AI search that answers with citations
HQ: San Francisco, CA, United States | Founded: 2022 | Employees: 1,350-1,450 (third-party estimates, mid-2026) | Stage: Late-stage private; reported talks at a $30B+ valuation | Website: https://www.perplexity.ai
Perplexity is best understood as a layer above the model. Most AI search products wrap a general-purpose chatbot around a search index. Perplexity went the other way: it built an index and crawler tuned for producing a specific, sourced answer, then borrowed whichever frontier model was best suited to each part of the job. That is why its product reads like a research assistant rather than a chat window — every claim is attached to a link, and follow-up questions keep the same citation discipline.
The company launched publicly in December 2022 and grew quickly, but the shape of the business has changed several times. Consumer search and the Pro/Max subscription tiers remain the core. On top of that sit two expansion layers. The first is distribution: Comet, a Chromium-based AI browser launched in July 2025 and made free worldwide that October, whose assistant can act on a user's behalf inside the browser rather than just answer questions. The second is labor: Computer, a cloud agent introduced in February 2026 that plans and executes multi-step work across research, code, documents and email, metered with a credit system rather than a flat fee.
Around those sit enterprise and developer products. Enterprise Pro and Enterprise Max sell cited search across the web, a company's own files, and its connected applications, with SSO, audit logs and data-residency controls. The API business includes the Search API, an embedding model, and the Sonar family — in-house models built on open bases — plus an Agent API that resells third-party models at pass-through rates.
Two commercial decisions define the current strategy. First, Perplexity ended AI-integrated advertising in February 2026, explicitly trading a revenue line for the accuracy and trust position it uses to differentiate from Google. Second, it moved to usage-based credits in early 2026, which lifts average revenue per user as agentic tasks consume far more compute than a search query. Management has said it plans an IPO in 2028 regardless of when OpenAI or Anthropic list.
The business is not without complications. Nine publisher and rightsholder suits were active as of May 2026, including New York Times, CNN, News Corp, the Chicago Tribune and Reddit. Cloudflare and Wired published findings alleging undisclosed crawler IP ranges and spoofed user agents, which Perplexity has strongly denied. Amazon won a preliminary injunction against Perplexity's shopping agent at the district level before the Ninth Circuit vacated it in August 2026, holding that a user-directed agent counts as the user. On the other side, the UK's Competition and Markets Authority has proposed letting Android and Chrome users choose an AI assistant such as Perplexity as their default search engine — a distribution channel that would otherwise be closed to it entirely.
Perplexity turned AI search into one of the fastest-scaling consumer products of the decade, but it is fighting on borrowed ground. The company does not own a frontier model, takes no advertising, and sits in the middle of a copyright fight with at least nine publishers, while the distribution it needs most — a default slot in a browser — belongs to Google.
What Perplexity actually owns is the layer above the model: a crawling and indexing stack built for answers rather than links, a citation product that users trust enough to adopt, and a distribution bet on Comet that sidesteps the fight over search defaults. Management says the business now runs past $750 million in annualized revenue, up from under $250 million at the start of 2026, and the product has widened into a browser, a cloud agent, an enterprise search tier, and a model family.
The unsolved question is durability. Nvidia sits on both sides of the table as investor and compute supplier, SoftBank is both backer and distribution partner, and the crawler-conduct findings from Cloudflare and Wired undercut the accuracy brand the whole company is priced on. A favourable regulatory development in the UK — where the CMA has proposed letting users pick an AI assistant as their default search engine — is the most concrete route to escaping the incumbent's grip.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki