Can rewriting one webpage change AI election visibility?
An offline benchmark using US and Japanese election data found that target-specific structural and argumentative rewrites changed party exposure more reliably than stylistic edits. It did not test live search, factual accuracy, voter persuasion or election outcomes.
The short version
3 key points
- 1.The benchmark assembled 27,550 US query cells and 650 Japanese query cells, representing 11,501 and 5,750 distinct webpages, then tested 11 rewriting strategies offline.
- 2.Target-unspecified rewrites did not remain significant after clustered multiple-testing correction, while structural and argumentative strategies produced the most consistent exposure gains when a party target was specified.
- 3.The study measured mentions and share of voice, not accuracy, sentiment, persuasion, voting behavior or live retrieval, and it deliberately kept rewritten political content off the public web.
The answer: in the benchmark, some rewrites changed exposure
Rewriting one election-related webpage could increase how often a named party appeared in an AI-generated answer, but the effect depended on explicitly choosing a target and on the kind of rewrite. Across the paper's offline tests, clearer page structure and stronger comparative argument performed more consistently than style changes or the addition of supporting evidence. Rewriting without specifying which party should gain rarely worked and none of those comparisons remained significant after clustered multiple-testing correction.
This is evidence of a potential manipulation surface, not evidence that an election was manipulated. The researchers fixed historical webpage pools and disabled live retrieval. They measured exposure—mentions, share of voice and which party had the most exposure—not whether the answer was true, favorable or persuasive. No voter saw the generated content and no vote changed. The practical conclusion is that generative-search safeguards should examine how retrieved page design shapes political visibility before an election makes that question urgent.[1]
What data went into VoteGEO-Bench
The National University of Singapore team combined historical data from two elections. The US component retained 260 voter-decision queries from the 2024 presidential election over 106 days from 23 July to 5 November. Historical Google search results produced 27,550 query-day cells with 9.15 usable webpages on average and 11,501 distinct pages. The Japanese component used 25 voting-advice queries across seven days around the February 2026 House of Representatives election and four web-enabled models, creating 650 query cells and 5,750 distinct pages.
For the United States, 109 queries from candidate-position, comparison, public-opinion, evaluation and vice-presidential categories entered the rewriting experiment. For Japan, the prompts covered 12 policy issues with opposing stances plus one control. The team reconstructed page content using the closest Wayback Machine snapshot, then live-page and browser fallbacks. Retrieval succeeded for 96.7% of attempted US URLs and 92.7% of Japanese URLs.[1]
Why the source mix raised concern
The researchers grouped domains into owned or self-published sources, open-platform user content, paid or earned press and independent institutions. In the Japanese generative-engine citations, news sources remained the largest category at 46.4% to 60.9%. Candidate or party-controlled sources accounted for 24.8% to 44.3%, open-platform content reached 22.2%, and independent institutions accounted for 1.4% to 3.0%. The US Google-search reference had much less owned and user content and 39.3% independent-institution material, although that cross-country and cross-engine comparison is descriptive rather than causal.
Some sources also remained visible across multiple days, especially for stable policy questions. Domain overlap exceeded exact-page overlap, meaning the same publishers could persist even as URLs changed. That matters because a low-cost source that is repeatedly retrieved offers a longer window for optimization. Current-event queries turned over faster. The benchmark therefore focuses on a plausible route by which external publishers can affect generated political information, without claiming that all cited owned content is false or improper.[1]
The eleven rewriting strategies
The experiment organized 11 strategies into four groups. Structural rewrites improved headings, ordering, key points and other presentation features or reformatted the page for machine readability. Evidence strategies added quotations, statistics or source attribution. Style strategies changed fluency or wording without adding substance. Two argumentation strategies strengthened comparison or responded to counterarguments. Some prompts named a target party and others left the target unspecified.
US pages frequently mentioned both major parties—92% of the retrieved pages in the experiment—so an unspecified rewrite had no clear beneficiary. Once the target was named, structure and argumentation produced the strongest average gains across engines and elections. Content Improvement and Comparative Distinction placed in the top three in 13 and 14 of 18 setting-metric combinations. The absolute gains were often modest, but similar point changes represented larger relative shifts for Japanese parties with low baseline exposure.[1]
What personas changed
The team also varied voter descriptions while keeping the page and target fixed. US testing included nine demographic identities, three explicit partisan identities and three profiles derived from pre-treatment questionnaire data from 2,306 participants in an earlier persuasion study. Japanese data included 44 demographic combinations. The rewriting gains were broadly positive across tested US personas, but persona conditioning did not consistently increase or reduce those gains after correction for clustered queries and multiple comparisons.
Explicit partisan identity changed baseline exposure more than rewriting one page. The gap in Democratic share of voice between Democratic and Republican personas reached 11.6 percentage points on one model, compared with roughly two points from the stronger single-page rewrite in that setting. That should not be read as a national estimate or proof of persuasion. The profiles were synthetic prompts, and the richer questionnaire-derived profiles were selected to form interpretable subgroups rather than represent every voter in a political category.[1]
Why the study stayed offline
A live experiment would require publishing rewritten political pages and waiting for search or generative systems to index them. The authors judged that content could reach real voters, persist in caches or be archived after removal. They therefore replaced one page only inside a fixed historical pool. That choice avoids introducing political influence into an active information environment and makes comparisons repeatable, but it also removes the retrieval stage where an optimized page must first be found and selected.
The paper explicitly identifies dual-use risk. Its strategy comparisons could help defenders design monitoring, but they could also guide campaigns or influence operations. The sensible policy response is not to repeat the experiment live during an election. Platforms can run controlled internal red-team tests, preserve source-level audit logs, detect sudden structurally coordinated page changes and diversify retrieval toward authoritative and independent information when users ask for voting advice.[1]
Limits that prevent a persuasion claim
Exposure was counted without determining whether a mention praised, criticized or merely named a party. The paper did not systematically verify the factual accuracy or semantic fidelity of rewritten pages. Some evidence strategies could introduce material absent from the original. Generated answers were not shown to people, so there is no measured change in belief, turnout, polarization or vote choice. The US dataset represents the 2024 election and the Japanese dataset a short 2026 window, not all electoral systems or languages.
The study is also an unreviewed arXiv manuscript with a placeholder conference citation and DOI. Its robustness analysis is a strength: clustering by base query and controlling the false-discovery rate removed all target-unspecified effects and several smaller comparisons. That makes the remaining qualitative pattern more credible within the benchmark, but external validity remains open. Real systems change their indexes, prompts, safety layers and citation behavior continuously.[1]
What would change the assessment
Confidence would rise with peer review, independent reproduction and controlled platform-side tests that include retrieval without releasing manipulative content publicly. Future work should score factual accuracy, sentiment, source diversity and semantic preservation alongside exposure. Results should be reported by language, electoral system, party size and time to election, with preregistered hypotheses and a clear multiple-testing plan.
A claim about democratic harm would require human evidence under strong ethical safeguards: whether people notice the source, whether exposure changes understanding or choice, and whether corrections undo any effect. For now, the study supports a narrower warning. Generative engines can be sensitive to how one retrieved political page is organized and argued, while partisan cues already shift baseline answers. That is enough to justify transparency and red-team work, not enough to infer that one rewrite can swing an election.[1]
What this means for people
- Voters may receive different party exposure from an AI answer even when the underlying query is similar, but this benchmark did not measure persuasion or voting behavior.
- Journalists and election officials need source-level visibility so they can distinguish official, partisan, user-generated and independent evidence in generated answers.
- Platforms can test manipulation internally and diversify retrieval without publishing optimized political material that could reach real voters.
Global context
The benchmark spans historical US presidential-election data and the proportional-representation component of Japan's 2026 House election. The settings differ in parties, languages, engines and time windows, so shared patterns are informative but not universal estimates of political-AI behavior.
What the evidence does not yet show
- Unreviewed preprint with a placeholder conference citation and DOI
- Offline fixed webpage pools with live retrieval and indexing disabled
- Exposure metrics do not distinguish positive from negative mentions
- No systematic factuality, semantic-fidelity or source-quality assessment
- No human participants, persuasion outcome or measured voting behavior
- Two election settings cannot represent all political systems, languages or models
What to watch next
- Peer review and independent reproduction of the surviving corrected effects
- Platform-side red-team tests covering both retrieval and generation without public release
- Election-answer interfaces that disclose and diversify sources
- Human studies measuring understanding and persuasion under strict ethics controls
Living evidence record
Impact record IAI-1D1KLCE
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
16 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 16 October 2026
Research question, topics and regions
Whether changing the structure or argument of one historical election webpage alters candidate or party visibility in generated answers
Topics: Elections · Generative search · Media visibility · Political information · Generative engine optimization · AI governance
Regions: Global · Asia · North America · Singapore · Japan · United States
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Society & Media
Why did AI become experts’ top future risk?
In AXA’s 2026 perception survey, AI and big data rose from fourth to first among 4,345 risk professionals, while 24,000 adults ranked it second behind climate change. The result measures concern, not the probability or cost of future harm—and 89% of the expert panel worked for AXA.
8 min · 2 sources
Society & Media
How did fake experts reach real newsrooms?
OpenAI says two covert influence operations used false journalist identities and a front research centre to place material in real outlets. AI mostly helped with drafting, translation and internal reporting; the harder failure was identity and source verification, and claimed reach remains only partly corroborated.
10 min · 2 sources
Society & Media
Do safer-driving AVs change human behaviour?
A peer-reviewed quasi-experiment found that Lyft automated vehicles followed more smoothly and conservatively, while human drivers left shorter gaps behind matched automated leaders. The study analysed real trajectories, but it did not randomise drivers or measure crashes, injuries or current vehicle systems.
8 min · 3 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.