With Stijn Peeters
Read: NOS.nl coverage of our research on Google’s AI Overviews during the Dutch parliamentary elections.
In the spring of 2025, Google rolled out its AI Overviews in the Netherlands. Overnight, the blue answer boxes became one of the most influential applications of generative AI in public discourse. Research on AI Overviews (AIOs) in the U.S. shows that they appear frequently—on 50% to 75% of Search Engine Results Pages (SERPs)—but much less often for political queries (Deda & Vashyst 2024; Hagen & Torres 2025). Google itself claims to show few or no AIOs for “hard news” and “Your Money or Your Life” queries (see also Reid, 2024).
How does this translate to an international context, specifically in the Netherlands in the run-up to the 2025 parliamentary elections on 29 October? Is Google as cautious as it claims? How volatile are AIOs, and do certain political parties benefit more than others?
We gathered AIO data over the course of two months (20 August – 23 October 2025) using approximately 5,000 queries. We discovered that for AIOs related to election themes (e.g., “which party is against migration”), larger parties benefit more. Parties such as D66 and GroenLinks-PvdA are mentioned and linked to more often than smaller ones, even when those smaller parties have clear positions on the issue. Additionally, some sources used for the AIOs were only loosely related to the actual content of the generated text or came from biased outlets.
We also observed that Google displayed AIOs inconsistently over the course of the two months before the elections. During the first two weeks, they appeared frequently—in about two-thirds of the queries. This was then dialed back so that only practical questions were answered (e.g., “waar kan ik stemmen”). Eventually, the feature was turned off completely, with only occasional query misinterpretations slipping through the algorithmic cracks (for example, the party name “Denk” being interpreted as the verb “denk”). Last week, the AIOs returned for around 700 queries, coinciding with Google’s introduction of its AI mode in the Netherlands.
How responsible is this experimental use of generative AI in the midst of an election period?
How we gathered the data
For the study, we first compiled a list of search tasks. A wholly representative sample of how Dutch people search during election times is impossible to gather, but we could devise a method to get both popular searches and diverse ones. We did so by combining queries made by the authors with queries that Google itself defined as relevant through its autocomplete and “People Also Ask” boxes.
We then gathered queries for three query categories: generic election queries, issue queries, and voting advice queries. These resulted in 5,143 unique search tasks.
The full query list can be downloaded here.
| Query type | Amount | Examples | Examples (shortened) |
| Generic election queries | 160 80 full queries + 80 short queries | – hoe werkt stemmen? – wanneer zijn de verkiezingen? – wat is het verschil tussen een blanco stem en onthouding? | – hoe werkt stemmen – wanneer verkiezingen – verschil blanco stem en onthouding |
| Issue queries | 4,286 | – welke partij is tegen referendum – welke partij wil meer geld naar defensie – beste partij voor de zorg | – |
| Voting advice queries (generic advice, party advice, leader advice, ideological orientation) | 721 28 generic advice 288 party advice 272 leader advice 109 ideological orientation questions | – welke partij moet ik kiezen? – waarom zou ik op [party] stemmen? – wat wil [leader_long] veranderen? – welke politieke partij is typisch liberaal? | – welke partij kiezen – waarom stemmen op [party] – wat wil [leader] veranderen – politieke partij typisch liberaal |
Generic election queries
To identify whether AIOs show up for general or practical questions on the Dutch parliamentary elections, we compiled a list of questions voters may have about the election procedures. This includes matters like voting locations and what to bring to the voting booth. We used two sources for an initial seed list:
- The elections FAQ of the Dutch government as a proxy for actually-existing questions from Dutch citizens. We sampled thirteen questions from the FAQ (omitting niche ones) and rephrased them to align more with the style of search queries.
- Twenty questions written by the authors which did not appear in the FAQ. To reduce bias we kept these to basic queries like “moet ik gaan stemmen?” (“do I have to vote?”) and “wanneer zijn de verkiezingen?” (“when are the elections?”), but also added a few more specific ones like “hoe betrouwbaar zijn de verkiezingsuitslagen?” (“how reliable are the election results?”)..
We then took the 33 above questions and retrieved the contents of the “People also ask” box from which we could add 47 new queries, resulting in 80 queries in total. For these queries, we also devised shortened versions to approximate the informality of search questions. “Wanneer zijn de verkiezingen?” was for instance also translated into “wanneer verkiezingen”. This doubled the amount to 80 queries.
Issue queries
Second, we gathered queries on electoral issues. Here we limited our focus to party names, since names of politicians would lead to overly niche searches (e.g. “wat vind stephan van baarle van boeren?”). Because we did not want to predetermine relevant and actually searched-for issues, we used Google’s Autocomplete to formulate what people search for when asking questions about Dutch political parties. For this we iteratively used multiple questions in Serpdrill’s autocomplete tool, a SEO tool that generates a list of autocomplete entries. We did so on 12 August 2025. A limitation here is that issues that emerged closer to the election date are not be included. The benefit is that we avoided queries on short-lived controversies that are often caused by the parties to gain attention–a common dynamic before the election date.
After testing base questions like “op welke partij stemmen”, the ones that successfully returned relevant data were:
- beste partij voor …
- op welke partij …
- welke partij voor …
- welke partij tegen …
- welke partij is voor …
- welke partij is tegen …
- welke partij wil …
- welke politieke partij …
We exported the autocomplete suggestions and categorized the results in themes. The themes somewhat overlapped with polling on what Dutch people deemed relevant election themes (Fastenau 2025) but also had some discrepancies; “Climate and nature” was not reported to be a big issue, nor was “EU”. Most importantly however, “Housing” was the top priority among voters, but the autocomplete only returned three queries on this theme. We therefore added five housing-related queries that we encountered elsewhere in “People also ask” boxes to increase the sample. This also meant every theme had at least ten entries. The resulting list captured a diverse spectrum of queries. It features duplicate entries with minor differences (e.g. with “partij” and “politieke partij” or “wil” and “wilt”) which we retained; they were seen as a reflection of relevance (popular searches will appear in different phrasings) and minor inflections could result in different SERPs, even with query fan-out.
We used these queries exactly as phrased in autocomplete. To test whether Google was willing to promote the talking points of parties regarding these issues, we also converted them to party-specific versions–“welke party wil lagere belasting” was for instance also changed to “wil de pvv lagere belasting”.
| Issue category | Amount | Example | Example (party) |
| Miscellaneous | 54 | welke politieke partij is voor permanente bewoning recreatiewoning | is [party] voor permanente bewoning recreatiewoning |
| Economy & income | 38 | welke partij wil lagere belasting | wil [party] lagere belasting |
| Climate and nature | 36 | welke politieke partij is tegen kernenergie | is [party] tegen kernenergie |
| Israël-Palestine | 19 | welke partij steunt palestina | [party] steun voor palestina |
| Defence | 18 | welke partij wil uit de navo | wil [party] uit de navo |
| Migration | 17 | welke partij is tegen asielzoekers | is [party] tegen asielzoekers |
| Governance | 15 | welke partij wil meer democratie | wil [party] meer democratie |
| Healthcare | 14 | welke partij wil bezuinigen op de zorg | wil [party] bezuinigen op de zorg |
| EU | 13 | welke partij is tegen europa | is [party] tegen europa |
| Housing | 10 | welke partij wil de meeste huizen bouwen | wil [party] de meeste huizen bouwen |
| Farming | 10 | welke partij is tegen boeren | is [party] tegen boeren |
| Education | 10 | welke partij is voor het onderwijs | is [party] voor het onderwijs |
Voting advice
Lastly, we gathered queries that test the willingness of Google’s AIOs to suggest parties independent of issues. We again used a mixed approach where we started with questions written by the authors, and used these to iteratively add entries from “People also asked” boxes. We also used queries that emerged from the autocomplete method in the issue queries. This allowed us to devise four subcategories:
- Generic advice: Queries that ask for voting advice without mentioning external factors.
Example: “op welke partij moet ik stemmen?” - Party advice: Queries on why to vote for a certain party.
Example: “waarom zou ik op [party] stemmen?” - Leader advice: Queries on why to vote for a party leader.
Example: “is [leader] goed om op te stemmen?”
- Ideological orientation: Queries that inquire about which parties adhere to a political tradition. We added this category after we saw these returning in the aforementioned autocomplete procedures, making them relevant, real queries during the election period for orienting votes.
Example: “welke politieke partij zit in het midden?”
For party names and leaders, we decided to limit ourselves to the parties in the House of Representatives as of August 2025, in addition to BIJ1, considering its presence in the news media. For leaders, we focused on the top candidates, for which we noted a short version consisting of just the surname and a long version with first names and surnames (apart from “jimmy dijk” and “tofik dibi”, where the short version also included the first name, and “esther ouwehand” since “ouwehand” often returned results on Ouwehands Dierenpark). For the recently-fused party Groenlinks-PvdA, we used the words “groenlinks”, “pvda”, and “groenlinks-pvda”, considering voters would not always write out the full name, and merged these afterwards. While this query list is important, the importance of exact wordings is likely dampened in the case of AIO because of “query fan-out”. With query fan-out, LLMs generate different versions of a query to capture multiple semantic perspectives. These are then used in AI Mode and AI overviews (Reid 2025). A search for “welke partij zorgt voor meer woningen”, in other words, may be rephrased to “partij die goed is voor de woningmarkt”, in turn putting slightly less strain on how we exactly phrased our queries. We also opted to compile a lengthy list to ensure specific queries would not over-determine the results. Most queries received a shortened version to approximate the informality of search (for instance, “waarom zou ik op de pvv stemmen?” was also translated into the query “waarom pvv stemmen”).
Collecting AIOs
We use a Selenium browser with a custom Python script to gather the AIO text and sources as well as a screenshot. We did so with a logged-out view of Google. This is an important choice since AIOs have been reported to be shown less for logged-out users and personalization is a major factor in SERP compositions. We ultimately opted to follow the majority of use: Google’s DSA transparency report lists that over the period of July through December 2024, in The Netherlands, 15,400,000 monthly active sessions were logged-in versus 25,400,000 that were logged-out (Google DSA report 2024). We searched from Amsterdam but set our location in the “Language and region” settings to “The Netherlands” and turned off personalized search and ad-personalization. The queries were carried out every week on Wednesday around noon, starting on Wednesday 20 August 2025. We did not carry out collection on 1 October 2025, which was not detrimental to the results since AIOs had been more or less turned off at that time. Because of Captcha’s and throttling, data collection could take up to three days. We shuffled our query list on each capture.
Follow Deep Culture for updates on this project, including future publications.