This post describes a preprint that has not yet been peer reviewed. It summarizes an investigation into a potential risk, not a demonstrated real-world harm.

More and more people ask a chatbot or an AI Overview about an election — who is running, what each party stands for, what the latest story is. Those answers are grounded in web pages the engine retrieves and cites. Meanwhile, generative engine optimization (GEO) has emerged as a practice of rewriting web pages so that generative engines are more likely to surface them. Our earlier work on GEM-Bench studied this in a commercial setting. VoteGEO-Bench asks a more uncomfortable question: if the same page-level rewriting were applied to political content, could it change which candidates or parties an LLM's election answers expose? We call this potential risk VoteGEO, and the paper's goal is to measure whether it is real and how large it could be — not to make it happen.

Why this is worth investigating

Prior work on LLMs and politics has mostly looked inward, at a model's own leanings, or at whether a model can persuade a reader in conversation. Much less is known about the outward channel: how externally controlled web content shapes the political answers an engine produces. That channel matters because web pages are cheap to publish and cheap to rewrite. If a single edited page could measurably tilt exposure, the barrier to influencing election information would be very low.

What VoteGEO-Bench contains

The benchmark is built on two real elections: the 2024 U.S. presidential election (election queries issued to Google Search over 106 days, with historical result pages) and the 2026 Japanese House of Representatives election (queries across four engines over 7 days). Together they give tens of thousands of query cells — a query, on an engine, on a day, with a fixed pool of web pages — and thousands of distinct pages. A third component, drawn from existing questionnaires of over two thousand U.S. participants, supplies voter personas that range from plain demographics to explicit partisan identity.

The core experiment is deliberately minimal. Take one page in the pool, rewrite it with one strategy, regenerate the answer, and compare it with the baseline answer. Exposure is measured per target party through three lenses: whether the target is mentioned at all, its share of voice among all parties mentioned, and how often the rewrite wins against the baseline. The gain is the change in exposure. Nothing else in the pool changes.

Two motivating observations

Before any rewriting, the retrieved sources themselves tell a story. Established news and press remain the largest cited channel, but a substantial share of citations — roughly a quarter to well over forty percent, depending on the election — comes from owned or self-published sources, and user-generated content on open platforms reaches about a fifth of citations in some settings. Independent institutions account for only a few percent. These are exactly the low-cost channels where a page can be written and rewritten at will.

Second, cited pages persist. For stable topics such as nominees, a meaningful fraction of URLs remains cited across days, so an edited page would not be a one-off: it would stay in the answer's evidence base for a while. For fast-moving topics like recent stories, overlap drops to zero. The risk surface is therefore uneven, concentrated on the durable, encyclopedic parts of election information.

What the rewriting study shows

We evaluated eleven rewriting strategies grouped into four categories — Structure, Evidence, Style, and a new Argumentation category — under an offline protocol described below. Three findings stand out.

Target specification is the hinge. Rewriting a page without specifying a target party rarely raised that party's exposure; across 36 settings it produced one significant positive gain and several significant negatives. Rewriting with a specified target produced significant positive gains in a majority of settings. The reason is structural: most election pages mention several parties, so an untargeted "improvement" of the page can just as easily amplify an opponent. The risk, in other words, lies in intentional rewriting, not in ordinary content polishing.

How the page is organized matters more than how it is worded. Structure and Argumentation strategies produced the most consistent gains, Evidence strategies fewer, and Style strategies almost none. Parties with low baseline exposure saw the largest relative gains, which means the smaller actors are both the most exposed to this risk and the most able to benefit from it.

Who is asking can matter more than what was rewritten. Across U.S. voter personas, rewriting gains stayed mostly positive. But an explicit partisan identity in the prompt shifted baseline exposure by more than a single rewritten page did: the Democrat–Republican baseline gap on one engine was several times the gain any strategy added. Questionnaire-derived profiles moved baselines too, by different amounts on different engines. Under stricter inference — standard errors clustered by query with false-discovery correction, reported in the appendix — the persona-specific differences in rewriting gains do not survive, while the target-specification and strategy-category results largely do.

What we did not do, and why

The study is entirely offline. Live retrieval was disabled, engines were run over fixed historical page pools, and no rewritten political page was ever published, because even a cached or archived copy could reach real voters. The metrics measure exposure only: whether and how prominently a party appears. We did not systematically verify the factual correctness or semantic fidelity of rewrites, and we make no claim about persuasion or voting behavior. Observed gains are interpreted strictly as changes in exposure. Persona analysis reuses existing survey data and reports aggregates only.

We also recognize the dual-use nature of these results. The same measurements that reveal a vulnerability could inform someone who wants to exploit it. We think the balance favors disclosure: engine developers, platform operators, and election-integrity researchers cannot defend against a channel whose size they cannot measure, and the benchmark gives them a reproducible way to do so.

Takeaways

• Election answers from generative engines draw heavily on low-cost, editable sources, and cited pages can stay in the evidence base for days.
• Rewriting one page can measurably raise a specified party's exposure; untargeted rewriting rarely does, so the risk is tied to intent.
• Structural and argumentative edits are the ones to watch; surface-level style changes have little effect.
• The user's stated identity can move answers at least as much as a rewritten page, which points to the engine side as a second place to look for safeguards.

Open questions include the sentiment and factuality of rewritten content, how such edits interact with many pages at once, and whether exposure differences translate into any change in what people believe. We hope VoteGEO-Bench helps the community study these questions carefully and ahead of time.