How Perplexity competed with Google by betting on better AI

Executive overview

Legacy search is optimised for ad clicks, not answers. When users stop clicking links, the ad economy that funds Google breaks.

Perplexity was built on a simple bet: skip the complex structured approach, feed raw web snippets directly to an LLM, and let the model improve over time. The bet paid off when instruction-following models got good enough to make the "dumb" approach work.

The core insight: the next generation search engine must orchestrate widgets, knowledge graphs, small models, and LLM answers — deciding which to use invisibly — and also fulfil the action, not just answer the question.

From research to company

• Ilya Sutskever told Srinivas early on: only unsupervised learning plus reinforcement learning leads to AGI; everything else is distraction.
• Reading In The Plex during a Google internship crystallised the ambition: build a company where AI research and product building reinforce each other.
• Two problem spaces fit that criterion: search and self-driving cars — both create a data flywheel where more usage improves the underlying AI.
• Co-founder Dennis was found through parallel research; they had written the same paper a day apart.

Early product iterations

• First product: a Twitter database organised into tables, queried via LLM-generated SQL using OpenAI Codex — built by three people in one month.
• The approach worked because social search (who someone follows, likes, unfollows) had never been queryable before.
• Attempted to replicate this for GitHub, LinkedIn, and other data silos; all refused to share data.
• Realised a general, unstructured approach — let the LLM handle reasoning at query time — was more scalable than building domain-specific structured indexes.

Building the web answer engine

• Prototype inspired by OpenAI's internal Truthbot (WebGPT): take top search results, use only cached snippets, feed everything into the prompt, generate a sourced answer.
• Key trade-off: no link-clicking or scrolling meant much lower latency, at the cost of depth.
• First version still took seven seconds; streaming wasn't implemented; verbosity had to be hard-coded to five sentences.
• First viral moment came from a hallucination: the model described a living academic in the past tense, confusing her with a deceased namesake.
• Adding follow-up questions doubled session engagement time and caused daily query counts to grow exponentially — the signal that the product was worth pursuing.

Competing with Google

• Perplexity's structural advantage: Google cannot easily replace its cluttered, ad-laden homepage with a pure-answer interface without destroying search revenue (~$200B/year).
• Microsoft's Bing Chat launch and Google's Bard announcement both happened within days of Perplexity closing its seed round; one investor extended due diligence from 30 to 45 days.
• Long-term thesis: as users talk directly to AI and agents execute tasks, search revenue will structurally decline — but Google's stock price makes that transition almost impossible to manage.
• Perplexity's edge over OpenAI and Anthropic: product taste and user obsession are in its DNA; it is not distracted by data centres, chips, or benchmark chasing.

Product philosophy

• "The user is never wrong": instead of asking users to write better prompts, the product should clarify ambiguous queries itself.
• Consumer products that feel magical take on user friction; enterprise software pushes friction back to the user.
• Primary metric: number of queries per day — reviewed weekly at all-hands with growth rates and decline investigations.
• Bug reports go directly to the responsible engineer without hierarchy; filing 50 bugs a day normalises quality expectations.
• Brutally honest feedback comes from X/Twitter; email is too polite; in-person demos produce only positive responses.

The future of search

• Current gap: Perplexity answers "what watch does Bezos wear?" but sends the purchase to Google; Google gets the monetisation credit.
• Goal: end-to-end experience — from problem to answer to fulfilled action — without the user leaving.
• Tension: a product card with a buy button is commercially necessary but perceived as an ad by early adopters.
• The hardest engineering problem is the orchestrator: a router that silently decides when to use a widget, a knowledge graph, a small fast model, or a full LLM reasoning chain.
• Whoever builds that orchestrator, operates it at billion-user scale, and solves the monetisation model for transactional queries becomes the next Google.