Aravind Srinivas on building Perplexity and betting on the browser

Executive overview

Search is structurally broken for AI: Google's ad model prevents it from giving direct answers, because doing so would cannibalize billions in ad revenue from booking, shopping, and travel partners. Perplexity was built on that opening — answers with citations, no ads — and has grown to the point where its infrastructure needs a 10x rebuild to keep up.

The next bet is a browser. Not a chatbot with tabs, but a cognitive operating system where a single omnibox handles navigation, queries, and agentic tasks running asynchronously in the background.

The browser is an abstraction layer above chatbots — and the only platform where persistent, multi-step agents can run without depending on third-party MCP servers.

Why Google can't copy this

• Google's live Bard demo failed; stock dropped 6%. One mistake costs them billions.
• Ad revenue depends on keeping users in the query loop — direct answers kill that.
• The same AI search feature has launched at Google IO multiple years in a row, with a different name and VP each time.
• Google had the fourth or fifth best models for most of 2023–2024; startups had access to better AI than Google internally — unprecedented.
• Innovator's dilemma is real: it's not incompetent people, it's incentive structure.

How Perplexity started

• Co-founders met in grad school; no grand plan, just interesting people talking about ideas.
• First product: natural language SQL over Twitter, LinkedIn, GitHub — a search tool over relational databases.
• Pivoted when they realised they couldn't turn the whole web into tables; bet on language models to handle unstructured data.
• Launched a Discord bot that gave answers with cited sources — seven days after ChatGPT launched, before ChatGPT had web search.
• The aha moment: New Year's Eve, 700,000 queries on a slow, hallucination-prone product with a confusing name — people were sharing screenshots anyway.

The browser bet (Comet)

• One omnibox: navigation, informational queries, and agentic tasks in one place.
• Browser as a "cloud": multiple tasks running asynchronously, pulling personal context — email, calendar, shopping, finance, real estate research.
• Each prompt becomes its own process, the way Chrome made each tab its own process.
• Mobile is harder: Apple and Google OS rules block third-party app access; a browser sidesteps that.
• MCP dependency is a risk — if you commit entirely to MCP, you rely on third parties to implement it reliably. The browser agent can just operate websites the way a human would, regardless of whether an MCP server exists.
• Switching browsers is a high-friction decision; that friction is a moat once users are in.

Concrete agentic use cases

• Schedule meetings, reply to emails you don't want to read.
• Filter event applicant lists by criteria (e.g., scrape LinkedIn URLs, check school, accept or reject).
• Pay credit cards, complete forms, buy items on your behalf.
• Go and do research as a background scout — periodic, recurring tasks.
• Access other tabs and browsing history as context.

Competitive strategy

• OpenAI will build a browser. Google already has Chrome. Speed is the only mode.
• Brand matters more than people expect: ChatGPT added search inside it and did not kill Perplexity; OpenAI is buying Cursor's competitor and Cursor still exists.
• Perplexity's identity: fastest time-to-first-token, most focused on accuracy, best answer presentation.
• Tens of millions of users earns you the right to keep playing — you don't need hundreds of millions to survive competition.
• No AI product has a true within-app network effect yet (unlike WhatsApp). The browser is the play to create stickiness: passwords, wallet, agent memory, recurring tasks, shared task libraries.

Partnerships and data integrations

• Hotel bookings: Selfbook (native checkout).
• Reviews: TripAdvisor.
• Maps integration.
• Restaurants: Yelp.
• Shopping: direct merchant integrations, Shopify, Firmly for purchases.
• Finance: FMP.
• Sports: StatsPerform.
• Distribution: Samsung pre-install discussions; Nvidia partnership for European AI model delivery.

Business model

• Subscription revenue is already at a scale nobody expected; targeting a few billion a year from subs alone.
• Usage-based pricing for agents: users pay per completed task, normalised against what hiring a person would cost.
• Transaction cuts: if users buy through Perplexity, take a percentage — but CPA margins are historically lower than CPC, which is why Google never became a transaction platform.
• Google's ad margins are likely the best business model ever built; the goal is not to match them, but to build a great business well below that ceiling.

On AI coding tools

• Mandatory at Perplexity: at least one AI coding tool (Cursor + GitHub Copilot).
• ML researchers upload pseudocode screenshots from papers; Cursor implements the algorithm, writes unit tests, runs experiments — cutting 3–4 days to ~1 hour.
• Design feedback loop: take iOS screenshot, annotate with arrows, upload to Cursor, get SwiftUI changes back.
• Do not vibe code infrastructure or production systems — distributed systems and infra still require trained engineers.
• AI coding tools introduce new bugs that engineers sometimes can't trace. Claude Code is noticeably smarter than Cursor for complex tasks.

Advice for founders

• Work incredibly hard. Strategy doesn't substitute for it.
• Assume any successful product making hundreds of millions will be copied by a model company — they need to justify $50B capex and are actively looking for new revenue.
• Your moat is speed, identity, and brand — not a feature that can't be replicated.
• Start with something and build it fast; don't change the idea every week, but don't be rigid either.
• Grad school is a good co-founder filter: you talk to people because they're interesting, not because you're calculating future partnerships. That's also what YC's network provides.

Hallucinations and accuracy

• Building internal benchmarks to track hallucination rates continuously.
• Better search index + richer page snippets + multi-step reasoning per query = fewer hallucinations.
• For objective facts (scores, weather): accuracy is non-negotiable and verifiable.
• For contested topics (COVID origins, political bias): surface all perspectives, avoid taking a position. Automated evals are hard here because there's no ground truth; human raters with domain expertise are needed but hard to scale.