TL;DR
- Generative engines answer election questions by citing web pages, and GEO rewrites pages to get cited. We investigate whether that could be used to shift which candidates voters see, a risk we call VoteGEO.
- Across two elections, 25% to 44% of citations came from sources campaigns control, independent institutions fell to 1% to 3% against 39% on Google search, and on stable topics a cited page stayed cited for days.
- Those three observations made the risk worth testing directly. Parts 2 to 4 report the experiments.
People increasingly ask a generative engine, rather than a search engine, who to vote for and what a candidate stands for. A search engine returns links. A generative engine retrieves web pages and writes an answer out of their content, so the pages it picks shape what the voter reads. At the same time, generative engine optimization (GEO) has become an ordinary marketing practice: rewrite a page so that engines are more likely to cite it. Put the two together and a question follows. Could GEO be used to raise one candidate’s or party’s presence in the answers voters get? We call this potential risk VoteGEO, and our paper VoteGEO-Bench is an attempt to find out how real it is.
This is the first of four write-ups. This one covers what we set out to investigate and the observations that convinced us the risk was worth testing. The other three cover the experiments: whether a rewrite has to name its target, which rewriting strategies move exposure, and whether the voter’s persona changes the effect.
What VoteGEO means precisely
A page about an election usually discusses several candidates or parties. So VoteGEO cannot simply ask whether a rewritten page gets cited more. It asks whether rewriting raises the exposure of a target, meaning a specific candidate or party mentioned on that page, in the engine’s answer. Exposure is measured three ways: whether the target is mentioned at all, its share of voice among all entities mentioned, and whether it is ranked first.
What we built to test it
We collected two elections. For the 2024 United States presidential election we kept 260 queries over 106 days on Google’s AI answers, which gave 27,550 query cells and 11,501 distinct cited pages. For the 2026 Japanese House of Representatives election we ran 25 voting-advice queries over 7 days on four web-enabled engines, which gave 650 cells and 5,750 pages. A cell is one query, one day, one engine, with the pages that engine cited. We also built 15 United States voter personas and 44 Japanese personas for the third study.
Observation 1: engines lean on sources that campaigns control
Before rewriting anything we looked at who the engines were citing. Earned and paid press was the largest channel in both elections, between 46.4% and 60.9% of citations. The next largest was owned and self-published content, meaning party sites, candidate sites and campaign posts, at 24.8% to 44.3%. Open-platform user content such as forums and social media reached 22.2% in some query groups.
These are sources a candidate or party can publish or modify at relatively low cost. If an engine draws a quarter to nearly half of its citations from them, then the people with the strongest motive to shape the answer already own much of the raw material.
Observation 2: the independent sources nearly disappear
Independent institutions, such as research centres and non-partisan voter guides, accounted for only 1.4% to 3.0% of what the engines cited. On a conventional Google results page for the same United States queries, those institutions held 39.3% of the slots. The generative engines were not simply reproducing the search mix. They shifted it away from the sources least open to outside editing.
Observation 3: on stable topics, the same pages keep being cited
A rewrite only matters if the page stays in use. We measured how many cited URLs repeated from one day to the next. Queries about recent stories repeated none. Queries about who the nominees are repeated 25.0% of their URLs, and Japanese policy-issue queries repeated between 2.7% and 8.8%, with social insurance at the top. Domain overlap was consistently higher than URL overlap, so the same publishers stayed visible even when the specific page changed.
For a stable topic, then, a page that gets in tends to stay in for days. A rewrite that works would not be a one-day effect.
Why these three observations led to three experiments
Taken together, the observations say that generative engines frequently use sources that campaigns can cheaply publish or edit, that they use fewer independent sources than search does, and that on stable topics a cited page persists. That is enough to make VoteGEO worth testing directly rather than assuming. The three questions we then asked were whether a rewrite needs to know which target it serves, which kinds of rewriting change exposure, and whether the persona of the person asking changes the effect.
How we tested without creating the risk
Every rewriting experiment ran offline against fixed pools of historical pages, with the engines’ web retrieval switched off, so an engine could only choose among the pages we gave it. We did not publish any rewritten political content, because a page published on the open web can persist in caches and archives and reach real voters. The metrics count exposure only and do not judge whether a mention is favorable, accurate or persuasive. We report the results in the hope that they inform safeguards against manipulation through retrieved web content.
Continue with Part 2: does the rewrite need a target?