For the past two decades, search engines like Google and Bing have been our primary gateways to the internet. Yet, over time, their signature “10 blue links” have been increasingly buried by advertisements and algorithm-optimized content. Driven by ad-based revenue models, these sites often feel more like auctioneers of users’ attention than trustworthy guides to the World Wide Web.
Aravind Srinivas and his cofounders at Perplexity.ai are building an alternative. Instead of entering keywords and sorting through a tangle of links, users pose their questions directly to Perplexity.ai and receive concise, accurate answers backed up by a curated set of sources. Powered by large language models (LLMs), this “answer engine” places users, not advertisers, at its center. This shift promises to transform how we discover, access, and consume knowledge online—and, with that, the structure of the internet as we know it today.
Wresting user loyalty from the reigning search giants is no small feat. Yet Aravind, armed with a Ph.D. from Berkeley and work experience at OpenAI and DeepMind, has long struck me as the ideal entrepreneur for this challenge. In this conversation, edited for clarity and length, Aravind and I discuss the origins of Perplexity, his approach to incumbents and interface design, and his advice for fellow AI founders.
Let’s start with your background. What led you to found Perplexity?
I was born in Chennai, India. Our culture values scholarship, even more so than financial success. Take cricket, for example, which is the most popular sport in India and is especially beloved in Chennai. For my family and friends, watching cricket is not just about entertainment. We geek out over the statistics, like knowing the averages, strike rates, and economy rates of each player by heart.
This drive to learn and understand shaped my education. I worked hard to gain admission to the IITs (Indian Institutes of Technology) and later dove into deep learning research, which led me to my Ph.D. at U.C. Berkeley. Movies like Pirates of Silicon Valley and The Social Network had sparked my interest in entrepreneurship early on. But once I arrived at Berkeley, I realized that most of the people doing startups were college kids in YC. I couldn’t take that path anymore. I was searching for examples of entrepreneurs who came from academia to take inspiration from.
A turning point came in 2019 during my internship at DeepMind in London. They had an incredible library and, in the evenings, when I was done focusing on my project, I read books about Google’s early days and got quite inspired by Larry Page. I was fascinated by the evolution of PageRank and how it led to a company that could create incredible advances like transformers, a new architecture for deep learning models. I reached out to the inventor of transformers, Ashish Vaswani, for an internship. Together, we worked on developing deep learning models for vision and making transformers more universal as a computational block.
While this work was fascinating, I lacked a clear startup idea. Deep learning still seemed very academic. Yet, by the summer of 2022, generative AI startups like Jasper and GitHub Copilot were beginning to make real revenues. The growing excitement among mainstream users convinced me that my interests were no longer only scholarly. It finally felt like the right moment to start a company. I was fortunate to have funding support from investors like Elad Gil and Nat Friedman, along with three great cofounders, Denis (Yarats), Johnny (Ho), and Andy (Konwinski). They all believed when there wasn’t really anything. That’s how Perplexity was born.
What drew you to the problem of search?
As I mentioned, I’m a big fan of Larry Page and Google. I’ve always had a drive to do something of the same scale and ambition as Google. Being scholarly, accurate, and truthful; having the answer at your fingertips and being able recall its sources: these are things that I value and strive to embody. Building a product that helps me and people around me become smarter every day and raises the knowledge capital of the planet holds deep personal significance for me. It’s much more than just a way to make money.
Before settling on search, you explored other problems.
Yes, our initial focus was on translating natural language to SQL. This was an idea proposed to us by our first investor, Elad Gil. Our goal was to build a copilot for data analysts. Being a data nerd myself, the idea of enabling more people to be data nerds really resonated with me.
Why did you not continue with that idea?
At the time, the technology was not sufficiently advanced. Codex was impressive, but it didn’t measure up to the capabilities of GPT-3.5 Turbo or GPT-4. In addition, the SQL market is very fragmented, which made it difficult to enter and establish a foothold there. Large warehouse companies that could afford our services were often locked into Snowflake or Databricks, while smaller companies didn’t have enough data to warrant moving away from basic spreadsheets. Each company also had its own unique data storage methods, which complicated our efforts to create a one-size-fits-all solution.
While this idea didn’t ultimately work out, the experience did help us grow. We shared our prototype, BirdSQL, on Twitter, and the response was overwhelmingly positive. It worked so well that Jack Dorsey, despite being inactive on the platform for a long time, suddenly came out of hibernation and tweeted about it. His endorsement brought a surge of attention and traffic to our product and helped us overcome the cold start problem of nobody knowing who we were. This early viral growth was key to our eventual success.
You’ve described Perplexity as an “answer engine.” Can you explain what you mean by that term, and how it differs from a traditional search engine?
Sure. The traditional approach to search was to return ten blue links, which users then had to comb through to find the information they were looking for. Over the past few years, this model has been evolving to provide direct answers to users’ questions. That’s what I mean by “answer engine”: users can ask any question directly and receive an actual answer, not just a list of web pages that may or may not contain the answer. Google began moving in this direction around 2020 using simple text extraction. Perplexity’s goal is to answer more complex questions that require synthesizing content from multiple pages and providing fast, accurate answers using LLMs.
For two decades, we’ve all been conditioned to use keywords to search the web because that’s how the leading search engines were designed. Today, LLMs are changing the way that we interact with computers to both find and consume information. In addition to providing direct, concise answers, LLMs can ask clarifying questions and act as your copilot while you browse the web. Over time, they’ll also be able to help you get things done by executing tasks. That’s our vision at Perplexity: to give everyone access to infinite knowledge and productivity, and to improve their lives by enabling them to interact with the internet in more intuitive, efficient ways.
Search engines have heavily influenced the existing economics of the internet, essentially optimizing it for advertising. How might answer engines allow us to create an alternative model?
LLMs will certainly change the existing advertising dynamics of the web—hopefully for the better! That’s because these models provide a more significant gain in relevance than any previous targeting technology. Reaching people will become even easier because queries will have higher intent. If I were an advertiser, I would focus on describing my product as accurately as possible on my website so that an LLM considers it citation-worthy. Instead of optimizing for clicks, I would optimize for high-quality content.
I’m still trying to wrap my head around what a sponsored citation means. As a former Ph.D. student, I’m approaching the problem from an academic perspective. Some journals will publish your article if you pay, but their reputation is much lower than peer-reviewed journals. In the next iteration of the internet, clear, trustworthy sources of information will ideally only cite content that other, verified sources have provided. In this scenario, the LLM will be the judge. There are also other ways that people will want to interact with and verify queries beyond citations. I’ve been thinking about this but haven’t quite figured it out yet.
Building an answer engine is no small feat. How did you get started?
We began by bootstrapping a database by scraping Twitter and powering search over that. That became BirdSQL. We also rely on an existing search index that pulls content from the web and organizes it. Perplexity operates an additional layer of abstraction on top of this content, which we synthesize and organize even further. As we’ve grown, we’ve started building our own index too.
How do you think about other players in the search ecosystem?
That’s a great question. Let me illustrate it with a story. We shook hands on a seed round with our venture firm, NEA, on a Friday afternoon. Afterward, I went from Sand Hill Road to hang out at Blue Bottle in Palo Alto with my cofounder, Denis, just to relax. Then I saw that someone had sent me a Verge article with leaked screenshots of Bing’s new chat UI. And Denis and I were immediately worried. We thought, “Oh man, our term sheet has a 30-day due diligence clause! NEA is going to back out of the deal.” Thankfully, our venture partner, Pete Sonsini, called us the next day and reassured us. He was like, “We trust you, you’ll figure it out.” That gave us the confidence to continue.
In the end, nothing bad happened. Our growth remained exponential, while Bing’s growth stagnated. This made us feel like they missed an opportunity, rather than us facing any issues because of their launch.
Why do you think that is?
In my opinion, Bing’s product was confusing. It was also limited to Microsoft accounts and the Edge browser, so it was not widely accessible. They tried to incorporate multiple features into one product, such as search, chat, and multi-turn conversations using GPT-4, which cluttered the user experience. By contrast, we focused only on creating an answer engine with citations and avoided freeform conversations. This clear focus helped us create a useful product, while Bing’s product, despite the hype, was less clear about its exact use cases.
Like Bing, Google also faces a dilemma because improving the generative search experience could undermine their ad revenue. For example, a query like “plan a trip to Tokyo” on Google, even with generative search enabled, will still display ads because the travel industry pays Google a lot of money to distribute their links. We realized that competing with Google would be manageable because they have a vested interest in protecting their ad revenue, which limits their ability to provide direct answers. Our main challenger was Microsoft, but that seems manageable now.
A typical complaint about ChatGPT is that its answers are not trustworthy. How does Perplexity ensure accuracy?
We have an incredible search and AI team, which Denis leads. The accuracy of answers relates a lot to search ranking, which is a notoriously challenging problem. At Perplexity, we use a more modern version of PageRank to build a trust map of the web. For example, sites like The New York Times are generally more reliable than Substack posts, which may be more opinionated. We use heuristics and data-driven learning from past queries to improve our results. If Google focused solely on providing correct answers, it could dramatically improve its search product with minimal effort. It’s business interests that prevent them from doing so.
With every new consumer tech product, there’s a scalability tipping point where it becomes the default for a subset of behaviors for people. How long do you think it is before alternatives to big-tech search reach mass-market awareness?
I hope it’s soon! For the moment, it seems like the world is still happy with Google, as their traffic has not materially changed. Just as Google and Facebook transformed how people consume news, a shift away from traditional search engines will happen eventually. Yet, until our product achieves full parity with Google search, and then gets 10x better, people will stick with the status quo. It might take a few months or a year to reach that point. Even then, many people will keep using Google out of habit, similar to how people still use Yahoo and MSN. Steady, mindful progress will help us become part of people’s day-to day workflows.
Let’s talk about the interface. ChatGPT was, at core, an interface innovation that spurred mass consumer adoption. How do you think about the user interface at Perplexity?
We’re fortunate to have great design and product teams led by very talented folks, Johnny and Henry (Modisett). Our general approach is to minimize the number of buttons and choices that users have to make during any given interaction with the product. Google and Apple are big inspirations in this regard. The core Google search is so simple, right? People sometimes complain that Google doesn’t add enough new features, yet they may not realize that doing so creates more responsibility and confusion for users.
Our philosophy at Perplexity is to do the hard work ourselves to make things as simple as possible for our users. Our bias is toward removing things rather than adding them. We push updates to production and observe interaction patterns. If a feature is not used, we remove it, even if some people complain. It’s essential to have a firm compass as a founder, which gives you the courage to say “no” to things.
Another key aspect of interface design is intuitiveness. Users shouldn’t need a manual to use your product. It has to be self-explanatory. That’s why when people find Perplexity on Google, they’re directed straight to our product without any long explanations or the need to log in. Removing as much friction as possible is another company philosophy.
Right now, the experience of Perplexity is primarily on the web. We’re focusing more on mobile and creating a truly mobile-native search experience. Features like swiping left or right to delete threads and voice-based interaction can only be done on mobile. The bar is really high, because everything has to work seamlessly.
What are one or two things about Perplexity’s product that keep you up at night?
I use our product a lot, so if it doesn’t work as expected, I know right away. Initially, I had the hacker mentality of suggesting a fix for every issue. Over time, after pushback from one of our cofounders, I’ve shifted to thinking about the bigger picture: namely, how can we address a wider range of issues with a more general solution? That’s the best way to ship scalable and sustainable improvements.
I always try to see things from our users’ perspective and talk to as many of them as possible. Larry Page had a philosophy when he started Google that “the user is never wrong.” It’s simple yet profound. When people complain about something not working for them, I reply to as many of their DMs as I can. And I try my best to inject this mentality into everyone at our company. Don’t blame users for not phrasing their questions correctly, or for not knowing there’s a button to share something. It’s our job to fix these things and make the product more intuitive.
Another big thing on my mind is making people aware that we exist and growing our user base. We want to achieve this by having a truly excellent product that users can rely on. Some folks say they think Perplexity is great, but they’re hesitant to switch from an incumbent because we’re only a few months old. Earning users’ trust so they feel comfortable making that switch is a significant challenge.
If you were to start over from scratch and found a company that leverages LLMs today, what would you build?
I would do the same thing again. There’s nothing more ambitious than changing the way people seek information and become smarter. Manuscripts, the printing press, libraries, search engines, and now answer engines and chatbots. This is a once in a generation moment.
Any final words of advice for fellow founders?
It may sound cliche, but start with what you love. It takes time to find the sweet spot between what you love and what the market is demanding. Eventually, you’ll converge to that point. The market will push you there. But it’s better to begin from a place of passion rather than going after what a VC or the market is telling you to do. So, prioritize what excites you most at the beginning. If you’re already doing that, stay patient. The competition is fierce, and anything you’re building is probably being attempted by many others. Stay focused, push yourself, and have a high bias for action. Don’t just strategize on a whiteboard. Get your product out there, gather feedback from users, and iterate on your idea in public.
