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

Shiqi Zhang; Xiaokui Xiao

cs.IROct 15, 2026

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

Authors: Shiqi Zhang, Xiaokui Xiao

Organizations: National University of Singapore, Singapore

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1. Source share by source-channel label over time.

U.S.			Japan
Subcategory	URL	Domain	Policy issue	URL	Domain
Nominees	25.0	25.0	Social insurance	8.8	21.8
Background	12.5	14.7	Nuclear power	8.7	17.9
Presidential candidates	12.1	16.4	Consumption tax	7.8	19.0
Election outcomes	10.9	16.7	Diet seat reduction	7.0	17.9
Issue importance	8.4	12.2	Espionage law	6.8	17.7

Table 2. Cross-day source overlap at URL and domain levels.

	gpt-4o-mini						gpt-5-mini						gemini-2.5-flash
Category	Mention		SOV		Win Rate		Mention		SOV		Win Rate		Mention		SOV		Win Rate
Category	U	S	U	S	U	S	U	S	U	S	U	S	U	S	U	S	U	S
Placebo	$-0.18$	$-0.09$	$+0.16$	$-0.26$	$+0.87$	$-1.13$	$-1.40$	$-0.44$	$-0.56$	$-0.24$	$-1.32$	$-0.87$	$+0.50$	$+0.29$	$+0.04$	$+0.16$	$-0.56$	$+0.18$
Argumentation	$-0.31$	$+1.94$	$+0.07$	$+1.48$	$+0.86$	$+2.57$	$-0.08$	$+0.74$	$-0.48$	$+0.53$	$-0.93$	$+0.52$	$+0.28$	$+1.23$	$+0.17$	$+1.11$	$+0.84$	$+1.58$
Structure	$-0.49$	$+2.81$	$+0.19$	$+1.30$	$+1.03$	$+2.14$	$-1.28$	$+1.76$	$-0.87$	$+0.89$	$-0.73$	$+0.73$	$+0.61$	$+1.89$	$+0.25$	$+1.47$	$+0.36$	$+1.84$
Evidence	$+0.32$	$+1.72$	$+0.27$	$+0.76$	$+0.94$	$+0.91$	$-0.54$	$+0.52$	$-0.65$	$+0.44$	$-0.92$	$+0.14$	$-0.04$	$+0.28$	$+0.04$	$+0.88$	$+0.31$	$+0.90$

Table 3. Effect of target specification across strategy categories and engines. Values are mean improvements in percentage points. Black values are statistically significant (

p<0.05

). U = target unspecified, S = target specified.

	Mention Coverage				Share of Voice				Win Rate
Category	U.S.	Top	Mid	Low	U.S.	Top	Mid	Low	U.S.	Top	Mid	Low
Baseline	85.86	72.41	72.43	28.55	54.41	12.03	10.10	3.73	58.88	13.72	9.25	2.34
Argumentation	$+0.52$	$+2.74$	$+1.75$	$+1.96$	$+1.29$	$+1.04$	$+0.69$	$+0.46$	$+1.97$	$+1.72$	$+0.74$	$+0.52$
Structure	$+0.78$	$+4.28$	$+2.55$	$+4.12$	$+1.54$	$+1.15$	$+0.74$	$+0.83$	$+1.94$	$+1.48$	$+0.79$	$+1.20$
Evidence	$+0.36$	$+1.76$	$+0.74$	$+1.03$	$+1.09$	$+0.35$	$+0.28$	$+0.20$	$+1.21$	$+0.20$	$-0.25$	$+0.15$
Style	$+0.08$	$+1.27$	$-0.26$	$+0.21$	$+0.83$	$+0.09$	$+0.01$	$-0.00$	$+0.45$	$-0.46$	$-0.48$	$+0.09$

Table 4. Target-specified rewriting across elections and target tiers, pooled across the three engines. Baseline is mean exposure (%), and other values are mean improvements (

\Delta

, percentage points). Black values are statistically significant (

p<0.05

Panel of Figure 2 — Figure 2. Ranking of target-specified strategies in 18 settings.

gpt-4o-mini

gpt-5-mini

gemini-2.5-flash

\Delta

Mod

\Delta

Mod

\Delta

Mod

Appendix

Category	Subcategory	Example query	#
Predictions	Election outcomes	How will South Carolina vote in the 2024 US presidential election?	105
Election issues	Candidate positions ∗	What is Kamala Harris’s position on immigration?	51
Candidates	Comparison ∗	Which 2024 US presidential candidate has the best chance of winning?	29
Election issues	Public opinion ∗	Which candidate do voters believe will handle immigration best?	17
Issues	Issue importance	What is the importance of and debate around the economy?	17
Candidates	Background	What controversies is Donald Trump the subject of?	12

Appendix

Table 6. Composition of the 260 U.S. base queries retained from Cen et al. (2025) .

\ast

marks the five subcategories used in Section 6.1 .

			$\Delta$		Mod
Tier	Persona	Base	CI	CD	CI	CD
Top	Persona-free	16.7	+2.64	+3.67	–	–
	Man	15.7	+4.88	+5.76	+2.25	+2.10
	Woman	15.9	+4.58	+5.23	+1.95	+1.57
Mid	Persona-free	7.5	+1.17	+3.26	–	–
	Man	7.4	+1.90	+2.76	+0.72	$-$ 0.50

Appendix

Table 7. Effect of gender personas on SOV and rewriting gains across three party tiers in Japan ( gpt-4o-mini ).

Mention Coverage ( $\Delta$ , pp)
		U.S.			JP – Top			JP – Mid			JP – Low
Category	Strategy	4o-mini	5-mini	Gemini	4o-mini	5-mini	Gemini	4o-mini	5-mini	Gemini	4o-mini	5-mini	Gemini
	Baseline (%)	90.3	74.4	92.8	51.5	65.7	83.2	25.5	75.0	87.6	4.3	24.6	39.0
	Placebo	+0.00	$-$ 0.39	$-$ 0.13	$-$ 0.27	$-$ 0.19	+0.76	$-$ 1.35	$-$ 3.24	+0.86	+0.81	+1.91	+0.19
Argumentation	Comparative Distinction	+0.83	+1.00	+0.74	+9.60	+1.52	+1.81	+7.86	+1.33	+0.57	+6.78	+0.38	+1.14
Argumentation	Refutational	$-$ 0.04	$-$ 0.09	+0.70	+4.00	+1.33	+2.10	+0.27	+0.57	+2.10	+2.17	+1.71	+1.91

Appendix

Table 8. Full target-specified rewriting results across strategies, elections, and response engines. Baseline reports mean exposure (%), and all other entries report mean paired changes from baseline in percentage points. Black values are statistically significant (

p<0.05

), while gray values are not. Placebo compares two generations without rewriting. For Japan, Top, Mid, and Low denote target-party tiers, and Gemini pools results over the grok-4-1-fast and grok-with-x webpage pools. 4o-mini, 5-mini, and Gemini denote gpt-4o-mini , gpt-5-mini , and gemini-2.5-flash , respectively.

gpt-4o-mini

gpt-5-mini

gemini-2.5-flash

Generative Engine Optimization

Generative Engine Optimization studies how to make web content visible, cited, and accurately represented in AI-generated search answers.

See all 40 papers

Who Would You Vote For? Auditing Political Alignment in LLMs: An Italian Case-Study
Aug 12, 2026Simone MungariLLM Auditing Political Bias in Language Models
Revelis s.r.l.
Position: Generative Engine Optimization Creates Underexamined Risks, Governance Must Target Concentration, Disclosure, and Academic Blind Spots
May 18, 2026Yizhu Wen, Nan Zhang, Haohan Yuan +3Generative Engine Optimization LLM Auditing
School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, USA · The Media School, Indiana University Bloomington, Bloomington, USA · School of Data Science, University of North Carolina at Charlotte, Charlotte, USA +1
Issue Bias in Generative AI Writing Assistance: Political Issues and LLMs in the Swedish 2026 Election
Sep 14, 2026Bastiaan Bruinsma, Annika Fredén, Paul Röttger +2Political Discourse Analysis Political Bias in Language Models
Chalmers University of Technology · University of Gothenburg · Lund University +1

Aug 12, 2026J/K move · Enter open · S save

Who Would You Vote For? Auditing Political Alignment in LLMs: An Italian Case-Study

Simone Mungari

Revelis s.r.l.

LLM Auditing Political Bias in Language Models Language Model Bias Evaluation

Open paper AskPDF arXiv

As users increasingly turn to Large Language Models (LLMs) for information and advice on political matters, particularly during election periods, the political preferences expressed by these systems have become a matter of public interest. Prior research has shown that interactions with LLMs can influence users' political attitudes and choices, raising questions about how these models themselves evaluate political actors. In this paper, we investigate whether and how LLMs express preferences toward political parties and political leaders. We introduce a systematic and reproducible auditing framework in which multiple LLMs are prompted to evaluate parties and leaders across nine criteria. Rather than attempting to infer the models' "true" political beliefs, we focus on their observable behavior, examining consistency across evaluations, differences between models, refusal rates, and sensitivity to prompt formulation. We further investigate how these evaluations vary when models are instructed to adopt different personas. We demonstrate the framework through an Italian case study, providing a systematic analysis of LLM-generated political evaluations on italian parties and leaders.

Abstract

Generative Engine Optimization

Explore similar work

Who Would You Vote For? Auditing Political Alignment in LLMs: An Italian Case-Study

Position: Generative Engine Optimization Creates Underexamined Risks, Governance Must Target Concentration, Disclosure, and Academic Blind Spots

Issue Bias in Generative AI Writing Assistance: Political Issues and LLMs in the Swedish 2026 Election