AI search engine

A traditional search engine hands you a list of pages and leaves you to open them. An AI search engine still searches the web, but a language model then writes the answer and attaches the pages it drew on. A 2024 paper calls this sort of product a generative engine: it retrieves documents and writes a response grounded on those sources, with citations a person can check¹.

Perplexity calls itself an answer engine. In a November 2023 post it described the method it was using then. Its own index pulls short excerpts, which it calls snippets, from web pages. A language model fine-tuned to use those snippets writes the reply. The post's example of a question this can answer, and an ordinary model cannot, is a basketball score from the night before².

The same pattern shows up in other products. On 31 October 2024 OpenAI launched ChatGPT search: ChatGPT searches the web when the question needs current information, or when you click the search icon, and the reply includes links to sources such as news articles³. An update on that page says that from 5 February 2025 this was available to everyone in regions where ChatGPT is available, with no sign-up³. On 4 April 2025 Microsoft introduced Copilot Search in Bing, which places a written answer with citations on the results page and still shows the ordinary links. It rolled out worldwide except Russia and China⁴.

A written answer is not a checked one. The same paper notes that language models can state things the sources never said, which is why each sentence is supposed to carry a citation, and why that citation is supposed to support the sentence¹. Google's own AI Overviews are a related but separate feature inside Google Search, not a standalone search product⁵.