General · EdgepediaSeptember 17, 20267 min read

Perplexity AI

Perplexity AI is a San Francisco-based artificial intelligence company whose main product is an "answer engine", a search service that generates a direct answer to a user's question with inline citations to the web sources behind it, rather than a list of links. Its chief executive and co-founder Aravind Srinivas has described the product as "like a marriage of Wikipedia and ChatGPT" that answers questions without the clutter of conventional search results.¹ The company competes with Google, ChatGPT, and other AI search tools, and has drawn both large investments and a series of copyright lawsuits from publishers.²

Key fact	Detail
Founded	August 2022, San Francisco; first product, the Ask answer engine, released December 2022³
Co-founder and CEO	Aravind Srinivas, computer scientist and former AI researcher at OpenAI and Google¹
Early funding	$3.1 million seed round (September 2022) led by Elad Gil and Nat Friedman; $25.6 million Series A led by Peter Sonsini of New Enterprise Associates³
Scale	Production search infrastructure processes 200 million daily queries; 780 million queries in May 2025⁴^•⁵
Valuation	$18 billion after a $100 million extension round in July 2025; about $1.5 billion raised in total⁶^•⁷
Browser	Comet, launched July 9, 2025, initially for $200/month Max subscribers, free for all users from October 2025⁸^•⁹
News-citation accuracy	Incorrect answer rate of 37% in the Tow Center's 1,600-query test of eight AI search engines, the lowest of those tested¹⁰

History and funding

The company was founded in August 2022 and released its answer engine, Ask, in December 2022. It reached 2 million monthly active users within four months of launch.³ Srinivas, inspired in part by a childhood love of Wikipedia, co-founded the company in the summer of 2022, shortly before generative AI products such as Stable Diffusion and ChatGPT drew broad public attention.¹

Funding grew quickly alongside the user base. After the seed and Series A rounds, the company raised $500 million in December 2024 at a $9 billion valuation led by Institutional Venture Partners, and in July 2025 brought in a $100 million extension that valued it at $18 billion, an increase from a $14 billion valuation a few months earlier.¹¹^•⁶ Srinivas told the Associated Press that Perplexity was not profitable but was more sustainably run than foundation-model companies because it does not train its own foundation models.¹

How the product works

Perplexity does not build its own large language models. It relies on existing models from OpenAI, Anthropic, and Meta Platforms and "post-trains" them, in Srinivas's words, to "be really good summarizers".¹ The company's own description is that it post-trains its models to search harder, find better evidence, and answer more clearly, with every answer sourced and cited.¹²

Pro Search and Deep Research. Pro Search, the advanced tier of the search feature, breaks a question down, runs multiple web searches across articles, academic papers, forums, and videos, and synthesizes the results into an answer with direct links to the original sources. Pro users can choose among models including Perplexity's own Sonar, OpenAI's GPT-5, Claude Sonnet 4.6, and Gemini 3.1 Pro, and can narrow searches to specific content types such as Academic or Finance. The feature also includes a code interpreter for executing code.¹³ Deep Research runs dozens of searches across hundreds of sources, checks the findings, and writes a report.¹²

Search API. Perplexity also sells its retrieval infrastructure to developers. Handling 200 million queries each day, the production search infrastructure merges hybrid retrieval mechanisms with ranking pipelines spanning multiple stages, indexing spread across distributed systems, and dynamic parsing, all unified by a content understanding module that improves itself. The company argues that overbroad retrieval can derail AI models regardless of how relevant the retrieved documents appear overall, so the API curates the context passed to a developer's model rather than simply returning documents.⁴

Comet and the Chrome bid

In July 2025 Perplexity entered the browser market with Comet, an AI-powered browser built around the Comet Assistant agent, which can summarize emails and calendar events, manage tabs, and navigate web pages on a user's behalf. At launch it was available only to subscribers of the $200-per-month Max plan, with a waitlist for others.⁵^•⁸ In early October 2025 the company made Comet free for all users.⁹ Srinivas has framed the browser as a step toward "an operating system with which you can do almost everything", arguing that becoming a user's default browser would translate into "infinite retention".⁵

In August 2025, weeks after Comet's launch, Perplexity made an unsolicited all-cash offer of $34.5 billion to buy Google's Chrome browser, more than the roughly $1.5 billion the company had raised to that point and nearly twice its $18 billion valuation. The offer included a commitment to keep Chromium open source, invest $3 billion in the project, and leave Google as Chrome's default search engine. It followed a Department of Justice proposal that Google be forced to sell Chrome after a court ruled the company had illegally maintained a search monopoly; the offer came to nothing when the judge decided against a Google breakup.⁷^•¹⁴

Accuracy and citation quality

Because the product's premise is cited answers, its citation behavior has been studied closely. An early academic audit by Stanford researchers Nelson F. Liu, Tianyi Zhang, and Percy Liang evaluated four generative search engines, including perplexity.ai, and found on average that only 51.5% of generated sentences were fully supported by citations and only 74.5% of citations supported their associated sentence; citation precision was inversely correlated with how useful readers found the responses.¹⁵

The Tow Center audit of 2025 tested eight AI search tools, including Perplexity and Perplexity Pro, with 1,600 queries asking each to identify the headline, publisher, date, and URL of excerpts from 200 news articles. More than 60% of queries received incorrect answers across the tools; Perplexity had the lowest error rate at 37%, while Grok 3 had the highest at 94%. Premium versions erred more often than free ones because they were more reluctant to decline uncertain responses.¹⁰ The same study reported that Perplexity's free version correctly identified all 10 excerpts from paywalled National Geographic content despite the publisher explicitly disallowing Perplexity's web crawlers, evidence that some tools ignore the Robot Exclusion Protocol.¹⁰

The unsearched Writing feature, which generates text without querying the web, is more error-prone. In an Associated Press test it produced a 465-word document resembling a news article, complete with fabricated quotations from real people; Srinivas called the feature a "minor use case" that is "more prone to hallucinations" because it is not tethered to web search.¹ A 2024 Forbes dispute over an AI-generated story closely paralleling a Forbes exclusive drew similar criticism that Perplexity was acting like a media company while presenting itself as a search engine.¹⁶

Publishers: partnerships and lawsuits

Perplexity's relationship with news publishers runs on two tracks. On the commercial side, it announced 22 revenue-sharing agreements with publishers in 2024, and chief business officer Dmitry Shevelenko said the company would not stop citing publishers outside its revenue-sharing deals nor show preference for paid partners.¹⁷^•¹⁸ In May 2025 the French newspaper Le Monde signed a multi-year, non-exclusive agreement covering the use of its content to generate answers, not to train models, and giving Le Monde access to Sonar, a version of the answer engine that responds using only Le Monde's content.¹⁹ In August 2025 the company announced Comet Plus, a subscription product that gives participating publishers 80% of the revenue from subscription fees, backed by a $42.5 million pool to compensate early partners.²⁰

References

Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › AI companies, people, and products › AI startups and application companies

Initially written Sep 17, 2026 · Reviewed: Oct 15, 2026 · Edited: Oct 15, 2026 · Last review: Oct 15, 2026

Notice something wrong?