What You Need to Know About Google AI Mode | AirOps & Mark Williams-Cook

Hello everyone, this is Josh Filker. I'm content and SEO at Airopps, and it is my pleasure to be joined today. Mark Williams-Cook, how are you?

Hello, very good, thank you. Very pleased to be here.

Yes, excited that you're here. We've had a lot of interest on AI mode, Google changes, all of that. So very excited for that. And a tradition here that we do at Airopps is if, if this is your first time here, please drop a Star emoji in the chat. So everyone look up the star. Star equals first time. And then if you've been here with us more than once, put a rocket in the chat. So 2x plus. All right, who we got? We got a lot of new folks. Thanks, Omer. Good to see you, Matthew Bushnell. Thanks for joining us. Oh, Jeff, 3 rockets. So that means you've been with us 3 times. That's awesome.

Vlade, thanks.

Geraldine, welcome back. I think I remember you from last time, if I'm not mistaken. So glad you are here. And this is a little early for our US folks on the West Coast, but Mark, you are over in the UK, right?

Yeah. So it's just gone 3 o'clock here.

Okay. And if you don't mind sharing, you were telling me a little bit about your after-work plans, which I thought sounded really fun.

Yeah, it's on the down low. So I'm boxing after work. I've got a bout coming up next month. A lot of people don't put that together with the nerdy side of me.

Yes, that's like such a fun activity.

I've always wanted to do like some training and stuff and just like haven't done it yet. Uh, but that's really exciting. Yeah, for sure.

Um, so we are talking today. Hello from Philadelphia. Hi, Robert. Uh, we are talking today about how to master the new search and with AI mode, lots of changes going on with different models along with Google and AI mode. So, um, Mark, if you would let me set it up here a bit. So we all know Mark Williams-Cook. He is the director at Candor. He is the founder of Also Asked, and he also is the co-owner of Top Dog Harnesses. What is the— what's Top Dog Harnesses?

Don't talk about that too much.

Okay, great.

Yeah, it's a small UK-run e-commerce business that sells dog harnesses, dog leads, like designer ones. We bought that business just after COVID because it had quite a good following, but Um, the SEO was terrible. So I was like, I know someone that can do that.

Yes, exactly. That's always like hearing about people's side businesses too. Um, and then of course, something that everybody needs to get on is the creator. He's the creator of the Core Updates SEO newsletter. Very helpful for keeping up with a lot of the stuff that we're going to be chatting about here today.

So that's, that's also a good signup.

Uh, we will drop that SEO newsletter in the, in the chat too. Um, Let's go set our agenda for today. Oh, before we do that, let's talk a little bit about AirOps. We want you to win AI search before your competitors. We have a new LLM visibility dashboard that shows you exactly where you are appearing and opportunities for you to take advantage of. You can track and audit your visibility footprint, act with workflows. You don't only get visibility, you also get an action layer with that that includes our grids, that you can take action with our content workflow grids.

I will be showing a little bit about that at the end.

And then along with that, you can prove pipeline to the C-suite. So we are rolling this out, our new visibility layer, to select folks request only. So to get early access, go ahead and drop us an email at growth@airops.com. You can put webinar in the subject line, you can put, um, LLM visibility in the subject line, and we will take care of you. Or you can scan that QR code. So Really happy to have that out. You can go to the next slide. Yes, our agenda today, myths, misconceptions, and maybes about AEO, GEO, AI search versus SEO, how LLMs are using schema, and then AI search and links. Talking a lot about keyword research and prediction and requests. Then we will have time for Q&A and takeaways at the end. Very excited to hear what you have for us today, Mark. Feel free to take it away.

Thank you so much. Yeah. So I don't know if people saw the news as well today, just before I got on this webinar, I saw that Google AI Mode is now rolling out to, I think it's like 180 countries with some extra agentic stuff. So really well timed on the subject, I guess, because it's not going away. And I think it's pretty clear that AI mode or similar will likely, I think, long-term become the primary destination for people on Google. I feel doing old-school Google searches will be a little bit like, for those my age or around my age, when you used to go to the library and you had to kind of crawl around on your hands and knees trying to find the right number of books, you know, just, it's just not going to happen. So I wanted to start just with, we're going to mainly talk about kind of keyword research and how that's changed with AI mode. I wanted to cover off a few misconceptions around AI mode, just some interesting things that I see people talking about. Obviously, there is like a massive geo SEO debate thing that's already pretty tiring. I have some fun with it. You've probably seen I have— I posted this meme like a couple of weeks ago about geo basically just kind of under the hood being SEO, which some people really agree with, and some people really don't, and some people really don't care. But I just did want to take a little second to talk about that. Because obviously, like with many things, there is some nuance to it. So if we think about what we do, how people interact with search in 3 kinds of blocks of things, which is how people interact with it, how we optimise for it, and how we measure it. So When I talk about interaction, I'm thinking about how do users actually do searches? What are their behavior patterns? And what paradigms does this establish? They're important to understand from a marketing point of view. The optimization part for me is what are the actual things I need to go and do for my job to optimize our exposure for this surface, whether it's Google, whether it's ChatGPT, whether it's YouTube, whatever. And then of course, measurement. How do we measure all the different aspects of this, whether it's visibility? How do we attribute revenue? Or how do we just measure success? And when I say I think geo is still mainly SEO, it's because I'm talking about the O part, the optimization part. In terms of interaction, there is very obviously massive changes. You know, the interaction has become conversational, and it's become multi-parameter in terms of if you do a classic Google search, you can pretty much only search for like one thing at a time. Whereas when you're interacting with a chatbot, you can specify, okay, you know, I want to buy this and it needs to be within this price range. And I like this color and this kind of thing. And it will try and, you know, mold an answer around all of those parameters. People can explore more of your brand, your content, your product, your services off your site as well. It happens inside the LLM, completely different way of interacting. And personalization, will become an even bigger thing. I mean, we have some elements of personalization we've had for a long time. But with LLMs, obviously, it's like next level. Huge changes as well with measurement. The big one is obviously people freaking out that we don't get as many clicks to count anymore, because we love counting clicks as SEOs. There's less search parity. And what I mean by that is in terms of what users type in. So trying to work out what are they actually searching for. Because the search box in Google is quite a primitive thing. You know, when you say people have like Google fu, like they know how to get a good result from Google, right? It's because they're good at translating what they want into a search query. That's not happening with LLMs. People are just having normal conversations like we're having now, right? Anyone can do it. So there's less parity. There's obviously no easy ranking in terms of we don't have a dashboard that tells us from ChatGPT or AI mode where we're coming up like we used to with SERPs. And the funnel's completely changed. It's not, you know, people discover you, they click on your site, and then you convert them. In my opinion, though, the optimization side is actually pretty similar. There's been a rebalance of the kind of tactics and things we do. But that has happened for years in SEO, right? Things become more and less important. So I'd say if we're talking about AI surfaces, like coverage on other sites is way more important than it used to be, but it's still the same kind of thing in terms of, you know, getting other people to talk about you, getting digital PR coverage. Some of the technical requirements are a bit firmer now. You know, we've seen it to death about LLMs not being able to render JavaScript, or some of them at least. But again, like, for me, that's been a pretty basic kind of SEO requirement would be I always love it if we can see all of the main content and links without JavaScript. It always makes me a bit nervous if we need it. And there's some other, again, I think smaller changes, like certainly for certain types of AI results, the recency of the content is more important than it used to be. I think sentiment will likely become more important. So, but again, these aren't outside the scope of SEO. So I will relent in some way that there are new things we need to be thinking about in terms of AI search. But the core of the actual actions you should be doing to improve your coverage, I don't feel have changed massively. And this leads me kind of neatly onto this schema question, right? Because the number one, I think, most common piece of advice I have seen in terms of optimizing for AI mode and optimizing for geo is put schema on your site because LLMs, you know, use it, don't use it. And I think there's a lot of confusion there. So Google and Google AI mode, I think, is quite different in many ways to contenders like Perplexity and ChatGPT, because they've got 20 years of metrics, databases, all sorts of things that they can use with Gemini. So, you know, absolutely, Google AI Mode leans on Google's existing knowledge graph. And one of the ways you can kind of help yourself get into that knowledge graph and be understood is with schema. However, I do think it's a mistake when people are saying that putting schema on your web pages, that that's used in the training data for large language models. And the reason primarily is because of the way they're literally working. So the process that we go through when LLMs are trained on content, some of you may know, is called tokenization, which basically means when they take all of this text, and they break it up into chunks and make unique tokens for these common chunks. So GPT-4o supports about 200,000 of these tokens. And I put in some— I think that's from, yeah, the 1984 novel, some text there. And it gives you an example of how that text is broken up. And each one of those is assigned a number that looks like this. And when the neural network training happens, it's basically taking big windows, trillions of these combinations of windows, and trying to predict which token comes next. And it's doing this, like, just a mind-boggling amount of times. And what you get from that is actually some order from chaos. It's amazing. We can actually, you know, within the graph of human language we've got, there is so much knowledge that we can extract. And it's a fascinating way of doing that. It's very different from the kind of model of old school crawlers that would go and look at URLs and HTML and try and piece things together there. The reason why I'm saying schema doesn't go in there is firstly, any course you do on like this kind of stuff, any tutorials you watch, if you look at the actual datasets like FineWeb, you'll see they talk about in pre-training they'll generally try and strip out the code, right? So any HTML, they just want the human text. That's like one of the parts of pre-training. So I think it's unlikely that even the kind of code parts of websites get into that training data because I don't think it would have a positive impact on what you're trying to achieve. But even if it did, The tokenization process sort of destroys the schema in terms of schema exists to make this key-value pair, um, relationship between things and to remove ambiguity during tokenization. Tokenization is lossy, so we're losing fidelity of the information. You can see here, like, um, content and type are broken into the actual words and tokenized, which means they're indistinguishable from the actual, uh, just words themselves. This isn't read as schema. What it means is that LLMs will come across schema on the web anyway, um, in tutorials, in documentation, in GitHub repositories. So they will be able to reproduce schema when you ask them to, but it's the equivalent of learning like, uh, something you hear on the radio in another language and being able to repeat it but not understand it. So the schema, you can generate it, but it's not being used at that training level, at least in that long-term memory. It's not being read in the same way that Google absorbs schema in the Knowledge Graph. It will, if it does go in, it will be in the kind of, in the parameters, in the tuning, but it's not going to be there saying, okay, this person works for this company, and this is their website, that information is lost. So I just don't see from a technical standpoint, unless there is a whole separate process, which there could be, but I haven't seen any evidence for that. It's just not possible to have it in at this level. So while I'm certainly saying, you know, I'm not saying don't use schema, because schema is helpful.

It's a great question.

You know, it can give you rich results in Google.

Yeah.

I'm just saying, if, okay, your mind is we want to do well in generative engine optimization, if you want to call it that, or AI services, it just isn't the top thing on my list because it's just not how they work. There's other things you should be doing. And the other kind of myth I hear is that AI search, whether it's again ChatGPT or AI mode, don't care about links. That, you know, it's just about language. And again, that to me just isn't— doesn't ring true. So firstly, from the declassified DOJ, Department of Justice documents, there's been loads of these surfaced from the Google trial. We can see where specifically they've sent emails talking about Gemini pre-training, which is when they're taking, you know, building that long-term memory, building the model. They're using the, what you'd call traditional Google link graph metrics and click data to help weight that model. So all of these metrics they've built, you know, and it makes sense, right? Because they've got all this data over many, many years. And they're applying that to their models that they're building. And I don't know how many people on this webinar will have seen. So last March, we reported a Google exploit. which they gave us the top bug bounty for. And this was a pretty big deal in terms of we uncovered 2,000 properties that surfaced in Google Search about websites, query classifiers. So we had things like site scoring metrics, information retrieval scores, SERP feature information, debug information, query classifiers. And we didn't just have the names for those things like we did in the Content Warehouse leak. We had the actual values. that Google was assigning to different sites. And when I went and gave this kind of in a presentation, because I did a talk about this Google exploit, a Googler watched it and said to me, it was the least wrong explanation he's seen of the systems I was talking about. And I'm very proud of that.

Love that.

So it's obviously still a bit wrong. But yeah, what I kind of want to communicate is this This is a terrifying image, right? But this was done by Dr. Maro Fischer. This is a pull together of everything we kind of have seen from DOJ, from patents, from leaks about how Google works. Basically, the point is, it's a whole bunch of microservices, right, that talk to each other and pass values between each other. It's not just like there's one algorithm that does stuff. And I feel that's almost what people think AI mode and these LLMs in terms of search are, you know, they are part of other systems that are feeding them information as well. So to massively oversimplify this for the amount of time we've got and how good I'm explaining things, in terms of when Google's building a SERP, we have kind of their core metrics, which take a long time to compute things like PageRank, where they have to go around the whole web to do a PageRank calculation, and it updates slowly. And they'll build out these classic, you know, the old school 10 blue links, like which documents should rank for these queries. That's what Google refers to, very oversimplified in terms of ranking, right? What we actually see on the front end happens after we've been through various processes like re-ranking, which can happen based on the query. And obviously, Google doesn't know the query until it's received it from the user. It's at this stage we deal with things like twiddlers. If you haven't read up on them, there's really great, again, leaked documentation on how Google uses twiddlers, which is internally— basically, twiddlers are like specific standalone bits of code that run conditionally. So if it's like, okay, this is a video result, there might be a video twiddler, which uplifts certain sites, for instance, for those types of terms. But they'll re-rank results based on queries. We go through re-ranking, we have feature generation, because all of that's done after all of this. So they determine based on the query whether they need featured snippets, PAAs, maps, or AI overviews is a feature. And then there's a whole system that literally glues all of that together and makes the SERP you see, right? So why is this important? This is important because one of the things we found in the Google exploit was they literally had a metric called site quality, which obviously I was super interested in because people talk about quality all the time. It was a metric that went from 0 to 1. It had a normal distribution, meaning most of the sites were in the kind of 0.4 to 0.6 bell curve. And that score was assigned on a subdomain basis, which again, I found super interesting, right? So if you have a www.yourwebsite and you have a forum.yourwebsite, they actually get different scores, right? Now, what we accidentally discovered based on this score was You are not eligible for things like featured results, like people also ask, like featured snippets. If your site quality was below 0.4, you're just not in the running, right? This is this kind of stuff applies to things like AIOs because Google has all this data, so they know if your site is not trusted, it's not of this quality, you're just not going to even be in the running. So I found it interesting, like. previously when we work with clients and they're like, oh, we haven't got any featured snippets, we're trying to rewrite our content to do this, this. It was nothing to do with the content. The content wasn't the reason you haven't got the featured snippets, it's because you've fallen at this hurdle that's 3 races back and you're not even in this race. So when it comes to AI mode, there's all this background apparatus that Google has that you need to qualify for before you're even going to be in that, in that running. So a lot of, especially the AIOs, are certainly built on these traditional ranking metrics. And I think, you know, Google even announced it today when they were talking about AI mode, using things like their Knowledge Graph, it all leans back on this other infrastructure, which is all SEO that you've currently been doing, should have been doing, assuming you're doing it in a good way. The other thing nobody talks about is The major LLMs are buying link graph data hand over fist from the big SaaS providers. Just talk to someone that works there. And hopefully, if you buy them a beer, they might tell you about it. But this is important because from a purely language graph perspective, there's no way to judge this is a good source, this is a bad source. So it's really important for them. So I've put together a little flow about kind of keyword. I'm just going to call it that. I know we're beyond keywords. We have been for a long time. Give me a break. I'm just calling it that because it's the term everyone uses. Yeah, it's about how we're approaching keyword research kind of in this AI era. And these are the problems we're going to tackle, which is what do we do with our big lists of traditional keywords? What do we do about personalization of assistant conversations? How do we map likely conversations? And how do we work out which searches we can actually impact? And this is particularly interesting because again, a few days ago, we had this interesting study from Louise from Ahrefs saying that only 12% of the AI-cited URLs that rank in Google's top 10 rank for the original prompt. So there's a difference there. If you do a Google search and then you search for that same thing in like an AI mode, you're not going to get the same results. And for me, it's kind of obvious why that's happening, right? So before we had AI mode, people did searches for traditional terms and everyone has their own way of searching, right? And there was money to be made understanding this part and understanding what you need to rank for. It was kind of hard in some ways because basic searches could mean different things to different people. And we see that a lot, especially with the broader terms, you know, we can type in the same query, but we want different things. And conversely, there are many, many ways to express the same intent, we can just word it differently. What we've got now is obviously people are having these full conversations with LLMs. And then LLMs are— LLMs have decided what searches to do in the background. And they do that with relative parity, meaning that they don't seem to be as wild as people in terms of choosing what to search for. I think of the LLM as almost like a universal intent decoder. So it doesn't matter how you ask the question, the LLM kind of gets it from its language graph. And it's like, okay, well, I need to search for this to kind of get that piece of information, then I can word it for you. And Essentially, the money-making part has shifted down to understanding what's happening here, because this is what's feeding the LLM. So how can we go about doing this? So this is a little flow that we're working on. There's a couple of different ways, several different ways you can do this. This is one of the easier ways. There is like a hardcore mode which involves training your own models. But I guess we can talk about that in Q&A. So if you have your traditional keyword list, which will look maybe something like this. It doesn't look like conversations. What we want to do is try and give them firstly some kind of personalization context. Why that's important is if you have a chat with ChatGPT, for instance, and you ask it for some vegan recipes because you're vegan, and then a week later, you ask it for some running shoes, it remembers that you're vegan. So the searches it's done then are specifically for like cruelty-free brands or brands that don't use leather. So there is this layer of personalization that we've never had before. Now, there isn't a perfect way to do this. It's very difficult. One of the ways that we've been doing is essentially describing your ideal customer profile to ChatGPT before asking it what kind of searches would someone maybe use with you in a conversation for this information. You can either do that with a custom GPT if you're programming at Verse, or you can just use the API and kind of bundle that information along. You could have 20 different profiles if you like to get a spread of the types of people you're dealing with. You don't have to have one monolithic, like, persona. But this will give you an idea of the spread of queries and the ways that people might type that into chat. GPT or Google AI mode. So you'll get normally like 4 or 5 prompts it will give you per query, and it'll be combinations of how people find that information. So like, what's the best running shoes for a beginner, or specifically for a vegan, or specifically I've told it it's a middle-aged person like me. The other thing you can do is try and map the rest of the conversation further down the funnel. So again, obviously kindly introduced, Also Asked is obviously our tool to do this, which maps People Also Ask data. So what People Also Ask data is doing is essentially giving you the next most likely question someone's going to ask from how Google has clustered that data. So if someone's looking at, they've got how much they want to spend on running shoes, you know, they're going to be asking questions like, How much quicker do good running shoes make you? So, you know, if I spend £200 instead of £50, am I going to be running faster? And then there's questions about specific brands like Lifespan. So these are all of the related questions that Google knows people are likely to ask. And these are the kind of things I'd expect to see in a, in like a fat, in a query fanout. Now, there are ways that you can build models to try and replicate what Google's doing with query fanout. At the moment, as far as I'm aware, nobody has actually done any kind of solid testing on those. So this is at the moment the way that we're approaching getting these queries. If you want to do this through the Also Asked API, it literally writes the code for— the documentation writes the code for you. So you just go and you choose what language you want and it generates it. So it's very simple. The ChatGPT stuff as well, you can vibe code that. It's very, very straightforward. It's all documented. What this does is it will give you a list of kind of candidate prompts, candidate conversations. So from these traditional keywords, if you had say 10 keywords, you might end up with like 100 or 200 possible conversations that people might be starting, um, with the assistant. And what we want to do next is find out if the LLM decides those queries need grounding. And by grounding, I mean if it needs to go and do a web search to check the answer. This is, to me, one of the most vitally important steps, because if it doesn't require grounding, you are going to get an in-model answer, meaning, meaning that it just is confident it knows the answer to your question. So it's just going to tell you with no external sources, which basically means you can't influence that result unless you somehow work your way over many months into the model and they release a new model. Fortunately for us, the majority of kind of queries that people are likely to be interested in do seem to trigger some kind of grounding. So that's either, I have no idea what information that is, or I need to check what the latest information is. Now, different models have different thresholds for when they go and do this grounding. Even ChatGPT, its grounding changes whether you're on a paid or free version. So the paid versions ground a lot more. I assume there's a cost involved in there. So we're trying to work out which conversations are not around what I would call, or Jono coined, the first person I heard it called, that was solve knowledge. So this is what we want to find. Now, we've known, so this is just a dive into the developer console. When we've done a search in ChatGPT, you can actually see I've highlighted it here. It will give you the probability of it needing grounding. So this is saying this query I did, I can't remember what it was like, oh, best SEO experts in 2025, it's saying, yeah, I'm like 3 quarters sure I'm going to need to ground this. And it seems that their threshold is around about 60% that they will go and do a search. We used to actually get the queries in the background that happened in ChatGPT. So I published that about a month ago. Since ChatGPT-5, though, they don't seem to be in there anymore. Sad violin. So what we want to do is we've got our candidate prompts, we want to do a grounding prediction. There are a few ways you can do this. Dan Petrovic, probably the leading person in terms of playing around with machine learning stuff in the SEO space at the moment, go follow him on LinkedIn. He's released a couple of tools. grounding.dejan.ai will give you the prediction that a ChatGPT query will need grounding. He's also got one for Gemini, which is, you know, Google AI mode. So you can put in these prompts, these search terms, and it will give you a confidence it needs grounding. So what do red blood cells do? Doesn't need grounding. I'm 98% sure of that. I know the answer. What happened in the news today? I'm incredibly sure I need to go and do a web search for that. Dan's also released these models. So if you're a bit more adventurous, you can actually go on to Hugging Face, which is an AI community, and download the model. And run it locally yourself. This is cool because it's basically a free way to do it. You can just run it if you've got even like a fairly modest GPU in your laptop PC, you can just run these yourself and compile your list. So firstly, that's super valuable. So immediately, we've found from these traditional topics, likely what people are searching for. And then we know which ones the LLMs are going to go and actually do a search for, which is Super important because quite frankly, if you're spending time now trying to get stuff to rank that LLMs and AI search is answering in model, you are wasting your time because they're not looking for you. They're like, we know the answer. You know, there's only so many beef lasagna recipes the human species needs, right? So the next thing we can do then is grounding requests. So we've got our list of conversations, prompts that we know need grounding. And one of the ways we can still get this data is Google does have a grounding API you can use with Gemini. Now, again, caveat, the Gemini model they use for AI mode is a custom model. I think it was Go Fish Digital did some work comparing the queries done or the sites at least that were found in the grounding API, which is this versus AI mode. And their conclusion was they're pretty similar at the moment. There isn't a huge difference. So this is, as far as I know, the best we can get. Again, there's full code on that URL. You don't need to be like a really good coder to do this. But all we're going to do is pass those queries to this grounding API. And one of the bits of information it will give us back is what are the web searches that have happened to ground this information. And this, to me, is the, at the moment, the kind of gold dust of, if you want to call it, geo or AI search, right? So this is what the assistant, the the chat assistant, the LLM, is doing a search for. And this is how it's going to— what it's going to model its answer around. Interestingly, what I found is we take our list of traditional search terms. Once we've kind of fanned those out, we get this massive list. Once we go back to this grounding data, the list shrinks. Because as I said, the AI agents seem to be fairly consistent in knowing what they want to translate into an actual search. So the target becomes smaller again. So this leaves us kind of at the end of our process here, which is, at the end of doing this, you will essentially have an almost traditional list of if you want to be seen for these kinds of conversations, this is the thing you need to traditionally rank for, whether it's in Google or Bing or traditional search. It's stuff we know how to optimize for because we know how to get search engines to rank us for these things. And then, of course, we won't go into it now. But you can do your kind of standard gap analysis of do we have that content? Is it up to date? I guess the final thing I would say about getting visibility there is the end goal is probably not, we have to get our site to rank for that query. It would be we need our product, our service, our brand mentioned on as many of those search results as possible. So if there are lists of, you know, these are the companies that do this, or if there are reviews or roundups or whatever it is, I would just, I would focus now more on being present, omnipresent across those results rather than worrying too much because your site basically doesn't matter anymore, right? It's just, you need to have your information retrieved in that search. And it doesn't matter if it's on someone else's site or whether it's wherever. that's what's going to feed the answer that the user's seeing. So I guess takeaways is, as I said, lean more heavily into being present in results than ranking your own site. I think traditional search in the way we've been talking about is going to stay for a long time with the current paradigm of how the LLMs work. We've kind of touched on that. They have, you know, by their very nature, they have limitations that make them not good at search. And finding those commonalities in the live background searches will be key. This is a basic approach. As I've said, there are more model-heavy approaches. I saw a couple of people in chat there talking about— Dan, there you go, he's been linked. So he's building like models to do query fan-out, which are trying to be faithful to how the patent works. Again, I don't think he's kind of at a production stage yet, but there are certainly kind of more tech ways you could do this. Hands up, I don't think we can fully account for personalization, not that we ever could really. One thing I learned about building Also Asked is Google's taking in longitude, latitude with every search we're doing. And that's how it works out the kind of localization on a city level. We're not accounting for that even in traditional search, and we haven't been for years. Same with devices. So I think there's an interesting conversation around personalization anyway. And yeah, I'd say watch this space because I think there's going to be a lot of changes. I do see the longer-term future, as I've said before, being more agentic in terms of you are just expressing your wishes to an agent and the agent knows you and your preferences. And people will outsource their decision-making to them, for better or worse. That's a whole topic, right?

Wow, that was incredible. Thank you so much. Lots of questions coming in for sure. Keep them coming. One that I had that came up— oh wait, yes, first, we will get to the questions here in a second. I wanted to talk a little bit about AROps, give Mark here a break. So a lot of what we're seeing with what Mark said is like, We're appearing in LLMs, we're trying to optimize for, for LLMs or for Google in AI mode, but how do you really take action with that? And so Airops is one of the ways that you can take action to get to your pipeline. And one of our customers, Descript, before Airops, they had a lot of declining organic traffic, low content quality, high operational overhead. And then after Airops, we saw their signups increase, we saw their non-branded SEO increase, we saw their LLM visibility increase. In only 30 days using AirOps workflows and visibility. If you could go to the next slide. And so AirOps is your way to see how you're appearing in LLMs, but then also take action on it, which makes us a lot different. So you can prioritize your opportunities. We show you opportunities where you can tackle those right away to help you appear in LLMs. You can take action by creating new content, refresh existing content, and then even engage. We show you opportunities to engage with like Reddit, some other external places to be mentioned. Like Mark said, it's not necessarily on your site, it's where you are in kind of the ecosystem of getting mentioned there. And then we're also rolling out some new measurement and visibility opportunities to show you how you can measure your success and then repeat. So if you are interested in that, email us at growth@airops.com for early access. And then here it is right here. So you can track and audit your visibility footprint, act with those workflows, which I'm about to show you one here in a second, and then you can prove your pipeline to the C-suite as well. One workflow I'm going to share, Mark, that I am excited about is—

let me share.

Here we go. This is a workflow that we have to help you refresh your existing content. And for example, we took this from MasterClass. Very— this is for a demo environment, but a very straightforward blog post. You can extract the article in all these different ways using our grid. Get the article body, and then we have some capabilities to help you improve it and refresh it for AEO, um, in search. So you can see the output right here, and there's a little bit more user-friendly version right here, and it shows you like how we think it scores based on evidence and source attribution, authority signals, and freshness. Gives you a list of fixes that you can start with right away. And then we'll even help you apply those suggestions if you would like, and then have an article comparison between those different ones showing you exactly what to keep and like what to update. And so you can use this across all of your URLs and then get priority updates right away. So this is part of our action layer. Once you kind of see some of your opportunities and visibility, you can then like take action right here. So, uh, we're very excited about the workflows and how it kind of combines with our LLM visibility. Let me go to our slideshow. Oh, I'm a bunch of slides away. Here we go. So now we're going to talk about our Q&A here in AI mode. Lots of questions coming in in the chat. Thank you for all of these questions. And first off, what I wanted to ask you, Mark—

I'm going to stop the share for a second because I have my questions down here.

One question that came in is that folks were asking a lot about how you blend, say, traditional metrics for gauging SEO with these new LLM volume metrics. How are you thinking about the ceiling of search traffic and opportunity? Oh, you're on mute.

There we go.

Yeah, start with the easy ones, right? Thank you. So the honest answer, I guess, is I'm not at the moment. And the reason is the total amount of commercial interest, the total amount of money flowing through search, whether it's through AI or traditional search, is kind of staying the same. Our ability to obviously measure it doesn't kind of impact that. Now, I was having this discussion a little bit with Josh before we started, I guess, in terms of tools, which I described to him as I'm taking the approach of spending longer sharpening my axe at the moment, because there are lots of tools out there with different methodologies for trying to estimate how many searches are happening. And I don't feel like anyone's got a good answer at the moment. Not saying it's impossible. I'm sure we'll find a solution. And the reason I'm not massively worried about it at the moment is just in terms of kind of the share is still very, very, very low realistically. So, The things that I know we need to do to kind of move the needle still need to be done. And they don't, they're not gonna, they're not going to change in terms of if you're doing, you know, good work, then those algorithms will be aligned to eventually reward that. But I don't have a, I don't have like a metric I use yet in terms of that. I think At the moment, I would still lean on kind of traditional search in terms of topic volumes. And you can say, you know, this many people are searching for these kinds of topics and assume that translates fairly well trace papers over into kind of into AI. But, you know, it's nobody's doing that well. The only way we could do it would be with some extremely invasive way of trying to figure out what people are actually typing in. On that, while we had the background query data for ChatGPT, we were on queryfan.com going to launch a Chrome extension that gave people access to a database of what web searches happened in the background. versus what people's prompts were. And the idea was that you got access to this database for free by letting the extension, with your permission, obviously submit your search, your query, and what happened in the background, with the idea that we could build up a database to actually look at this. But now we've lost that background query data. I don't see how that's possible without basically spyware. And I think that's something we have to accept. And I just feel quite strongly about this. So, um, John O'Anderson had a great kind of point about this, just talking about SEO in general, which was he said, um, SEO stopped becoming a performance channel when we lost query data. You know, because there was a time when in Google Analytics, we could see what people had typed into Google when they arrived on the site for. We lost that a long time ago. And because of various privacy laws and such, it's very hard to accurately attribute specific searches to actions. And that's only going to get harder. So I think the kind of obsession with these very fine-level metrics, to me, is a lot of work for very little payback.

So then it's really more of like your brand visibility, like just appearing in all these different spots and it kind of becomes more of an awareness channel in some ways?

Well, I think, I mean, unless I misunderstood the question, I think we're talking about like trying to figure out how many people are searching for specific things in AI search, right? Yeah. So I just think, yeah, I wouldn't even try at the moment. Like, again, we spoke about this to begin with. And I've had, I've been under very, you know, pressure from various companies to provide some kind of numbers, because I do feel like, to speak very honestly, there are people I know at companies that are kind of giving various reports, and they know the numbers are kind of quite wishy-washy and not accurate, but people want that security blanket of saying, here's the narrative with the numbers, this is the data we've got. And while we can do that in broad strokes, when it comes down to the nitty-gritty of Well, how many people search for this? You know, it hasn't been a relevant question in my mind, even in traditional SEO for a long time. Also, as a great example of that, right, I did a whole, I did a whole talk in 2023 about zero volume keyword research, which was this whole idea that people are just discarding trying to rank for queries because the actual volume came back as zero. But the actual volume might be 5, 10 searches a month. But the point is, I can rephrase that same 8-word question in 50 different ways. So the actual search volume is in the hundreds. And that fans out again to all kinds of queries. That's one thing we've seen over many years is queries have got longer and more complicated. I've got my AltaVista book when AltaVista was the place to be. And the example searches are like 1 or 2 words long, all of them. So they—

I mean, I think we're seeing that in all of the data that ChatGPT queries, for instance, are much longer than traditional Google keyword searches. Another question that we did have was around site quality. I forgot who asked this question, but they said, where can you find and see site quality? Is it within the console or where can it be reproduced, the site quality score?

It can't be because that's what they paid us the $13,000 to patch. And so when I spoke to an ex-Googler about that and he said to me, oh, he said, I'm really surprised that when I heard you'd managed to find that, because that was a bit of— that was quite secret because it was something they'd developed to help fight spam, basically. So no, I can put you— like, share a link. I'll share it on my LinkedIn later where I did a talk about that exploit and exactly how we found it. But yeah, we gave that money to charity in the end as well. Because Google do a doubling thing. So we tried to do that, but it had to be a US charity. So in the end, we just took the money and gave it to a local charity here. But yeah, unfortunately, that's not available anymore, but someone might have saved millions of rows of it. Yeah.

Well, I think someone dropped in the chat, I guess the podcast maybe, where you chatted about it. So that's good for everyone to kind of revisit there.

You had the very huge, overwhelming infographic of all the different systems that Google uses to show their results. Somebody asked, do you think over time LLMs will change their approach and align more closely with Google's results, which are heavily filtered by multiple algorithms?

Yeah, so I think the— I mean, this is related to the— I saw a question about this, the LLMs kind of buying up link graph data. So the Google's got these few decades of experience of fighting spam. And I know we all give Google a hard time. But, you know, we don't see 99.9% of what Google is actually holding back for us. You know, the web is full of junk and spam, right? And I think the language graph approach, which is essentially all text is treated equally, and we just look for these averages doesn't work as well as a system where, for instance, one thing that's represented in the link graph, and obviously the link graph is the thing that made Google what it is, right? The things that are represented in the link graph are commonalities like generally popular sources of information are linked to regularly. Generally, you will find clusters within a link graph of things that are the same topic. And within that same topic, there'll be a commonality on who they link to. The way, the way I would describe it in an analogy, I guess, that any non-technical person could, could kind of grasp would be think about kind of your friendship circle, right? And general people, you wouldn't take a recommendation from a random person off the street about what restaurant to go to over like your best friend who knows you, right?

Right.

So that's like a representation of maybe like trust. A representation of like topical authority would be, if my car broke down, I probably wouldn't ask my best friend how to fix it, because I know they don't know anything about cars. But I've got that car friend. You know, that's what Google can do with the link graph. It can work out those connections. That's harder to do from a kind of text analysis point of view. So it makes sense to me that LLM kind of companies are interested in weighting that initial training. So, you know, for instance, not saying they do this, but a basic way might be, okay, well, if we get information from like Wikipedia, for instance, we might weight that slightly heavier because we know it's a good source of information. The link graph helps them do that so they can adjust the weightings on which sites, which URLs they want to listen to, because that's the data the link graph has encoded, URL kind of relationship data. Language Graph is kind of separate from that.

One question that— that was great, thank you.

One question that did come in from— got some plus ones here from Paige was, do you think prioritizing and optimizing for pre-conversion and product pages are a higher priority right now with AI referrals being more likely to convert? So I think she's referring to I've seen this and had this in some conversations that people are kind of doing their research like offsite, uh, and ChatGPT is giving them a lot and Perplexity is giving them a lot of background information. And then they, when they do land on your site, they're a little bit more willing or ready to buy. Are you seeing that as well? And do you think that's worth prioritizing for?

Yeah. So I would, I would lump that with on the kind of initial slides I did around what's changed in optimization. I was saying, Being present on more sites, having more coverage is definitely more important. I mentioned sentiment being more important. So more people saying more positive things about you. And again, being related to what's getting fetched in terms of the background queries that are happening, which, as I said, isn't focused necessarily just on your website. So I do think that's becoming more important. The way I find it interesting to think about is, In my mind, one of the things that made Google successful was they took power away from webmasters. Because initially, when we had like AltaVista, Hotbot, Lycos, we were heavy— the search engines heavily relied on information provided by the webmaster, what is in their meta keywords, for instance, right? And obviously, it's in your interest to spam that the heck out of it and just try and get, you know, get your site to rank. In terms of Google, they democratized that slightly because they were like, well, we're going to put a lot of weight on what other people link to you. You know, that's harder for you to control. So we're democratizing it. People have to tell us you're good through the link graph. I think what will happen with LLMs is even further, it's you're taking away that power of the individual site and surmising kind of more of the important voices on the web. So I think there's potential that results in general, the experience could get better. Because you're, you know, absolute power corrupts, right? When you have these concentrations of power, you tend to get people doing stuff in their own interest. So yes is the answer to that. I mean, we're certainly going to focus more on stuff like digital PR being reviewed by other websites having people talk about us more. And again, I think it's great because it means SEO links up with different job roles that previously we've had a little bit of kind of, you know, conversations with. Like, for instance, I'm much more interested now in speaking to people running social media than I ever have been.

Hmm. Yeah, again, going back to those offsite opportunities and appearing in those places.

One question I did have was, you know, we've And you mentioned this in your SERP there, is we've kind of been thinking about rankings for a long time.

Do you think the concept of rankings is dead, or is that something that we should still kind of think about when we have these very personalized and broader answers?

Yeah, so I think it's just moved down the chain. So I can't see how, without some new technology, how LLMs are going to function well without a traditional kind of search apparatus in the background. Because the— there's a— this is related actually to— there's a question about that it seemed weird to just optimize for like RAG stuff. Should we not optimize for in-model as well? Like, yes, the in-model is the long-term goal. But the— You can't just update your site and appear in model because what's represented in the model is the average across the language graph. So to appear in the model consistently, you need a lot of people talking about you. And that's not just like a let's press the optimization button. That's like a company-wide, months or years-long effort to achieve that. Versus RAG, which can be done in days or weeks, you know, we've managed to impact results. So I think the traditional paradigm of search will remain, but it will be in the background in terms of it's not the searches people are doing, it's the searches that the robots are doing to help the people because the robots are acting as, as I said, kind of intent decoders, encoders. So, okay, I think this person needs this information. So here's how I talk to this other machine to extract that information.

And then I got one more technical one and then maybe one broader one. Maddie asked, If we are targeting queries that require grounding, wouldn't schema still be important in helping those sources understand our content?

I guess going back to your schema point from earlier.

Yeah. So, I mean, don't not do schema, but I mean, if let's say for argument's sake that the results are coming from a Google index, right? Google has told us many times that structured data schema does not impact rankings, right? And I believe that I've never seen any measurable difference. What schema does do is allows you, it gives Google the confidence to allow you to have like rich results, right? And yeah, I've generated loads of extra traffic from getting rich results. I mean, people were abusing the heck out of the FAQ schema, right, for ages. So much so that Google basically would take the toys away from you now. But it's not helping you kind of rank in terms of that. Arguably, it helps you get in the knowledge graph, which by proxy then might help AI mode understand you. But again, it's just not a high priority thing for me. I mean, Gary Ish from Google specifically said like he would like to see Google be able to function without schema. Because schema just represents technical barriers that Google hasn't fixed yet. And from an LLM point of view, the main way their models work, they don't need it. So yeah, I think feeds and stuff will become more important. There's a lot of talk with ChatGPT partnering with Shopify and providing that experience and an agentic experience as well. So I think there will be definitely more stuff for us to do with structured data and like MCPs and ways for robots and agents to interact and understand our content. But schema to me feels structured data is a very web kind of thing. It's run, you know, it's a volunteer-run thing. It's not that active anymore. We haven't, we're not really seeing new features come for schema. So yeah, I think it's achievable without schema. Schema can impact CTR. Someone said, they know Elijah, Elijah Blue said they're absolutely right if you get rich results. Yeah. And then it's brilliant.

Yeah. Thank you. I hope that was helpful, Matty, for that.

And then maybe taking one step back as we close here, you know, we're all going to go back to our team leads, our CEOs, our CMOs, our clients, and they're going to say like, What I heard, AI mode rolled out some new stuff today. Like, what do I, what do I need to know? How is this maybe different than, than traditional Google here? So, uh, maybe any last words about what we need to know about AI mode in, in a few sentences there?

Yeah, I mean, I'd say as, as always, like, don't, you know, don't panic. Um, AI mode's still a selectable, uh, vertical. You've got some of the, um, kind of call to actions for people to experiment in Air Mode. The last data I saw for the US adoption was still in the low single-digit percentages, like 2, 2.5%, I think was the last data I saw, which is still a lot of searches. It's still like, you know, if they did that globally, it's still like 100 billion searches or something. But I wouldn't be too worried about it at the moment. I feel it's a stopgap between the the endgame, which is the, um, the more agentic stuff. So I would be— what I like doing at the moment is like Perplexity's got a browser called Comet that you can get. Um, you can go on a waiting list, or if you pay for Perplexity, their Pro Max, whatever it's called, you get it. Um, and you can tell it to do things like go buy blue paint from this website and see if it can actually do that on your website. That's longer term where I've been talking to clients, making sure that their websites are technically kind of accessible enough for robots to do things reliably, because I think that's what's going to come quickly on the heels of AI mode. I still stick to my optimization part, is mainly the same stuff we're doing for SEO with different weighting. So again, stuff we talked about, a bit more time on digital PR. a bit more time on getting other sites to talk about you, sentiment, making sure your site's technically accessible.

Love that. That's a great takeaway to end on.

And I wanted to thank everybody for joining.

Be sure and email us for early access to Visibility Layer.

Here it is on our website, growth@airops.com, and we'd be happy to set you up.

We have put— whoops, lost Mark's info.

Oh, here it is on the non-slide show version. But be sure and follow Mark on LinkedIn. And with Candor, Also Asked, uh, and then his SEO newsletter too, which we've dropped in the chat.

Mark, appreciate your time. Like, this was incredible. I think we're all going to go back and like make sure we digest this and review it. Uh, the recap will be out later today. Uh, so really appreciate this.

Thanks everyone.

Yeah.

Have a good evening. Have a good morning. Have a good afternoon. Talk later.