How Aravind Srinivas built Perplexity into a Google challenger

Executive overview

Most search engines return links; Perplexity returns answers with citations. Aravind Srinivas left OpenAI in 2022 with no product and no team, cold-emailed investors, and launched a week after ChatGPT with no sign-up and no wait list.

The core insight: incumbents can't cannibalise their own ad revenue, which is exactly the gap a startup can exploit.

From IIT Madras to Berkeley to OpenAI

• Studied at IIT Madras (India's MIT equivalent); ~1 million applicants compete for a small number of CS places
• PhD at UC Berkeley in the same AI lab as OpenAI's research lead Ilya Sutskever; shares thesis advisor Peter Abbeel
• Interned at DeepMind in 2019; read How Google Works in the DeepMind library — sparked his obsession with search
• Inspired by Larry Page's principle: only academia or founding a company gives you the freedom to pursue ambitious goals
• Interned at OpenAI in 2018; returned full-time after his PhD before quitting to found Perplexity

The founding moment

• Cold-emailed investors Nat Friedman and Elad Gil at idea stage — before any product existed; both agreed to invest
• Co-founder Johnny Ho was world #1 in competitive coding; co-founder Dennis was a former Bing engineer
• The entire team was search-obsessed: morning design discussions, dinner conversations — always about search
• First prototype: a Slack bot using GPT-3.5 to answer internal company questions (health insurance, admin tasks)
• Problem with the bot: it hallucinated; Dennis's fix — plug it into a live web index — became Perplexity's core architecture

The Twitter search demo that attracted top investors

• Built a search tool over Twitter's full graph: query follower networks, find liked tweets, surface social patterns
• Jack Dorsey tweeted it unprompted; the viral "AI knows everything about me" moment drove early growth
• Demo attracted Yann LeCun (Meta chief AI scientist), Andrej Karpathy, and Jeff Dean (Google AI head) as investors
• Twitter search was shut down after Musk's API rule changes

Launch strategy: the contrarian bet

• ChatGPT launched November 30, 2022; Perplexity launched December 7 — one week later
• No sign-up, no wait list, no demo video — just a search bar on perplexity.ai
• No chat interface at launch: a single query box returning answers with inline citations
• Citations came from the team's academic background — treating the product as if a researcher or journalist wrote it
• Michael Dell messaged Aravind directly on LinkedIn after launch to say it was a great app

Why Google can't easily copy this

• Google's ad executives internally pushed for more ads per search page to hit quarterly revenue targets (cited in antitrust trial evidence)
• Replicating Perplexity's UX would cannibalise Google's cost-per-click revenue model
• Startups can serve expensive models at a loss, iterate fast, and learn — incumbents cannot afford to make those mistakes
• Section 230 protections don't cover AI-generated content, creating legal risk for large incumbents that startups can absorb more easily

Growth, metrics, and business model

• At time of interview: 3 million+ queries per day, ~10 million monthly active users
• Users in US, UK, Canada, Japan, Germany, France; product handles any query language and responds in kind
• Tens of thousands of paying subscribers (paid tier unlocks more "copilot" queries)
• Not yet profitable; infrastructure costs scale with model usage
• KPIs tracked weekly: query volume, retention, user growth — monetisation is a secondary indicator, not the primary target
• Monetisation strategy: not Google's ad model; exact approach undisclosed but explicitly designed to disrupt it

Product principles and defensibility

• Five axes that must all work simultaneously: accuracy, reliability, latency, UI delight, personalisation
• Aravind's estimate: 1 in 100,000 startups can excel at all five; add iteration, and it's 1 in 1 million
• Retention requires trust built over time — users need to believe Perplexity will still exist before switching default search
• Added follow-up question suggestions because knowing what to ask is the harder skill; AI shifts critical thinking to question formulation
• Product roadmap: continuous accuracy and index-freshness improvements (citing Google's 20 years of silent backend work as the benchmark)

Content moderation and fake news

• Two-part approach: (1) strong LLM reasoning that can synthesise across multiple sources without being misled by any single one; (2) a domain trust/authority score analogous to PageRank
• Srinivas is sceptical of over-restriction: if information is freely available on Google or YouTube, blocking it on Perplexity is inconsistent
• His stated role: make the product work well and educate users on responsible use — not act as a legal or moral authority