Hidden Playbooks: How B2B SaaS Companies Dominate LLM Results

Hello everyone and welcome to today's webinar which is on the uh secret topic of hidden playbooks and we're going to be talking about B2B SAS companies and how they actually dominate in LLM results how they do well in chat GBT Gemini perplexity and all the other AI chat bots that you hopefully have played with know about and are thinking about getting your business featured in. We have a fantastic guest speaker today. His name is Multi and he is CPO and CMO, a double role at Peak AI, which is a search analytics platform focused on AI search as well. And he's going to be taking us through a lot of their firsthand data because if I if I catch it rightly, Moly, you guys have a lot of data about businesses from the tools and the analytics that you have to go through what these hidden playbooks look like. some successful B2B SAS companies and how they're actually doing it with some real case studies as well. So without further ado, Multi, I'm going to hand over to you and anyone who has questions because there's going to be questions and we can't wait for them. Please just drop them in the chat. You can drop them in any time. We're going to make sure that we've got at least 10 minutes at the end to answer your Q&A on anything relating to B2B SAS and LLM AI search strategies and anything you want to know about how to get your brand more visible in Chat GBT and other AI platforms. All right, let's go. No, I can't hear you. Oh, it looks like we've got a sound problem. I think Multi's dropping in and out to get his mic back working. Here we go. So, can you hear me now? Yeah, perfect. Perfect. Perfect. Then, uh, sorry for those technical difficulties and let's get started. Um, yeah, I'm Malta from PKAI. I'm the chief productive marketing officer there. um have spent the last five years working at ideal law the largest price conference website in Europe doing SEO and then at some point also AI search and before that I spent five years at search metrics building um SEO software and now I'm at PKI it's a software that helps brands or also agencies to uh track how visible they are in AI search and to uh do activities that help with more visibility and today we will talk about why stuff like this exists. What are we looking at? It's an article from Super Office on the Superoff website and it's about the 10 best sales CRM platforms. And surprisingly, number one on the list is Super Office CRM. And we'll talk a little bit about why such articles are currently flooding the internet and what's happening there. But uh oh yeah, this was the zoom into that slide in case I would not be able to find the juicy text parts. But let's start with this right we have we have a prompt in in a typical LLM based search and answer engine something like tell me about the CPO of PKI as we already heard that's me so it's very easy to detect the entity multilanfare in there um and whenever I speak about entities or brand mentions or visibility I mean this that an LLM in a answer suggests or recommends a brand that is often not actually accompanied by a link or anything. It's just a just a recommendation. And uh we now said that my name is the relevant entity in that prompt. But actually just as well, it could be peakai, ideal, and search metrics could be the relevant entities in that brand. So it's in that prompt. So it's a little bit um it's a little bit different from SEO where we have these traditional uh 10 blue links that are sorted where it's very clear what is actually the answer, what are we looking for. Here everything depends a little bit on how we want to interpret the answer. And in addition to these mentions there are also sources or citations. So when the LLMs conduct web search or their grounding process they uh of course use documents as sources or actually they use certain text chunks from these documents and these are also there and then we have actually links but nobody is clicking on those. Um and uh yeah, one my first piece of advice is you have to be the source. Like if you can make your website be the source during the websites process, the grounding process, you can greatly influence the answer. Uh why am I saying this? We have a case study is from one of our clients. Um when they joined us, they realized that for the prompts they're monitoring with us, they are only uh the source or their website is being cited in only 3% of the chats. And then they published expert content with over the course of a couple of weeks and they got that up to 61%. So huge increase and uh what happened to their brand mentions those also increased from 3% to 33% so 11x um and in exactly the same time frame. And why did that happen? Of course because they were cited as the source. they mentioned their own brand in these documents on their own website and that also that influenced what brands the LLMs are mentioning and then I have a second data set we can look at I have to anonymize it unfortunately uh but I can reveal that it's in the uh German price comparison uh space and so who who paid attention to the first or second slide might make a guess what kind of brands are mentioned here and if you look at the absolute top brands here by the way we always look at first visibilities So is the brand mentioned in the answer and then we look at is the website of that brand uh cited and we look at the percentage uh how many chats how many chats each of these is true and that's we did that for a couple hundred prompts run over a couple of weeks every day. And if we look at the top brands, we can see that there seems to be a very obvious correlation, right? The more you get cited, the more visible you are. Or maybe it's the other way around because for now it's just statistics we're looking at, right? So we can only see a bit of a correlation. We can't actually from this data know the the causation. But there are other brands where this seems to be not true. So there's a brand that is mentioned in 15% of answers, but their website is never cited. And when this happen usually two conditions have to be fulfilled. One is something positive must be written about this brand somewhere on the internet like this must be true otherwise it would not happen that they are recommended in the in the answer unless they are super relevant in the training data and the LLM just recommends them without a web search but usually you can expect that there are websites that write positively about this brand and the second condition there are a couple of different explanation either these brands are blocking the bots of the LLMs I would personally not recommend that but some brands are doing or they just don't have any content that is worth citing like they have a very bad content strategy or they have a technical SEO problem. So there's something goes wrong uh either with the initial rendering with the initial crawling with the rendering with the ranking maybe everything is hidden behind JavaScript we can't know based on this data but usually this is the explanation and there can also be cases where a website is cited a lot but the brand is invisible and this is the kind of case I want to also look at a little bit because that's easy to fix um because I just told you right you have to be the source but now the second thing is being the source is not enough. [snorts] Um why? Because here is something happened to another client of ours and unfortunately I can't reveal the the brand but it's a financial services business in the United States. Basically they when they joined their visibility in perplexity for a set of prompts was zero. Their brand was never ever mentioned but they are ranking super well in Google. They are ranking super well in Bing. So, how could this happen? How could Perplexity never mention them? Well, Perplexity answers looked like this. You got a long list of competitors. Uh, there are two on the bottom like Wells Fargo, Morgan Stanley. They are more well-known. I didn't have to censor them for the rest I had to censor to not reveal who's our client. But basically, long list of of brands. And then our client's website was being cited as the source, but their brand was not recommended. So what was happening here since they were ranking so well in search engines because of traditional SEO of course when Perplexity did their web search and grounding process they found their uh their website and found relevant doc found it treated as as a relevant document and then on their website they had a little headline and some paragraphs that said well we are not the right solution for everyone. If you want to work with a different brands, here are five other companies that are active in the same states, offer similar service to ours, and they are trustable. And since the LLMs always look for like a nice chunk of text to use for their grounding process, they thought, hey, this list is amazing because it is a list that mentions entities and it's in one of the most relevant documents, so it must be good. And then what they did is they um simply uh changed the the paragraph a little bit, put their own name on top and overnight they became visible. Um I I have the whole process also in a few more details but I think I I just explained what is happening. There's no need to go through go over all of this. The important thing is once they changed their website basically overnight they became visible. So suddenly instead of being mentioned in 0% of the answers, they were mentioned in 23% of the answers and no competitor was mentioned more often. So just by making a small change to the website, they became uh one of the market leaders in perplexity. And did that affect the other platforms multi or was this mainly perplexity for this specific client that you were looking at? This very very strong effect was only visible for perplexity. the other LLMs managed to mention them from the beginning. Um, and there was too much there's too much noise in the data to exactly pinpoint what percentage increase this brought them in the other LLMs. Just perplexity was super clearcut from 0% to 20%. Overnight crazy. Such a such a massive uplift so quickly as well. I think many SEOs who are watching this are probably thinking like oh my goodness when we implement things on websites it can take like a week couple of weeks [laughter] for things to really start moving and in AI search we've seen it ourselves with some of our clients that you make a small change and actually it can almost be like next day next couple of days you start to see uplift in results so it's pretty amazing to see yeah there's very little caching going on right now so many many times just for every single prompt answer that's being generated the websites are fetched again. So if there's a change in the source, the answer can change immediately. Um yeah, now a question might be how do we get cited? And uh there also brought some data. So very important this is just correlation data that I'm showing you. Yeah, there is this is not causation necessarily. But if we look at documents that get cited by CHC GPT and compare them to documents that rank on Google, what we can see is documents that get cited by CH GPT, they are much more likely to have a correct semantic headline and sub headline structure. So they have an H1, then H2, then H3. um they are using lists, they are using multiple schema types and very specifically they are using FAQ schema. Now I said it's not necessarily causation, right? we just we only can observe there's a correlation. But if we see what a lot of successful websites are doing and copy it, often times you're doing enough things right, more things right than you're doing wrong. Um, of course it could happen that in the coming months or years we learn that one or two of the two of these are complete crap and they just don't matter at all. Sorry, I hope I can say crap on this podcast. I didn't ask first. Um, but I just want to share that this is the the state of correlation data that we have at the moment. And there are two things where I'm very sure they work because I've repeatedly used them myself and seen other use them. One is summaries and the other is questions. Why are summaries relevant? I I think I mentioned earlier that LLMs when they do this grounding process, they don't take the whole document, right? They take certain chunks of text from these documents and writing summaries forces us as humans to very very precisely and in a condensed manner write about something specific like in a very short way. So if you have a very long text article, it can really help to write a quick summary. By the way, even if you have a very long paragraph, it can sometimes help to write a one or two sentence summary of that paragraph and very important tables and charts. Now people always tell me malt the current level of LLMs they are multimodal what are you talking about they can just understand the table. Yes this is true but in the moment I ask uh Google something simple and it generates an I overview or I ask JTP something and it was to quickly answer it's not going to interpret some graphic or something. It's going to take a text snippet to generate the answer. So what can really help is to not describe the table but put the key message of the table into one or two sentences or do the same for a chart and similar things. And why are the questions and of course when I say questions I always mean questions and answers. Why is that powerful? Again because of the same thing. Often the LLMs use fan out queries that might contain questions or something very similar to a question. And if I write an answer, I'm very likely to have the exact text chunk that they can just use for the citation. So either using sub headlines as questions or using the FAQ schema at the end of an article can really help forcing yourself to write in this manner. Now the this thing that I want to talk about, it works amazingly well. We're going to look at some real examples, but first I want to say I'm embarrassed to show this. I hate the fact that it works so well, but it works so well that I have to show you this and I have ton of examples that show it's working. What I'm talking about are listicles. They are very very often here in this case I asked JGPT what's the best CRM and in the answer they were documents from HubSpot and Zoho. Now both of these are CRM and if you were a journalist and wanted to look for independent information about the best CRM, we would probably pretty quickly come to the conclusion that docu that documentation directly from the CRM providers can have value but it's probably not the number one primary source for information which CRM is the best. However, with LLMs currently, uh the this kind of content can work very very well. And I we already looked at this one here from from Super Office. Um there's also a great one, Nightw Watch wrote an article about the best S10 best SEO tools. And the number one best SEO tool is Nightw Watch. And the second best SEO tool is the browser plug-in from Nightw Watch. Who would have thought that? Um Monday wrote an article about the best the 10 best CRM systems. And of course, the Monday CRM is the best. And FlatLogic wrote an article about the best VIP coding tools. And of course, Flat Logic is the best Vipoding tools. Now, I'm not saying that any of these articles are wrong, right? All of them might be true. I'm sure the authors are convinced that they are true, but this is a bit of a don't hate the player, hate the game situation. Um, these brands are basically giving themselves a medal. And why are they doing it? because listicles are often a really really good match for what the LLMs are looking for with their fenot queries. And these listicles actually work really really well. I'm going to say what everyone's thinking here, which is this is probably short-lived, right? And you're thinking it too. I know multi like how long do we actually have where listicles like this continue to work? So I I find predictions very difficult especially if it's predictive about the future. Um I'm very sure that at every large LLM lab, there is at least already one product manager or one engineer or data scientist thinking about we have to reduce this. And in my 20 years of doing SEO, whenever something worked, was cheap, everybody could do it, it worked reliably, and it worked in almost every industry. People spammed the spammed it a lot and uh this this is also happening here. It's already happening in some industries. I mean, ask perplexity who are the best SEOs in the world. You will find a couple of prominent SEOs, which is often in the system the proxy for good SEOs, which I think is also valid. And then you will find a couple of people who wrote an article about the 10 best SEOs and put themselves on top or on number two. That also sometimes happens, but then usually somebody like Matt Cuts is is number one who maybe shouldn't be number one on these lists. Um, and so yeah, it's the situation is too bad. And if this becomes more more mainstream knowledge and I hope every time I'm on stage somewhere I talk about it it becomes more mainstream knowledge at some points the the LLM providers have to have to fix it and it's very easily fixable. Um it's already less impactful in chat GPT in comparison to perplexity and I overviews where sometimes you have just 10 listicles from 10 CRM companies or 10 SEO agencies. Um but for now it's working. Um but still I wouldn't recommend to bet on it. Um no. And I think a lot of businesses while it's working it people will be like it's worth a crack but eventually the internet will be flooded by listicles and they're not all going to be able to win in AI search as well. So it feels like a very short-term strategy for now. Yeah. And I already know of the first cases where two SEO agencies wrote these listicles and then wrote each other like made an agreement to put each other on the second position. Like of course everybody puts themselves on the first position. Um and I know that there are brands that are making an exchange like one says, "Hey, can't you write top 10 CRM article, put us on number one?" And then we write a top 10 uh payment provider article and put you on number one. uh like this is already being gamed to a degree where something has to happen and it has to happen quickly. Yeah. And it def it definitely will. This feels like very much uh reminds me of the days of keyword stuffing when we're talking about like SEO that's getting a little definitely gray if not a little bit black when it gets there as well. Okay. I can't wait to hear what you've got coming up next about how to avoid not getting cited. Yes. Uh so coming with the double negative to make it maximum confusing. Um when you create content there are and you want to be cited there are two concepts you need to be aware of. One is semantic lettering. If you look at the kind of documents that LLM site it's usually very focused documents like they can be very in-depth or as the essay also like to say holistic. Yeah it can be very long with great details but it has to be focused on one specific topic one specific entity uh not be all over the place. So if you think of uh like recipe bloggers for example, recipe bloggers often write a long article about how they traveled to Italy and reconnected with their grandmother and then they write about the actual recipe. This is not the kind of content that is great for citation SEO in in in LLMs because of the semantic cluttering, right? It's all over the place. And the second thing to avoid are embedding collisions. What does that mean? Um, 20 years ago, you could take a Wikipedia article, you would rewrite every paragraph a little bit, and then you would rewrite every sentence a little bit, and then you would replace some words with synonyms, and tada, you had unique content. And if we compare this character by character, word by word, string by string, it is unique content. But if we build embeddings out of it, if we vectorize it, it actually can become very, very similar. And what you can think about it in a non-mathematical more abstract way is does it contain the exact same information in a various very similar structure in a very similar tone of voice. If yes then you having an article that is identical to a Wikipedia article or another article on a very popular website will always lead to the LLM deciding okay which of these two do I use? ah Wikipedia I don't need you right it will pick in most cases the article on the more popular website so try to always have a unique point of view or some unique information so there's something in your article that makes the LLM decide okay from all of the candidate documents that I could site this is one of the articles I actually want to site and want to use to generate the answer so now you might ask yourself is all of this working and I can tell you Yes, it is. Um, this is an example where I'm tracking a couple of prompts in this like uh website generator vip coding website space and um the the first uh uh table shows us the most visible brands and the second table shows us the website that are being cited the most and you can see that for example web flow what a coincidence they are the most visible brand in the space and they are being cited all the time. Um, and when you look at it, they have articles like 18 no code apps and tools to help build your business. And of course, there's web flow number one. Or if we look at uh this is I don't know how to pronounce them. Blaze tech. I think these are all the listicles from blaze tech in this data set. And they are a lot right and they only appear in my data set if they are being cited by the LLMs. And you can maybe see that the um fav icon from from blaze is also always the first entity that's mentioned in these listicles. So they have uh uh at least 10 plus listicles where they are always number one and these are always being cited and of course that's that's helping them as well. Here's also the highlight for that. Um well this is one from from rippling who are writing a lot of articles where of course always rippling is the first entity that's being mentioned. Um, so that also works well. Um, or here are a couple of like [snorts] the most cited uh listicles for for this data set that I tracked. And you can always see like the brand for that created listical is the number one listed brand. And roughly roughly 40% of of sources in this B2B ZAS space that I'm tracking in one of my research projects are uh from some of the provisticles from these providers. um themsel sorry I formulated that wrong 40% of the listicles that are being used as sources are from the providers themselves of course there also listicles from from independent websites so this this format this content format is really really dominant among the sources now is this the only thing you can do to get site to get recommended more like do you always have to uh have to be cited with your own website no you can of course also look at the current sources and then get active there. Right? So you the idea here is you pick a couple hundred prompts, you run each of them a dozen times, you write down all of the sources and then you look at them. And if you see all your competitors are constantly being cited as a source, then you should look at their content and should think about can you produce similar content in good quality? Can you add something unique? Is it something that fits your brand? If yes, that can actually work very very well. Um, we have some clients who literally take these lists of the most cited content from their competitors, throw that in one of the typical AI content generation tools, uh, generate content, publish it on their website, and it actually works for them. It's the same thing as with statistical. Yes, it works right now, but there's a risk that it stops working. You should also be very careful if this actually fits your brand. So, I don't want to recommend this. I can just say it's one of the things I'm currently observing in the in the market. Now, if most of the sources that you see are editorial websites, of course, the most obvious thing is digital PR, right? You you tell your digital PR agency, please target these five news publishers. They are often used as sources for my prompts. Very easy, very straightforward. But sometimes you can find articles like this that are being cited. It already says advertisement in the title and then when you click on it you will see this is ad actually an advertorial like you can buy an advertorial on some news websites and it will get indexed by Google and Bing. It has a normal URL it's ranking and it can end up as a source in the LLMs. At the beginning of the year I think this was like the number one trick to get more visibility. Nowadays they tend to not work that much anymore but still from time to time these listicles are being picked as the source and of course in a this sorry these listical these advertorials are being used as a source and of course in an advertorial you can just say I'm the only brand and you can write lots of superlatives that you are the best the market leader the best solution for everyone and of course if that gets cited it can have a huge impact on the LLMs um I have seen the first cases where the LLMs then actually say according to a paid article on I don't know CNN or something um so that would of course devalue it but it's still better that the answer says according to a paid article in CNN than citing your competitor right so especially for those marketers that have a big budget this is one of the things that are worth trying out even though it doesn't work as great anymore as it picked in the beginning of the year so we have now two attack vectors for editorial websites, two kinds of interventions to the PR and the advertorials. And there's a third thing because sometimes you will find articles as the source where something is written about affiliate commissions. What is happening here? Most large news publishers, they don't just have a team of journalists writing content and maybe an AI. They also have a content commerce team that is writing articles, creating content. And what you can do is when you see such articles being used as the source, you can give them a call, offer them really really good affiliate commissions and often magically if you pay the highest affiliate commission, you are also the brand that they write about the most or ranked number one in their uh own top lists. And you're not paying them the affiliate commission to get so much traffic from them. you're get very well placed in their content that is then cited by the LLMs. Um, of course you always have to do some research, find the actual articles, publishers, affiliates that are being cited a lot, but then if you can get on those, this can also have a huge impact on your visibility in the LLM answers. So we have actually three different interventions we can run on news and media publishers. Another type of source you will often find regularly is you generate content in social media. Um especially in the US you will have YouTube and Reddit amongst the top sources almost always. Um but there are also spaces like um Gen Alpha Skincare where you will see that actually Tik Tok and Instagram are the number one number one and number two sources. For other topics, you will not even find these in the top hundreds. Um, in the B2B space, LinkedIn is often a source and especially LinkedIn pulse articles can also stick around a really long time as uh like one of the most cited documents. So, for all of these, you of course need to decide, do you want to create your own content? Do you want to work with creators? Um, especially Reddit is a difficult one because most subreddits either hate commercial content or what's even worse, sometimes one of your competitors might already be a moderator and always inject his own brand in answers and delete yours. Um, and of course uh like with Tik Tok, Instagram, if you want to create your own content, that is actually a major effort and you need to do it well. You can't just uh do this as a side project. That that has to be a main effort. It's not something you should do just for visibility and LLMs, but if you have a team that's dealing with it, I think as an SEO, you need to be able to talk to them, recommend some keywords for them to use in the titles, in the descriptions. Um, speaking of keywords in YouTube, um, the whole transcript is available to the LLMs. So, anything that is said in a YouTube video can be cited by the LLMs. But um what is in the title and in the description has much much much more impact. And the LLMs cannot see what is in the video. So they only use the um the text transcript. They don't actually listen to the video within two seconds to to write their transcript. That's that's not happening during normal grounding and web search um processes. And you can find a couple of YouTube videos of me where I ramble like some nonsense and mention some words and then there are different words shown in the video because I specifically tested this. Uh so if you encounter any such videos just know I did it for science. Uh I didn't go crazy. Uh and I hope you will not find them but I fear some people will find them. Um then corporate websites if a lot of corporate websites are amongst the sources um you have to get a bit creative. uh because most companies don't have any incentive to write anything about your company unless you have some kind of relationship like maybe you convince their SEO to write one blog article on their blog and then you write one of one in your blog about each other um or you have some [clears throat] kind of partnership but otherwise these are like super super difficult. Um then e-commerce is a bit of a special case. Of course, for most topics, e-commerce websites are not among the most site domains, but sometimes they are, of course, if you're selling products, right? And what can be very very interesting here is many many online shops are nowadays offering retail media. So, buying advertising on the shop. Anything that's served via an ad server, via JavaScript, we don't care about doesn't help with LLM visibility. But sometimes you can get a blog article as part of a retail media deal or you can influence what is written on a category page in the so category pages often have this SEO text right on the on the bottom and you can maybe get your brand mentioned there for the relevant categories. So it's worth checking out what the retail media options are. Some of them might be beneficial for your AI search visibility. And then yeah, with reference websites, I mean, sometimes you just pay to get like the premium entry to add more details. Sometimes you can motivate your customers to leave uh positive reviews. Um, here it's really just it's about checking are any of these reference websites relevant sources for my prompts? If yes, it's probably always worth it to invest a little bit of time and money there. And now, what metrics should we look at when talking about sources? The number one thing I like to look at is what I call the citation share. So out of all the citations that are happening for my set of prompts, how many of them are coming from my own website, right? This is what I need to need to optimize for and if I can increase this, it is also very very likely that I can increase my brand mentions. And the second is source coverage. So here it's not about me being the source. Here it's about if I look at the tops cited documents in my set of prompts is my brand mentioned there or not and then the second step would be is my brand mentioned uh very prominently and if it's a listical is it mentioned at the beginning but if you can if you can track this and can increase this value every week I'm also very very very sure you will increase your visibility the LLMs will start recommending you more and more often And in this case, there's an article that's cited like 186 times um in in this data set much more often than any other article. If you are not listed in that article, you are immediately at a disadvantage and this is the article you need to get into. And then what can you do to optimize to operationalize this? Um, I c I I was always recommend to look for the kinds of articles where a lot of competitors are mentioned and then get yourself also mentioned in there because often if a lot of competitors already mentioned, you have a very very easy time getting in there. We just saw an article with this example with a Wikipedia article. Um, here's a more general example. And here you can also see like half of these are again uh listicles. So I'm not joking when I say that listicles um are very very common and um yeah this is usually the easiest. If like eight of your competitors are mentioned somewhere it's very very likely you were just forgotten about or if you write them they're like oh yeah thank you for making our article more complete we are going to include you. The interesting thing about this list that you're showing here as well, this one's clearly specifically about web browsers, is so many of these titles are really similar, like the best web browsers 2025 with different variations of that in the title. And a lot of the work that we do with our clients is both what you're sharing here like the reverse engineering basically all of the articles listicles that get cited how do we get into these but also in an example like this where you can see really common like topical themes. I think it gives SEOs and anyone who works in content and PR a bit of inspiration to the kinds of articles you probably could start publishing that might also get cited. And I'm kind of thinking about the new ground that might be trodden there. Actually, some of it will be how do we get the ones that are already cited and how do we get new ones that AI might site because what does your data show multi? What I imagine is actually this list changes like it's not always static. There's going to be new articles, new listicles, new contenders coming in here. Yes, [snorts] absolutely. Um, this list is very far from static and I'm pretty sure if I run this list a month later, there will be three articles will be out of the top 10, three new articles will be in there. Um, some of it is of course because more and more people are starting to realize, hey, this is something we can track, we can measure, we can optimize for. But some of it is also that LLMs have this uh bias for recency. So they want recent documents with recent information which is why many people now update their article once a month and then write the month and the year into the title. And this will get super interesting in January because right now the LLMs very very often when I s when I ask the LLM what's the best browser one of the fan out queries will be best browser 2025 and what we don't know yet is will this change on January 1st right is this hardcoded will it immediately change to 2026 or uh will it take a while for the LLMs to learn this and I'm I already see now people publishing the articles uh best AI search uh SEO 2026 like these articles are already published so that they already rank for these future fan out queries. Um, and I could imagine that around the year change, we will see a lot of changes in the sources and then also a lot of changes in the visibility. And I think this could be like a really crazy period this year. And then of course, somebody will write about it, will blog about it. The people running these these LLM based search and answer engines will think about it. Okay, we got to do this better next year. But I predict that around that time there will be some some changes. And personally, if I worked at OpenAI, I would keep track on everybody who's publishing these best SEO 2026 articles. And these are the domains, the domains. So, I would give them a huge penalty and almost never site them because I can already know they are only publishing this to manipulate the the results. It's interesting though. I think even in SEO, people update their best of 2024, best of 25, best of 26 articles on a fairly regular basis, at least annual. And it feels like this is a very similar sort of uh content refresh that's going to happen, except now when people start to realize this, it's going to be like you say, everyone all at once going after the 2026 ones. [snorts] Yeah, it's going to be going to be a bit crazy, I think. All right. And then I could have made a cut here, but I think we have enough time for one more topic. So I'm going to dive into that one. And that's a question if all of the LLMs give us the same answers based on the same sources. And what I did here, I looked at 10,000 prompts in the B2B ZAS space in the US and I tracked it for 30 days. And we are seeing the top 10 sources from CHPT, from Perplexity, from Google AI overviews. And first thing, there's one article, it's number one or number two in all of these systems. So we could now say, okay, it's all the same. Easy. And by the way, there are other articles that are like in the top 10 for each, but this one is twice in the top four, once barely in the top 10. So maybe it's not all exactly the same. And then there's an article here. It's the second most cited article by CHPT. It's not in the top 10 of the most cited articles for the other two LLMs. So this article has huge impact on CHBT. Very little influence on the others. Maybe zero. Probably it's somewhere below the top 10. And this is not something that is CHB specific. There's also an article that is the number two most cited article by AI overviews. It's not in the top 10 for JBD and perplexity. So there really are differences between these um systems. And then the the very last the very big difference and the very last slide I want to show on this topic is Grock. Um most people ignore Grock but actually it has hundreds of millions of users. Um and grock uses in addition to this web search grounding process also a uh X or Twitter grounding process. So posts on X have a huge impact on Grock answers. Like here I have an example where I ask a question where Grock found one single tweet written by me and uh I believe 25 yeah 25 websites and the one tweet had the biggest impact because it was cited almost word for word to generate the answer. And it's not like they found like 20 tweets saying this, right? found one single post on X and they decided to trust that one as much as the 25 websites that they found. So writing posts on X is the number one thing if you want to influence Grock. Yeah. And that was it from my side. Um if you found any of this interesting, if you don't want to track hundreds of prompts by hand and copy and paste the sources, I happen to work at this company called PKI. We have a software that can do that for you also locally with proxy servers around the world and we can then also help you identify and find for example all the articles where at least four competitors are being mentioned where you are not uh mentioned. Amazing. Thank you so much Multi. This was uh amazing to see some of that actual data and these hidden playbooks as well. And I like that you didn't keep anything uh secret here even though you have some doubts about how long they might last. Uh we've got quite a few questions. Anyone who has more questions, feel free to put them in the chat and we'll do our best to get to them. Um the first one is does CHBT or Perplexity read PDFs on websites and do you see them cited very often in the data? I have seen PDFs used as sources. Um especially Perplexity loves to site PDF documents on university websites. Like in general, Perplexi is a bit of a bias towards university websites, governmental websites, nonprofits, um, and often sites what I would think are rather irrelevant PDF documents like a bachelor thesis or something that yes has something related to the topic, but it's not like it has any kind of actual scientific big breakthrough. It's not being cited a lot in a scientific sense of citations. Um so yes these can be sources actually any document that can rank on Google or Bing can be a source also a txt file can be a source or MD file. Um I have no statistic right now how often PDFs are being cited but on perplexity um especially in topics that can go a little bit in this like uh scientific area regularly. Yeah, I've also seen PDFs cited, but I have to say much less compared to just seeing plain text, like plain text articles cited. So, if anyone's thinking about whether they should be creating more PDFs, I don't think that's necessarily the route to go. But if you happen to have lots of PDFs, it's not going to be a terrible situation. That's for sure. Um, got another one which was about the point you were making about listicles. So earlier when you were speaking about how many listicles there actually are whether those listical pages, this is a sneaky question in itself, whether those listical pages actually need to be publicly visible, whether they need to be indexed, I guess, or whether you've seen any hidden ones actually being referenced. I mean, they need to be index like the URL needs to be publicly accessible, but they don't have to be in the main navigation of your website, right? you can just put them in your in your um XML sitemap and if that is enough if your website is strong enough that that makes them rank that can work. Yes. Okay. So they can be somewhat hidden. I love this question because it's really relatable and actually this is the top upvoted question as well. Uh someone is the first SEO hire at a company. Congratulations to you because that's a big job and that company has years of unoptimized content. Um, is it smarter to publish new content while cleaning up or pause publishing and focus on pruning and fixing all of that old content first? Where would you start if you're thinking about optimizing for AI, even though this is also an SEO question? So if it was purely an SEO question, I would for the first couple of months focus on pruning, deleting, and updating content. Um because if there was content just published for years, you likely already have like I once had a client that had this this exact case. They had eight blog articles about how to tie a tie. And I don't mean simple articles. I mean they casted a model. They got them in the studio. They took pictures. They sometimes took a video. Two or three of these times they produced a YouTube video out of it. Like this cost thousands of dollars every time. And they had done it every single year because every single year someone had the idea, let's do it. And of course like I visited them once. It was like the first thing I showed them and the next day they deleted seven of those and and merged everything into one and were a bit embarrassed uh uh that they did it. Um, so I think from a pure SEO perspective, content pruning, updating it, getting rid of stuff you don't need anymore, combining stuff that should be combined should be the focus. Because of the AI search uh, effort, I would probably still try to publish a couple of articles specifically to be cited in AI search. Um, just to build that muzzle and see if it's working. But I would do like maybe five of those and then do an audit of the next 100 existing articles and see if you can get rid of something or delete something. Yeah, I would say especially if we're talking about potentially thousands depending on how big this company is, thousands of blog posts or old articles as well. There's going to be a lot in there I would guess that probably isn't performing and actually you probably want to get into your analytics data to find the ones that aren't driving anything. And then there's going to be a lot that you are thinking about refreshing anyway, particularly if anything is a 2025 guide or previous that might be coming up for 2026 as well. Um, okay. Next question we have is about video and a company who makes a lot of thought leadership videos, which is awesome to hear. Does AI pick up those kinds of videos and would it be able to relate it back to that company if these videos are published on YouTube? Yes. And the LLMs also understand what's being said in the video like from the uh from the captions. Um if you publish the videos anywhere else, especially self-hosted, I would have huge doubt that the LLMs understand that. And then I would put the the transcript uh on on a document on your website where the video is then embedded. Otherwise, there's a huge risk that this is just being lost. And if it's short form videos like reals or Tik Tok, make sure the important stuff is in the in the title or the caption and you should be good to go. I have also very rarely seen videos coming up in AI answers as well. I think I can count on one hand the number of times I've seen a video surfaced in a chatbt answer in like the last four or five months. So when we're when we're thinking about what's more important to publish like lots of videos for AI or write more text, I think because it just takes so much less resource for AI platforms to read text, they're going to do that. That's just much less intensive for them to compute it. The one video I did see actually was a Greenpeace video and I think it must have been so relevant to the search query that I'd put in that it the LLM must have felt like it had to surface that piece of content where it was quite a niche niche query. I don't know how often you've seen in B2B SAS though videos actually being surfaced in in chat GPT very rarely in perplexity and I overviews and also AI mode all the time like YouTube is one of the top three domains in both Perplexity and I overviews um with AI overviews maybe there's like a Google connection uh behind it and Perplexity really seems to love um YouTube um I guess they like these transcripts files where they can easily get access to a lot of text all at once. Perfect. And then I'm going to end us on a final question. I promise I haven't planted it. Is there a tool that you might recommend for tracking a brand's visibility in LLMs? I can't think what it could be. So, you're going to not believe this, but yes, I can actually recommend one. Uh, it's Pai uh where I'm working [laughter] uh and where I'm try to build the the best tool for exactly this use case. Okay. Amazing. That's perfect. Uh, we also had a whole flurry of extra questions off the back of some of those. Um, have you seen slideshares being cited? Um, not to my knowledge. I would have to check. Okay. I also haven't seen any slideshares being cited myself either. So, uh, sources are often easily manipulated. This is kind of what you've been uh, getting at with the listicles point as well. How would you recommend preparing for the next LLM updates? UGC feels like a clear place to go, but how important do you think content marketing, digital PR are going to be? I think digital PR and content marketing are very important because if you can get reputable websites to write about you positively, that must have an impact, right? And if it's a thousand users on Reddit always upvoting your stuff because they love you or if it's journalists writing about you um that must have an impact like LLMs cannot ignore that and uh they will try to find ways to identify where this is because of manipulation but if it's because there's actually a consensus in the market that you are the best brand the best solution they will want to understand that and then these websites are the place where they can understand that. So community management, content marketing, GPR are here to stay and probably going to be more important in 2 years than they were 2 years ago. I would completely agree from some of the case studies that we've done for AI search work for our clients as well. Digital PR has literally increased the amount of traffic coming to a website from chat GBT and Plexity through specific published articles within that niche. So, it's a very tangible increase in results when those articles are really well placed, which goes back to exactly what you were speaking about earlier, Moltu, with looking at reverse engineering what is being cited and then trying to find the topics that are really interesting as well. Um, you're going to like this one. What's the best way to build a list of prompts that are actually likely to be searched by people? Yeah, that's a bit of a difficult one. Um, of course at PKI we are also trying to solve this for our clients. Um, we because we have access to a big data set of real user prompts where people knowingly or unknowingly have a browser plug-in installed that reads what they are prompting and then they are sharing that and there are companies that aggregate and sell that data and we are using that to understand what real prompts look like and then make prompt suggestions that are as good as possible. But almost every single prompt is unique. If I look in these data sets and look for prompts that have actual volume, it's like, "Thank you. Please try again. Hello, by almost everything else is unique." And um because people tend to write prompts that are much longer than traditional searches, people make typos, have different ways to express themselves. So even if you knew the exact prompt that one of your customers used yesterday, probably the next 20 days, it will never ever be used. maybe next 20 months it will never be used again. But LLMs are very very good at doing something that for us humans gives the impression that they understand the topic, the goal and the context. So it doesn't really matter to have the right the right prompt word for word. In the B2B space, I always recommend to look at your call transcripts from prospects and customers. Um, in the US this is usually usually easier than the European Union. But if you have tools like granola, uh, fixer, gong, io, etc., etc., just look at what are the exact words that your potential customers or existing customers are using. What is the language that you're using in your in your intercom support tickets, Zenes tickets, and build your prompts based on that because your customers are your best source of how do they express themselves. If you have a very large B2C website, look at your internal search. What are people typing into that search bar can also be super valuable as as an SEO? You should look at internal search bars anyways. They are one of the best source to generate landing pages, but especially also for AI search and for generating these prompts. Try to understand the language of your users and Google search console is a good source as well. Of course, always love Google search console. I've got a final two questions. One is uh off the back of what you've just answered there. Biocsychology. How helpful is that for thinking about the prompts that you want to track? So, I cannot give a numeric score how helpful this is. Uh, I would answer with yes, it is helpful. But yeah, I would say so too. The more the better you understand your personas and what they're likely to be asking, I think the more accurate your prompts are going to be, there's always a little bit of uh personal feeling about it. And a lot of the work that we do at Exposure Ninja when it comes to like bio psychology and who those personas are is actually having conversations with your customers and finding out what they're asking because it gives you a much better idea of the wording that they would use when they're writing prompts which kind of speaks to exactly what you were saying multi that actually people write in their own language their own way of saying something for sure. And then our very final one. Okay, it's a bit long. From an SEO and an AI perspective, they've got a lot of articles published on this website in 2025 and 2024. Would you update publication dates and add summaries even if the new date isn't in the title? So even if it's not a best of 2025 list. So this was recently tested by someone and just changing the publication date already works with increasing visibility in AI tools. This is again of one of these things that I hate. Yeah, this shouldn't work but right now it works. So if you just go through all of your articles and change the publication date to yesterday, you will immediately get cited more often and you should immediately see more visibility of your brand. Of course, this cannot work forever and of [snorts] course this is something that is not just seen by CH GPT and this could lead to for example Google putting less trust in your publication date. So personally if I work on a website with like an established brand that is supposed to rank long term I would not lie about my publication dates. Now if you add a summary I think it is totally fine to change the publication date or the last updated date and then you hide the publication date or whatever. Uh so I think that is fine. That is not spam. I would probably not add the summary to my 10,000 articles all on one day. That could again give uh weird impressions. Um, unless Google is crawling your website very slowly, then you can do it all on one day because it's going to take Google a months to actually rec crawl all these pages and see it. Um, and yeah, this this of like would I do it like as I said, there are benefits to it. There are also risks to it. Um, every market has to decide for their own. One of the clients that we worked with in the financial space as well, we updated all of their articles with actually personal summaries from a specific spokesperson in the business who gave almost like an executive summary, like an opinion-based piece at the top of the articles. But we did that over time, not all just in one week through them up, but actually started updating the content, then adding that summary from that that person's perspective and saw that it really influenced AI surfacing certain chunks of the content in there. So I think you can also do this in a way that feels like relevant to the business and relevant to the article especially if you happen to have something new to say which quite a lot of the time if the article was published six months or more ago most businesses tend to have an update that's going to be relevant and I think that's a much better opportunity for making sure that content is likely to be surfaced in AI rather than just like switch the publication dates all to the newest date all the time which is definitely going to start to look look suspect. And I don't think anyone wants to do that. Um, that's it for all of our questions. We had so many questions and if you have more questions or anything comes to mind that you would like to ask either myself or Malti, you can find us on LinkedIn and we're always more than happy to have a chat about anything that you need. If you need an AI visibility tracking tool, then Peak AI is the place to go. Uh, you therefore don't have to track all of your prompts manually. And it also suggests for you the kind of prompts that you're going to want to track for your business as well to give you a starting point. Plus, you can reverse engineer all of the citations relating to the prompts you do care about as well. If you are thinking, "Oh my goodness, there's a lot to do here and I'm not sure where to start." Then I work for Exposure Ninja and we build and implement AI strategies for our clients. We use tools just like Peak AI, which we absolutely love to do all of these things across technical SEO, website content and structure, deciding the positioning that your business needs to have to show how you want to be shown in AI and the PR and AI citation analysis trends and building side to make sure that you have the visibility and the sentiment that you want AI to have about your business. Uh you can find us at exposioninja.com. You can also find us on our YouTube channel where we talk about AI search all of the time, including featuring tools like PKI and exactly how they work in our workflows as well. Uh, thank you so much for joining us today and Malti, thank you so so much for sharing all of your data and expertise with us as well. It was amazing to have you. Thank you for inviting me and thanks for all the interesting questions. Amazing. See you later everyone. Thanks for joining. I buy.