Mark Williams-Cook: AI Overviews & The Impact On search – IMHO

Hello and welcome to IMHO. Today I'm talking AI overviews with Mark Williams-Cook. So SEO is still trying to figure out just how much search behavior is going to change and how much this might impact organic search traffic. The SERPs are evolving on a daily basis and we're all watching very closely just how much organic traffic is possibly going to be taken away by AIO. And more importantly, how are we going to adjust to these changes? So over the last few weeks, I've been speaking to some really smart people in the industry, and today I'm following that theme. I'm speaking to Mark. Mark is well known in the industry as the director of the agency Cando in Norwich, and he could possibly be one of the busiest people in search apart from Alayda Solis, of course. Mark organizes Search Norwich Meetup with his colleague Jack Chambers-Ward. He is the founder of AlsoAsked search intent tool. Mark publishes the Core Updates newsletter. He has also got a top-rated SEO course on Udemy, and Mark has built and flipped several sites on the side. He's been a top-rated speaker at Brighton SEO, and Mark still finds the time to offer his daily unsolicited SEO tips on LinkedIn. So I'm really pleased that Mark's found the time to squeeze me in all that. So Mark, I'm going to ask you, in your humble opinion, What do you think about AI overviews? How do you think they will impact the industry? And where do you think this is all going?

Fantastic question. And I will definitely take second position to Ilayda in terms of busyness. She does absolutely amaze me in where she's going. So yeah, AI overviews, I think, firstly, from a consumer searcher point of view, I think we're in a really weird space right now. We see, I feel like we're sitting on the fence between 2 technologies. And the reason I say this is I was actually doing an interview the other day with someone talking about bias in search. Okay. And they were using the example of saying, asking Google, was Trump a good president? Was Trump a bad president? And their kind of argument point of view was, well, if I ask Google, is Trump a good president, it will just return articles mostly showing me the good things he's done and why he was great. But if I ask the question, was Trump a bad president, it will highlight mainly bad things he's done. So it's just reinforcing, you know, my opinion, which is, you know, alluded to with the question, right? And the point I brought up here was, now we have LLMs, and we have AI overviews. We have introduced this new way and delivered a new expectation to users about how they can interact with the technology in terms of, you can ask, say, ChatGPT a question exactly as you and I would converse now, and it would give you an answer. Okay. Obviously, if you think back to roughly how search engines work, and especially early days, you know, you had to think very hard about the couple of words you typed in to try and get answers. And then Google's been progressing things saying, oh, now, you know, we've got BERT. So we understand how the query affects what you're searching for. But ultimately, if I do a search on Google at the moment for something like, Is Trump a good president? What I'm still asking Google to do is return documents that are looking for you know the entity of Donald Trump and good. So it's still going to; it's still doing its job in terms of Google's like well you've asked me to show you articles where Trump's good right? But there's this disconnect I think between the user's expectation in that you know over ten years ago Google. when they started talking about their knowledge graph, said, we're on the road to becoming an answers engine, I think was the language they used, right? Rather than a search engine.

Yeah.

And I feel this slightly muddied AI overview, LLM water now has made the general population think we're there. Okay, so they think now you can just ask questions to Google, So we have this thin layer of AI overviews kind of draped on top of the normal search results, which has kind of given some pretty wild results. So firstly, I think the underlying kind of large language model technology is really, really interesting. I saw a paper the other day that claimed they've mathematically proven there's no problem that can't be solved with enough layers of— with enough tokens and layers of transformers. Because there is that conversation around with large language models, you know, they don't know anything. We're talking about token probability with some generative spice, which is why, you know, you get very weird things happen, like it can't tell you how many Rs are in the word strawberry. It tells you to put glue in the cheese in your pizzas. You know, we've seen all these examples, right? So firstly, in terms of how I think AI overviews are impacting search, because as SEOs, I feel we're always led by not necessarily the technology, but consumers' behaviour to that technology is the important thing. Firstly, we're in a dangerous place, because I think if you stopped 99.9% of people in the street and asked them, just basically describe how ChatGPT works, how an AI overview works, how a large language model works, Most people have no clue.

Right?

Google's got 20-odd years of trust earned as a consumer brand of, I type questions in, you give me good answers. That's why you're my go-to search engine. And I think there's been a bit of a kind of hot swap in terms of Google is now, as has been demonstrated, sometimes giving some very interesting and just sometimes flatly incorrect answers with AI overviews. Right. And my concern for consumers is they're not aware of the danger of that. They're not— their defenses aren't primed. Right. So if I read an article written by someone, maybe, well, as a human, I know the person that's written it can be biased, and they may have an agenda. And my defenses are naturally up. when I'm reading something maybe someone's written, say, about a president, and I know their political alignment. So I kind of, you know, flavour that. If you don't know how large language models can hallucinate and make mistakes, your guards are totally down. So when it gives you incorrect information, and we've seen it even with medical, you know, we get AI overviews for medical questions.

Yeah.

I think people's guards are completely down because they see AI, think it must be clever because it's AI and it's Google. You know, who doesn't trust Google? So firstly, I think there's that big issue as an industry we've got to deal with. And I think that's going to be limited by the type of technology, i.e., LLMs. I can't see without another big step how we're going to get over that problem. I saw Gary Ish from Google talking about how they use a technique called grounding with their AI overviews, which is essentially, as far as I understand, I may be completely wrong, but as far as I understand, like an intermediary when they're generating these answers, trying to fact-check it internally, like with Google's Knowledge Graph.

Yeah.

So we've got some, you know, Knowledge Graph stuff there. So in terms of how it impacts us as an industry, while it is this kind of as I said, thing that's plastered at the top of the SERPs. I've heard people call it a tacky version of a featured snippet, which isn't, in my opinion, too far off how it's seen. We've got AI overviews now in the UK, actually not for so long, you know, it's only been a few weeks, I've seen them, you know, everywhere.

Yeah.

I have seen some impact in terms of clients, in terms of click-through rates on some terms. But nothing, you know, major, the same as when featured snippets came about, everyone was like, oh, this is, this is the end, nobody's going to click on websites anymore. So I think we're, we're, we're past that in terms of I don't think it's the end of the current SERP, because there is value in rather than saying this is the answer, to saying, here is a collection of answers from trusted people, because that's the gamble Google makes, right, with AI overviews. It's, here is the one correct answer or overview. And the issue with that is, if you actually discuss it with people, if you say this answer is wrong, and you actually discuss it, you get a room full of people, most things are subjective. Say if we take, is Trump a good or bad president, right? People those terms, good and bad, anyway, they're human constructed, they're based on morals, principles, people may perceive them differently. So while you may have a very strong argument one way or another, people won't agree. So how do you get an AI overview to give you a correct answer? The only way we're seeing that is like what ChatGPT does, is it gives you the most vague, boring, middle of the road, well, you can look at it both ways kind of answer, which nobody, I think, really wants. We want to see the spectrum of opinions, which they struggle to do. So lots of challenges. And certainly, you know, that barely even touches on how Google monetizes the whole thing, right?

Yeah, I mean, I actually call AIO featured snippets on steroids.

There we go.

Pet name for it. In terms of how they're going to impact, I think for me, it's that I think people are going to change how they're actually searching and their whole search behavior because as AI is becoming integrated into their mobile phones and more AI assistants, I think naturally people will actually use Google less. They'll actually go to a browser less. And I think there'll be sort of a more of a— it's going to become more of a seamless, frictionless experience. Whereas, you know, they're looking at images and then you can click on an image and it'll take you to like a shopping product, etc. So, I mean, in those kinds of terms, you know, in your experience, how are you going to adapt? How do you actually think search is changing and how are you going to adapt to that?

Yeah, that's a really good point. So, I mean, I've used LLMs and AI overviews for instructional things quite a lot, which is where I'm trying to do something and I'm stuck on a particular step. And the kind of classic route would be to Google a guide to that. you know, try and navigate around all the adverts that they've used to monetize that guide and then try and find the right step. Whereas I've had a lot of success with AI Views LLMs just being like, I'm trying to do this, I've got this error message, or I'm at this stage, what do I do next? And it's worked really, really well. So it's actually a concept I heard John O. Wilderson talk about before, which is solved knowledge spaces, which I think was a really interesting way he summarized it, which was there are various kind of human knowledge spaces, which you could class as solved in how to make this specific kind of lasagne, for instance, right?

Yeah.

That's kind of known human knowledge. We're probably not going to improve on that massively. So that would be something I think that would fall very comfortably inside an AI overview or an LLM as like, this is stored kind of bit of human information, and you can interrogate it as you like. You don't really need the guide that was published a week ago for this. Of course, there's limitations to all of that. But I think there is— the biggest area of impact will be these solved knowledge bases, right? And that is a shame for people who have business models where they monetize those kinds of things. Although I struggle to shed a tear, because when I think about the kind of things and why people are publishing, you know, things online, I just think it would be great if we were pumping out less of this kind of solved content online. There just seems to be an endless, you know, stream of it.

Yeah.

And the experience of accessing what is fairly basic information in 2024 is quite painful. When I compare it to 10 years ago, like, sure, you know, we had ads and stuff on the internet 10, 15 years ago, but you could access a lot of information, it seems, with less friction now than you can, you know, you can in 2024. So, I think there will be a readjustment for certainly the kind of content types of sites that are, you know, giving guides, instructions, recipes for things that aren't actually changing that much and where it's easier for the user to interact in a conversational way. So, I mean, how would you adapt to that? I mean, if you're in one of those spaces, I think it's honestly maybe one of those times in technology where you're just like, this has changed. And this isn't how people are going to use this anymore. And you can't crowbar the old way of doing it into the new way. And that's, I think, something you can consistently see throughout history, whether it's technology or anything else, which is that eventually people do things in the way that offers the least resistance.

Yeah.

Like, you know, same task, this was slightly less effort. Eventually, you know, there's great images online where you see where they've built the path around the grass, and then there's just the muddied footprint, right? That's what happens with technology. And in terms of Okay, for the blended, what's left of the SERP, how do we, how do we, how do we deal with this? And maybe the LLMs, you know, the new version of LLMs that are kind of live using what's on the web. I think that's a kind of more interesting question for SEOs. The data I've seen at the moment seems to suggest that there's a big correlation between the kind of things that are appearing and being surfaced by LLMs and stuff that's ranking quite well organically at the moment. My opinion would be at the moment, it's probably more to do with how, where and how often you're mentioned across the web than classical like page optimisation stuff. You know, which is, you know, links have always been important. But I think this is more about now coverage and people talking about you and the sentiment. So almost more like a traditional PR doing good things, making good products, getting people to talk about it online and making noise. That's essentially what these AIs are hoovering up. So the bigger digital footprint you have in those things, you're going to naturally be surfaced more by these kinds of tools. So that's, and anyway, that's from an SEO point of view. I think if you've been doing SEO, like really kind of well over the last few years, you have been focusing more on, you know, entities, brand recognition, digital PR, and maybe less on, you know, stuff we used to very long time ago think about more, which was more about, you know, thinking about keywords instead of intent, and thinking about, you know, just links and anchor text rather than coverage. So it's kind of been going that way. I've always looked at, like branded search as the gold standard, if you like, because, you know, that's eventually what I want people to do, right? If I'm selling running shoes, ideally, it's great if I can rank number 1 or whatever for running shoes. But what would really seal the deal for me is people Googling, you know, Mark's running shoes, because they're not looking for anyone else.

Yeah.

So that's where I want to push everyone. And I think that's how you will find success with AI overviews and LLMs as well.

Well, that's a really interesting opinion from you, Mark, and lots of really— I think one of the things I'm really enjoying about doing this series so much is, and especially asking different people about the same topic, is you're getting such a diverse range of answers and some really interesting information. So I'm going to wrap up there because we are just about out of time. So, Mark, thank you very much for your time, and thank you very much for offering your Humble opinion.

Thank you so much for having me. I really enjoyed it.