Founding and Origins (2022)

Perplexity was established in August 2022 by a team of engineers and researchers with backgrounds in large language models and search systems ²²²⁹³⁰. The founding group included Aravind Srinivas, formerly of OpenAI and Google; Denis Yarats, a research scientist at Meta; Johnny Ho, an engineer from Quora; and Andy Konwinski, a co-founder of Databricks ²²³⁰³¹. Headquartered in San Francisco, the organization initially focused on developing natural language interfaces for database interactions ²²³⁰.

The company transitioned toward developing a conversational "answer engine" that utilizes retrieval-augmented generation (RAG) to synthesize information from the web ²²². This approach was designed to process real-time information from across the internet into responses with inline citations ²¹⁵.

Early Growth and Funding Milestones

In March 2023, Perplexity secured $25.6 million in a Series A funding round led by New Enterprise Associates (NEA), with participation from Elad Gil, Nat Friedman, and Databricks ²²³⁰. This capital was used to expand the platform's search capabilities and refine its citation-first architecture ²². The organization emphasized a "source-centric" approach, which prioritizes factual traceability by grounding model outputs in verified web data ¹⁵²².

By early 2024, the organization experienced an increase in its valuation and public profile ³⁰. In January 2024, Perplexity announced a $73.6 million Series B round led by IVP, which included investments from NVIDIA and Jeff Bezos through Bezos Expeditions ¹²²²³⁰. This round valued the company at approximately $520 million ³⁰. Investors cited the platform’s user growth and its potential to compete in the search market as primary factors for their support ¹²³⁰.

Transition to Unicorn Status (2024–2026)

In April 2024, Perplexity reached a private valuation exceeding $1 billion following a $62.7 million funding round led by Daniel Gross ³⁰. During this period, the organization expanded its product offerings to include "Pro Search," a feature intended for multi-step reasoning, and "Pages," a tool for generating structured reports based on search results ⁴⁶⁷.

To support its computational requirements, Perplexity entered into a multi-year agreement with CoreWeave in early 2026 to power its AI inference workloads ¹⁹. According to the developer, this partnership granted the organization access to NVIDIA GB200 clusters for running its "Sonar" family of proprietary models alongside third-party models such as GPT-4 and Claude ¹³¹⁴.

Technical Evolution and Product Milestones

Perplexity utilizes a modular architectural strategy for its search engine ². While the organization provides access to third-party models, it also developed proprietary "Sonar" models optimized for web-grounded summarization and real-time citation workflows ⁸¹⁴. The system employs a pipeline that includes semantic ranking, evidence deduplication, and multi-turn planning for research tasks ²¹⁵.

In late 2024 and 2025, the company focused on expanding its "Deep Research" capabilities, which use sequential planning to address complex queries ⁴. According to the developer, these workflows were designed to gather targeted evidence for different facets of a query before synthesizing a final report ⁴. The organization also introduced an API platform, allowing developers to integrate its citation-centric search engine into external applications and knowledge management systems ¹⁰¹⁵.

Perplexity offers a suite of AI-native search and productivity tools centered on a conversational 'answer engine' designed to provide direct responses with inline citations rather than traditional search results ¹². The organization states that its platform processes over 1.5 billion questions per month as of early 2026 ¹².

Consumer Search and Discovery

The primary product is the Perplexity search interface, which provides a conversational bar for user queries. The system utilizes a 'Best' mode that automatically selects the most appropriate underlying model based on the complexity and type of query ¹⁵. For more intensive research, the platform offers 'Pro Search,' which performs multi-step internet searches and provides up to three times more sources than standard searches ¹³.

Perplexity integrates several third-party large language models (LLMs) alongside its proprietary models. Pro subscribers can manually toggle between different models for their searches, including:

• Sonar: Perplexity's in-house model, based on Llama 3.1 70B, optimized for real-time web search and summarization ¹³.
• GPT-5.2: A model from OpenAI noted for reasoning and reduced hallucination rates ¹³.
• Claude Sonnet 4.6 and Opus 4.6: Models from Anthropic used for technical reasoning and complex agentic tasks ¹³.
• Gemini 3.1 Pro: Google's multimodal model optimized for code generation and analysis ¹³.
• Specialized Reasoning Models: This includes NVIDIA’s Nemotron 3 Super 120B and Kimi K2.5 Thinking, which are designed for step-by-step logical processing ¹³.

Content Creation and Browsing

In May 2024, the company introduced Perplexity Pages, a tool that converts research findings into structured, visually formatted articles and reports ¹⁶. Pages allows users to customize the tone for different audiences, such as educators or subject matter experts, and automatically includes citations for each section to maintain credibility ¹⁶¹⁸. The tool also generates AI-based visuals and allows for the arrangement of sections to create comprehensive guides or study materials ¹⁶.

Perplexity also develops Comet, described as an AI-native web browser ¹². Comet integrates the 'Comet Assistant,' an AI agent designed to execute tasks within the browsing environment ¹². Another offering, Perplexity Computer, provides multimodal orchestration across tools, files, and code creation, utilizing persistent memory to manage complex user workflows ¹².

Enterprise and Developer Services

For organizational use, Perplexity offers Enterprise Pro and Enterprise Max. These versions allow employees to search both the open web and internal company knowledge bases ¹². According to the company, these versions support multi-step research, data analysis, and the use of advanced reasoning models within a secure environment ¹². In March 2026, CoreWeave announced a multi-year agreement to power Perplexity’s inference workloads using NVIDIA GB200 NVL72-powered clusters ¹².

Developers can access Perplexity’s search capabilities via the Sonar API. The API is structured into three search modes—High, Medium, and Low—allowing developers to balance depth and cost-efficiency ¹⁹. Perplexity claims the Sonar API provides factual accuracy that approaches or exceeds competitive models like GPT-4o while maintaining lower costs ¹⁹. To simplify billing, the company does not charge for citation tokens in responses for most Sonar API tiers ¹⁹.

Subscription and Pricing Tiers

Perplexity operates on a tiered subscription model:

• Free: Provides basic search capabilities and limited access to Pro Search ¹⁴.
• Perplexity Pro: Priced at $17 per month when billed annually, this tier includes access to advanced models (GPT-5.2, Claude 4.6), increased file upload limits, and enhanced image generation ¹⁴¹⁵.
• Perplexity Max: Priced at $167 per month when billed annually, this tier is designed for high-volume users. It includes 10,000 monthly credits for Perplexity Computer, priority access to new features, and the 'Model Council' feature for comparing responses across different AI models ¹⁴.

Perplexity is managed by a leadership team with backgrounds in artificial intelligence research and engineering from organizations including OpenAI, Meta AI, and Google ¹². The executive team is led by Chief Executive Officer Aravind Srinivas, who previously worked at OpenAI and Google ¹². Other members of the founding group include Chief Technology Officer Denis Yarats, a former research scientist at Meta; Johnny Ho, an engineer formerly with Quora; and Andy Konwinski, a co-founder of Databricks ¹². Dmitry Shevelenko serves as the organization's Chief Business Officer ⁸.

The organization is headquartered in San Francisco, California ¹². While specific employee counts are not publicly detailed in provided documentation, the company maintains a technical workforce focused on the development of its AI-native search infrastructure and proprietary model family, Sonar ⁴, ¹².

Strategic Partnerships

Perplexity maintains a multi-cloud operational strategy and has established several strategic partnerships with hardware and infrastructure providers to support its compute requirements ⁸. In March 2026, the company entered into a multi-year agreement with CoreWeave to power its inference workloads ⁸. Under this partnership, Perplexity utilizes dedicated clusters powered by NVIDIA GB200 NVL72 hardware to support its Sonar models and Search API ecosystem ⁸. Additionally, Perplexity uses Weights & Biases (W&B) Models to facilitate the training, fine-tuning, and management of its internal AI systems from experimentation through to production ⁸.

In addition to hardware infrastructure, Perplexity maintains commercial relationships with multiple artificial intelligence laboratory partners. The organization’s modular architecture allows it to integrate a variety of large language models (LLMs) into its search interface ⁴. According to the developer, these partners include OpenAI, Anthropic, Google, and xAI, providing access to models such as GPT-4, Claude, Gemini, and Grok for premium subscribers ⁴.

Corporate Operations

Perplexity's corporate structure supports both consumer-facing products and enterprise-level services. In 2026, the company expanded its enterprise reach through a pilot program with CoreWeave, which deployed the 'Perplexity Enterprise Max' platform across its internal organization for web search, data analysis, and multi-step research tasks ⁸. This deployment involves the use of the Comet Assistant and the Comet Browser, which the organization describes as an AI-native web harness for its agentic systems ⁸.

Perplexity's research and development efforts are primarily focused on retrieval-augmented generation (RAG) and the optimization of low-latency inference for conversational search ⁴, ¹². The organization’s research philosophy centers on a "grounded generation" methodology, which reorients the response process toward an evidence-first approach ⁴. Under this framework, the system initiates a real-time search across web indices to collect and rank relevant data snippets before the language model begins generating text ⁴. According to the company, this architecture is designed to minimize hallucinations by ensuring that every claim is tied to an explicit source marker embedded in the generation logic rather than appended post-hoc ⁴.

The organization maintains a proprietary line of models called the Sonar family ⁴. These models are purpose-built for web-grounded summarization and real-time citation, serving as the default engine for the platform’s core search functions ⁴. While the system integrates third-party models from providers like OpenAI and Anthropic for complex reasoning, the Sonar architecture is optimized for speed and factual traceability during web-based research tasks ⁴.

Advanced research capabilities within the organization's products, such as "Pro Search" and "Deep Research," utilize agentic workflows and sequential planning ⁴. These systems decompose complex or ambiguous queries into multiple research steps, executing sub-queries and dynamically incorporating intermediate results into a final context window ⁴. This multi-pass retrieval method is intended to mirror expert research methodologies, allowing for systematic literature reviews and cross-document analysis that exceed the scope of single-turn prompting ⁴.

To facilitate the technical requirements of real-time indexing and high-scale inference, Perplexity entered a multi-year agreement with CoreWeave in 2026 ¹². This partnership utilizes NVIDIA GB200 clusters to power the organization’s inference workloads, specifically targeting the computational needs of agentic AI ¹². This infrastructure is designed to provide the performance necessary for iterative retrieval and reasoning processes at scale ¹². Additionally, the organization provides research infrastructure to external developers through its Sonar and Agentic Research APIs, which expose granular controls over retrieval configuration, model routing, and citation metadata ⁴.

Perplexity utilizes a retrieval-augmented generation (RAG) architecture designed to prioritize factual accuracy and mitigate the occurrence of large language model (LLM) hallucinations ⁴. According to the organization, its safety and ethics framework is centered on an "evidence-first" methodology, where the generation of a response is conditional upon a prior real-time search and ranking of web indices or trusted knowledge repositories ⁴. By performing automated filtering, semantic ranking, and deduplication of candidate documents before prompting the model, the system aims to bind the AI's output to verifiable, external evidence ⁴.

Hallucination Mitigation and Grounding

The organization states that its architecture reorients the answer process away from pure generative text, treating the LLM as a controlled synthesizer rather than a primary source of information ⁴. In this workflow, citations are not appended after generation but are structurally embedded into the generation prompt and response logic ⁴. This metadata framework includes provenance data, source URLs, and summary snippets, which the developer asserts allows for deep due diligence and verification of claims by the user ⁴. For complex or high-stakes research, the platform employs sequential planning and multi-pass retrieval to collect evidence for various facets of a query before synthesis ⁴. This multi-step process is intended to ensure that answers are coherent and traceable to a multi-document knowledge base rather than relying on the internal weights of a model ⁴.

Data Governance and Privacy

For corporate and institutional users, the organization provides "Perplexity Enterprise Max," a service that allows for search and research within internal knowledge bases and the public web ⁸. The system is designed to support data analysis and deep research within a platform that scales through dedicated infrastructure ⁸. To maintain the reliability and performance of these enterprise workloads, the organization has entered into strategic partnerships with infrastructure providers such as CoreWeave ⁸. These agreements involve the use of specialized NVIDIA GB200 clusters to ensure consistent inference performance and low latency for production environments where reliability is a critical safety requirement ⁸.

Bias and Content Policy

Perplexity's approach to objective search results relies on the separation of retrieval and composition ⁴. While the underlying language model—whether the proprietary "Sonar" series or third-party models such as GPT-4 or Claude—handles the tone and formatting, the factual boundaries of the response are dictated by the retrieval pipeline ⁴. The organization claims this separation ensures that the system remains responsive to user needs without sacrificing the traceability of the information provided ⁴. The proprietary Sonar models are purpose-built for web-grounded summarization and are designed to prioritize real-time citations ⁴. Although the system leverages various external models for reasoning, the developer maintains that the retrieval stack serves as a primary defense against misinformation by grounding every sentence in a ranked context window ⁴. Specific public disclosures regarding the internal governance of the organization's safety teams or exhaustive lists of prohibited topics are less detailed in available technical documentation, which instead focuses on the architectural safeguards against technical inaccuracies ⁴, ⁸.

Perplexity has received critical attention for its user interface and the perceived efficiency of its information retrieval model. The platform's "answer engine" approach has been noted for providing synthesized responses that offer greater traceability than traditional large language models due to the inclusion of inline citations ⁴. Organizations including Zoom, SAP, and Doximity have adopted Perplexity's API for business and clinical applications, with partners citing the platform's speed and its ability to summarize real-time information as primary reasons for integration ⁵. Perplexity asserts that its retrieval-augmented generation (RAG) architecture prioritizes factual accuracy by grounding every generated sentence in explicit source markers ⁴.

However, the organization has faced significant controversy regarding its web scraping protocols and the ethics of its data collection. In June 2024, an investigation by Wired alleged that Perplexity's crawlers bypassed the Robots Exclusion Protocol (robots.txt), a standard used by web servers to indicate that certain pages should not be accessed by automated bots ⁶. The report claimed that a Perplexity-controlled server accessed Wired content despite being explicitly blocked, and the publication characterized the resulting AI-generated summaries as a form of plagiarism that reproduced original reporting without sufficient transformation ⁶.

These concerns have escalated into multiple legal challenges from major media publishers. In October 2024, Dow Jones and NYP Holdings—subsidiaries of News Corp—filed a lawsuit against Perplexity in the Southern District of New York ⁷, ⁸. The complaint alleges "massive amount of illegal copying" and claims that Perplexity’s responses often reproduce entire paragraphs or articles from The Wall Street Journal and The New York Post without permission or compensation ⁷. The lawsuit further alleges trademark infringement, claiming that Perplexity's system produced "hallucinations" that were falsely attributed to News Corp publications, potentially causing brand dilution ⁷, ⁸. News Corp CEO Robert Thomson described the startup's practices as "content kleptocracy" ⁷.

Other major publishers have issued formal demands for the organization to cease its data collection activities. The New York Times sent a cease-and-desist letter to Perplexity in October 2024, citing unauthorized scraping of its content, including paywalled material ⁷. Similar legal pressure was reported from Condé Nast, which issued a cease-and-desist order following allegations that the AI tool was infringing on its copyrighted works ⁷. In an effort to address these criticisms, Perplexity launched a revenue-sharing program for publishers in July 2024, signing agreements with organizations such as Time, Fortune, and Der Spiegel to share advertising revenue generated from content citations ⁷.

Perplexity’s development of an "answer engine" represents a shift in the internet economy from a keyword-based search paradigm to one focused on direct information synthesis ¹². Unlike traditional search engines that provide a list of external links, Perplexity provides direct, conversational responses that incorporate inline citations from web sources ¹². According to the organization, this model is designed to "power the world's curiosity" by providing a more efficient method for discovery and research ¹². By early 2026, the platform was processing over 1.5 billion queries per month, reflecting a significant adoption rate and a potential shift in the way users consume digital information ¹².

The organization's technology influences professional and academic research habits through tools such as the "Comet Assistant" and "Enterprise Max" ¹². These systems are designed to perform "deep multi-step research," visualize data, and manage persistent memory, which allows users to move beyond simple queries into more complex analytical tasks ¹². For example, the organization’s partnership with CoreWeave involves the deployment of these tools to help employees research both internal knowledge and the open web within a single interface, indicating a shift toward integrated AI-native workflows in corporate environments ¹².

Regarding the impact on the business models of traditional digital journalism, Perplexity emphasizes a retrieval-augmented methodology that draws from "credible sources" and "deep research" ¹². The organization asserts that its systems prioritize factual accuracy and provide traceability for the information retrieved through its citation model ¹². Furthermore, the growth of the "Sonar and Search API ecosystem" suggests that this answer-based paradigm is being integrated into various third-party applications and software sectors ¹².

Perplexity also aims to influence the future of web navigation through its "Comet Browser," which it describes as an AI-native web browser ¹². This product, along with the "Perplexity Computer"—an orchestration of AI across tools, files, and code—suggests an increasing displacement of manual search-and-compile tasks with automated AI agents ¹². These initiatives reflect the organization's stated goal of creating accurate AI-native products that function as a harness for powerful AI agents across the open web and private data repositories ¹².

• Anthropic
• Claude Sonnet 4.6
• Gemini 3.1 Pro
• Google
• GPT-4o
• GPT-5
• GPT-5.2
• Kimi K2.5
• Meta
• NVIDIA
• OpenAI
• Sonar
• xAI