A German research group just tested what Google’s AI Overviews actually say about candidates, using a legal channel that did not exist for most of the past decade. AlgorithmWatch, an advocacy organization focused on automated decision-making, obtained access to Google’s Search Researcher Result API under Article 40(12) of the EU’s Digital Services Act, the provision that lets vetted researchers query platform data directly rather than relying on scraping or leaks. The group used that access to run 4,480 search queries tied to three 2026 state elections in eastern Germany: Saxony-Anhalt, Mecklenburg-Vorpommern, and the Berlin House of Representatives vote.
The access matters as much as the findings. Most prior research on how AI answers handle political content has relied on manual sampling or third-party scraping, both of which platforms can dispute as unrepresentative. A DSA-authorized pull of 4,480 queries is a different evidentiary standard, and it is one of the first studies built on this specific access mechanism since the regulation took effect.
On frequency, AlgorithmWatch found that Google displayed an AI Overview for 39.1 percent of election-related queries, compared with 65.3 percent for non-political ones. That gap was not uniform across topics. Questions about poll standings almost never triggered a summary, while questions about parties or individual politicians produced one in 40 to 50 percent of cases. Queries about the far-right Alternative for Germany party (AfD) returned an overview in 24 percent of cases, below the 45 to 58 percent range AlgorithmWatch measured for other parties. Google told the group its summaries are built to stay neutral and surface only where the company sees genuine added value, but did not specify the criteria that trigger that judgment, according to AlgorithmWatch.
On sourcing, the study found that nearly half of all links cited across the overviews came from just ten domains. The single most-cited source was YouTube, which Google owns. AlgorithmWatch reads the concentration as evidence that a small set of properties, including Google’s own platform, dominates what gets cited in these summaries. Google disputed the characterization that this reflects favoritism toward its own services. Among traditional outlets, German public broadcasters (NDR, rbb, MDR, ARD) together with Zeit Online supplied three quarters of every media link counted, and AlgorithmWatch noted that party-run websites appeared among the cited sources without any label distinguishing them from independent outlets.
On language, AlgorithmWatch sampled how the overviews described each party in Saxony-Anhalt and rated the tone of each description. In that sample, 81.5 percent of descriptions of the center-right CDU read as clearly positive, compared with 50 percent for the center-left SPD, 10.7 percent for the Greens, and zero percent for the AfD. The group also found that overviews used vague, favorable adjectives such as “pragmatic” or “close to citizens” that echoed how the parties describe themselves. It flagged one AfD candidate whose political positioning was described inconsistently across similarly timed queries, sometimes as far-right and national-conservative, sometimes as holding radical far-right views, sometimes as right-wing populist, with no discernible pattern behind the variation.
AlgorithmWatch’s study measures whether overviews appeared, what they cited, and how they read; it does not establish why Google’s system produced these particular patterns, and the group did not claim otherwise. It traced the favorable phrasing to a known failure mode in language models, sycophantic wording that mirrors a subject’s own self-description, rather than to any stated Google policy, and it called for transparency requirements and clearer liability rules. Google had committed in 2024 to blocking Gemini-generated answers on election questions; AlgorithmWatch said it could not determine whether that safeguard still applies in 2026, and Google did not respond to the question.
For any team building on AI Overviews, AI Mode, or similar answer engines, the operating lesson is that DSA researcher access now makes this kind of query-level audit repeatable and citable, and more of these audits should be expected across the 2026 European election calendar.
Reported by Maximilian Schreiner for The Decoder on September 1, 2026.
