Building Search for AI Agents with Exa

0:00To influence GEO, you need to influence two things. So you need to influence first. How do you get in the model weights, which is harder to control? The other thing is those models, basically the way that they're working and generating answers is still going to the web, right? They're going and searching. So if you're not ranking in those top 10 results, you're not going to be cited either. Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI.

0:32Search is being rewritten for agents and not humans. While LLMs are amazingly capable without live, reliable web context, they hallucinate or they just fall hopelessly out of date. At the same time, we have this mass inflection point where consumers are abandoning the 10BlueLink Google search routine and asking ChatGPT, Perplexity, and other AI tools to find answers. And that shift is completely changing SEO and it's spawning a new discipline, which some are corny, coining generative engine optimization or answer engine optimization. Pick your favorite, we'll see who wins. Because what matters now is how content is interpreted by LLMs and AI agents and not just how it ranks.

1:15And our guest Thias and her company, Exa, sit at the center of this upheaval with a massive goal to create a different type of search. and one that's actually understanding the meaning behind all of the websites on the internet and one that's tailored for agents. Welcome to the show, Thais. Thank you so much. Super excited to be here. Yeah, this is going to be fun. I'm going to hit you with a big question out of the gate. Is SEO dead? Everyone keeps asking me this. SEO is not dead. And I guess we'll get into all the details of it, but I think what a lot of people miss is where are all of these AI chats getting their actual data?

1:56And ChatGPT is still searching Bing if you think about it. So if you're not winning the SEO game, you're not going to win the GEO game either is a little bit of the answer. It's almost a prerequisite. We can jump into all the details later, though. Yeah, I like that point because our CEO keeps going to this. because we keep getting, we keep showing up in AI searches. And he's like, what are you doing? And I'm like, I don't know. Because I think the principles of like trust, authority, relevance, like this thing still counts. But I'll give you a couple interesting use cases that I've heard lately.

2:29One is 10K's companies, annual reports and things like that. Companies are literally starting to write those tailored for AI and agents because they know, what do you do? Nobody's reading through that huge document. They're going to go dump it into some AI tool so it can research it. Right. So those live on the web. So they're trying to tailor it for that. And then I saw another interesting use case, a large bank, and I mentioned the name, but they were having, somebody just asked, what do you think about X bank? Right. Just to see what would happen in an AI search. And it came back with all these negative comments.

3:03And what they found is it was going and referencing reviews, review sites and things like that. So it was like their smaller branches, one-off bad reviews were causing them to show up bad in the AI search, which is like just an interesting shift, I think. I think what a lot of, no, you're so right. And I think what a lot of companies will have to realize is they're not going to be so in control anymore, right? Because those AIs are smart and obviously like researching the web in all sorts of ways that are not so controllable by you. So they will be fetching information that sometimes you don't put out or that is a news articles or reviews to your point.

3:42And you won't have control over that information. So what can you control? And I don't know if it would be helpful to give a quick overview of the differences between SEO and GEO. But I think. Yeah, if you want to hit that now. Yeah, that'd be a good thing to hit. And then I think, too, for context, what is XF for people? Because you'll have an API that's like helping do this. Maybe hit that piece first. So we have the context there and then let's like dive into this new discipline that's emerging. That sounds great. For those not familiar with XSAT, we're basically building the first search engine that is made for AI agents.

4:14So Google, Bing, we're all built for consumers and for an age in which us people were going on the web and searching. And obviously we're lazy. So we search in keywords. We don't write full complex paragraphs explaining what we want. But now we're moving towards a world in which AI agents and think ChatGPT or even more complex products will be doing the search on our behalf. So we're actually asking ChatGPT for an answer and then they will go ahead and research the web. So how do we optimize the search engine for that, basically? And the whole principle behind Exa is we're the first search engine that can actually understand the meaning behind all of the documents across the web and can return content that's super relevant for those AI questions.

4:55and in a way that's much simpler for the AI to basically process and be able to absorb all of that information. So instead of our search being optimized for SEO and for consumers, like Google has been, we're way more trying to be as relevant as possible for those AI so that they can be grounded and they don't hallucinate and they have access to kind of the best data. Quick break in the pod. If you're listening to this podcast, chances are you've been thinking about how to actually use AI inside your business. And that's exactly why we built the AI Opportunity Finder. It's a free tool that helps you uncover high impact, tailored AI use cases based on your business, your goals, your pain points, and your industry.

5:32No fluff, no generic use cases, just real ideas that fit your business and the ranked by ROI potential. It takes about three minutes to run and it's like having your own personal AI strategist for free. If you want to try it for free, check out the link in the show notes or go to hatchworks.com backslash AI dash opportunity dash finder. yeah and there was some cool i'm trying to think i think it was open router was one of your customers using your product and some different ones like how are they using it maybe some other interesting use cases i think this is like a new it's almost like a new discipline tool whatever you want to call it for companies in the future and how did how do you use it as one thing people need to be educated on so basically our main products in api which so basically we plug into to company's products, become that source of data from the web.

6:24So for OpenRouter, for example, what's super cool is they have over 400 LLM models. And some of them are not usually connected to the web. We basically partner with them. And now all of those models can access the web via EXA. So they can basically send a prompt to EXA. And then we go fetch that data, return it to that model. So now any LLM really can be grounded with web data, which is super cool. We also have some other really interesting customers. So from large search engines such as Ecosia, helping power a lot of the AI answer generation for them, or even big consulting firms, big banks, tech companies that want to not only pull from your internal data, let's say, think productivity tools or even CRMs.

7:05Not only do they want to be able to pull from internal data, but also enrich that data. Imagine if you have a CRM filled with customers and if you could already have written there, okay, who are their competitors? What are recent news about them? So we're partnering with also like sales enablement tools, recruiting tools, and a lot of tech companies have basically become that data source that's infused throughout their product or in AI features within their product. Nice. That's cool. Okay. We got the context now. So let's go back to SEO, GEO. I've heard AEO. I've heard there's some other acronyms I've heard.

7:38First of all, which one do you like the most? And then give me the breakdown of like how it's different. Yeah, I've been using GEO because I think it also it goes beyond just answer generation in a lot of ways. I think we're going to start seeing AI be infused through all sorts of products that sometimes are not that clear answer format always. Right. Sometimes it's baked into a larger feature. So I like GEO, but I think for those less familiar with it, the TLDR is GEO is essentially how do you optimize your content in your website so that it gets picked up by AI answers and gets cited by them.

8:11So I think the major difference is instead of just trying to show up in the top results of Google or Bing, you're trying to be a source that's actually directly mentioned and quoted by those AIs. And we're already starting to see a bunch of companies mentioning how a significant portion of their traffic is actually coming from cloud chat-to-bitty mentions. I think Guillermo from Bursell actually mentioned recently that I think over 4 % or 6 % of their traffic was already coming from chat-to-bitty mentions. 46 % or 4 % to 6 %? No, I think it's 4 to 6 % is coming from Chachupiti. I was about to say, wow.

8:43But increasing every month, like that's a lot. Yeah. So I think it's something that we're obviously in the early stages and still somewhat fully unclear all the things that come into play to influence that. But I think some things we already do know. What's interesting is to influence GEO, you need to influence two things. So you need to influence first. How do you get in the model weights, which is harder to control? So things like Chachupiti, for example, the cutoff is October 2023. unless you were relevant in October, 2023, have to get around that. The other thing is those models, basically the way that they're working and generating answers is still going to the web, right?

9:19They're going and searching. In the case of ChatGPT, they're using Bing, for example. So if you're not ranking in those top 10 % of, sorry, top 10 results, you're not going to be cited either. So there are things that you need to do specifically within your content to make it super like short and clear and have headlines that that AI will be able to snap up. But if you're not showing up in those top results in the first place, forget it. You're not going to be in the game to begin with. Yeah, it's like you mentioned, that's the key thing. It's OK, if I can just get on the first page, I'm good.

9:53Somebody's going to scroll. But like with this, like you're lucky to get in a list because sometimes it's just, hey, here's my response. But that's the interest. You know, it's one thing that and I'd say this past quarter, it's crazy. Like I think 20 % of our inbound, because we have just an open field in our inbound form. That's like, how did you hear about us? And about 20 % is some AI tool. And we're trying to understand, okay, what did you ask the AI tool? But some of them I've noticed, it's like the cross between I'm looking for X in this geography. And there's things we have that put us squarely in that spot.

10:30So I think that's one angle. But you mentioned the point around like making content that's easy for LLMs to consume. Like one thing that always comes to mind is like question answer type of stuff. That's somebody has a question they're going to ask, make it very clear and explicit somewhere on your site. So it's very easy for the LLM to go grab it and then just return that back to the customer. I guess that's one example. Cool. Any other, because I'm thinking like for teams that cover content and stuff like that, like where do they begin to optimize so they do start showing up in these different search results?

11:08Yeah. So I think for content, for let's say thinking about the content within those pages, I think the question answer format works great. Also very clear headlines and then short maths, like three lines that kind of contain that core piece of information that you want. Another thing that I'm hearing a lot is actually numbers. So including data and including almost like micro case studies within your content can help a lot because AI likes to pick on a piece that it considers like media information. it's the opposite of sometimes what i think traditional seo blogs used to be which were very fluffy very repetitive repeating the same keyword over and over again it seems like for geo it's better for you to have actually a breadth of words that are related to that topic but that feels slightly distinct and that the content actually feels meaty concise direct and has like good data points good little analogies like things that you as a human would also want to quote, right?

12:05If you come out of this podcast, like what is a quote you're going to remember? AI works in a similar logic where it has to pick like something that's self-contained and has a lot of meeting information. No, it's perfect. It's funny as you're talking, I'm like asking chat GPT on the side. It's giving me the key things for like GEO and AEO and whatnot. A couple of interesting ones that it said, let's see, optimizing for feature snippets, zero quick answers. That's interesting. So it's like direct, easily accessible, and then speeding clarity over long form SEO bloat. So like in the past, it's okay, let's stuff a million keywords into this thing.

12:41It's super long. It's painful for humans to read. And I guess it's also painful for LLMs to read. So like that may be the opposite of what you want in the future. But the one that's interesting to me is they say using schema markdown where possible for structured data. It was some video or something I was looking at related to image generation. And that's the tough thing. It's like, how do you get consistent images? And the way this person was doing it was they were reverse engineering it. They had an image, they fed it to whatever AI tool, chat, Gbt Cloud, and said, give me the JSON describing this image.

13:17So then it created JSON that very clearly defined the image. And then you had the attributes that you could tweak that you could feed back in and do that. That's super cool. I wonder in the future, and you mentioned this when we were chatting before, you wonder if the internet's going to shift to being less for humans and more for AI and agents. And we're just, okay, I need this. And then it's going off doing the thing. What does that future look like, I guess, in a sense? Yeah, so that's a great question. I think we're definitely betting on that future. And I think obviously as humans, we're always going to gravitate to things like social media.

13:56And there's always going to be places within the internet that are still very human first. But in terms of content, it almost will be intermediated by an AI, right? We're likely going to be wanting something that's more personalized or more curated to our tastes or to our questions. So we don't want to go read a whole gigantic article to then go find the one sentence that matters to us. We're likely going to ask an AI agent to then go fetch that on our behalf. So I think you're spot on. And I think the Internet will start being optimized for AI agents. And we're seeing this even now with things like MCPs.

14:27Right. I think we talk about this internally a lot, but I think for the dotcom boom, if you miss the mark on SEO, like you lost huge opportunities. And I think for agents, you also need to get your products to a state that they're very easily utilized by AI agents. So MCPs are one way for that to happen. Geo is another way. But I think, yeah, how do you make it easily navigatable and easily actionable also by AI agents? I think we'll shift a lot in the next years. And MCPs, if anybody's curious, what is that? And that's been a very hyped up thing as of late. So we have a whole Hatchworks AI lab going deep in that topic.

15:06But in general, like it's just a way to make it easy for LLMs to use tools to access things. Right. It could be any number of things. It's just, it almost feels like the, it's like the connective tissue, almost like the APIs for LLMs in a sense, even though LLMs use APIs too, but it's like. No, but that's exactly right. It's basically an API that requires less integration and that they can proactively basically seek out. So you're almost exposing it and making it easy for them to fetch it. And this just triggered something in my head. I think it was maybe the Greg Eisenberg podcast. He was talking, I think it may have been Riley Brown and they were on this topic, but that's, I think the thing that's going to change too is things are going to require less physical code and more natural language in a sense.

15:53It's almost like giving context to the LLM. Here's how you should use my website. Here's how you should navigate my website. And you can give that in markdown, which is just natural language, easy to read. But I think that's the next step too. It's just if somebody's coming to your house, it's like, hey, here's where the bathroom is. here's burgers are outside exactly i think that's another thing that's gonna emerge i don't know any thoughts on that or is there we haven't done this personally what's the effective way to do it almost feels like meta tags and things like that in the seo world this is like the new version of that yeah for sure i think a that obviously allowing the crawling to happen so like allowing robots to talk text to actually crawl your site is imperative but i think yeah having that mark down and like really structured content is I think what we're going to gravitate towards.

16:46I think a lot of those websites that used to be almost like overly designed and where text wasn't displayed in such a clear way might actually hinder a lot of geo and a lot of the way that these AIs are searching. So I think that's one piece. But what's interesting is for, so this is more for, let's say, consumer use cases, right? Where you're seeing a person going to Chachaputi asking a question. I think what's interesting is our customers tend to actually want to use XA because it's not gameable necessarily by SEO and geo. So if you're a business and you want the best data to be funneled into your systems or into your product, XA can almost, it breaks the logic of SEO because we actually embed the whole document.

17:27So we can understand the full context of all of that document and give you those matches based on relevance. so there's almost like no way for people to to game our system and try to optimize this for seo let's say because we are already taking the whole context into account obviously the things that work for geo will also be great for that because it just makes it clear but the tldr is really you have to be as clear and concise and direct as ever and ais tend to love that so i think we're definitely going to see a movement in that direction you just made enemies with a lot of marketers all around the world trying to game the system.

18:03I know. And fans of businesses that are trying to avoid that. So I guess the other point to, actually, hold on one second. And by the way, maybe I can give an example on that actually while you pull this up, just in case it's helpful. So for example, how would a search traditionally work on Google versus how does it work on access? So let's say you're trying to find interesting AI startups based in San Francisco working on robotics. If you type that on Google or Bing right now, you'll basically get like an article about companies and then a Reddit thread and then a blog post, but you're not going to get the list of companies when what you're asking for is actually the list of companies.

18:38So X is very good at understanding that intent and returning to you exactly what you want. So we would basically be able to read those companies' websites, understand that they're matching what you're asking for and returning those company homepages. So it's just much cleaner. And for an AI, they need that primary data source, right? A blog, an article, a Reddit thread is almost like overly processed information instead of being the primary source of it. Excel works great to return that primary source and more of that meaty content that we're talking about. So I think that's just one example query of how that would work differently.

19:11So more of a valuable result. And I'm curious, how do you view the competitive landscape right now? Because it's insane. Like you'd mentioned that the Bing search API is getting deprecated. You have that going on right now. Obviously, there's tools like Perplexity that are trying to own AI search. They're more on the consumer side of things, obviously. But how do you view this competitive landscape that's starting to emerge? Yeah, so I think there's two sides to it, right? You have the consumer search and then you have B2B search. So I think for us, basically the Bing API got deprecated. Google doesn't really have a public API.

19:51So it was actually interesting because we saw a huge inflow of customers following the Bing deprecation, which still all need access to web data. And I think the move that Bing did was a lot of they basically removed their public API and moved people onto more of this black box type of search within Azure where people can already request the search, but it has a bit of that LLM preprocessing. So it just doesn't give developers that much control. And I think Exa moves a lot more in the direction of, OK, let's give you a lot of control. Developers can choose whatever they want. So I think for B2B, we're seeing actually a lot less options happening in the market.

20:28And very few companies are investing in foundational search infrastructure, like training and building their own models like we are. On the consumer side, I think it's really interesting what's happening here. And I'm very curious on how this will evolve because the Chachapetes of the world, Perplexity and Claude are all stealing search from Google. And even within Google, I'm actually dying to hear the discussions between the head of Gemini and the head of Google ad because they're directly competing for traffic. And I think as these companies evolve, we also need to see how do their business models change and how you create incentives for people to go towards these AI products without taking huge hits on revenue and not being able to monetize a lot of those consumer requests.

21:11I'm super curious on how that will evolve. It's funny too, on the Google side, it was a part of their early product research program or whatever with their AI mode. And you could see they're trying to like encroach on this chat GBT feel. But the interesting point you made is like they are facing the biggest innovators dilemma, probably of all time, because their whole business is obviously it's a very profitable business across tons of business units, but their ad revenue is. huge and the margins are insane. And if you all of a sudden say, hey, I'm just going to return this AI return answer and you can't really do the ads, what happens there?

21:54I'm sure they're going to come up with some new version of ads and whatnot, what that'll be. But I think that's a really interesting point because one thing that, so obviously in the last, let's say two years, is when these AI chat apps really blew up. And I think they started from a place of building a lot of trust, actually, with consumers, right? I think consumers right now really trust HatchPT, Claude, Gemini, et cetera, because it doesn't feel like they're incentivized to sell them anything or push anything specific. And there are no ads. For now. For now, but that's the thing. I think I'm also super curious on what will happen to consumer trust if these apps have to start pushing ads because on Google, it's super clear, like what's an ad versus what's not an ad versus for an answer, it might start feeling a little bit weird for consumers to then trust what that AI model is telling them if they're incentivized to recommend specific things and not others.

22:52So I don't know how they're going to solve that. It's obviously a really tough problem. I think for Google especially, they have one of the most beautiful business models of all time with ads, but it seems like they're going in a direction of self-disruption, which I think is smart because it would happen anyway. So it might as well be to themselves and then figure out the revenue or ad piece within Joe and I later. Yeah, they almost have to. But I'll give you my like analogy for it. And it's based on two movies. It's like The Truman Show and I think it's Minority Report. And if you remember The Truman Show, like they have ad placements and Truman has no clue, right?

23:25They're trying to sell something. He's like, why are you holding this can opener so weird or whatever? And the Minority Report, it's like when he's walking in somewhere, it knows who you are. It's like trying to sell you this exact thing. So who does? Maybe that'll be our future one day. But I think the interesting thing also is not only how companies are thinking differently about search and how they appear in search. The consumer behavior is changing to not only where they go to search, but what they're searching for and how they're searching. Because like for me, the way I use AI is completely different than how I use Google.

24:05Like I can't tell you how many conversations I have where it's about thinking through something strategically, planning something, outlining something, doing a lot of deep research. Like we're doing, and this is a segue into the websites tool you have, but like research on finding prospects in a specific niche, or maybe you're in recruiting and doing things like that. there's so many different types of searches that are beginning to emerge. How do you think that's going to evolve over time? Great question. So first, I think, by the way, I think you're in the minority of users still. And you're ahead of the curve, which makes total sense.

24:43But the majority I saw - Which I got to remind myself of, and you're probably the same way, right? Because I'm in it every day, so I just assume everybody else is too. And then I'll go do a training or workshop and I'm like, oh my God. It's like people just still haven't figured this out yet. Exactly. And I think right now we're in this phase of retraining, of mass consumer behavior. I think people have been trained over a decade on how to prompt Google and intuitively how to prompt Google at this point. And I think now people are going through this mass retraining phase of how to actually prompt HPT or Claude or Gemini and get the most out of it.

25:23And I think the style of searching, obviously, it's a lot more longer form. And I think that is a big shift. You're going to start seeing way longer queries, way more context. So you don't need as much of that guesswork as Google has to do with keywords, which is why Google leans so heavily on keywords, because people type in a very keyword-y way as well. So I think with these tools, consumers are going to learn to give more context and give more depth to what they want and more direction. So these AIs can then, I think, give more informed answers, but also the way that then these AI agents will go and search the web is also with longer queries, which goes a little bit back to what we're talking in the beginning of even how geo is different than SEO.