Google’s AI Overviews are showing up for election-related searches less than half the time, and when they do appear, they’re pulling from a startlingly narrow pool of sources. That’s the core finding from AlgorithmWatch, a German advocacy group that used EU Digital Services Act research access to run 4,480 search queries tied to upcoming state elections in Saxony-Anhalt, Mecklenburg-Vorpommern, and Berlin.
The numbers paint a picture of selectivity without explanation. AI Overviews appeared for just 39.1 percent of election questions, versus 65.3 percent for non-political queries. Questions about polls almost never triggered one, while party and politician questions did 40 to 50 percent of the time. Queries about the far-right AfD got overviews at just 24 percent — roughly half the rate of other parties. Google says overviews appear when there’s ‘high added value’ and that the system is designed to be neutral, but declined to explain the actual criteria.
Source diversity is where things get uncomfortable. Nearly half of all links pointed to just ten domains, with YouTube — Google’s own platform — cited most often. Six German media outlets, mostly public broadcasters like NDR and MDR plus Zeit Online, accounted for 75 percent of media links. Party pages appeared alongside Wikipedia and government sites with no labeling to distinguish partisan sources from neutral ones. AlgorithmWatch reads the YouTube dominance as self-preferencing; Google denies it.
The tone problem may be the most troubling finding. AlgorithmWatch rated 81.5 percent of overviews about the center-right CDU as clearly positive, compared to 10.7 percent for the Greens and zero percent for the AfD. The same AfD candidate was described alternately as ‘far-right,’ ‘radical far-right,’ and ‘right-wing populist’ within short time spans, with no discernible pattern. The group chalks this up to sycophancy — language models defaulting to flattering, agreeable framing. Back in 2024, Google said it would block Gemini from answering election questions; whether similar safeguards still apply in 2026 remains unanswered. The flattering adjectives echo what we’ve seen across generative AI: models trained to please users often mirror the language of whatever source they’re drawing from, which gets dangerous when those sources are the politicians themselves.
