---
title: "Target-specified webpage rewriting raises a party's AI-answer exposure far more reliably than target-unspecified rewriting, in VoteGEO-Bench (Zhang & Xiao 2026)"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/pdf/2610.21330"
source_author: "Shiqi Zhang, Xiaokui Xiao"
source_date: "2026-10-15"
source_title: "VoteGEO-Bench: Measuring the Impact of Generative Engine Optimization on Electoral Visibility"
source_venue: "arXiv:2610.21330 [cs.IR] (cross-listed cs.CL, cs.CY), National University of Singapore"
source_quote: "In the U.S. setting, 92% of retrieved webpages mention both major parties, so rewriting without an explicit target does not indicate which party's exposure should increase."
source_quote_2: "Among the 36 category, engine, and metric settings, target-unspecified rewriting produces only one significant positive gain, while four show significant negative changes and the remaining 31 are not significant."
source_quote_3: "In contrast, target-specified rewriting produces significant positive gains in 24 of 36 settings, including 10 of 12 for SOV."
source_tier: 1
source_sha: "c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a"
source_url_2: "https://arxiv.org/pdf/2610.21330"
source_sha_2: "c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a"
source_url_3: "https://arxiv.org/pdf/2610.21330"
source_sha_3: "c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a"
source_quote_4: "For RQ1, 18 of 36 target-specified comparisons remain significant after clustering and correction, compared with 24 originally, while none of the 36 target-unspecified comparisons remain significant, compared with 5 originally."
source_url_4: "https://arxiv.org/pdf/2610.21330"
source_sha_4: "c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a"
audits: ["2026-10-19 claude-opus-5"]
audit_status: "capture-verified — all three quotes checked against the extracted PDF via quote_check at capture time (2026-10-17), all grounded: true || APPENDED 2026-10-19, scheduled cross-model audit (auditor claude-opus-5; writer was claude-sonnet-5) — the PDF was re-fetched with extract_pdf (tls verified) and returned sha256 c7f6091c8127…, identical to the capture-time sha; all three original quotes are verbatim (the 92% figure appears twice, in the introduction and again in section 6.2, in the wording quoted here). Two changes applied this date, old wording preserved: (1) CORRECTION — the body said '36 tested combinations of rewriting strategy, generative engine, and exposure metric'; the paper's unit is the strategy *category*, not the individual strategy ('Among the 36 category, engine, and metric settings'), i.e. 4 categories x 3 engines x 3 metrics, where the 11 individual strategies are ranked separately across 18 settings in Figure 2. (2) ADDITION — the note reported only the uncorrected significance counts, so the body now carries appendix A.4's clustered and Benjamini-Hochberg-corrected figures as source_quote_4; they widen the gap the claim rests on rather than narrowing it (18 of 36 against none of 36). Single-source concentration cap (sources.md): the vault holds three claim-notes resting on arXiv:2610.21330, an unrefereed preprint, which is the cap; the next finding from this paper goes to 50-questions as a corroboration question rather than becoming a fourth note."
provenance: "Promotion from 10-inbox/raw/2026-10-17-outside-arxiv261021330-votegeo-bench-measuring-the-impact-of.md, 2026-10-18 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-10-17-outside-arxiv261021330-votegeo-bench-measuring-the-impact-of.md"
date_created: "2026-10-18T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["GEO","generative-engine-optimization","elections","LLM","information-retrieval","arxiv"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-19 — source_quote verbatim (normalized) in the capture-time archive (sha256 c7f6091c8127…) [reading-order], same-night model-free sweep"
verified_verbatim: "2026-10-19 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Zhang and Xiao's VoteGEO-Bench tests [[entity-generative-engine-optimization|generative engine optimization (GEO)]] — rewriting a webpage to improve its visibility in AI-generated answers — against election queries from the 2024 U.S. presidential race and the 2026 Japanese House of Representatives election. They compare two rewriting conditions: "target-unspecified," where the rewriting model is not told which candidate or party to favor, and "target-specified," where it is told explicitly. Across 36 tested combinations of rewriting-strategy *category* (the four categories, not the eleven individual strategies), generative engine, and exposure metric, target-unspecified rewriting produced a significant positive gain in that party's measured exposure in only one setting, with four significant negative changes and the remaining 31 not significant. Target-specified rewriting, by contrast, produced significant positive gains in 24 of the 36 settings, including 10 of 12 for Share of Voice (SOV). Those are the uncorrected counts; the paper's own robustness appendix widens rather than narrows the gap between the two conditions — "For RQ1, 18 of 36 target-specified comparisons remain significant after clustering and correction, compared with 24 originally, while none of the 36 target-unspecified comparisons remain significant, compared with 5 originally."

The paper attributes the gap to how election content is actually written: in the U.S. data, 92% of the webpages in the benchmark's retrieval pools mention both major parties, so a rewrite given no stated target has no signal for which party's exposure it should be increasing. A rewrite that explicitly names its target removes that ambiguity and the gains follow.

> [!note] Seek's commentary:
> The mundane-sounding mechanism here is the interesting part: this isn't an AI being persuaded by one argument over another, it's an AI being *told what the task is*. Omit the instruction and the model has no way to know which of two parties sharing a page deserves the lift — so most of what looks like "GEO doesn't work for elections" in a naive test is actually an underspecified prompt, not a resistant model.
> — Seek