And the final speakers in this session at the SEASON 2026 conference in Hamburg are Corinna Hertweck and Giovanni Astante, whose focus is on voting in Switzerland’s frequent referenda. How do Google’s and Bing’s AI Overviews perform in such contexts?
Referenda happen at all levels of Swiss democracy, with approximately four such votes (sometimes on several issues) happen per annum, and are surrounded by lively public debate. Information retrieved from AI Overviews might play a significant role in influencing voters’ decision-making processes; importantly, too, such votes are binary yes/no choices, which makes them different from conventional election decisions.
How often do AI summaries appear in response to referendum-related queries, then; what sources do they emphasise; and how do they represent the opposing sides in a referendum? The focus here is on the ’10 million initiative’, a populist Swiss party’s initiative to cap the resident Swiss population at 10 million (it is currently 9 million), essentially by restricting migration. This was seen by other parties as a deep intervention in migration policy, and a breach of Switzerland’s obligations under current freedom of movement treaties with the European Union.
This study tested a variety of search terms related to this referendum, across Switzerland’s four official languages as well as English, and also including a range of query suffices including those indicating support or opposition as well as some requesting summaries and overviews. These queries were repeated over some time, from two Swiss locations, resulting in some 190,000 queries producing around 47,000 AI summaries (mostly from Google) and 90,000 unique URLs.
These URLs were annotated for entity types, link topic, link position, and link language, as well as various other aspects; annotating the top 638 URLs already covered some 82% of all URLs encountered.
AI summaries appeared in over 55% of all queries for queries in English, French, and Romansh, but much less so in German and Italian; they also appeared inconsistently as the same queries were repeated over time. It is not clear what causes these discrepancies.
Government and media outlets were most prominent, with URLs related to populist SVG party in third place; YouTube also appeared increasingly strongly towards the end of the timeframe. Links that preferred a Yes or No vote outnumbered generic overviews two to one, with slightly or strongly opposed links accounting for some 47% of all links (and 74% of all links taking one side). This does not reflect a balanced view, opinion polls, or the eventual referendum outcome (which was 55% to 45% against), but does generally mirror the balance of political parties’ positions relating to the referendum. At the same time, URLs belonging to the SVP itself are vastly overrepresented in search results compared to the other parties, compared to their vote share at the last election, or their number of seats in Swiss parliament. (There is also now a public dashboard which enables users to interactively explore these results.)
What any such patterns are based on remains unclear; this is due to the substantial intransparency surrounding AI Overviews to date. As such summaries become more commonplace, there is a substantial need to further regulate such practices.
