During the 2024 European Parliament and US presidential elections, Damião et al. (arXiv:2603.23474) deployed bots across five EU countries and 15 US counties, issuing approximately 4,360 election-related queries to four search engines and two language models. The methodology was privacy-preserving and standardized — no personalization, no history, no cookies.
European search results disproportionately surfaced far-right political entities. Not proportionally to their polling numbers, not proportionally to their historical election performance, not proportionally to their mainstream media coverage. Beyond all three baselines. The search engines amplified fringe political positions that the electorate, the electoral record, and the press all agreed were marginal.
In the US, the bias was partisan but engine-specific. Google's results skewed toward Republican-aligned topics; competing engines skewed Democratic. The disagreement between search engines means the bias isn't reflecting an underlying reality — it's reflecting the engine's training data, ranking algorithm, or source selection.
Language models were more balanced. They still overrepresented far-right and Green party content relative to polling, but the magnitude was smaller and the pattern less systematic than search engines.
The through-claim: search engines don't reflect political reality — they distort it, and the distortion disproportionately amplifies the fringe. The finding that different engines amplify different parties proves the amplification isn't demand-driven. It's an artifact of the ranking system itself, and the artifact is strongest at the political margins.