PaperlayerView exact revision on arXiv

Source-linked AI summary

VoteGEO-Bench: Measuring the Impact of Generative Engine Optimization on Electoral Visibility

Shiqi Zhang, Xiaokui Xiao

arXiv:2610.21330v1cs.IRcs.CLcs.CY

TL;DR

Research has paid less attention to how external webpages shape political information generated by LLMs. VoteGEO-Bench evaluates source visibility and webpage rewriting across two elections, engines, and voter personas; target specification and strategy type matter, while persona conditioning shows no consistent effect on gains.

  • Problem

    Although LLM political research examines model bias and voter persuasion, less is known about how external web content shapes political information in generative-engine responses.

    Source: §1 · p. 1

  • Method

    VoteGEO-Bench combines election queries, historical webpages, and voter personas from the 2024 U.S. and 2026 Japanese elections to examine source visibility and evaluate 11 rewriting strategies.

    Source: §1 · pp. 1–2

  • Results

    Target-specified rewriting more often increases target exposure than unspecified rewriting, and Structure and Argumentation strategies show stronger gains than Evidence and Style across engines and elections.

    Source: §1 · p. 2

  • Takeaways & Limitations

    For the two strongest strategies, gains remain broadly positive across evaluated U.S. personas, while explicit partisan identities can shift baseline exposure more than rewriting one webpage.

    Source: §1 · p. 2, §7 · p. 8

    The offline rewriting evaluation does not examine live retrieval or citation selection, and measures exposure without assessing mention valence, factual accuracy, or voter persuasion.

    Source: §2.1 · p. 2, §8 · p. 8

Abstract

from arXiv · show

Large language models (LLMs) are increasingly used for election information, yet little is known about how external web content shapes the political information they present. Generative engine optimization (GEO) rewrites web content to improve its visibility in generated responses. To investigate its potential risks in voting scenarios (VoteGEO), we introduce VoteGEO-Bench, a benchmark covering the 2024 U.S. presidential election and the 2026 Japanese House of Representatives election, combining election-related queries, historical webpages, and voter personas. We first analyze sources cited by generative engines and find that many citations come from lower-cost owned and open-platform sources, while some webpages remain visible across days. We then evaluate 11 rewriting strategies in 4 categories, examining target specification, effectiveness across engines and elections, and the role of voter personas. Rewriting without specifying a target rarely increases target exposure, whereas target-specified rewriting more often does. Strategies that improve page organization or strengthen a position through argumentation yield greater gains than adding supporting evidence or changing wording. For the two strongest strategies, gains remain broadly positive across evaluated U.S. personas, but persona conditioning shows no consistent effect on these gains. Explicit partisan identities can create baseline exposure gaps larger than the gain from rewriting one webpage. When partisanship is implicit in the voter profile, baseline shifts depend on how strongly the engine responds to that signal.

1 Introduction

The paper studies how external webpages shape election information generated by LLMs, a question less examined than political bias in model outputs or the effects of generated messages on voters. VoteGEO-Bench evaluates webpage rewriting and electoral exposure across two elections, engines, and voter personas.

1 Introduction

  • VoteGEO-Bench combines the 2024 U.S. presidential and 2026 Japanese House elections with election queries, webpages, and three categories of voter personas.It contains 28,200 query cells and 17,251 webpages, including questionnaire-derived profiles built from 2,306 U.S. participants.Source: §1 · pp. 1–2
  • Generative engines cite many lower-cost owned and open-platform sources; some webpages recur across days for stable topics, while current-event sources turn over faster.These findings motivate examining source channels and persistence alongside rewriting strategies.Source: §1 · pp. 1–2
  • Target-specified rewriting more often increases target exposure than target-unspecified rewriting, which rarely does.In the U.S. setting, 92% of retrieved webpages mention both major parties.Source: §1 · p. 2
  • Structure and Argumentation strategies show stronger gains across engines and elections than Evidence and Style strategies.Structure improves page organization or LLM-friendly presentation; Argumentation strengthens a position against plausible alternatives or opposing arguments.Source: §1 · p. 2
  • For the two strongest strategies, gains remain broadly positive across 15 evaluated U.S. personas, with no consistent persona-conditioning effect on gains.Explicit partisan identities can create baseline exposure gaps up to 11.6 percentage points, larger than a single webpage's rewriting gain; implicit-partisanship shifts vary by engine response.Source: §1 · p. 2

2 Preliminaries

The paper frames VoteGEO as rewriting one webpage in a fixed query-specific pool to change a named candidate's or party's exposure. It distinguishes this target-level question from prior work on political outputs, voter persuasion, and source or item visibility.

2.1 VoteGEO

  • In the fixed-pool setting, one webpage is rewritten while the other pages remain fixed, and the engine generates a response from the replaced pool.The exposure gain compares the target's exposure in the response before and after replacement.Source: §2.1 · p. 2
  • VoteGEO studies whether rewriting a webpage can increase the exposure of a specified candidate or party mentioned on that page.The target matters because voting webpages may discuss multiple candidates or parties, unlike webpages centered on one entity.Source: §2.1 · p. 2
  • The paper measures target exposure textually without distinguishing favorable from unfavorable mentions, motivated by prior findings that repeated exposure can increase perceived accuracy even for false information.This measurement focus does not assess whether mentions are favorable or factually accurate.Source: §2.1 · p. 2

2.2 Related Work

  • Prior political-bias and persuasion studies examine model outputs or their effects on voters, whereas this work asks how modifying external webpages changes candidate or party exposure.Prior GEO work mainly optimizes source visibility or item ranking; VoteGEO instead considers target exposure within pages that may discuss multiple candidates or parties.Source: §2.2 · pp. 2–3

3 Data Collection and Preprocessing

The benchmark combines election-query and webpage data from the 2024 U.S. presidential and 2026 Japanese elections with U.S. questionnaire data for constructing richer voter personas. Historical search results and citations support query-specific webpage pools.

3 Data Collection and Preprocessing

  • The benchmark comprises three datasets: U.S. and Japanese election query cells with webpage pools, plus U.S. questionnaire data for constructing voter personas.The election settings cover the U.S. presidential election and the proportional-representation component of Japan's House election.Source: §3 · p. 3, Table 1 · p. 3

3.1 Data Sources

  • The U.S. data derive from a longitudinal election corpus and 2,306 participants' pre-treatment questionnaires; the Japanese data include 25 voting-advice queries, 44 demographic personas, and five web-enabled LLMs.The Japanese responses were collected daily from February 2 to February 8, 2026.Source: §3.1 · p. 3

3.2 Query Cells and Voter Personas

  • The U.S. query set retains 260 decision-relevant queries, while Japan retains 25 voting-advice queries; the benchmark includes demographic, partisan, and questionnaire-derived personas.The U.S. personas include 9 demographic identities, 2 partisan identities, an Independent identity, and three questionnaire-derived profiles; Japan contributes 44 demographic identities.Source: §3.2 · p. 3

3.3 Webpage Pools

  • Query-specific webpage pools are reconstructed from historical search results and citations, using archived snapshots when available and live-page or browser fallbacks otherwise.Pools average 9.15 usable webpages per U.S. query cell and 63.1 per Japanese query cell, with 11,501 and 5,750 distinct webpages respectively.Source: §3.3 · p. 3

4 Motivating Insights

The section motivates VoteGEO by examining which sources generative engines cite and how persistent those sources are across days. It compares Japanese generative-engine citations with U.S. Google-SERP results as a reference.

4.1 What Sources Do Generative Engines Use?

  • Earned or paid press is the largest generative-engine source channel, but owned and open-platform sources also account for substantial citation shares.Earned/paid press accounts for 46.4–60.9% of citations, owned sources 24.8–44.3%, and open-platform UGC reaches 22.2%; the U.S. Google-SERP reference shares for owned and UGC sources are 5.1% and 4.5%.Source: §4.1 · p. 4, Figure 1 · p. 4

4.2 How Long Do the Same Sources Stay Visible?

  • For relatively stable topics, more of the same sources persist across days than for event-driven topics, and domain overlap exceeds URL overlap.In Japan, URL overlap is 7.8–8.8% for Social insurance, Nuclear power, and Consumption tax, versus 4.1% for China relations and 2.7% for Constitutional amendment.Source: §4.2 · p. 4, Table 2 · p. 4
  • The U.S. reference shows the same broad stability pattern: Background queries have 12.5% URL overlap, Comparison queries 0.8%, and Recent stories have no URL shared across all sampled days.The authors suggest rewriting may remain useful longer for stable-topic queries, while current-event queries require fresher content.Source: §4.2 · p. 4

5 GEO Rewriting Strategies

The paper evaluates 11 GEO rewriting strategies across four categories, adapting existing structure, evidence, and style approaches and adding two argumentation strategies for voting webpages.

5 GEO Rewriting Strategies

  • The study organizes 11 rewriting strategies into four categories: Structure, Evidence, and Style contain nine strategies adapted from prior GEO research, while Argumentation adds two strategies for voting scenarios.The first three categories’ strategies are adapted from prior GEO research [1] [25]; Argumentation comprises two strategies for voting scenarios.Source: §5 · p. 4

5.1 Existing GEO Strategies

  • Structure strategies reorganize pages without changing facts, while Evidence adds quotations, statistics, or attributions and Style revises wording without adding substantive content.Structure follows C-SEO [25]; Evidence and Style follow GEO [1].Source: §5.1 · p. 5

5.2 Argumentation Strategies

  • Comparative Distinction strengthens a case against plausible alternatives, while Refutational adds substantive opposing arguments and directly rebuts them.Both strategies preserve the original factual information and overall position; their designs draw on contrast advertising and two-sided argumentation.Source: §5.2 · p. 5

6 VoteGEO-Bench

VoteGEO-Bench evaluates target-specified rewriting across elections, engines, and voter personas. Structure and Argumentation generally show stronger gains than Evidence and Style, while target specification matters and persona conditioning does not consistently change rewriting gains.

6.1 Benchmark Setup

  • The benchmark covers 2,939 U.S. and Japan query cells across three engines, using fixed webpage pools and treatment-minus-baseline gains averaged over three replications.The queries, historical webpages, and voter personas support three research questions about target specification, strategy effectiveness, and persona effects.Source: §6.1 · pp. 5–6, §6 · p. 5

6.2 RQ1: Does Target Specification Matter?

  • Target-specified rewriting yields significant positive gains in 24 of 36 settings, including 10 of 12 for SOV; unspecified rewriting has one significant positive and four significant negative changes.In the U.S., 92% of retrieved webpages mention both major parties, leaving the intended target ambiguous without specification.Source: Table 3 · p. 6, §6.2 · p. 6

6.3 RQ2: Which Rewriting Strategies Work?

  • Structure and Argumentation show stronger gains than Evidence and Style across engines and elections, with similar absolute gains yielding larger relative improvements for lower-baseline parties.In Japan, Structure gives Top and Low tiers Mention Coverage gains of 4.28 and 4.12 percentage points (relative improvements 5.9% and 14.4%); Win Rate gains are 1.48 and 1.20 points (10.8% and 51.1% relative).Source: Table 4 · p. 6, §6.3 · p. 7
  • Content Improvement and Comparative Distinction lead Figure 2’s 18 settings, with mean ranks of 2.61 and 2.67 and Top-3 finishes in 13 and 14 settings, respectively.Strategies are ranked by improvement over baseline within each setting.Source: Figure 2 · p. 6, §6.3 · p. 7

6.4 RQ3: How Do Voter Personas Affect GEO?

  • Across evaluated U.S. personas, rewriting gains remain broadly positive without consistent persona-conditioning effects; Comparative Distinction is significantly positive for all 15 personas on two engines.On gpt-5-mini, 29 of 30 persona–strategy gains are positive; persona-conditioned gain differences show no consistent direction.Source: Table 5 · p. 7, §6.4 · pp. 7–8
  • Democrat–Republican baseline SOV gaps reach 11.6, 3.4, and 3.0 points across gpt-4o-mini, gpt-5-mini, and gemini-2.5-flash, exceeding single-page rewriting gains under the Democrat persona.Comparative Distinction gains under that persona are 2.20, 1.76, and 2.16 points across the same engines.Source: §6.4 · p. 8
  • Questionnaire-derived profiles are selected subgroups, not representative voter samples; baseline shifts vary by engine, while rewriting gains are positive in 17 of 18 combinations.Comparative Distinction has significant positive gains in 8 of 9 engine–profile combinations.Source: §6.4 · p. 8

7 Conclusion

VoteGEO-Bench examines GEO’s potential risks in voting scenarios, finding that source visibility and rewriting can shape electoral exposure across engines and elections.

7 Conclusion

  • Generative engines frequently cite lower-cost source channels, and some webpages remain visible across days.These findings motivate evaluating how webpage rewriting affects electoral exposure.Source: §7 · p. 8
  • Target specification matters: target-specified rewriting more often increases target exposure, while target-unspecified rewriting rarely does.Structure and Argumentation strategies show stronger gains across engines and elections.Source: §7 · p. 8
  • Rewriting gains remain broadly positive across voter personas, while partisan information can strongly shift baseline exposure.Future work may examine ethically controlled end-to-end evaluation, contextual sentiment, factual quality, semantic fidelity, and human persuasion.Source: §7 · p. 8

8 Limitations and Ethics Statement

The study uses offline rewriting experiments and measures exposure, not information quality or voter persuasion; it also identifies dual-use risks.

8 Limitations and Ethics Statement

  • Rewriting experiments use fixed historical webpage pools with web retrieval disabled, leaving effects on live retrieval and citation selection unexamined.The authors avoid publishing political rewrites because content could persist in caches or archives and potentially reach voters.Source: §8 · p. 8
  • Exposure measures do not distinguish favorable from unfavorable mentions or assess factual accuracy and voter persuasion.The authors interpret gains only as exposure changes, not as improved information quality or evidence of persuasion.Source: §8 · p. 8
  • The persona analysis reports aggregate results from previously collected datasets, but released rewriting strategies and comparisons carry dual-use risks.The findings could be used to influence political exposure in deployed systems.Source: §8 · p. 8

A.1 Data Preprocessing

The preprocessing appendix describes query and date selection, URL retrieval and extraction, and source-domain labeling for the election datasets.

A.1 Data Preprocessing

  • For cross-day source overlap, U.S. queries use seven shared dates; the rewriting benchmark instead samples seven available dates independently per query.The Japanese overlap analysis uses all available days, while the rewriting benchmark uses all available days for its 25 queries.Source: §A.1 · p. 9
  • The study attempts retrieval for 11,895 U.S. and 6,202 Japanese URLs, with extraction success rates of 96.7% and 92.7%, respectively.URLs are normalized before deduplication, and main text is extracted with trafilatura or a readability-lxml fallback.Source: §A.1 · p. 9
  • Domains are assigned to four source channels by publishing or modification cost, using model-assisted labeling followed by manual review of ambiguous cases.The channels are owned/self-published, open-platform UGC, earned/paid press, and independent institution.Source: §A.1 · pp. 9–10

A.2 Party attribution

Party attribution maps tagged and detected mentions to party entities using curated aliases, with some spans left unmapped.

A.2 Party attribution

  • A curated longest-match alias dictionary maps party mentions; running mates and third-party candidates are excluded from U.S. mapping.Japanese aliases include party names and unambiguous abbreviations, while Japanese candidate names are not mapped.Source: §A.2 · p. 10
  • The same mapping procedure supports both target selection and exposure measurement, while 7.5% of U.S. and 0.8% of Japanese tagged spans remain unmapped.U.S. unmapped spans mostly consist of third-party candidates.Source: §A.2 · p. 10

A.3 Questionnaire-Derived Profile Construction

Questionnaire attributes were used to construct high-purity voter profiles, while robustness checks tested whether the main rewriting findings persisted under clustered inference and multiple-comparison correction. Full-results tables document strategy, election, engine, and persona comparisons.

A.3 Questionnaire-Derived Profile Construction

  • The profile construction used 2,306 participants’ pre-treatment questionnaire responses and excluded ideology and candidate-evaluation variables from subgroup discovery.The study used lean_bidentrump_1 only as a subgroup-discovery label, with scores below 40, from 40 to 60, and above 60 defining lean-left, swing, and lean-right groups.Source: §A.3 · p. 10
  • Table 7 reports gender-persona SOV and rewriting gains across three Japanese party tiers using gpt-4o-mini.The caption identifies the table’s scope but does not state specific outcome values.Source: Table 7 · p. 10
  • Decision trees selected high-purity leaves for each preference group, and GPT-5.5 converted their selected rules into concise voter profiles.The selected tree used entropy, depth 4, and minimum leaf size 20; its five-fold balanced accuracy and macro-F1 were not the primary selection objectives.Source: §A.3 · p. 10
  • The resulting profiles describe high-purity subgroups within the sample, not the entirety of their corresponding preference groups.Their Lean left, Swing, and Lean right labels were used only for reporting.Source: §A.3 · p. 11

A.4 Inference Robustness

  • None of 36 target-unspecified comparisons remained significant after clustering and correction, compared with 5 originally; 18 of 36 target-specified comparisons remained significant, compared with 24 originally.The checks clustered by base query and applied Benjamini–Hochberg correction at a 5% false discovery rate.Source: §A.4 · p. 11
  • After clustering by base query and correcting for multiple comparisons, the strategy ordering remained unchanged in the U.S. and all three Japanese tiers.The ordering was Structure ≥ Argumentation > Evidence > Style.Source: §A.4 · p. 11
  • None of 90 persona Mod contrasts remained significant after clustering and correction, and the originally significant contrasts had mixed directions across engines.The robustness checks therefore found no consistent tendency for persona conditioning to increase or decrease rewriting gains.Source: §A.4 · p. 11

A.5 Full Results

  • Tables 8 and 9 provide full target-specified rewriting and U.S. persona results, respectively, including baseline and placebo references.Table 8 covers strategies, elections, and engines; Table 9 covers the two strongest strategies across U.S. response engines and personas.Source: §A.5 · p. 11, Table 8 · p. 12, Table 9 · p. 13
Loading 2610.21330v1…