Models Are Commodities. Perplexity's CEO Says the Index Is the Moat

Nathan Labenz welcomes Aravind Srinivas back to *The Cognitive Revolution* six months after a March appearance, when Perplexity still sat on other people's APIs. Labenz says the product is the first new app in twenty years to meaningfully displace Google in his own daily work, that academic papers have started using it as an accuracy baseline, and that when he asked Replit's VP of AI, Michele Catasta, who else belonged on a list of serious AI players, Catasta named only Perplexity. The company is still about one-thousandth Google's size, yet it now serves millions of queries a day. Last time Srinivas hoped Google and Bing would not cut off index access. This time Labenz hears a different plan: work around the clock on a crawler, a search index, and Perplexity's own language models, because the giants will eventually notice generative startups eating at the core business and raise the drawbridge.

Asked, half in jest, whether search giants are already trying to buy them out, Srinivas is dry. There is only one giant, Google, and Google is in an antitrust case, so it would be the last to come after them. Microsoft he does not treat as a threat in the same way. On adoption he will share what he can. Perplexity serves multiple millions of queries every day and has grown six to seven times since the last conversation. If that rate holds for another six months, he thinks they would be a genuinely significant consumer product.

Against ChatGPT he points at Similarweb: Perplexity is about 40 times smaller, or roughly 2.5 to 3 percent of that traffic, and he says the ratio is a fair picture even on mobile, which Similarweb does not fully track. The more surprising comparison is Google's own Bard, where Perplexity is already in the 15-to-20-percent range of traffic — a consumer giant's product only about five times larger. Headlines about ChatGPT visits declining he calls misleading. What he cares about is retention. In consumer, he says, retention is the lifeblood of a product: really good retention is infinite life, and nothing can kill you, whereas a flashy surge with poor retention is something nobody wants to invest in or keep using. Perplexity's retention, he claims, is already strong, because the company nailed one use case — orchestrating search and language models together — instead of stuffing every feature into one app.

Labenz now splits his own searches: Google for locating a thing he already knows, Perplexity for a novel question, especially if it needs to be current or cited. Srinivas is less romantic. They are still largely a weekly-usage app. Getting to weekly was hard; getting from weekly to daily is the biggest hill a consumer company climbs. It requires reliability, speed, accuracy, constantly better answers, new features that are actually valuable, and ways for people to share what they learned — all at once, on a small team, which he says at least brings focus and adrenaline. Diffusion is still early. On a New York subway almost nobody uses ChatGPT; in a New York café maybe one laptop in fifty has it open; in San Francisco it is more like twenty in fifty. Among AI enthusiasts some already have a Perplexity or Claude tab open. That is not enough. The next two or three years are about being used because the product is useful, not because AI is cool. Search has an unfair advantage: curiosity is a fundamental human need, so you do not have to educate a market. He wants the 80/20 answer — what to eat in this neighborhood without posting on Twitter, or twelve minutes with a lawyer instead of an hour.

Srinivas's favorite origin story is also a product demo. His first founding engineer asked for health insurance. Srinivas had never been a CEO, had not bought a plan for himself, and his co-founders were covered through their spouses. Justworks, HIPAA, coinsurance — none of it meant anything. The company's Slack bot, wired to GPT-3.5, hallucinated with confidence. They bolted on web search, still Bing at the time, and he was able to choose a real plan. That is the unlock he keeps coming back to: if you know nothing about a topic, you should not need an expert. Hours collapse into minutes. You still click through the links. You just get the job done much faster.

The rest of his personal usage is a tour of a CEO who context-switches all day. Before a meeting he pulls a person's bio and what their company actually does, so he does not burn the first ten minutes on small talk and still looks prepared. He needed to download someone else's tweet as a video; Google offered twenty links with spam in the first few, Perplexity handed him tweetvideodownloader.com. After the Series A he asked how to park the cash so inflation would not eat it, with risk and reward side by side — a decision he would not trust ChatGPT to make, because a YouTube stunt is entertainment and a real allocation is not. When a mobile engineer said a problem was hard, he did not want to keep interrupting, so he taught himself SwiftUI components and native rendering versus WebView and came back with a different angle.

Labenz adds two of his own. A corrupted GitHub Codespaces environment threw a package error too new to be in any training set; Perplexity found other people who had hit it and turned the fix into commands he could run. Separately, he wanted to go to an Ashnikko concert and had been told he was too old. Google had almost nothing useful. Perplexity triangulated fan demographics in a Billboard-like way and gave him a fact-based reason to go. The failures, they agree, are usually coverage. When the index cannot find the page, the product would rather say there is not enough information than hallucinate, which feels like getting stuck. Default search still sometimes returns first-page-of-Google quality when the user wanted something deeper. And many people cannot turn a vague itch into a well-formed question. Google remains king, Srinivas says, because it told the world that keywords are enough. Asking a good question is a human skill. Perplexity still requires it.

Copilot exists because default search often does not go deep enough. It is an interactive companion that asks clarifying questions, queries the web in real time like an agent, and scores multiple pages before answering. Srinivas uses it too, but it is off by default and capped for free users, because people like free. To keep more daily Copilot uses affordable they switched the router from GPT-4 to a fine-tuned GPT-3.5. On the web they are testing a retrieve button so you can retry any query with Copilot; if it works, the phone gets it too. The honest problem is that even the builders do not always know when to use which mode. If the answer says the web lacks information, retry with the agent. If it feels hallucinated, retry with a smarter model. That is a lot to ask of a user. A good product, he says, should think for you.

They also moved sources above the answer so anyone in a hurry can click out without waiting, and they want to unify the two interfaces so free versus Pro is mostly a query budget, not a different brain. Default landing uses the index and does not retrieve live; Copilot does the on-the-fly search. If default search with GPT-3.5 were always fast and always right, he says, there would be no bigger alpha in the world — it would almost be criminal for that property to exist. It does not exist yet. Harder queries still want GPT-4, and live questions still want the agent. Until the product can hide that choice, Copilot remains useful and confusing at the same time.

Have they moved off search APIs? They use a bunch, and they do not rely on any single vendor shutting them off. They are building their own search index, and so are OpenAI and Anthropic, because everyone can see the same future. In a world where large language models are a commodity and the training recipes or weights slosh around in open source, the edge goes to the data markets — the people who own the best data in the world. There are on the order of a trillion pages. You cannot index all of them, and you do not even want a hundred billion in the index. Quantity is the wrong axis. You want the best pages, probably a billion or ten billion, the ones that matter to the knowledge worker, the researcher, the curious mind. Capture that distribution and you already have a moat. Few companies can even attempt it, because it is a chicken-and-egg: you need a product to build the index, and an index to have a product. Somehow, he says, Perplexity broke that loop, which is why they can now dream of being important when many companies, not just one, have good models.

Labenz asks how the web itself is changing — Reddit possibly walling off crawls, licensing deals, and a stunt by Nat Friedman, who hid white-on-white text on a page telling AI agents what to say, which then showed up in a Perplexity answer. Srinivas does not linger on adversarial content. His larger claim is that today's web becomes an API, the way disks became the cloud after the Mosaic browser moment. Links as we click them may lose their reason to exist. Advice to go build a browser misses the point: what is a browser in this era? Neeva already tried a Google-identical UI with a summary on top, same fonts and all, and it failed. You cannot win by copying the old shell.

The interface he imagines is opening a laptop to a search bar or a chat box that routes you into the right workflow. Chrome may not need to exist. Assistants become first-party citizens, they speak a protocol to each other, they live on operating systems with new security layers. That is why Perplexity is not building a browser. An AI-first OS is a distribution problem for Apple and the OEMs. Perplexity is comfortable living as the assistant on every platform that already exists, and waiting for someone else to ship the AI-native device.

The economics follow. Platforms will wall their data. If Amazon can answer any question about its catalog, it may not need to be indexed by Google. Google's leverage has always been that it routes the traffic without which you cannot be discovered. Facebook and Instagram were never deeply indexed anyway; Elon Musk is making a bet that not being on Google is survivable. Srinivas sees islands of data, each with its own assistant, and a global aggregator AI — Perplexity's seat — for research and knowledge, while commercial-intent search gets done inside each walled garden. Google is the middleman between buyers and sellers. If you can discover your own content, you do not need that middleman. Young people already look up restaurants on TikTok; Google is embedding itself there in response. Directions still need Maps. Everything else is the Yahoo move in reverse: still useful for a few things, losing most of the real value.

That vision is why monetization stays a subscription, not an ad network. Srinivas comes back to Jeff Bezos: keep shareholder interest and customer interest aligned, and you can keep making user-friendly decisions as you scale. Google's shareholder value comes from advertisers, not from searchers. Leaked internal email, in his telling, had an ads executive writing to a search executive to demand more ad slots because they were missing Ruth Porat's quarterly targets. Larry Page and Sergey Brin used to spend their cognitive bandwidth fixing bad queries, with small display ads on the side. Copying that machine is the wrong game. Think for the user, then invent a business that does not fight them.

He will not close the door on advertising, but he wants it to feel like Instagram, where you barely notice native ads. He even floats the idea that the question itself could be the ad: search a concert, see questions about other dates by the same band. Publishers might pay to be one of the URLs in the index. Bezos, again, found a second business in AWS and used it to subsidize the core. Perplexity Pro is not pushed hard. Srinivas would rather people sign up because they feel the value, so churn stays in control and revenue is real. Pro today is unlimited Copilot on GPT-4 with a fast, reliable router that searches the live web, GPT-4 or Claude 2 as the default model on every query, and unlimited file uploads for people who want to ask questions about their own documents. More features are coming, and they will land first on Pro.

The company is 25 to 30 people. They raised a $25 million Series A. Labenz asks whether compute spend is as high as payroll, or higher. Srinivas says yes, and that this is the correct allocation. GPUs are expensive; everyone prices off them. If a GPU costs X an hour, a hosted model on top is 2X, or more. The irony of infrastructure is that you have to spend in the short term in order to save later: make a serious capital investment, utilize it, commoditize it. Almost nobody in generative AI is printing their own cash — Midjourney is the exception — so the work is still VC-bankrolled, which is why they raised again. The original product was Bing plus the best available GPT. The strategy now is to replace every piece of that stack with their own version: scraper, index, models. Using GPT-4 and Claude 2 in the meantime is fine. Self-sufficiency on the key layers is the point.

How does a team of thirty ship this much? Srinivas does not reach for culture theater. The other option is not shipping, and that can kill the company. Momentum is everything in a startup; the CEO's job is to keep the company growing. Kinetic energy decays with friction unless you keep injecting potential. A true flywheel is rare. Google has one, which is why Larry and Sergey can sit on top of it. Meta has not: if Mark Zuckerberg takes his foot off the pedal, TikTok takes the ground. OpenAI and Anthropic are not there yet, and Google, having taken it easy, now lives in a deeply competitive space. Generative AI startups, he has tweeted, are a war zone. Incumbents are shipping because they know an AI wedge can steal the users they spent years acquiring. Zuckerberg putting an assistant and an image generator into WhatsApp and Messenger is the same logic: delay, and ChatGPT or Midjourney take the distribution.

Perplexity's one advantage is the search-plus-LLM combination Google most needs to protect and is least free to ship. Why not just replace google.com with Search Generative Experience? They cannot. The business model will not let them. That is why Srinivas is more bullish on a game that is hard for the incumbent to play: you cannot wake up one morning, flip a switch, and be done with Perplexity, especially when even Perplexity still fails plenty of queries. Nat Friedman has a line he likes: the more mistakes you make per unit time, the more you learn, and by the time a competitor learns X you have learned 10X. Keep moving, because the business model is still unclear and the risk is still high.

Labenz closes on a Silicon Valley joke Srinivas has overheard — people who would rather not think, and let the AI manage everything — and asks whether human-AI interaction is already unhealthy. Srinivas is not interested in replacing the feeling of talking to a person, even if a *Her*-style attachment is imaginable. For intellectual questions, especially in domains you are bad at, he thinks people will prefer the AI, because you can ask dumb follow-ups all night without shame. You might get an hour with Ilya Sutskever. You will not keep DMing him. A model that answers almost anything about neural nets with almost no hallucinations, you will bother for hours. The same is true of medicine, law, chemistry, physics, or sitting with your kids so nobody has to look dumb. Zuckerberg's video of asking an AI how to cook steak for his parents is the everyday version. Empathy, therapy, and Character.AI-style companions will take a real share of screen time. The question is not whether humans are replaced. It is whether the time gets better. He even likes the picture of two or three people in a group chat with an AI in the room: fewer fights, because "go do your research" can happen on the spot.