Perplexity AI

Contribution to AI

A chatbot that invents things is a curiosity; a search product that invents things is useless, and that distinction shaped what appeared on 7 December 2022, a week after ChatGPT. The answer arrived with numbered citations set beside each clause, an academic habit transplanted into consumer software, so any sentence could be traced to the page it came from without leaving the result. Retrieval first, generation second, sources visible — that arrangement hardened into the default expectation for AI search, and the answer panels the large search companies shipped afterwards carry source links because users had learned to look for them. The machinery was then sold separately: the online models released through pplx-api in November 2023 were the first commercial LLM endpoints grounded in live web results rather than a frozen training corpus, letting developers buy retrieval instead of building a crawler, an index and a reranker. The company also demonstrated that the application layer was a viable position, routing queries to whichever outside model suited them and never training a frontier system. The third thing it produced was a problem. When the answer removes the reason to click, the publisher who paid for the reporting earns nothing. The Publishers' Program of July 2024 offered cited outlets a share of advertising revenue; News Corp sued in October, Condé Nast sent a cease-and-desist, and Wired documented crawling that disregarded robots.txt. The web's old bargain between search and publisher is being renegotiated around those cases.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).