Data donations is a promising new method for accessing digital traces that has recently emerged and is facilitated by the 2018 GDPR in the EU and United Kingdom. The GDPR grants individuals the right to access their collected data and the right to data portability.²⁰ In recent years, various jurisdictions worldwide have introduced legislation that grants similar rights.²¹ As a result of this legislation, all data processing entities are required to provide their data subjects a digital copy of their personal data upon request. This legislation applies to large online platforms such as Facebook, TikTok, YouTube, and Google, but also all VoD services. These digital copies typically come in the form of .zip files containing various file types, such as .json, .csv, or .txt files, and potentially also image and video files depending on the nature of the platform, which are commonly referred to as Data Download Packages (DDPs). While data donations face challenges with participation rates, it offers key advantages, such as the ability to collect data retrospectively and across multiple devices.²²

How VoD services handle data access requests in practice varies substantially. Some platforms have fully automated data request procedures for their users, as is the case for global VoDs such as Netflix and Amazon Prime. With a single click of a request button, users receive a file in their inbox within 30 days, containing a copy of their personal data collected by the processor. In contrast, other VoD services require that their users send a letter or an e-mail to their privacy officer. We typically see this for VoDs that serve a local market, such as NPO Start or Videoland in the Netherlands. This is likely due to their scale, resulting in less need to invest in an automating process. While data donations are exciting in theory, in practice, platforms do not always fully comply with data access legislation. Moreover, there are sometimes issues such as unclear data access procedures, incomplete DDPs, and ambiguity regarding what the data in the DDPs exactly represent.²³

Since data donation relies on legislation that only exists since 2018, it is a relatively new approach for data collection. First studies that made use of data donation asked participants to share their complete DDPs. An example is Project AWeSome, which collected complete Instagram DDPs to investigate the relationship between adolescent well-being and social media use.²⁴ Another example is a study which examined musical preferences by analyzing data donated by Facebook users. However, an important challenge was that the DDPs collected contained a lot of sensitive data, not only of the research participants but also of their contacts who did not give consent. To address this issue, data donation studies were designed so that participants either visited a research facility, where they immediately applied a de-identification procedure to the Facebook DDPs they received, or participated online, where the de-identification process was carried out immediately after receiving the DDPs.²⁵

The presence of sensitive data in DDPs poses significant challenges. First, participants may be hesitant to share this information. Second, not all data contained in a DDP is typically needed to answer a specific research question. Sensitive data can also be present in DDPs from VoD services. This includes direct identifiers such as names, email addresses, and IP addresses. Additionally, user profiles may indirectly disclose household or family structures, while viewing histories and preferences could reveal special category information such as sexual orientation, religion, or political preferences. To tackle issues of sensitive data, Boeschoten et al. developed an alternative workflow.²⁶ It consists of the following steps for a research participant:

The participant requests their personal DDP at the platform that is of interest to the research project.

Multiple software tools exist that facilitate the conduct of a data donation study. For instance, the tool OSD2F²⁷ allows for the donation of .json files while locally removing various identifiers. It has been used to collect Facebook, Instagram, Twitter/X, and YouTube data to understand digital news use,²⁸ and in a study that collected Google Browsing, Searching, and YouTube histories to understand media consumption behavior.²⁹ Alternatively, the tool Designerly Data Donation³⁰ focuses on data visualization and was used for data collection from menstrual tracking apps³¹ and Google Assistant.³² Both tools focus on understanding how users perceive the intimacy of such data. The Data Donation Module³³ tool, in turn, facilitates a graphical user interface to prepare data donation studies and has been used for example, for collecting Google search histories to understand what participants search in relation to upcoming political referenda.³⁴ Additionally, the tool Port offers researchers the flexibility to design studies that collect various types of data from any platform that complies with GDPR.³⁵ Port has supported several studies, including the following:

As discussed shortly, Port is also what we used for our Netflix study.

Importantly, while people can donate their data, they are not always willing to do so. Numerous studies have focused on what types of participants are inclined and also able to successfully donate their data for research. For example, a study that collected WhatsApp data donations in the LISS panel, a probability-based panel in the Netherlands, found that younger and higher educated users are more likely to donate.³⁹ Similar conclusions were drawn by a vignette study on hypothetical willingness of data donation in the probability-based German Internet Panel.⁴⁰ In addition, a study with participants from an online panel company concluded that participants with higher digital literacy skills were more likely to donate in a study with participants from an online panel company.⁴¹

Data donation is a user-centric approach to collect digital trace data,⁴² meaning that users need to be actively recruited to participate and guided through the process to share their personal digital traces. In practice, participant recruitment and guidance can be organized in different ways. In recent data donation studies, two main approaches can be distinguished: a fully online approach or a more personalized approach. In fully online approaches, participants receive an invitation to an online system for data donations and are guided through the process via instructions in their web browser. This approach must accommodate a diverse range of participants with varying backgrounds, device types, operating systems, language settings, and levels of familiarity with the service being studied.

For studies with a more qualitative focus and/or where researchers have the resources to provide direct support, a personalized approach can be beneficial. Setting up a physical lab environment, conducting on-site visits (e.g., at schools), or offering individualized assistance (e.g., through Zoom meetings) allows researchers to explain the procedure and to provide real-time troubleshooting. In such cases, it is common practice to compensate participants for their time, which is often done by monetary reimbursements.

When the goal of the research is to make inferences about a certain population or conduct experiments, researchers typically aim to collect data donations from over one hundred participants. In such cases, they often collaborate with panels such as the LISS panel⁴³ or commercial panels like Ipsos I&O. Alternatively, opt-in panels such as Prolific can be used. The costs associated with different panels and the incentives offered to participants can vary significantly due to cultural differences. Which approach is most suitable depends on the purpose of the study, the financial means, and capacity in terms of support staff.

In 2013, Netflix commissioned an online survey by Harris Interactive, involving nearly 1,500 TV streamers. The survey found that binge-watching was common, with 61 percent of respondents reporting that they regularly binge-watch, defined in this context as viewing two–six episodes of the same show in one sitting.⁴⁹ Netflix subsequently framed binge-watching as “the new normal.”⁵⁰ Wayne observes that Netflix frequently promotes binge-viewing as a preferred mode of audience behavior, positioning it as an improvement over traditional television’s live broadcasts and linear scheduling.⁵¹ But how common is binge-watching in practice?⁵²

This is where data donations become useful. Yet even then, since there is no consistent definition of binge-watching, the answer remains far from straightforward.⁵³ Pierce-Grove notes how discussions about binge-watching frequently highlight the number of episodes viewed rather than the total duration (e.g., minutes or hours) spent watching.⁵⁴ It has been found that most definitions focus on the number of episodes watched in a single sitting.⁵⁵ For our analysis, we stipulated that for a session to be considered a binge, it must

Table 1 presents our analysis of binge-watching based on the data donations, showing the proportion of respondents who binge-watched (or not) at some point during the data collection period. Here, we can see that 92 percent of participants would qualify as binge-viewers if we define binge-viewing as watching two complete episodes within a single session. The results also show how adapting that threshold changes the proportion of the population that are classified as binge-watchers. Harking back to Netflix’s 2013 commissioned survey, we find that, over a decade later, actually more people seem to engage in binge-watching. Even with a threshold of five episodes, more than half of the users have engaged in such viewing activity. When looking at the percentage of total sessions, however, there is a sharp decline as the threshold increases, indicating that while binge-watching is popular (number of users) and rather common (in terms of frequency), “extreme” binge-watching is relatively rare and certainly not the norm. In the end, most sessions constitute non-binge-watching behavior (over at least 67 percent).

The concept of popularity has long been central to television analysis in cultural studies, yet researching the cultural power of VoD services and their content is challenging due to the lack of available consumption data.⁵⁷ As Wayne states:

whatever shared sense of collective audience-hood was discursively produced by widely accepted albeit thoroughly flawed ratings systems like Nielsen is being replaced by the industrial discourses associated with black-box audience data within which claims of popularity and cultural significance cannot be substantively challenged.⁵⁸

Compounding this problem, Netflix’s method for calculating a “view” has changed over the years, which has had implications for how popularity is understood and what content is popular. Initially, viewing a title meant completing 70 percent of its total runtime. However, in 2020, the company changed its definition: A view was now considered as watching just 2 minutes of a title. This decision conveniently increased their average viewership by 35 percent compared to the previous metric.⁵⁹ In 2023, the definition of a view changed once again. It was calculated as the total hours viewed divided by the total runtime. According to Netflix, this method “offer[ed] a more easily comprehended number to represent popularity and also helps level the playing field for shows and films with shorter running times.”⁶⁰ However, it also includes repeat views, which inflates viewership figures. As Mittell explains, Netflix profits more from perceived popularity, linked to attracting subscribers, than from actual viewer numbers.⁶¹

To determine what TV shows were popular among Netflix viewers in the Netherlands in 2023, we first identified the number of unique titles and calculated how many unique users watched these. In this way, it becomes possible to detect TV shows that seem to be watched by a relatively large share of unique users and what titles receive less attention. In total, users engaged with 1,196 different TV shows. This also includes interactions where users watched only a fraction of an episode. When considering only interactions where viewers completed at least one episode, this number drops to 779 unique titles, accounting for 65.2 percent of all TV shows. This number suggests that in circa 35 percent of cases, users started episodes but did not complete them, likely reflecting browsing behavior. When focusing on TV shows where users watched complete episodes, the analysis reveals a typical long-tail distribution (Figure 2).

Table 2 highlights the top 10 TV shows, ranked by the number of unique users who watched full episodes. The Night Agent emerges as the most-watched show of 2023, with forty-five unique viewers, representing 37 percent of all users in the dataset. Other top titles such as Lupin or Sex Education attract around 25 percent of unique users. However, only seven shows (0.8 percent of all unique titles) attract at least 20 percent of all users. Expanding the lens to the top twenty titles reveals a sharp drop-off, with many TV shows watched by fewer than 20 percent of all users. So, while the top twenty shows represent a large portion of our participants’ attention, there are signs of fragmentation in viewership. Still, 113 unique users (92 percent) watched one of the top twenty shows.

Examining the long-tail, we find that 86 percent of all unique TV shows have fewer than five unique viewers, while over 52 percent are watched by just a single unique viewer. This suggests that most TV shows are consumed by very small groups of unique users or even individual viewers. However, engagement with titles in the long-tail is not limited to just a few unique users. On the contrary, 91 percent of all unique users watch shows in the long-tail. This indicates that most users also watch TV shows that are not widely popular based on audience share. In other words, many users watch both popular and “niche” content.

These observations are further supported when visualizing titles and users in network graphs (Figures 3–5). Each circle (node) represents a show title and a unique user. The larger a circle, the more users watch that title (indicated by an edge). Only show title nodes can differ in size based on how many users are connected to it. For example, a title such as The Night Agent is noticeably larger than others, due to the relatively high number of unique users who watched it. Accordingly, users are connected through the shows that they watch. Figure 3 shows the network graph for all shows, regardless of how many users watch them (but at least one). There is a dense cluster in the center, with several shows as large nodes, underscoring their prominence. However, the graph also indicates that many shows are watched by few or even only single users at the periphery of the graph (Figure 4). The colors in the visualization mark communities, clustered based on similar content consumption, identified via the modularity class algorithm in Gephi. TV shows that serve as a link between multiple user clusters can appear in multiple communities.

The size and complexity of the network graph change noticeably when the threshold for including show titles–based on the number of unique users watching them–is increased. For example, the number of nodes for show titles drops from 779 to 239 (30 percent) when including only shows that have at least five unique users who watched them. It drops down even further to 105 show titles (13 percent) with a threshold of ten unique users (Figure 5).

These exploratory analyses illustrate potential applications of similar methods to identify and investigate matters such as taste communities, content popularity, or diversity of content consumption.

In the EU, the Audiovisual Media Services Directive (AVMSD) aims to promote cultural diversity by requiring global VoD services such as Netflix to include at least 30 percent European content in their catalogs and ensure these titles are prominently featured. While the quota system is intended to support local productions, there have been concerns that these content quotas lead to “quota quickies”–low-quality filler content that fails to attract viewers.⁶² Moreover, local content may be in the interest of these services. Theories of cultural proximity, originally developed to explain transnational television consumption, suggest that audiences tend to prefer content that is familiar to them in terms of language, cultural references, and ethnic appearance.⁶³ However, whether this preference still applies to VoD services is unclear.⁶⁴ In the meanwhile, Netflix uses its branding of diversity to appeal to transnational audiences.⁶⁵ As Ted Sarandos has repeatedly stated in interviews: “We’re entering into a new era now where content and great stories can come from almost anywhere in the world.”⁶⁶

While numerous studies have examined Netflix in terms of catalog composition and prominence/discoverability of European titles, a key question remains: Are users watching European content? The earlier mentioned 2022–2023 report by the European Audiovisual Observatory gives some useful clues.⁶⁷ Based on survey data and covering a selection of titles, their findings suggest that 62 percent of television viewing time in the Netherlands is spent on US productions, followed by 11 percent on UK content. Additionally, it observed that the share of EU-produced TV content in viewing time (29 percent) roughly aligned with its share in the catalog (33 percent). Data donations provide valuable insights into the role of European content while also offering a broad perspective on consumption diversity.

Our dataset included 1,170 unique TV shows from 104 different production countries. Notably, 46 percent of all titles were produced–or co-produced–by the United States (Table 3). With considerable distance, this is followed by the United Kingdom (11 percent) and South Korea (5.3 percent). Content from the Netherlands accounts for only 2 percent of all titles watched by users in 2023.

These findings suggest that anglophone productions also dominate the consumed content catalog in our dataset. However, when clustering the production countries into regions, 31 percent of all titles have a European origin (Figure 6).⁶⁸ Asian productions are in the third place, accounting for circa 10 percent of all content and spearheaded by South Korea and Japan. Productions from the Middle East and Africa are only a fraction of the catalog (1.5 percent and 0.9 percent, respectively).

While the catalog of consumed shows seems biased toward productions from North America, further exploring what TV titles users interact with implies that underrepresentation of production countries does not mean less engagement per se. Table 4 shows the number of unique users watching content from different countries, in this case the top 10 based on user counts.

While anglophone titles appear very popular, a considerable proportion of unique users also engage with Dutch content: 55 percent of them check Dutch titles at least once, with 36 percent watching at least one full episode of Dutch TV shows. So, while Dutch titles are only a small fraction of the catalog, they attract over half of all unique users in the dataset. The visible presence of South Korean and Spanish content in the top 10 shows that Dutch audiences are open to non-anglophone titles, aligning with trends seen in the success of K-dramas and Spanish-language series.

Taken together, these findings indicate that anglophone productions dominate the consumed TV show catalog on Netflix in the Netherlands–at least among our participants. However, this does not mean that quantitatively less represented local and regional productions are irrelevant, as engagement with them remains significant. While concerns of US dominance and cultural imperialism are not unwarranted, a substantial amount of content from other countries is also being consumed. While we don’t know if and how these productions are featured on the platform, users are discovering and engaging with them.