Google Showed AI Overviews on 39.1% of Election Queries and 65.3% of Others

What the 4,480 queries actually measure

The query set is 200 election-related and 100 non-political templates, run across three state elections, collected at the end of July 2026. Parties in scope were the ones polling above the threshold to enter at least one parliament at collection time: AfD, CDU, Bündnis 90/Die Grünen, Linke and SPD. FDP, BSW and smaller parties were left out. Candidate queries were restricted to each party’s lead candidate per state.

Most queries were run twice, once bare and once with an election suffix (“Landtagswahl” for Saxony-Anhalt and Mecklenburg-Vorpommern, “Abgeordnetenhauswahl” for Berlin). That is where 300 templates become 4,480 recorded queries.

Figure	What it is a percentage of	The paper’s own sentence
39.1%	queries about the 2026 state elections that returned an AI Overview	”Bei Suchanfragen zu den Landtagswahlen im Jahr 2026 zeigte Google in 39,1% der Fälle eine KI-Übersicht”
65.3%	non-political queries that returned an AI Overview	”während es bei nicht-politischen Anfragen in 65,3% der Fälle so war”
about 26 points	the difference between those two rates	”ein Abstand von etwa 26 Prozentpunkten”
24%	AfD-related queries that returned an AI Overview	”24% bei der AfD, 45 bis 58% bei anderen Parteien”
45% to 58%	queries about the other parties that returned an AI Overview	same sentence

The AfD figure holds across all three states and across different query forms, which the paper states directly: “Dies gilt für alle drei Bundesländer sowie für unterschiedliche Suchanfragen.” Consistency across three independent elections is a stronger result than a single large gap in one place, and it is the finding here that rests on the full dataset.

What the study does not do is explain any of it. There is no mechanism attached, no claim about intent, and no way to separate a deliberate election guardrail from a quality threshold that election queries happen to fail more often. The paper documents a pattern. It does not identify a cause, and neither does this post.

Google’s only on-record answer, and its exact boundary

AlgorithmWatch asked. Google replied by email on 21 August 2026, and the paper quotes the substance: Google confirms that AI Overviews are not shown on every search, but only when the company believes they offer users added value and the quality of the answers is assured.

Read that carefully, because it is doing less than it appears to. It confirms the existence of a threshold. It does not disclose what the threshold is, how “value” is measured, whether election content is treated as a distinct category, or whether the 39.1% figure reflects a guardrail or a side effect. A company can answer a transparency question truthfully and still leave every operational detail unanswered. That is what happened here.

The paper says the same thing about the wider picture: “Es ist unklar, wann Google eine KI-Übersicht anzeigt.” It is unclear when Google shows an AI Overview. That sentence is the section heading for the strongest result in the study.

Where the citations come from

Two findings on sourcing, both from the full dataset.

Nearly half of all links displayed came from just ten providers: “Fast die Hälfte aller angezeigten Links stammten von nur zehn verschiedenen Anbietern.” For a surface that presents itself as a synthesis of the web, that concentration is the finding practitioners should sit with.

And for candidate queries specifically, the single most-cited source is Google’s own: “Die konzerneigene Plattform YouTube ist das meistzitierte Angebot in KI-Übersichten bei Suchanfragen über Kandidat*innen.” The paper reads this as evidence that Google favours its own platform, and notes in the same breath that Google denies it: “was das Unternehmen auf Nachfrage bestreitet.” Alongside the social platforms, the most-cited domains are media sites — public broadcasters above all — plus Wikipedia, government pages and party pages.

The number everyone will quote comes from 115 overviews, not 4,480

Here is the part the coverage will flatten. The study also hand-coded the language AI Overviews used about candidates, and produced this:

Party	AI Overviews coded	At least one party-owned domain cited	Share with a party domain	Coded as clearly positive
AfD	30	6	20.0%	0.0%
CDU	27	27	100.0%	81.5%
Grüne	28	27	96.4%	10.7%
SPD	26	18	69.2%	50.0%

The paper’s own summary of it: while the share of especially positive or praising AI Overviews sits at 81.5% for the CDU, for the AfD it is zero.

That is a striking table, and there are four reasons it does not carry the same weight as the 39.1% figure. All four come from the paper itself, not from us.

First, sample size. The whole table rests on 115 AI Overviews: “Die Stichprobe umfasste 115 KI-Übersichten.” Not 4,480. The row totals above add to 111 across four parties.

Second, one state. The coding was restricted to lead candidates in Saxony-Anhalt. Saxony-Anhalt was picked at random, which is fine — but it is one of three states, and the paper elsewhere reports that Saxony-Anhalt returned noticeably fewer AI Overviews than Berlin or Mecklenburg-Vorpommern in the first place.

Third, and this is the one that matters most, the question subset was not chosen at random. The paper’s own words: “Das Subset an Fragen wurde gewählt, da wir beim ersten Lesen dort besonders viele lobende Formulierungen beobachtet hatten.” The subset of questions was chosen because on a first read we had observed particularly many praising formulations there. A sample selected because it looked interesting cannot then be used to estimate how common the interesting thing is. The researchers disclose this in the methodology section rather than burying it, which is to their credit, and it still means the 81.5% is an illustration and not a rate.

Fourth, a confound the authors flag in a footnote: Berlin’s CDU replaced its lead candidate shortly before the collection period, so they recorded both the old and the new candidate.

The paper is also careful about what the table implies. It concludes that a higher share of party-affiliated domains goes together with a stronger tendency toward positive, partisan phrasing, but not equally for all parties — and then explicitly leaves open whether the choice of domains actually determines the wording.

Who paid for it

The study was funded through the project Auditing Algorithms for Systemic Risks by the Alfred Landecker Foundation, with additional support from Luminate. Neither is Google, and neither sells an AI-visibility product whose value depends on the result.

This is worth a sentence because most of the AI-citation numbers in circulation this year come from companies selling AI-citation tools. That does not make vendor data wrong. It does mean the funding line is part of the methodology, and it is one of the few parts you can check in five seconds.

What we would want before treating any of this as settled

• A replication outside Germany. Three states in one country, one election cycle, one language.
• The coded sentiment analysis re-run on a random sample rather than a subset selected for looking notable, so the 81.5% becomes a rate instead of an example.
• A pre-registered version of the display-rate test, so the query set is fixed before anyone sees which way the gap runs.
• Any answer from Google more specific than “value” and “quality”.

The study’s own cover language for itself is the right register for all of it. It calls its findings first observations from an exploratory study. Coverage that keeps saying so is coverage that is reporting the study rather than mining it.

The 39.1% against 65.3% is real, sits on 4,480 queries, and was collected through a legal access right rather than a scrape. The 81.5% is real too, and sits on 115 overviews from one state that were pulled precisely because someone noticed them. Both numbers came out of the same PDF on the same day. Only one of them is a finding.