Answer engines

Retrieval-augmented generation pairs parametric memory, the weights of a pre-trained sequence-to-sequence model, with non-parametric memory, a dense vector index the system queries at answer time^[5]. That second half is what separates an answer engine from a chatbot working purely from training data, and it is why training data versus retrieval is a live distinction for anyone trying to be cited. The retrieval step is rarely one lookup. Google's documentation states that both AI Overviews and AI Mode may use a query fan-out technique, issuing several related searches across subtopics and data sources before composing one response^[6].

Perplexity was founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho and Andy Konwinski, and launched its search engine on 7 December 2022^[1], built around synthesising one answer from current web content and citing the sources it drew on^[7]. Microsoft shipped Bing Chat inside a revamped Bing on 7 February 2023 and folded it into the Microsoft Copilot brand on 15 November 2023^[2]. Google unveiled the Search Generative Experience at Google I/O on 10 and 11 May 2023 as an opt-in Search Labs experiment^[3], rebranded it as AI Overviews and made it a default part of United States results in May 2024^[8], then started testing a fully generated results page as AI Mode in March 2025^[9]. OpenAI deployed ChatGPT Search, a web-grounded mode inside ChatGPT, between October and December 2024^[4].

OpenAI reported more than 900 million weekly active users of ChatGPT by February 2026^[10]. Audiences of that size are why AI referral traffic is now measured as its own channel.

A controlled study of 252,000 trials across six large language models found that topical relevance and a source's position in the retrieved list were the biggest drivers of being cited first, that an explicit price and a recent timestamp helped consistently, and that formatting-only edits had little impact^[11]. Relevance and position are both settled before generation, which is why generative engine optimization targets much the same signals as ordinary search work; see SEO and GEO for the overlap and AI visibility measurement for how presence in these answers is sampled.

Citing a source is not the same as citing the right one. A Columbia Journalism Review audit of eight AI search tools put Grok 3 last on source attribution, with 94% of its queries answered incorrectly^[12]. The spread between tools is wide enough that an answer engine's citation behaviour has to be read per engine; citation accuracy covers the audit and its method.