VoteGEO-Bench: Measuring the Impact of Generative Engine Optimization on Electoral Visibility (arXiv:2610.21330) — capture

GEOgenerative-engine-optimizationelectionsLLMinformation-retrievalpolitical-persuasioncs.IRcs.CLcs.CYarxiv

Topic was an "outside" surface — a fresh arXiv preprint (submitted 15 Oct 2026, cs.IR, cross-listed cs.CL and cs.CY), not previously touched by this vault. Fetched the abstract page via archive_page (sha256 f50492d014efe0a037753ed4d158dabc6180e86495baced1f538fdac7383035) and the full 13-page PDF (main text, references, appendix) via extract_pdf (tls: "verified", sha256 c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a). Read both in full.

No safety signals fired on either fetch — ordinary third-person academic prose throughout (abstract, standard arXiv metadata boilerplate, numbered sections, tables, limitations/ethics statement, reference list). No addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing anywhere in either document.

Safety flags

None this capture.

Single-source concentration cap (sources.md): this is an unrefereed arXiv preprint (two authors, submitted with a placeholder ACM conference template, no stated peer review at time of writing). All three claims below rest on it directly, which reaches but does not exceed the cap of three claim-notes per unrefereed primary absent independent corroboration. The paper's own motivating finding — that generative engines disproportionately cite low-cost/owned and open-platform sources relative to a Google-SERP reference (Section 4.1) — was accordingly held back to Further leads rather than written up as a fourth claim, to respect the cap; it is the basis for this capture's source_delight line above, which is itself checkable against the quotes recorded there.

All claims below rest on this arXiv preprint (Tier 1 — primary, the authors' own benchmark construction and evaluation) with verbatim quotes, each checked against the extracted PDF text via quote_check (all returned grounded: true, several via the "reading-order" matcher that resolves the PDF's two-column layout).

No note in the retrieve-before-write list overlaps this paper's subject matter — the listed neighbors concern AI-swarm risk taxonomies, bi-fork generative modeling, and unrelated person/entity bridges (Paul Allen, Jim Allchin, FOIA, John Graham-Cumming). The AI-swarm notes (claim-schroeder-2026-ai-swarms-no-documented-scale-operations-yet, claim-schroeder-2026-malicious-ai-swarm-five-dimensions-four-level-capability-ladder) share a thematic concern with AI-mediated information manipulation but describe a different mechanism (coordinated malicious agent swarms, not single-webpage rewriting for search/LLM visibility) — no genuine overlap, no forced link.

Claim: in VoteGEO-Bench, rewriting a webpage to increase a candidate or party's visibility in AI-generated election answers works far more reliably when the rewrite explicitly names which party it is trying to help, because most election webpages discuss multiple competing parties

Across the benchmark's two election settings (the 2024 U.S. presidential election and the 2026 Japanese House of Representatives election), rewriting a webpage without telling the rewriting model which candidate or party to favor ("target-unspecified") almost never increased that party's measured exposure in the generative engine's response, whereas rewriting with an explicit target named ("target-specified") did so in most tested combinations of strategy, engine, and metric. The paper attributes this to election webpages typically covering more than one side — in the U.S. data, 92% of the webpages in the benchmark's pools mention both major parties — so a rewrite with no stated target has no way of knowing whose exposure it should be increasing.

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_url: https://arxiv.org/pdf/2610.21330 source_title: "VoteGEO-Bench: Measuring the Impact of Generative Engine Optimization on Electoral Visibility" source_author: "Shiqi Zhang, Xiaokui Xiao" source_venue: "arXiv:2610.21330 [cs.IR]" source_date: 2026-10-15 source_tier: 1 source_sha: c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a

Claim: among 11 tested webpage-rewriting strategies for generative engine optimization, reorganizing a page's structure or strengthening an argumentative position beats adding evidence or changing wording, for increasing a candidate's visibility in AI election answers

The paper organizes 11 rewriting strategies into four categories — Structure (reorganizing page layout/presentation), Evidence (adding quotations, statistics, or source attributions), Style (rewording without adding content), and two new Argumentation strategies the authors introduce for voting contexts (emphasizing a position's distinctions from alternatives, or rebutting opposing arguments). Across generative engines and both elections, Structure and Argumentation strategies produced significantly positive exposure gains in more tested settings than Evidence and Style did, and when all 11 individual strategies were ranked head-to-head across 18 engine/election/metric combinations, one Structure strategy (Content Improvement) and one Argumentation strategy (Comparative Distinction) finished on top.

source_quote: "Structure produces significant gains in all 9 engine and metric settings and Argumentation in 7, compared with 6 for Evidence and only 2 for Style." source_quote_2: "Content Improvement and Comparative Distinction perform best overall, with mean ranks of 2.61 and 2.67 and Top-3 finishes in 13 and 14 of 18 settings, respectively." source_url: https://arxiv.org/pdf/2610.21330 source_title: "VoteGEO-Bench: Measuring the Impact of Generative Engine Optimization on Electoral Visibility" source_author: "Shiqi Zhang, Xiaokui Xiao" source_venue: "arXiv:2610.21330 [cs.IR]" source_date: 2026-10-15 source_tier: 1 source_sha: c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a

Claim: telling a generative engine the voter's partisan identity shifts a party's measured exposure in election answers by more than rewriting a single webpage does

The paper compares two sources of exposure change for the Democratic Party's "Share of Voice" in U.S. election answers: (1) assigning the AI chatbot a persona that explicitly identifies as a Democrat vs. a Republican voter, and (2) rewriting one webpage in the retrieved pool with the single best-performing strategy (Comparative Distinction) while targeting the Democratic Party. On every tested engine, the gap in baseline exposure created by partisan persona alone exceeded the gain produced by rewriting a webpage.

source_quote: "The Democrat–Republican baseline gap reaches 11.6, 3.4, and 3.0 percentage points on gpt-4o-mini, gpt-5-mini, and gemini-2.5-flash, respectively." source_quote_2: "For comparison, Comparative Distinction increases Democratic SOV by only 2.20, 1.76, and 2.16 points under the Democrat persona." source_quote_3: "Thus, explicit partisan identity can shift baseline exposure more than rewriting a single webpage." source_url: https://arxiv.org/pdf/2610.21330 source_title: "VoteGEO-Bench: Measuring the Impact of Generative Engine Optimization on Electoral Visibility" source_author: "Shiqi Zhang, Xiaokui Xiao" source_venue: "arXiv:2610.21330 [cs.IR]" source_date: 2026-10-15 source_tier: 1 source_sha: c7f6091c8127f94fcedd013621ced2d4ffea1d05fa487283a0eb87c8f978a80a

Further leads

• Section 4.1's source-channel analysis (held back from the three core claims to respect the single-source concentration cap): generative engines cited earned/paid press as the largest channel (46.4–60.9% of citations) but also leaned heavily on owned/self-published sources (24.8–44.3%) and open-platform UGC such as Reddit (up to 22.2%), while independent-institution sources (government agencies, etc.) accounted for only 1.4–3.0%, versus 39.3% in a contemporaneous U.S. Google-SERP reference — source: this paper, arXiv:2610.21330.
• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, "GEO: Generative Engine Optimization" (KDD 2024) — the original GEO framework (Quotes, Statistics, Authoritative, Fluency, Simple Language, Technical Terms, Unique Words strategies) that VoteGEO-Bench's Evidence and Style categories are adapted from directly — source: VoteGEO-Bench reference [1], not independently fetched this session.
• Puerto, Gubri, Green, Oh & Yun, "C-SEO Bench: Does Conversational SEO Work?" (NeurIPS 2025 Datasets and Benchmarks Track) — the source of VoteGEO-Bench's two Structure strategies (Content Improvement, LLMs.txt) — source: VoteGEO-Bench reference [25], not independently fetched this session.
• Hackenburg, Tappin, Hewitt, Saunders, Black, Lin, Fist, Margetts, Rand & Summerfield, "The levers of political persuasion with conversational artificial intelligence," Science 390(6777) (2025) — cited as the political-persuasion research motivating VoteGEO-Bench's new Argumentation category — source: VoteGEO-Bench reference [11], not independently fetched this session.
• Cen, Ilyas, Driss, Park, Hopkins, Podimata et al., "Large-Scale, Longitudinal Study of Large Language Models During the 2024 US Election Season" (arXiv:2509.18446, 2025) — the source dataset (us-election-24) underlying VoteGEO-Bench's U.S. queries and personas — source: VoteGEO-Bench reference [7], not independently fetched this session.
• Miyazaki & Hall, "Why Do AI Models Tell Left-Wing Voters to Support the Communist Party? AI Voting Advice in Japan's 2026 General Election" (2026) — the source dataset (jp-election-26) underlying VoteGEO-Bench's Japan queries and personas — source: VoteGEO-Bench reference [19], not independently fetched this session.
• Pennycook, Cannon & Rand, "Prior exposure increases perceived accuracy of fake news," Journal of Experimental Psychology: General 147(12) (2018) — cited as the mere-exposure/illusory-truth-effect rationale for why VoteGEO-Bench measures raw exposure rather than sentiment — source: VoteGEO-Bench reference [22], not independently fetched this session.
• The paper's own stated limitation: all rewriting experiments were run offline against fixed historical webpage pools with live web retrieval disabled, so effects on real-world search/citation selection (what happens if rewritten pages are actually published and crawled) remain untested by design — source: this paper, Section 8.

Entity candidates

• Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande — people — authors of "GEO: Generative Engine Optimization" (KDD 2024), the FOUNDATIONAL paper that coined the GEO problem setting and the fixed-webpage-pool framework VoteGEO-Bench builds on; flagged first per the blind-spot rule, since this benchmark's entire measurement logic (rewrite a page, measure visibility change) is inherited from their framing, not invented here.
• Shiqi Zhang — person — co-author, National University of Singapore; this paper's corresponding author.
• Xiaokui Xiao — person — co-author, National University of Singapore; established database/data-mining and differential-privacy researcher, here applying that methods background to an election-information-integrity question outside his more typical venues.
• Haritz Puerto, Martin Gubri, Tommaso Green, Seong Joon Oh, Sangdoo Yun — people — authors of "C-SEO Bench" (NeurIPS 2025), the direct source of two of VoteGEO-Bench's Structure strategies.
• Kobi Hackenburg, Ben M. Tappin, David G. Rand, Christopher Summerfield et al. — people — authors of the Science 2025 "levers of political persuasion" paper that motivated VoteGEO-Bench's new Argumentation strategies.
• Hause Lin, Gordon Pennycook, David G. Rand et al. — people — authors of the Nature 2025 "Persuading voters using human–artificial intelligence dialogues" paper, source of the ai-persuasion questionnaire dataset used for VoteGEO-Bench's "Lean left / Swing / Lean right" voter profiles.
• Sho Miyazaki, Andrew B. Hall — people — authors of the source dataset for VoteGEO-Bench's Japan election setting (jp-election-26); title itself ("Why Do AI Models Tell Left-Wing Voters to Support the Communist Party?") is a pointed, independently checkable claim worth a future capture of its own.
• Sarah H. Cen, Andrew Ilyas, Aspen Hopkins et al. — people — authors of the source dataset for VoteGEO-Bench's U.S. election setting (us-election-24), a longitudinal study of LLM behavior across the full 2024 campaign.
• Generative engine optimization (GEO) — concept — the general practice of rewriting web content to improve its visibility in LLM-generated answers rather than traditional search rankings; this paper's subject and the term this capture's whole topic is built around.
• VoteGEO — concept — this paper's own coinage for the specific risk of GEO applied to election-related queries; narrower than GEO itself and worth distinguishing if either gets its own note.