AlgorithmWatch ran 4,480 election-related searches through Google and found the company generates an AI Overview on 39.1% of them, against 65.3% for non-political queries — and will not say what makes the summary appear. When it does appear, the sources are narrow: nearly half of all links came from ten domains, and the single most-cited was Google's own YouTube. The queries covered three eastern German contests: the 2026 state parliament elections in Saxony-Anhalt and Mecklenburg-Western Pomerania, and the Berlin House of Representatives election. The group got the data through the Google Search Researcher Result API under research access granted by Article 40(12) of the EU Digital Services Act.
The trigger rate moves sharply by topic. Questions about polls almost never produced an overview. Queries about parties and individual politicians produced one roughly 40 to 50% of the time. Searches about the far-right Alternative for Germany returned an overview in 24% of cases, while other parties sat between 45% and 58%. Saxony-Anhalt drew fewer overviews overall than Berlin or Mecklenburg-Western Pomerania.
A Google spokesperson told AlgorithmWatch that AI search features are meant to be neutral and objective and are not obliged to appear on every query, and that a summary is shown when the company expects it to give the user additional value. Asked which criteria apply in practice, Google did not answer.
This is a retreat from a position the company took two years ago. In 2024, Google said it would block Gemini from answering election-related questions, a measure announced around the US and Indian elections, and AlgorithmWatch notes that it worked at the time. Whether an equivalent restriction is in force in 2026, the researchers could not establish. Google declined to comment on that too.
The source concentration is the part that should worry anyone who works in German media. Beyond YouTube — which AlgorithmWatch reads as self-preferencing and Google denies is anything of the kind — the most frequent outlets were the public broadcasters NDR, rbb, MDR and ARD, plus the news site Zeit Online. Six media domains accounted for 75% of all links to media. Wikipedia, government sites and party pages also turned up in the summaries, and party pages carried no marking to separate them from sources offering neutral information.
The framing findings are the ones that will travel. Political positions were often rendered through vague positive descriptors — "pragmatic", "close to citizens", "stable" — language that echoed the parties' own self-descriptions. In the Saxony-Anhalt sample, 81.5% of overviews about the center-right CDU were explicitly positive. For the center-left SPD it was 50%, for the Greens 10.7%, and for Alternative for Germany 0%. The more party-linked domains appeared among the sources, the more positively the political force was generally described, though the same relationship did not hold for every party.
Individual politicians got treated inconsistently. Ulrich Siegmund, the Alternative for Germany candidate in Saxony-Anhalt, was described by the overviews with varying degrees of sharpness: as a representative of the far-right and national-conservative wing, as a politician with radical far-right views, or as a right-wing populist. AlgorithmWatch found no pattern explaining the contradictions, and some of the variants appeared within short intervals of each other.
The 81.5%-to-0% spread is the number that will get quoted, and it is also the one to handle most carefully. It does not by itself show a system tilted against the Alternative for Germany; the correlation AlgorithmWatch reports between party-linked sources and positive descriptions points somewhere less dramatic and more damaging — that the overview is reading a party's own web presence back to the user in the party's own adjectives, and that parties with more of it get more of it. The unlabeled party sources are the real finding here. A summary that repeats campaign copy without marking it as campaign copy launders it into what a reader takes for a neutral answer, and no amount of accuracy testing catches that.
Which is the other thing this study quietly changes. Unlike earlier work, this time the researchers rarely found outright factual errors, and they attribute the flattering language to the tendency of language models to agree with the user rather than to any deliberate slant. Every election-AI safeguard built so far, including Google's own 2024 switch-off, is aimed at hallucinated facts: wrong dates, wrong polling places, wrong results. The failure mode here is tone, and tone survives every benchmark the industry runs.
AlgorithmWatch is calling for transparency requirements and clear rules on liability. The starting point is smaller than that and already on the table. Google was asked which criteria decide when an election summary appears and did not answer; asked whether its 2024 restriction still stands, it declined to comment. Two refusals, and 39.1% of election queries got a summary regardless.