The Search Engine for the Agentic Web — Will Bryk, Exa

How's everybody doing?

Good? Woo! All right. Pretty crazy times we live in. Oh, my microphone's open. Okay, I am very excited to tell you all about Perfect Search, built for AI agents. And I'm going to say a lot of crazy things in this talk, so bear with me. They are all true. And I will tell you what is EXA, the story of EXA, so how we got here, and then where we're going. Okay, cool. If you take away anything from this talk, it is this slide. This is showing web searches per day over the past 30 years from humans and AIs. Obviously, it was all humans until around 2020s, and now we're in 2026. And actually, this year, we expect the number of searches from AI systems to exceed that of humans.

Pretty crazy. And then in the next few years, it should be 1,000 times more. So AI systems, AI products, whatever AI system you use, some together will search 1,000 times more than humans. That's a pretty crazy world that we're getting into, and the entire ecosystem of search is changing because of it. So EXA is the search engine for AI agents. We are the first search engine for AI, and now things are getting wild. We now serve over 5,000 companies, over 400,000 developers. I see some customers in the audience. We serve a very diverse set of agents from coding agents like Cursors.

If you use Cursor, at some point, the Cursor agent decides to search for the latest technical documentation or the news. It will be using EXA under the hood. We serve go-to-market agents like HubSpot, so we help their users get really high-quality lists of companies to sell to. We serve all sorts of financial agents. I was in New York a few weeks ago, and pretty much everyone there is now building financial agents, and they need the best financial data. Yeah, and really just a huge diversity of agents from labs to YC startups. Okay, but how did we get here? So what's the why of EXA? To me, that's always been the most important.

And there are a lot of ways to say the problem, but the short way is just misinformed anarchy. This is what we're trying to avoid. We want to create the opposite of this. So what does this mean? Well, this is the Internet, or it's a visual depiction of the Internet. You've got a bunch of pages, you've got blog posts, you've got company websites, you've got images, you've got tweets, you've got all sorts of things. And the web is really, really big, right? This is showing a few thousand pages. The web is on the order of a trillion pages, so a million times bigger than this. And it contains a huge amount of the world's information.

And that means that if this is just readily available, if you could go to any link and just get all the world's information, then I'm sure we all walk around with deep understanding of everything, right? Obviously not. It's messy. And it's crazy, and it can't fit in our heads. So we need information tools that can help synthesize this or filter it into the things you need to know. And we have a tool called Google, and it's a pretty solid tool. It could get you things like the Costco homepage or information about Taylor Swift or whatever you want to search. But it's not perfect, right?

And I love this example where you type shirts without stripes, and if you notice, you get shirts with stripes. Why is it doing that? Well, it's not trying to be a database of the world's information that gives you exactly what you want. It's a recommendation engine. And you can see this with more examples. So, find me everyone in Singapore who works on AI search and any blog post or research paper they've written. I bet you've never typed in anything like this to Google because you know it's not going to work. You're going to get links and documents that contain some of those words, but not actually a database result of all 412 people who match.

And it gets pretty serious, right? Like, I'm a citizen, I'm trying to be informed about what's going on in the world. I want to find the most important US news across all media. Or, you know, US news articles, whatever it is. You just don't trust Google to give you that, right? It's just going to recommend some things. It's almost akin to social media in a sense. It's a recommendation engine. Okay. So that means that no one really has a complete understanding of anything. I actually, when I walk around like SF or wherever I'm walking around and I see people, I often think like, no one knows what's going on in the world. Everyone's in this situation, including myself.

You know, you can imagine this person on the right on their phone, trying to find a new job. They're looking for biotech companies to work for. Are they going to get all the possible biotech companies that match? No. So there's always going to be this unknown of what's out there. Or, you know, maybe this other person with the headphones, maybe they're trying to understand what's going on in some region of the world. They're just not going to have a deep understanding. They can't, not only can they not find the information, they might not be able to trust the information.

So we basically are in a world where we live without this key information infrastructure that is so critical. And I think this is extremely important. So important that if we don't fix this problem, I believe we get a world that looks like this, a dystopia. I'm not kidding. This is AI generated version of San Francisco in 2035. And it's basically a world where no one, no person really understands what's going on. If people don't understand what's going on, then as we have this crazy AI technology that we all are talking about today, that's coming and the world is getting way more powerful, there's going to be conflict, all these things.

If people don't know what's going on in the world, this is very bad. We will be manipulated. We will make really bad decisions as individuals and as a society. And I think this, if we could fix this problem, it would be way better. And when I think about all the possible problems that are really important and neglected, to me, solving information is the most important and neglected problem. Okay. So that's the why of EXA. Quick story of how we got here. So I've been thinking about this problem for a very long time, way before even 2021 when we started, even since high school.

But I think what was really cool is that in 2021, it suddenly became possible in our eyes to build a new type of search engine because Transformers had gotten really good. So this is a time when GPT-3 had recently come out. GPT-3 was magical. You type in a paragraph of text and it fully understands you. At the same time, as we saw with Google, it doesn't fully understand you. And so what if you could combine the power of GPT-3 with a search engine? And maybe you could have perfect search over the world's information.

And actually the thought experiment that always drove me was, if we take a query, a complex query, and a document, and we run GPT-3 over it and we say, does this match? It'll do a really good job of saying it doesn't match. Now do that over a trillion documents for every search and you get a perfect search engine or near perfect. The problem is that would cost like $10 million per search. So it really becomes an interesting optimization problem. How do you billion X or trillion X reduce the cost of that? So that's the idea that started EXA. It's actually the first day of EXA, 2021. It's actually, by the way, our five-year anniversary as of yesterday.

So yeah, it's been a great time. I wish I took a better selfie here, but this is the first day. And EXA was basically built on the idea that traditional search engines use keywords. Keywords are very efficient and they can handle simple queries. But if you want to handle more complex queries, you just need to use neural networks. And particularly embeddings are a way of encoding. You can't run a neural network over every document for every query. But you can pre-process every document into some sort of structure like an embedding. And then you could use that, and that captures a lot of the intelligence of a neural network. And then you could use those embeddings.

Of course, embeddings have all their own problems. And often you want to combine embeddings and keywords. But certainly embeddings are a big part of the picture. And that's how you can handle shirts without stripes. We can handle this clearly. Another way of saying it is just stack more layers. Very bitter lesson-pilled. We were very early on to being bitter lesson-pilled. I don't know if you know this meme. If you don't, it probably looks really weird.

But if you want to handle more complex queries, you just need to use neural networks. And particularly embeddings are a way of encoding. You can't run a neural network over every document for every query. But you can pre-process every document into some sort of structure like an embedding. And then you could use that, and it captures a lot of the intelligence of a neural network. And then you could use those embeddings. Of course, embeddings have all their own problems. And often you want to combine embeddings and keywords. But certainly embeddings are a big part of the picture. And that's how you can handle shirts without stripes. We can handle this kind of clear. Another way of saying it is just stack more layers. Very bitter lesson-pilled. We were very early on to being bitter lesson-pilled. I don't know if you know this meme. If you don't, it probably looks really weird. But anyway, so we were very bitter lesson-pilled. So we did some crazy things, right? We raised a couple million dollars. We spent half of it on a GPU cluster. That was crazy at the time. We did a huge amount of research for really years, just heads down. And we did a lot of work. We were very new to search, to be honest. We were just really obsessed with the problem. And so we invented a lot of new stuff that I still haven't seen even today. So, history of Exa. In 2022, a year and a half after starting, we were called Metaphor at the time. Some of you might know it. We launched our first search engine to the world. And it was pretty exciting. It was a new way of doing search. A lot of people were really excited about it. The next big thing that happened two weeks later was ChatGPT came out. And that really changed the world. And this is what San Francisco looked like at the time, if you remember. And this is the Exa team at the time. Thank you. But everything was saved when we got this message on Twitter from this person who wanted API access to our search engine. And that was really weird, because we were never thinking that this was going to be an API. We were just trying to build a better search engine than Google. We'll figure out how to make money later. Then someone asked us for an API. We were like, no, we don't have an API. Sorry. But then we started getting more requests for API access, including my roommate who lived downstairs. And then we very quickly realized, wait a second. There's a business model. What we realized was AIs need search. Because the argument is basically, look, even GPT-5, a gigantic model, it's tiny in comparison to the internet, right? So these systems always need to search. You're never going to have GPT-6, GPT-7. It's not going to be able to just know everything about the world. It needs to be connected to a retrieval engine. And that was a really interesting insight, because these things now need a search API, right? And so we pretty quickly realized, okay, wait. AI's going to search the web. In fact, they're going to search the web way more than humans. And they're going to search in very different ways. So this is an example of what humans search like, right? They search simple queries. This is what Google was made for. Like "SpaceX News." It's good at that. But an AI system is very different, right? It kind of looks like this information guzzler creature that's insane. And it would be crazy if the same search engine that was optimal for humans was also optimal for these AI systems. So anyway, we realized, okay, if we build a search API for these AI agents—it wasn't called AI agents at the time, it was just AI or LLMs—then we could make money from that. That's a nice business model. And we think it's going to grow really fast. And also, the beautiful thing is it matches our initial mission, which is perfect search, right? Like AI systems really want perfect search. They don't want SEO. They don't want ads. They just want almost like a database of the world's information, which is what we were always trying to build. So we built the first search for LLMs. And yeah, in 2023, we said soon AIs will search more than humans. You know, three years later, it's now happening. And so then the next couple years, we built a lot of really crazy stuff. It's way more complex than just embedding search. It combines all sorts of systems, some of which are included here. And now we're a much bigger team. And that's how we got here. So just quick, what can you do with Exa? And then I'll talk about where we're going, how we're going to get the perfect search. Right now, we're the highest quality information for AI agents. And you could do all sorts of things. So for example, a lot of people like really complex queries. So you want to find every startup funded by YC working on AI, give me their batch and status. You could do that with Exa now. And you could make this arbitrarily complex. A lot of people aren't aware of this, but you could just use Exa and find data at any level. You could have a list of companies, people, blog posts, news articles that you want. It will take some time. Maybe not seconds. It might take minutes. But you'll get the information you want. At the same time, we also have the fastest search API in the world. So we have a 200 millisecond search endpoint. And that's what it feels like. It's super fast. It's way too fast for humans, right? But we're not serving humans. We're serving AI systems. So, for example, we serve some voice agents. And if you're a voice agent and you talk to the voice agent and it wants to do a search underneath the hood, every millisecond counts. You want it to do a search really fast so that it could process it with an LLM and then output the best audio back to the customer. We also have things like super efficient token extraction. Everyone's talking about the compute crunch and how everyone's spending way too much on LLMs. Well, Exa could help there because when the LLM makes a query, it wants to get just the information it needs, just the tokens it needs. And so we take the documents. We give you 10 documents. And then we'll give you only the most important, 100 tokens from those documents. And that will save you a lot of downstream LLM costs. We also have something where people don't necessarily want snippets from each document. They actually want structured output. So, say you're building a recruiting AI agent and you want to find all the engineers who recently left their big lab job. Give me the most cited paper that they've written. Give me the college they went to and the year they graduated. And we'll just give you that as structured output. It makes it really easy. And I think one takeaway here is we're not building one search engine. Perfect search is not one thing. It's actually we have 5,000 search engines for each of our 5,000 customers, right? We want to build our system so it's super flexible because we want every business to be super optimized. And that's a beautiful thing because we don't want to declare what is the perfect search. We want you to tell us what exactly do you want. Do you want super fast? Do you want the highest possible quality even though it takes minutes? Do you want to search only over these 1,000 domains? Do you want to never search over those 1,000 domains? Do you want to search within this time window? Do you want to never get product pages? Some people ask us for that. So there's all sorts of things that you could do. With Exa it's very flexible, customizable. And our search quality is really good for these AI agents because we've spent years doing research into how do you build a new type of search engine for agents. It's even better than Google, which was built for humans, which makes sense. A cool thing that we released recently was Exa Connect. Agents don't really care whether the information is from the public web or from private data sources. They just want the truth, right? And so it's always been obvious to us that we want to assemble all the world's information. It's perfect search over all the world's information, not just the web.

Do you want to search within this time window? Do you want to never get product pages?

Some people ask us for that. So there's all sorts of things that you could do. With XSET it's very flexible, customizable. And our search quality is really good for these AI agents because we've spent years doing research into how do you build a new type of search engine for agents. It's even better than Google, which was built for humans, which makes sense. Okay, a cool thing that we released recently was Exa Connect. So agents, they don't really care whether the information is from the public web or from private data sources. They just want the truth, right? And so it's always been obvious to us that we want to assemble all the world's information.

It's perfect search over all the world's information, not just the web. And so now we have a system where data providers can actually partner with Exa so that developers can then get the data from those providers. So we're basically creating a new market, a new economy for agents where if you have high valuable data, you could get paid from all the developers who want their agents to access that data. So we're creating this beautiful marketplace. I think it's really cool. It's a free market. The data providers can decide how much they think their content is worth, and then developers can decide what data they want. And then you could do cool things like this.

Research AI in for companies, monthly website visitors, similar web. Oh, I guess it went too fast. But you could get basically this is combining information from the public web and also information from similar web, which is not publicly available. So you could do queries like that right now. Okay. So that's where Exa is right now. But the future has always been super exciting to me, and the goal has always been perfect information. And now we know it's for AI agents. So instead of a world like this, no, bad. We want search to feel like this. It's actually really hard to describe what perfect information feels like.

The best way you could say it is literally any information query you have, it just works. No matter how complex that is. Another way you could think about it is it's as if you did a year of research in a second. So imagine no matter what you're looking for, whether it's people or companies or news, imagine you spent a whole year, you spent all of 2026 just doing research for it. You get that in a second. That should give you a sense of what perfect information feels like, and you do that for every search, all the crazy number of searches that AI agents are going to make. And so we want to move really fast at Exa.

We're moving extremely, basically every quarter or two quarters now, we have as much progress as we did the past five years. And it keeps being like that. It's exponential growth. And so our ambitions for 2027 are pretty crazy. We want the world to be like this, where everyone walks around with deep understanding of what's going on. I think it's particularly important because things like the 2028 presidential election are coming soon, and I would love for the entire world to have access to near-perfect information so that everyone is very informed going into that election.

You kind of can feel the gravity of what we're doing here, of perfect information, and I encourage others to try to do it, too. It's very important for the world. It's key information, it's key infrastructure. And that slide I showed at the beginning, it's not the whole picture, right? If you play it out, we're talking a thousand times more searches from AI systems than humans. You can't even capture that on a graph. That's 20 times. A thousand times would be all the way up there, top of a building or something, right? So it's crazy what's coming, and it's really happening.

Humans on average search a few times a day on Google, but when everyone has AI systems and every software product you use has AIs in it, every interaction you're doing is going to be grounding itself in search. So it's going to be a huge number of searches, and if each of those searches are as true as possible, as near-perfect, then the world looks like this in 2035. Yeah, I do think that if we're basically getting into a world where our technology is so good, it's really just a matter of can we coordinate and just decide together on sensible things, right?

If you look at San Francisco, there's amazing things happening here, and there's really stupid things happening here at the same time. That's just coordination problems, and coordination comes from the information we consume. So that's what we're working on. You all have a role to play in also getting to this world, so thank you all for working on whatever passion you're working on, and thanks for listening to me. All right, thank you. Thank you. It needs to be connected to a retrieval engine. And that was a really interesting insight, because these things now need a search API. Right? And so we pretty quickly realized, okay, wait. AI's going to search the web.

In fact, they're going to search the web way more than humans. And they're going to search in very different ways. So this is an example of what humans, this looks like when humans search, right? They search simple queries. This is what Google was made for. Like SpaceX News. It's good at that. But an AI system is very different, right? It kind of looks like this information guzzler creature that's like insane. And like it would be crazy if the same search engine that was optimal for humans was also optimal for these AI systems.

So anyway, we realized, okay, if we build a search API for these AI agents, or it wasn't called AI agents at the time, it was just AI's for LLMs, then we could make money from that. That's a nice business model. And we think it's going to grow really fast. And also, the beautiful thing is it matches our initial mission, which is perfect search, right? Like AI systems really want perfect search. They don't want SEO. They don't want ads. They just want almost like a database of the world's information, which is what we were always trying to build. So we built the first search for LLMs. And yeah, actually like in 2023, we said soon AI's will search more than humans.

You know, three years later, it's now happening. Okay. Yeah, and so then the next couple years, we built a lot of really crazy stuff. It's way more complex than just embedding search. It combines all sorts of systems, some of which are included here. And now, you know, we're a much bigger team. And that's how we got here. Okay, cool. So just quick, like what can you do with EXA? And then I'll talk about where we're going, how we're going to get the perfect search. So the present. So right now, yeah, we're the highest quality information for AI agents. And you could do all sorts of things. So for example, a lot of people like really complex queries.

So you want to find every startup funded by YC working on AI, give me their batch and status. You could do that with EXA now. And you could make this like arbitrarily complex. Like a lot of people aren't aware of this, but you could just use EXA and just find really like any level of data. You could have a list of companies, people, like blog posts, news articles that you want. It will take some time. So it'll take, you know, maybe not seconds. It might take minutes. But you'll get the information you want. At the same time, we also have the fastest search API in the world. So we have a 200 millisecond search endpoint. And that's what it feels like. So it's super fast.

It's way too fast for humans, right? But we're not serving humans. We're serving AI systems. And so like, for example, we serve some voice agents. And if you're a voice agent and you talk to the voice agent and it wants to do a search underneath the hood, you know, every millisecond counts. You want it to do a search really fast so that it could go, you know, process it with an LLM and then output the best audio back to the customer. We also have cool things like super efficient token extraction. So everyone's talking about the compute crunch and how everyone's spending way too much on LLMs.

Well, actually, Exa could help there because, you know, when the LLM makes a query, it wants to get just the information it needs, just the tokens it needs. And so we take the documents. We give you 10 documents. And then we'll give you only the most important, like, 100 tokens from those documents. And that will save you a lot of downstream LLM costs. We also, you know, some people, they don't necessarily want, you know, snippets from each document. They actually want structured output. So, hey, you know, like, let's say you're building a recruiting AI agent and you want to find, you know, all the engineers who recently left their big lab job.

Give me, like, you know, the most cited paper that they've written. Give me the college they went to and the year they graduated. And we'll just give you that as structured output. It makes it really easy. And, yeah, I think one takeaway here is we're not building one search engine. Like, perfect search is not one thing. It's actually we have 5,000 search engines for each of our 5,000 customers, right? We want to build our system so it's super flexible because we want every business to be super optimized. And that's, I think, a beautiful thing because, like, we don't want to declare what is the perfect search.

We want you to almost, you know, tell us what exactly do you want. Do you want super fast? Do you want, you know, the highest possible quality even though it takes minutes? Do you want to search only over these 1,000 domains? Do you want to never search over those 1,000 domains? Do you want to search within this time window? Do you want to never get product pages? Some people ask us for that. Like, so there's all sorts of things that you could do. With XSET it's very flexible, customizable.

And, yeah, I mean, our search quality is really good for these AI agents because we've spent years, you know, doing research into how do you build a new type of search engine for agents. It's even better than Google, which was built for humans, which makes sense. Okay, a cool thing that we released recently was Exa Connect. So, you know, agents, they don't really care whether the information is from the public web or from private data sources. They just want the truth, right? And so it's always been obvious to us that, you know, we want to assemble all the world's information. It's perfect search over all the world's information, not just the web.

And so now we have a system where data providers can actually partner with Exa so that developers can then get the data from those providers. So we're basically creating, like, a new market, a new economy for agents where, you know, if you have high valuable data, you could get paid from all the developers who want their agents to access that data. So we're creating this beautiful marketplace. I think it's really cool. It's like a free market. Like, the data providers can decide how much they think their content is worth, and then developers can decide what data they want.

And then you could do cool things like this. Research AI in for companies, monthly website visitors, similar web. Oh, I guess it went too fast. But you could get, like, basically this is combining information from the public web and also information from similar web, which is not publicly available. So you could do queries like that right now. Okay. So that's where Exa is right now. But the future has always been super exciting to me, and the goal has always been perfect information. And now we know it's for AI agents. So instead of a world like this, no, bad. We want search to kind of feel like this.

It's actually really hard to describe what perfect information feels like. The best way you could say it is, like, literally any information query you have, it just works. No matter how complex that is. Another way you could think about it is, like, it's as if you did a year of research in a second. So imagine no matter what you're looking for, whether it's people or companies or news, imagine you spent a whole year, you spent all of 2026 just doing research for it. You get that in a second. That should give you a sense of what perfect information feels like, and you do that for every search, all the, you know, crazy number of searches that AI agents are going to make.

And so, yeah, we want to move really fast at Exa. Like, we're moving extremely, like, basically every, basically a quarter or two quarters now, we have as much progress as we did the past five years. And it keeps being like that. It's exponential growth. And so, yeah, our ambitions for 2027 are pretty crazy. We want the world to be like this, where everyone walks around with, like, deep understanding of what's going on.

I think it's particularly important because, you know, things like the 2028 presidential election are coming out, are coming soon, and I would love for, you know, the entire world to have access to near-perfect information so that everyone is very informed going into that election. You kind of can feel the gravity of what we're doing here, of perfect information, and I encourage others to try to do it, too. It's very important for the world. It's, like, key information, it's key infrastructure. And, yeah, you have to, like, that slide I showed at the beginning, it's not the whole picture, right?

If you play it out, we're talking a thousand times more searches from AI systems than humans. You can't even capture that on a graph. Like, this is not, that's, like, 20 times. A thousand times would be all the way up there, like, top of a building or something, right? So it's crazy what's coming, and it's really happening. Basically, like, you know, humans on average search a few times a day on Google, but when everyone has, you know, AI systems and every software product you use has AIs in it, every interaction you're doing is going to be grounding itself in search.

So it's going to be a huge number of searches, and if each of those searches are, you know, as true as possible, as near-perfect, then the world looks like this in 2035. Yeah, I do think that if, I do think we're basically, our future is limited by ourselves. Like, we're basically getting into a world where our technology is so good, it's really just a matter of, like, can we coordinate and just decide together that, like, on sensible things, right? Like, if you look at San Francisco, there's amazing things happening here, and there's really stupid things happening here at the same time.

Like, that's just coordination problems, and coordination comes from the information we consume. So that's what we're working on. You all have a role to play in also getting to this world, so thank you all for working on whatever passion you're working on, and thanks for listening to me. All right, thank you.

Thank you.