The direct democracy paradox: Microtargeting and issue ownership in Swiss online political ads

This dataset contains the raw and annotated data supporting the findings of the research paper: "The direct democracy paradox: Microtargeting and issue ownership in Swiss online political ads." Switzerland's system of direct democracy, with its frequent popular referenda, offers a unique context for studying online political advertising. This study presents a large-scale analysis of approximately 40,000 political ads published on Facebook and Instagram in Switzerland between 2021 and 2025. The research examines demographic microtargeting, partisan strategies in referenda and. federal elections. Data Files The repository includes the following four files: ads.zip: A compressed archive containing the raw text and metadata of the collected advertisements. The data is organized into three sub-folders by language: German (DE), French (FR), and Italian (IT). federal_elections_authors_annotation.csv: A manually curated list of authors who published at least one ad in the 30 days preceding the 2023 Swiss federal elections. Each author is labeled with: Political affiliation. Relevance to the federal election. Entity type (Party, NGO, Newspaper, Committee, Union, or Association). Specific linkage to a political party or individual politician. federal_gpt_answ.json: Results of an automated annotation process using GPT-4o for ads published within the 30-day window of the federal election. It identifies whether an ad is election-relevant and provides up to three thematic keywords describing the ad's topic. referendum_gpt_answ.json: Annotated data for ads published within 30 days of the 42 referenda held during the study period. This file identifies the ad's relevance to a specific referendum and its stance ("Yes", "No", or "Neutral") as determined by GPT-4o. referendums_topic.json: Official dataset containing the title, date, and description for each referendum. refimp_with_results_and_participation.csv: Official dataset detailing the results and voter turnout for each referendum. topic_mapping.json: A classification of referendums into 12 primary categories, clustered using GPT-4o. The data collection was initiated via the Meta Ad Library API using election-related keywords in three national languages. For full transparency, manual label verification on a random sample of 100 ads confirmed a 100% accuracy rate for referendum stance and 96% for federal election topic related keywords. Terms of Use & Attribution If you use this dataset or the associated code for your research or data projects, you are required to cite the following paper: Capozzi, A. (2025). The direct democracy paradox: Microtargeting and issue ownership in Swiss online political ads. arXiv preprint arXiv:2512.14564. Additionally, please cite this Zenodo repository using the DOI provided on this page.