GEO paper — 15–30% relative improvement on the Subjective Impression metric (LLM-rated answer quality from the source)
Created 2026-06-21
Summary
Claim: Same top methods achieved 15–30% relative improvement on the Subjective Impression metric — an LLM-rated measure of how favourably the source comes across in the synthesised answer.
Source: Aggarwal et al., arXiv:2311.09735 — Jun 28, 2024.
Confidence: Verified.
Caveat: Subjective Impression uses an LLM as judge, so the metric has the usual circularity caveats of LLM-as-judge methodology. Pair with GEO paper — Cite Sources / Quotation / Statistics methods achieved 30–40% relative improvement on the Position-Adjusted Word Count metric (the harder metric) for a balanced citation.
Why this matters for Candid: The "softer" of the two paper metrics — useful as a corroboration but not as the headline number.
Related entries
Referenced by (2)
- reference Research brief: the website as a working surface of the business — four capabilities, AI-citation decoupling, freshness as a real signal (June 2026) relates-to
- rule R3 — Favor body-text citations, quotations and statistics over schema markup as the AI-visibility lever; the peer-reviewed lift is in body text depends-on