didn't describe this here, but, like, the technical components of making sure the content is accessible and then creating more content that aligns with those semantic gaps that we talked about. So looking at query fan out, figuring out where they're missing things, and then plugging those holes so that they have more opportunity to appear as a part of these results. The way I look at it is that AI search is effectively like a raffle. And we don't have a lot of control over what happens in the synthesis side of things, like, once the pages are selected. But where you do have control is how many of those synthetic queries do you rank for. So think of each of those queries in your ranking being a raffle ticket. And so the more raffle tickets you have, the more likely you are to win the raffle. And that's basically how we do it. And here's another brand. So this is in the telecom space. We grew their AI overviews by 253%. Again, a lot of the same work that we're talking about, just making sure that the content is aligned in the way that we're we need it to be and, you know, doing the adjustments. This one, a financial services client we work with. So this was, like, very focused on the various, content engineering metrics that we have and, you know, just really improving that semantic relevance for these pages. And we saw, you know, obviously, the results are the most important part as far as, like, for the business. We saw 121% increase in sign ups, 52.6 increase in organic search traffic. Because another thing I wanna kinda highlight here is there's no real difference between search and AI search. Like, all search is AI search because all of the things that are part of Google's technology are being used for both things. But, yeah, we saw 17 x improvement in relevant scores, 24% increase in AI overviews, and then 72% increase in AI overviews per page. So, like, yes, what we're doing is working. But my ultimate conclusion here is that until our software catches up, it's not realistic that everyone is doing that idealistic form of SEO that we talk about when we say it's just SEO. You know, most people are doing a very checklist focused version of SEO where it's like, okay. I'll I'll fire up this tool. This tool tells me everything's on fire, and then I do these things. But a lot of the things that that these tools are looking for are so out of date with the state of the art. Yeah. So just to wrap it up, like I said, up from my full rank, we also have this AI search manual, which is 20 chapters of purifier, you know what I mean, of everything that you need to know about this. And we're actually in the process of updating it as well. So, you know, it's a great source for you to check it out. You grab it here. is the only time I find it acceptable to use AIR. And, also, there's SEO week. So, you know, come through. I'm me. Oh, Mike. That was that was incredible. I'm sure you saw some of the comments that were just blown away and amazed. Thank you. wow. What masterclass in how to do this incredibly hard and challenging thing. We have got a very healthy number of, questions, and I'm gonna hopefully get through as many as we can. For folks who need to drop off the webinar, no problem. The recording will be sent to you, including the next fifteen minutes of q and a with Mike. That recording is also available to anyone who registers. So if you're like, oh my god. My boss needs to see this. Oh my god. My client needs to see this. Oh my god. My friend needs to see this. You can send them the recording link. They will get the last forty five minutes and the next fifteen with Mike. So alright. Let's go through oops. Most alright. Kristen asked in here, are blog post updates considered recent if you update the content to show recency? You know, the the recent stats, the content with this time timestamps, or does the original publish date on on so called evergreen content still affect or or harm you? So that's another place where it's going to be different in Google than it is in these AI search platforms. So the way that Google measures recency is that they basically have, like their index works like the wayback machine. They have various copies of your pages over time. And it also drives, how often they crawl you because they basically score how often you truly update things. So they don't trust what you put in the, sitemaps or anything like that. Now on the AI cert side, because there is no index, they are heavily reliant on whatever date you put there. And even with Google, they do look to associate a date with all your content. Like, even if you don't explicitly have one in there, but you say something like, oh, five days ago, they will try to use that implicit date that you use to associate that with your content. So whole point here is, like, yes, update the date and update the content, and you'll see better performance. Awesome. Colin asked about the query fan out issue. Does does query fan out solidify the the sort of long held ideal of writing for more than just a primary keyword set, like like that you might focus on a a topic or or a primary search keyword, Yeah, but then you try and work in a lot of secondary and tertiary keywords. Like, is that classic, you know, SEO from ten, fifteen years ago? Is that still relevant? that's exactly it. And again, I think that kind of went away with when passage indexing was introduced because they're looking at things on a passage level, they can be like this is the greatest passage on the Internet for this query that is not your target primary query. So you can think of every page having many opportunities to rank for things rather than just your one to three primary queries that you were thinking about. Right. Okay. Awesome. Mikhail asked, and and Darren Shaw actually asked about this too. With these differences in in sort of how AIs do things and and now Google AI overview and AI mode, how important are metadata accessibility and and specifically schema? Yeah. All of those are really important. And and that's another argument that's been happening in our space. That one, I don't really engage in, like, whether or not you should use structured data. I mean, we just keep seeing more and more evidence that these LLMs can consume all of the structured data. They're not just using the things that Google gave rich results to. So one of the things that we are doing is, like, going back through and saying, like, well, what vocabularies were we not using that are worth considering for a given page? The metadata, incredibly valuable, you know, because, again, like, that's what's being used to determine whether or not they're either even gonna fetch your page. And then the accessibility, yes, really important because I mean, it kinda speaks to, like, the the lack of rendering content. Right? Because, like, most of these platforms are not using JavaScript. So what you're probably doing as a function of accessibility is also gonna support them understanding your content as well as possible. Okay. Okay. Yeah. Cool. How how important are are is the URL string itself? Should should people be focused on that more, or it's just a, hey. This is a small little signal. If, you know, if you have it structured in a particular way, you don't need to worry. Is it problematic if you use, you know, slash blog versus slash articles or have a does does subdomains versus subdirectories still matter? All that kind of stuff. Yeah. I would not go back and change your URLs because, again, that could be catastrophic for classic search. But moving forward, I would probably think a bit more about what's in that string or within that what's what's in that slug. Because the way it's interpreted isn't just gonna be like, oh, it doesn't have a high cosine similarity. Like, the language model is gonna read that. It's gonna read that text and be like, oh, yeah, this is highly relevant to that user's prompt. So I would, moving forward, probably have, like, longer slugs. So, like, what you have on your blog post where it could be the full title of the article rather than, like, reducing it to just the target keyword would probably be a better approach moving forward. But I wouldn't go back and change URLs. Cool. Okay. That is that is very helpful. And also, some work off our plate. Can you can you speak to the so Becca Peace commented that, like, for example, when she asked questions to ChatGPT or or or Gemini, Google AI mode, that kind of stuff, she's seeing that a lot of the time, they will cite things, especially on Reddit from years ago. rather than more recent but maybe, like, less, you know, fewer response threads that have happened in the in the more recent past. And I see the same thing. Right? Like, when I a lot of the times, if I search classic Google, I will see the most recent Reddit threads. Mhmm. If I if I use AI mode, I'm seeing these, like, older ones, a year, two years, four years. What is that a oh, that's that's just random. Like, that's an anecdote. That's not that's not data. They do prefer recency, or is that because of something that the AIs are doing differently than than classic Google search? So in the example you just gave, it may be a function of the fact that AI mode does more queries in the query fan out. Like, AI mode can do twenty, fifty. In some cases, I've seen, like, a 100 queries that it runs. Because the way they think about it is, like, if you're in that environment, it can be a slower response. Whereas in search, it needs to be almost instant. So it may be a function of the fact that they are, running more queries, and some of those Reddit posts come up for different things. And being that they may be longer posts, they may be more relevant, whatever that query is. So it's not necessarily about the, like like, authority or or having built up links or anything like that. It's basically there's a whole lot more content on a lot of different topics in this old thread. Yeah. Let's go we're gonna excite that one rather than this new thread, which is hyper focused and has only four responses about this specific thing. Right. So, I mean, there generally is still a recency bias for Reddit specifically, but it can be overpowered by the relevance bias. Got it. Got it. Very helpful. Okay. That that fits with everything you said before too. Okay. Melissa asks, how are you tracking AI visibility? You know I have many thoughts, but. I'm I'm I'm gonna put my thoughts aside. So actually I want to touch on what I think I've seen your thoughts are. So the primary concern that people have about measuring measurement of visibility in these environments that it's highly probabilistic. Right? And my counter is that rank tracking has always been probabilistic. Not to the same degree, but people are are calling rankings a source of truth when it's not, which is reflected directly in what you see in your Google Search Console data where they're giving you an average of how you show up. So we've already we were already using imperfect data for rank tracking to begin with. Now the problem is further compounded in these AI search environments because all three of us could do this exact same prompt and get wildly different information because of the context that we have with the machines and so on. And then the different tools are doing it differently. So some tools are getting the responses from the APIs, which is gonna be very different from what you get from the interfaces. As I understand how ProFound does it, they're scraping the interfaces, and they're also scraping it multiple times a day. So the answer that you're getting or the the the value that you're getting is the average of how you're showing up, which to some degree is not much different from the value that you would get from Google Search Console. So I'm not saying that any of this is perfect or it's the way to do it. I'm just saying that of the approaches that I've heard, that one makes the most sense to me. But at the end of the day for any of these tools you can't really expect accuracy. You can only expect precision. Okay. Yeah. Yeah. I I've always been I've been deeply curious about, like, hey. If, you know, a thousand people were to run this same prompt, Mhmm. how how many different, you know, brands would show up in a thing? Which ones would be first, second, third, fourth, fifth? You know, what whatever. And then, like, how does that map to how any particular collection or tracking tool might monitor it? And is there enough overlap to where you could you could basically say, like, This methodology is directionally accurate or, like, wow. You know? Whatever. You know, maybe maybe ProFound takes a prompt and runs it five times in a day. And it's like, oh, it turns out, actually, if you run it less than 500 times, you won't get enough statistical significance to be able to really say how often a brand shows up. And so they're they're just throwing random stuff at you. Like, it it. you know, it's meaningless. That that I can't speak to. Like, I don't know the nuances. I just understand that they do it more than once. So it's like, at least. it's better than just doing it once and being like, here's who you. are and where you stand. But I think the other part of it and I and, again, I don't know if they are doing this, but I think that because so many of these tools are also pulling in clickstream data, they may be comparing it against that as well. That that, I think, would be a a real like, a science y solution. Right? Like, that that would be a provable solution. I would love that. Okay. Marlise asked, where where do you begin with the content restructuring process? Like, you know, you work with a new client or you're recommending to to an SEO out there. They've got a page. They feel like, hey. This page provides a lot of very important information. We want to make sure that it is getting cited in, you know, AIs. We wanna make sure that that information is being reflected in the answers. What do I actually do? Yeah. And that's another place where I feel like SEO software is really falling short right now because you should be able to put in a URL, your prompts, and then it should pull the query fan out. And then it should go through and score each of your passages and say, like, hey. This is covering too many different topics. Split this up in this way or whatever. And, you know, the reason I can clearly describe that is because we have equivalents to what I just said in house. And, you know, like, again, I I intend to open source a whole bunch of stuff that we're using because I feel that strongly that the SEO space is so behind. And I hope that, you know, some of these tools will just grab my code and throw it in into their tools because it it it's just such a miss right now. But where I would start if I had no tools is I would just read the content, and then I would say, okay, which of these paragraphs is just covering too many things? Which of these paragraphs is not, you know, clearly explaining the topic? And where can I inject more clear extractable meaning? And that'll be basically, like, what are the the self sufficient sentences or paragraphs, and how can I split things up? And then also, how can I add headings that are, like, very clear, like this section is about this thing? So you have clear, like, informational boundaries in your content. And then after that, I would start going through and saying, well, where can I put, like, data points that reinforce what I'm talking about and make it sound more authoritative? Also, how can I improve the readability? Like, there's plenty of tools out there that'll give you readability feedback. And, you know, I I just start with those things and then see how that performs. And then by the time you've launched your stuff, all my free stuff will be out. Amazing. Some quick lightning round ones. Should brands avoid using acronyms or shortened versions of their brand name? Because, No. essentially, like, you know, AI can't recognize it or so or make the correct associations. No. Because, you know, all this stuff is built built on basic natural language processing, and the whole entity, you know, resolution thing, I don't wanna say it's been completely solved, but, like, it's really good. So as long as you define that acronym somewhere, use it. Okay. Cool. Very helpful. How do you decide this is a little longer than lightning round. How do you decide where LLM optimization, you know, appearing in AIs should sit on your list of initiatives and priorities versus versus any other sort of growth marketing or SEO thing that you're doing? Yeah. I can answer that one fast, actually. These platforms are more branding channels than they are performance channels. So I would think about it more about, like, if this is a brand effort that we need to care about, then, yeah, it it gets in your order of priority based on that. Right? But if your goal is to drive performance, I would not put AI search on the top of your list because, you know, they it drives just far less referral traffic. It tends to be more performant referral traffic. It's more qualified because the users spent all that time educating themselves in the channel. But, you know, your your ROI is gonna be reasonably low, at least right now, from a performance perspective. Super helpful. What about PR? Is PR sort of the the quick way? Like, is that the cheat code to get mentioned by and covered by AIs? The cheat code is YouTube videos. It's the second most used or cited source in these platforms. But PR is super helpful because, again, the more your message is in more places, the more sources to generate that consensus. So it's a big part of the play, but so is user generated content because because Reddit is still the number one cited source across all of these platforms. Yeah. Reddit, YouTube. That feels big. Do you do you think Instagram, threads, Facebook, maybe even, I don't know, Twitter if someone buys it and stops making it a underage porn site? Like, will those will those platforms also become sources, or do you think that, like, Meta is intentionally trying to keep AIs out of them? So I I think it's more difficult because a lot of that content is, you know, multimedia, and they don't have the. level of access that they want to be able to ingest it in real time. So Google has the advantage there because they index all of YouTube, and so they have the transcripts. They have more than that because they can understand multimodal content natively. So they can understand, like, what each frame means and so on. So, you know, it just speaks to Google just having more advantages than anyone in this space. In long term, they're gonna win this game. But to your point, I don't know that we're gonna see an immediate usage of those platforms because of how difficult it will be to use them in real time. Yeah. Yeah. Yeah. I think this is gonna have to be our last question, but, Mike, Matthew Silverman's curious. What's the most overrated advice that you see in optimizing for AI? That is just SEO. Fair enough. Fair enough. Perfect. Perfect. And, Mike, thank you so much for joining us. This was absolutely outstanding. Amanda, thank you for being here despite getting over the flu. I hope you feel better soon. And to everyone who joined us, remember Mike's got a bunch of great resources on his website at iPullRank. He runs this amazing conference called SEO week in New York that you should check out if you can be there. And, of course, make sure that you send the GoldCast registration link to other folks. They will get this recording as will you and the slides. Amanda, I'll make sure that you have all of those in the next few days. Thanks so much for joining us, and take care, everybody.