Answers are grounded in this paper and cite their sources.
Observational study assesses grounding rate of citations in AI answers, indicating potential inaccuracies in citation support.
Pre-registration for E080, an observational (Path B) study extending the GEO Lab citation funnel with a third stage: grounding. The prior two stages are "retrieved" and "cited"; E080 measures, of the citations an AI answer engine gives to a page, the fraction where the specific claim attributed to the page is actually supported by that page's content (attribution fidelity). The primary measure is a human-adjudicated grounding rate over citation-claim pairs, each labelled SUPPORTED, PARTIAL, or UNGROUNDED against the full owned-page content. The confirmatory hypothesis (H1) is that the grounding rate is below 100% — a non-zero fraction of citations are not grounded — scoped to Perplexity and pooled across three citation-eligible T1 pages (/geo-stack/, /extractability/, /retrieval-probability/). The cross-platform contrast (H2) is exploratory; ChatGPT and Google AI Overviews are pre-registered NA-by-construction (below the 30-pair power floor) and Gemini NA pending a Vertex-redirect resolver. The query set is frozen prior to data collection: 59 unique (query, target-page) pairs with prior observed Perplexity citation, fixed by SHA-256 checksum (a1ae2ae08d46aa0079d709dcc9e39354aa4fec490202a275bfb0b2cb8fd72822). Measurement window: 1–5 July 2026. This document is deposited before any data is collected; any departure is logged as a numbered deviation. Regulatory context: the Munich ruling of 28 May 2026 (case 26 O 869/26).
No takes yet. Share an insight, caveat, or question.
Artur Ferreira (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: