And the next speaker at the SEASON 2026 conference in Hamburg is Patrick Parschan, presenting a paper that some of of my colleagues have also had a little input into. The focus here is on what sources are present in AI search results on information relating to elections; the focus here is especially on a number of recent German state elections.
Google and Bing have of course recently rolled out AI-generated search summaries, which have displaced the conventional organic search results; how such summaries are produced remains largely opaque at this point. This produces new risks: search is already a problematic influence on how people inform themselves about key matters in society, and AI Overviews further muddy the waters. Additionally, AI Overviews also substantially reduce click-through rates to the source information, increasing users’ reliance on such information.
The focus of this paper is on the Google and Bing search engines; it focussed on the Rheinland-Pfalz and Sachsen-Anhalt state elections in 2026 and queried the search engines for a battery of some 100 queries per election, on key election candidates, parties, issues, and whom-to-vote-for questions. Results were collected using the Result Assessment Tool (RAT), capturing both the AI Overviews and organic search results.
For Rheinland-Pfalz this produced some 12,000 to 13,000 URLs from around 1,000 unique domains in the organic results, and between 5,000 (Google) and 2,500 (Bing) URLs in AI Overviews. Notably, for some 50% of queries no AI Overviews were provided at all; this is especially the case for queries about candidates. Notably, AI Overviews for questions on whom to vote for were suppressed by Google, but not by Bing.
The study then examined the variety, balance, and disparity of search results; disparity was measured as source frequency – inverse results-page frequency (SF-IRF), which is analogous to the TF-IDF measure in text analysis. Balance and variety in Bing AI Overviews are substantially more restricted than for Google.
Top domains are broadly comparable for organic results; in Google AI Overviews, YouTube is highly featured, where Bing AI Overviews feature statistics sources quite highly. Overall, Google AI Overviews feature social media quite prominently, while organic results emphasis government sources; Bing’s organic results emphasise commercial media, while its AI Overviews include a substantial percentage of ‘other’, uncategorised sources.
A sample of AI Overviews were also manually coded, focussing especially on the URLs and their fit for the query, their party-alignment, and their agreement with the AI summary itself. Party alignment was generally low, and relevance to queries was strong throughout. In 25% of Bing AI Overviews, at least one claim in the summary was unsupported by the URL provided, and there were a range of (in fairness, comparatively minor) errors which were not extreme hallucinations, but nonetheless incorrect assertions.
These results are somewhat concerning, but not especially disastrous: AI Overviews make mistakes, but do not fundamentally mislead their users. More work is to be done on two more elections (Mecklenburg-Vorpommern and Berlin), and the analytical approaches also still need to be developed further.
