‘On Google, You Get Links. On Perplexity, You Get Answers’: An Interview With the CEO of Perplexity, the AI Startup That Wants to Kill Its Biggest Inspiration

I recently met with Aravind Srinivas, the CEO and co-founder of Perplexity. The previous day had been intense for him and his company. The AI startup had launched Perplexity Assistant for Android, a direct competitor to Google Assistant and Siri.

Srinivas’ career is impressive. After graduating from the Indian Institutes of Technology in Madras with a degree in electrical engineering, he earned a Ph.D. in computer science from the University of California, Berkeley. He excelled at both institutions, and his interest in deep learning led him to accept a fellowship at OpenAI in 2018. He then worked at DeepMind and Google for two years before returning to OpenAI as a research scientist. After less than a year, he decided to embark on a new adventure.

That was when he co-founded Perplexity, a startup dedicated to creating a conversational search engine–an “answer engine” rather than a traditional search engine–with artificial intelligence as a fundamental component. Perplexity became popular shortly after the launch of ChatGPT, and it was clear even then that the AI company’s model posed a significant threat to Google’s search engine.

More than two years have passed since its creation. Despite some controversy, Perplexity has established itself as a clear example of how AI development is poised to change many aspects of our lives.

Srinivas highlighted this during the interview. It was surprising to see how calm, serious, and composed he was when answering questions. He didn’t always respond immediately. He often took a few seconds to process the questions, much like how Perplexity functions. Emulating his answer engine, the clarity of his answers was truly remarkable. I had the opportunity to discuss the AI landscape in general and Perplexity specifically, and his message was consistently optimistic and promising.

How Perplexity Works Internally

Perplexity is an AI-powered search engine that stands out because it functions differently from standard conversational chatbots. While ChatGPT can occasionally hallucinate and make mistakes, Perplexity is much less prone to this issue. This reliability stems from the fact that the information provided by Perplexity is based on referenced sources and links, which, in theory, contain trustworthy information.

I asked Srinivas how Perplexity determines a source’s reliability to find out if the company has a system similar to Google’s PageRank. He replied, “Each of those links has a snippet, and sometimes the snippet is very detailed. Sometimes the snippet is pretty minimal, depending on how much we cached.”

Perplexity’s links go through several stages. The first stage is the initial information retrieval phase, “where you look for keyword matches.” The second stage is what Srinivas refers to as “n-grams,” which are “groups of words that match” the search criteria. Finally, the third stage “looks at semantic similarity.”

Srinivas explained that the next steps involve embedding “all words into vectors” to analyze “vector similarity.” Then, Perplexity applies additional filtering stages, which consider factors such as a domain’s number of visits, authority, and reputation.

The AI model uses all this information. Based on the user’s query, it decides “which sources are useful for writing the answer as part of the answer itself.”

Additionally, there are “human evaluations to measure ranking quality.” These evaluations can help identify spam domains. The Perplexity CEO explained that, in some cases, “once some users point out some hallucinations, we decide to lower their trust scores.”

However, Srinivas also pointed out that “most of the time, the trust scores are learned in an algorithmic way.”

“I would say it’s a mix of humans and algorithms,” Srinivas concluded, though he added that the influence of human engineering is less significant compared to that of algorithms.

Perplexity vs. Google

Comparing Perplexity and Google is inevitable. In my conversation with Srinivas, I wanted to understand how he would persuade an average person to try Perplexity so they could see the differences compared to the traditional Google search engine.

His answer was pretty straightforward:

“On Google, you get links. On Perplexity, you get answers.”

“I think that’s the simplest explanation,” he added, although it may not be 100% right. “It’s not like Google never gives you answers. Sure, if you type in how many kids Elon Musk has, it will just give you a good answer box, but it won’t be able to answer questions like, ‘What’s the average age of Elon’s kids?’”

According to Srinivas, this highlights a key difference between Perplexity and Google. He believes it’s preferable to develop a comprehensive solution that can answer any question rather than relying on Google’s small information boxes.

“Don’t you think that sometimes the box is better?” I asked, giving Srinivas an example. Users don’t always need lengthy answers to every question. For instance, a concise, straightforward answer is enough if you ask, “How old is Tom Cruise?”

However, Srinivas pointed out that this format may not always provide accurate information. If, for example, you were to ask about someone’s age on their birthday, the data might not be up to date. “It relies on the knowledge graph, the knowledge graph relies on a database from Wikipedia, and Wikipedia doesn’t update that fast,” he explained.

Srinivas emphasized that traditional search engines like Google can be affected by “imperfections in some databases” they depend on, making it essential to “take the information from multiple sources.” In contrast, Perplexity often combines answers from several sources, making it somewhat less likely (though not impossible) for the information to be incorrect.

I also posed an inevitable question regarding the evolution of search engines that have taken place in recent months. “Do you see the traditional search engine dying?” I asked.

Srinivas replied:

“I don’t think it’ll die for the next two to three years, at least. I think for informational queries, I already would consider it dead, except for certain widgets like sports and flight status, local maps, and maybe a little more accurate weather predictions–these sorts of things.”

Perplexity Poses a Significant Threat to Web Traffic, Too

As a search engine, Perplexity operates in a way that could change how users interact with the Internet. If the responses provided by Perplexity are high-quality, you might wonder why you should visit the original links and the sources from which the information is derived.

Srinivas explained, “I would say that definitely, the click-through rate on Perplexity answers is lower than traditional search engines because it’s a different way of interacting with information.” However, he believes that comparing it to traditional search engines like Google isn’t entirely appropriate.

“You’ve got to look into why people come into Perplexity. They aren’t coming here to read the news. They’re coming here to understand the news in the context of other things.”

For instance, if Xataka On posts an article about new regulations, users may want to understand how the policies can affect their plans and businesses. “Perplexity can take your article… and do some reasoning on top and give an answer,” he said. “That’s complementary to the article you wrote. It benefits from it, but it’s also observing a different value.”

Srinivas further detailed how Perplexity aims to support publishers and media content creators: “We are encouraging publishers to build these kinds of products within their site also using our APIs. We have this program called the Perplexity Publisher Program [that outlets can join]… and we are going to share revenue with them.”

Under this program, Perplexity can provide APIs for building products like suggested questions and offer other tools or free access to Perplexity Pro. Srinivas believes that his product can also “help journalists write better articles faster, with better research, [and] make their existing websites even better for the AI native world.”

How Perplexity Makes Money

Many AI platforms offer their features for free, but some also provide advanced versions through a subscription model. I wanted to explore Perplexity’s business model, how it currently generates revenue and its future plans for monetization.

According to Srinivas, “A lot of companies are really successful business models where they combine subscriptions and ads in a very, very impactful way.” Examples include Amazon, Netflix, YouTube, and Spotify.

Srinivas believes that AI companies like Perplexity are focused on proving to users that these options are worthwhile. According to him, “if AI is very useful to them on a daily basis,” paying for it won’t be an issue.

As such, Perplexity’s goal isn’t only to answer questions but also to enable deep research, effectively providing users with access to an expert financial analyst or a highly capable executive assistant right at their fingertips. “I think $20 a month feels really like a bargain if it works really well,” he concluded.

The Perplexity CEO envisions a future where the company can achieve this and persuade millions of people to pay the subscription fee. He pointed out that it’s feasible to see “100 million people paying for the true value added in their lives… That’s a lot of revenue, and it can be very profitable.”

Moreover, Srinivas indicated that this revenue could be reinvested to enhance the Perplexity AI, making it cheaper and more effective. “I’m pretty excited about it,” he said.

He’s also exploring advertising opportunities. In our conversation, he explained how Perplexity aims to find “a way where we can monetize both the free users and the paid users without compromising on the quality of the service.”

Europe and the Dilemma of Regulation vs. Innovation

We also discussed Europe’s role in AI development. I asked the Perplexity CEO what he thought about the current situation and what Europe would need to do to compete with the U.S. or China in AI progress.

The answer was clear to him: “I think move fast, remove some regulations, and move faster.” He said funding local startups would also be necessary: “The government should try to encourage them, remove blockers, create companies, and give them more visibility.” Providing tax incentives for startups would further accelerate growth.

Srinivas also recommended encouraging the development of open source projects. However, he wanted to clarify that while “some applications do require regulation, most of them do not.”

The engineer and entrepreneur stressed that “there’s tremendous talent in Europe. A lot of people go to America to work.” As he put it, “Figuring out ways to retain the talent by encouraging entrepreneurship would be a big example.”

“It’s not a good thing that any time you think about regulation, people always think about Europe. It’s established itself an image. So, I think [it’s necessary] to try to change that a little because it will feed into the narrative more and more.”

Let’s Talk About OpenAI

OpenAI is undoubtedly one of the major players in the AI field and a formidable competitor to any startup in the industry, including Perplexity.

OpenAI isn’t just developing LLMs and chatbots like ChatGPT. Recently, it launched Search GPT, positioning it as a direct competitor to Perplexity. However, OpenAI has its own family of models and its own “answer engine,” similar to Perplexity’s. Unlike OpenAI, Perplexity doesn’t develop its own LLMs.

The company has a Llama-based model called Sonar Large, which you can test in its “lab.” It’s also available as an API. Perplexity’s answer engine leverages multiple models—including OpenAI’s—to generate responses, with Sonar Large being one of them. Is this a challenge for Perplexity? Does the company envision a future where it no longer relies on OpenAI or other AI model developers? Srinivas didn’t seem concerned:

“We’ll use their models as long as they allow us to use them. And I feel like there are so many players in this game that nothing holds us to them in any manner. I think it’s not a big deal to use any model. I feel like it’s very natural to benefit from other people’s progress and build along the way and compete with them simultaneously.”

As a former OpenAI employee and now a competitor, what does Srinivas think about OpenAI becoming a for-profit company? On this point, he was cautious, noting that “there’s clearly some conflict” but admitting he’s not an expert in the field.

Still, he sees this shift “as a positive step for OpenAI to become a profitable company.” However, he acknowledged the process is complicated and expects experts to determine whether the transition and its conditions are appropriate.

The Rise of Open Source Models

The emergence of Meta’s Llama has demonstrated that open source models can be a strong alternative to proprietary LLMs. Perplexity has its Sonar Large model based on Llama, but the open source landscape is rapidly evolving.

For Srinivas, these models are already as good as proprietary ones. According to him, “DeepSeek V3 is almost as good as GPT-4,” and the same applies to DeepSeek R1 versus o1. In his view, “the story has changed.” Open source models are now competitive with those from OpenAI, Google, and Anthropic.

He believes that when Meta releases Llama 4, it “will show the world that open source is amazing, and I think this is really good” because it will drive down the cost of foundational models.

However, Srinivas sees a vicious cycle at play: “When something is open, and you have to go around, inference will become cheaper. Chips will become better. And so people will be able to build even more applications on top of models. That will be a lot more contribution from the open source community to make these models better.”

Perplexity Assistant Takes On the Challenge of Accessing iOS and Android

During the interview, Srinivas gave a small but interesting demonstration. It’s not every day that the CEO of a Silicon Valley company personally showcases his product, but that’s exactly what happened when he introduced Perplexity Assistant, the company’s new AI assistant for iOS and Android.

Srinivas picked up his Android phone and, after holding down the power button for a moment, launched Perplexity Assistant. He then said, “I’m going to be in Zurich on Monday. Could you set an alarm five minutes before sunrise?”

Perplexity Assistant displayed an animation as it “thought” and searched for the sunrise time in Zurich on Jan. 27, 2025. Moments later, it confirmed that the alarm had been successfully set.

However, things took a strange turn when he tried the same request with Siri on his iPhone. Siri responded that it couldn’t set alarms for dates beyond the next day—the first issue. Then, when Srinivas asked it to set an alarm for five minutes before sunrise the following day, Siri still failed, instead prompting him to manually select a time. ““[Siri] doesn’t know. It doesn’t know that it has to go on a search to the web [to find out the time for sunrise],” he explained.

At this point, he clarified, “I'm not saying they can't fix these things.” However, in the demo, Perplexity Assistant was clearly superior. “It can do a lot more cool things. It can book an Uber. It can send emails.” He claimed these features weren’t possible with Gemini or Apple’s assistant.

According to Srinivas, Perplexity has been working on this assistant since October, and “the models have gotten better.” He noted the difficulty of optimizing latency, speech recognition, app selection, web searches, and function calling—all critical components of the assistant’s performance.

But Perplexity now faces a major challenge: convincing users to switch from their default assistants on iOS and Android. Historically, default options—like Internet Explorer on Windows or Google Search on Android—have made it difficult for competitors to gain traction. However, Srinivas sees things differently:

“I feel like the ones with all the distribution advantages (Google and Apple) have the biggest brand reputation risk, also from making mistakes and hurting their user trust, and especially with Apple, like, they’re way more cautious about these things.”

He remained optimistic about Perplexity Assistant’s potential. “Assistant is this thing where the more user data you have, the more personalized it can get. Once you know that the assistant is truly just working for you, you're not going to switch away.” If assistants offer similar capabilities, users won’t switch, “but if the difference is quite radical, like how I showed you, then they are going to use the superior product.”

AI Advances and the Threat of Increasing Loneliness

AI development is revolutionizing industries, but its social impact is significant. Services like Replika can help people feel connected while also increasing isolation by reducing real-life interactions.

I asked Srinivas about this potential risk. Although Perplexity isn’t a conversational chatbot, he acknowledged both sides of the issue: “There are obviously dangers in people getting too attached to AIs. And if AIs stop doing what they want, or people getting frustrated with it and potentially emotionally getting attached, and all this stuff, definitely there are some consequences of making mistakes there.” But he also pointed out another angle:

“The other side of the coin is loneliness, one of the most prevalent mental diseases. And you could argue that having these AIs that you can talk to is doing something to address that problem, too, rather than having no one to talk to.”

Srinivas doesn’t believe AI chatbots should be banned but emphasized responsible design to avoid harmful consequences.

“I’m very happy that we are not in that space, honestly, because almost every query you ask on Perplexity goes through the web and the sources. So, you get boring—the educated, uncool response.” Srinivas made it clear: Perplexity would stick to its answer engine and leave conversational AI to others.

Trends in AI: Agents, Reasoning Models, and AGI

In light of these fascinating topics, several exciting AI trends are emerging for 2025. We explored AI agents, reasoning models, and the industry’s ultimate goal: artificial general intelligence (AGI).

Srinivas began with an insightful observation: “Well, they’re all quite related.” He then shared his perspective, starting with AGI: “AGI has too many definitions out there, and there are too many uses for this word. But it’s just like a remote digital knowledge worker, like an intern you can hire.” According to him, current models already demonstrate impressive capabilities:

“But the nice thing about hiring an undergrad intern is that they’re not going to stop there. They’re going to work hard and improve. At the end of the internship, they should be better. When they join full-time, they’ll be even better. I would say we’re still not there at AGI. And the reliability is still very low.”

The same applies to AI agents, he explained. While some programming agents perform remarkably well, “we don’t have an agent that can just debug any bug. Honestly, whenever I use our product and test it, and I find a bug, I would love it if there’s an AI that can take the screenshot and the link I send, know exactly how to triage the bug, know exactly which part of the code base to go and fix it, and know exactly what unit test to write to test it, stage it, and then push it to production.”

Srinivas acknowledged that the current capabilities of AI agents are still limited. However, he expects a major breakthrough in the next two to three years. He believes these agents “will have a huge impact on the world in terms of increasing productivity,” noting that they’ll inevitably affect some jobs. However, the transition may happen so seamlessly that “you wouldn't even feel like an agent is there.”

Srinivas placed particular emphasis on reasoning models, which he described as “the answer to this puzzle.” He explained that with these models, if “something doesn’t work, the model knows to go and seek help or figure out, like alternate paths.” He further elaborated:

“Ambiguity, uncertainty, is only resolved through good reasoning, right? It’s the case in humans, too. Most people will solve problems that are very precise and clear. For problems that are unclear and precise, you require higher-order thinking skills to break them down and collect more input or data from the world or other people and then use that new information to make deductions. Then, you end up solving the problem. That is the part that’s lacking today.”

AI Is Like “a Roadster” for the Mind

I concluded the interview by reminding Srinivas of former Apple CEO Steve Jobs’ quote: “Computers are like a bicycle for the mind.”

I believe AI goes beyond that. To me, it’s more like a motorcycle for the mind. When I mentioned this to Srinivas, he jokingly suggested it could even be compared to “a roadster.”

However, some argue that AI could potentially wipe out humanity, a narrative often depicted in Hollywood films. When asked about this concern, Srinivas said, “Obviously, movies are more fun. If that’s a situation where we’re going to die, and then someone comes and saves us, it makes for a good story.” He doesn’t believe these kinds of apocalyptic scenarios are likely, but he also doesn’t view the future as entirely bright.

“I do think people are going to lose some jobs, [but] it’s always happened,” Srinivas added. He recalled professions like stenographers and accountants that dwindled after the introduction of calculators. While computers indeed made us more productive, they also led to job losses, a trend that mirrored the rise of the Internet.

Srinivas compared the situation to the arrival of AI. He stressed this point but added an important caveat: “I think the effects will be felt faster in this case.”

The Perplexity CEO is particularly optimistic about the future, saying, “The marginal cost of creation and creativity, and content and research [is] going to rapidly approach zero.” He pointed out that this shift will provide numerous opportunities. However, like any other significant change, it’ll also present risks “like every other paradigm shift.”

He concluded, “We should just embrace it and try to make the best out of [AI] and push the human race forward.”

Image | Xataka On

Related | I’ve Been Using ChatGPT Search as My Default Search Engine Thanks to a Chrome Extension. Google Has a Big Problem