Gunning for Google with Perplexity CEO Aravind Srinivas
Summary
Perplexity had reached multiple millions of queries per day, up 7x since the prior interview, while remaining roughly 40x smaller than ChatGPT at 2.5%–3% of its traffic. Amjad Masad argues retention matters more than launch spikes — “you have infinite life if you have really good retention” — and says Perplexity already draws traffic equal to 15%–20% of Bing’s, leaving “so much more alpha” if it can sustain growth.
The defining consumer challenge is turning a weekly research habit into daily use without sacrificing speed, accuracy, or trust. San Francisco’s AI bubble — perhaps 20 of 50 café laptops on ChatGPT versus one of 50 in New York — is not mainstream adoption. Masad’s next two-to-three-year goal is for people to use Perplexity because it is “really useful,” not because AI is cool.
Perplexity’s differentiation is orchestration: live search supplies current evidence while models convert it into answers, executable commands, or a decision-ready 80/20. Health-insurance selection, Docker troubleshooting, meeting preparation, and investment research show the wedge. But asking users to choose between default search and Copilot remains UX debt: “The product should think for you.”
Masad’s strategic call is that as LLM weights and training recipes commoditize, “the edge goes to the data markets,” with a moat around a curated search index. From roughly a trillion web pages, the valuable target might be the best 1 billion or 10 billion. Perplexity is diversifying its search APIs while building its own scraper, index, and models, having “broke that asymmetry” between needing a product to fund an index and an index to build a product.
Google’s advertising economics create a structural opening because its shareholder value is tied to advertisers, whereas subscriptions can align Perplexity directly with users. Masad is not categorically anti-ad: “What if the whole question is an ad?” points toward sponsored questions or paid inclusion that feels useful rather than interruptive. For now, the company prioritizes retention and controlled churn over pushing harder paywalls.
Vertical integration carries an unusually heavy capital bill: after a $25 million Series A, the 25–30-person company was spending at least as much on compute as on its team. Masad sees no alternative to training its own models and making short-term infrastructure investments that lower future serving costs. The company also aims to remain resilient if an individual search provider shuts it off.
The broader end state is an AI-first operating layer where assistants, rather than pages or browsers, become “first-party citizens.” Walled commercial assistants could erode Google’s middleman role, while a global service remains valuable for research and knowledge. Masad expects humans to embrace effectively unlimited expert help because they can “keep digging” without embarrassment, including inside group conversations rather than only one-to-one chats.
Deep dive
1. Strong retention gives a small search challenger time to compound
Masad reports “multiple millions of queries every day,” 7x the level discussed at his previous appearance. Similarweb’s rough comparison still makes Perplexity 40x smaller than ChatGPT, or about 2.5%–3% of its traffic, but its volume already equals roughly 15%–20% of Bing’s.
His preferred health metric is retention, not a flashy acquisition spike: “You have infinite life if you have really good retention.” Perplexity’s combination of search and LLMs has established a differentiated weekly-use product; converting that behavior into daily use is “the biggest hill to climb.”
Labenz’s own habit illustrates the wedge: Google remains useful for quick lookups or locating something already known, while Perplexity gets novel questions requiring current information and links. Masad cautions that this is still an enthusiast pattern: New York cafés might show ChatGPT on one of 50 laptops, against 20 of 50 in San Francisco.
2. The product earns its place by compressing unfamiliar work
Perplexity’s founding use case came when Masad’s first engineer requested health insurance. Masad had no understanding of the compliance language, while the company’s GPT-3.5 Slack bot hallucinated; adding web search let him determine the right plan and finish in minutes what otherwise required hours or an expert.
The recurring CEO use case is context switching. Masad researches people before meetings, learns unfamiliar SwiftUI or rendering concepts without repeatedly interrupting engineers, and finds specific utilities such as a tweet-video downloader without sorting through “the first three or four” likely-spam Google results.
His practical target is the “80/20”: preliminary research might reduce a lawyer’s paid hour to 12 minutes. For consequential investment decisions, however, he wants options, risks, and returns to research rather than blindly following ChatGPT — the latter belongs in a “MrBeast-style YouTube video,” not real capital allocation.
Labenz offers two stress tests. Perplexity found the recent web evidence behind a broken GitHub Codespace and converted it into runnable commands; for an Ashnikko concert, it triangulated audience demographics to answer whether he would feel too old, despite there being no single factual truth.
3. Copilot improves depth but exposes an unfinished product decision
Masad says ordinary search sometimes returns “not sufficient information” precisely because Perplexity refuses to invent an answer. Copilot instead queries the web in real time, on the fly, like an agent. Labenz describes it as asking clarifying questions, issuing multiple live queries, and scoring pages.
Copilot was not turned on by default and had limited free uses. Perplexity was testing a “Retry” button that reruns any query with Copilot and a fine-tuned GPT-3.5 router to reduce its cost, while harder questions could still require GPT-4. The long-term goal is one interface where free and Pro differ mainly by query allowances.
Labenz’s pushback — worth keeping: he almost always uses Copilot and assumed it was strictly better. Masad concedes neither company nor user reliably knows when deeper retrieval is necessary: if people must classify a question before searching, “that’s not a good product. The product should think for you.”
A second obstacle is articulation. Many users prefer keywords because “asking good questions is a human skill,” and many people cannot articulate the question they have in mind. Masad credits Google’s durability partly to tolerating keyword-style intent, a tolerance Perplexity still needs to reproduce.
4. A curated index becomes the moat when models commoditize
Perplexity still uses several search APIs, but Masad says no single provider can shut it off. The company is building its own search index because OpenAI, Anthropic, and others are pursuing the same independence.
His core thesis: when model weights and training recipes become broadly available, “the edge goes to the data markets.” The web may contain roughly a trillion pages, but indexing everything is neither feasible nor desirable; value lies in identifying perhaps the best 1 billion or 10 billion pages for knowledge workers, researchers, and curious users.
That distribution is difficult to copy because of a chicken-and-egg loop: a strong index requires usage and a product, while a strong product requires an index. Masad believes Perplexity “somehow broke that asymmetry,” giving a company of its scale permission to “dream of being important.”
Self-sufficiency extends beyond crawling. Asked about dependence on external models, Masad’s answer is categorical: “We have to train our own models and make them good. There’s no other option.” Perplexity began by stitching together Bing and the best available GPT, then started replacing those layers with its own scraper, index, and language models.
5. Assistants could reduce today’s web to an underlying API
Masad imagines the existing web becoming infrastructure rather than the primary interface. Instead of opening Chrome and navigating pages, a user might state an objective and be taken directly into the appropriate workflow: “Does Chrome even need to exist?”
This is why he rejects the standard advice to build a browser. The larger opportunity is an AI-first operating system and device, with assistants as “first-party citizens,” protocols and schemas for agent communication, and new security layers. Perplexity lacks OEM distribution, so it is content to remain an assistant across existing platforms.
His honest answer on future content economics is “I don’t know.” His prediction is that valuable platforms will wall off data: Amazon could answer product questions itself, while social networks and commercial catalogs become independent islands connected through assistants. A global aggregator would concentrate on research and knowledge rather than every commercial transaction.
Google’s middleman position then weakens because sellers no longer need it for discovery. TikTok already captures restaurant discovery while Google retains directions; Masad sees that fragmentation as analogous to Google’s earlier displacement of Yahoo. Labenz’s related question about adversarial content remains materially unresolved.
6. User alignment matters more than maximizing early conversion
Masad favors charging consumers because customer and shareholder incentives then coincide. His counterexample is Google: advertisers fund the business, and he points to leaked internal emails in which ad leadership pressured search leadership to place more ads to meet quarterly targets, diverting attention from fixing bad queries.
Labenz interprets this as a subscription-led, anti-middleman strategy; Masad answers “Absolutely” on direct charging but leaves room for advertising that does not feel like advertising. Instagram is his model, and search could surface sponsored follow-up questions: “What if the whole question is an ad?”
Perplexity Pro provides unlimited interactive Copilot use, GPT-4 or Claude 2 as the default model even outside Copilot, and unlimited file uploads. The company deliberately avoids forcing conversion after one free question because it prioritizes retention and wants users to sign up proactively, making churn and revenue easier to control.
The economics remain compute-heavy. Perplexity had raised a $25 million Series A, employed roughly 25–30 people, and spent as much or more on compute than staff. Masad’s paradox is that “if you really want to save on infrastructure, you actually have to spend on infrastructure” upfront, then utilize and commoditize it over time.
7. Speed compounds learning while AI expands access to expertise
Masad calls generative-AI startups “basically a war zone”: incumbents must ship quickly to protect distribution, while startups lose momentum if they relax. Google’s constraint is that replacing its familiar search interface with a generative experience is difficult and risky.
The operating doctrine is to maximize learning per unit time. Borrowing Nat Friedman’s framing, shipping faster means making mistakes faster, but it also means learning “10 times as much” before competitors repeat the same errors; Perplexity’s narrow search-plus-LLM focus is the wedge that keeps this pace strategically coherent.
On human dependence, Masad expects AI to handle many intellectual questions because it can be questioned indefinitely without embarrassment. An hour with an expert such as Ilya Sutskever is scarce; a hypothetical GPT-4- or GPT-5-level expert can be “bothered” for hours, join families learning together, or settle factual disputes inside group conversations with “do the math” immediacy.