Capturing the Form of the Catalogs while Browsing through Their Pages

One facet of my interest in the cinephile catalogs led me to study how their websites were structured, how long it took to reach the movies, and what kinds of visual and textual elements were used to build what Lobato et al. aptly call the VoD interface.¹³ Taking a descriptive and qualitative approach that was also sensitive to implicit aspects, I studied the blocks of content presented on a given page and the intended or accidental effects of their selection and arrangement. I drew inspiration from Luc Pauwels’ multimodal framework for the analysis of websites.¹⁴ The study of SVoD catalogs must also consider two distinct types of circulation: how users navigate the catalogs and how data (and films as digital copies) circulate through those websites. For the former, I drew on Elisabetta Adami’s work on the preexisting notion of pathways seeking to consider the different paths and interactive features the editorial teams of those catalogs provided to their subscribers.¹⁵ I also worked to unpack how content retrieval and salience algorithms work, highlighting how part of users’ browsing was modulated by machinic interventions.¹⁶ For the latter, I studied not only the selection of movies presented by each catalog (more on this below) but also how content was aggregated around the movies. In addition to movies to watch, catalogs offer visual and textual content that literature and fan studies call paratexts to highlight how they introduce and complement the movies.

Cinephile catalogs help extend and perpetuate the cinema experience by producing cinephile and programming content, such as short movie reviews, periodic newsletters, in-depth essays, video introductions by famous critics, or clips from associated film archives.¹⁷ This part of their work sets them apart from other subscription video services, such as those linked with linear TV broadcasters or even pure-play streamers, such as Netflix. Although paratexts can be produced in-house, they are often taken from other online and offline sources and then reformatted, transformed, and adapted to fit a new editorial use.¹⁸ Retracing their circulation and assessing their relative novelty seemed a good way then to approach the added curating work done by catalogs. I started to follow the journey of the short synopsis, teaser paragraphs, and even blog posts featured through the catalogs’ pages, tracing back their intermittent showing through the web in a chronology reconstructed with the help of Google search.

Following the trajectory of visual content proved a bit more difficult. As I conducted my research, tools to automate visual analysis of the web became more reliable. After considering different solutions, I opted for TinEye, a solution used by scholars in the humanities to study both online and offline image circulation.¹⁹ TinEye is a very accessible, efficient, and free tool that can accept whole image files. It compares the entry image “signature” with a very large index of pictures built from online sources. When matches are found in this index, TinEye displays them in chronological order with links pointing to their webpages or, more precisely, pointing to the pages that hosted them when visited by TinEye’s robots.²⁰

Although less efficient and versatile than other services, such as Google Images, TinEye returns exact matches only, making it particularly useful to find previous or subsequent publication of the same image.²¹ Here I used the tool to look for the presence of identical images in past publications, downstream if we will, a usage reminiscent of methods journalists and activists use to authenticate visual content in open-source intelligence (OSINT).²² Besides search engines that find images on specific sites, like Reddit or Imgur, Microsoft Bing now offers a similar most effective solution.²³ A caveat of this method is that the search is indeed limited to open sources, to images that are available online, thus excluding private image databases and non-digitized material, such as festival programs.

After navigating the catalogs’ websites in a very systematic manner, to assess the amount of curating work and locate the different places where subscribers could interact with the content or engage with others, I made two observations: one almost existential and the other methodological. The first point alludes to the once-popular feature of allowing subscribers to provide feedback. Apart from Mubi, which offers its lively community ways to express thoughts and share movie lists, other catalogs, some after experimenting with comment sections, choose to abandon them. I realized that subscribers’ comments and interactions were happening elsewhere: on social media, such as on Letterboxd, or on dedicated platforms, on SensCritique for instance.²⁴ Still, outside of the parts fueled by curating teams, the methodology I had designed revealed vast spaces inhabited only by machinic activity–numerous pages populated only by the aggregation and shaping of movies’ technical metadata and paratexts. What this means for online cinema distribution goes beyond the scope of this article. To mention it very briefly, browsing the numerous pages generated for each movie, arguably in a more systematic manner than the one intended, I came to realize that even the more curated side of the streaming cultural offer felt very machinic. This affective reaction is similar to the first impressions that Luc Pauwels encourages researchers to keep track of in the initial phase of studying websites.²⁵ I would argue that it is also relevant to keep track of these reactions during the data-gathering process. Further exploring this path, for instance, enabled me to develop an argument about the catalogs’ production of eclectic cinephile paratexts as a way to conceal the presence and centrality of machinic processes within their websites.²⁶ Broadly speaking, the deployment of large online libraries of titles seems to come inevitably with the feeling that we are, perhaps as never before in the history of film distribution, very often interacting and dealing only with machinic forms and processes. To come back to the main point, it is relevant here to mention that this observation stems from manually browsing the many catalogs’ pages, their busiest sections, and their more neglected corners. An automatic analysis could very well overlook this aspect. Or, to avoid that, such an automatic approach would need to be complemented with a more detailed reading of samples to fully grasp this experiential side of film databases’ usage and viewing context.

The second observation, methodological in nature, points to the need to build a small archive of the catalogs’ design and features. As I began to work through their sections, the catalogs started to move. Some changes were modest, new rows of thumbnails appeared to highlight thematic or editorial groupings, while others were more substantial. Mubi and Filmatique completely altered their website layout in 2020 and 2021, respectively. To be able to come back to a particular moment of this continuous transformation, I needed to find a way to capture and archive snapshots of the catalogs’ states in time.

Google’s Chrome browser extension GoFullPage allowed me to create scrolling screenshots, some as long as the large numbers of content rows stacked vertically on webpages (see Figure 1). Because of their stacked visual form, I metaphorically called them ice carrots in reference to the ice core produced by scientists to study layers of ice-bound dust. Here, the aim wasn’t to drill into the past of the web but to build an archive of the way different elements and modalities (text, image, video, and graphical elements) were used and laid out in the catalogs. Once archived, those screen grabs helped me systematize the description of the various websites’ sections. They retain a visual trace of the state of the catalogs at a given moment in their history. I then could track the changes applied to the layouts and the fate of dedicated sections and features, such as pages reserved to known festival or, as I mentioned earlier, comment sections. I applied this method to the capture and analysis of the desktop version of the catalogs’ websites.

Figure 1

Screen Captures of BFI Player, Sundance Now, FilmStruck, and LaCinetek.

If SVoD consumption largely takes place on TV screens and mobile devices, there are several reasons to consider desktop layouts rather than the ones created for mobile or TV applications. In line with industry standards, the cinephile SVoD services started deploying mobile applications and software developed to reach specific platforms, such as tablets, smart TVs, and other OTT devices (Apple TV, Chromecast, Roku, etc.). Until recently, this development work was outsourced to third-party companies. Mubi collaborated with a small agency named MonkeySource. The Criterion Channel relied on the Vimeo OTT service. You.i TV, the company that was behind FilmStruck’s multi-platform interface during the service short run between 2016 and 2018, also provided technical solutions to major media groups (Sony’s Crackle and Rogers’ Shomi, for instance) and worked on behalf of Twitch, Amazon’s subsidiary dedicated to live streaming of video game sessions. Tënk and LaCinetek used a proprietary content management system created by Kinow, an OTT platform provider bought in 2022 by a larger international player, Alpha Networks.

It is worth noting that catalogs, either due to limited resources or because they are part of large media groups aiming to rationalize expenses, often rely on ready-made solutions in this domain. The gain of adopting this somewhat out-of-the-box solution measured in terms of presence on multiple platforms (which required a great deal of development work), but it came with related constraints. When film catalogs choose to outsource their apps development, the video routing function was often the only one retained. Likely due to similar factors, paratextual content and other extras are still not included in the apps version of those catalogs today.²⁷ This is also meant to facilitate movies’ navigation and broadcasting. Still, the lack of paratextual content and the relatively flat structures found in those applications make them less appealing to study in an approach that is interested in the formal and cultural aspects of cinephile SVoD interfaces.²⁸ The fact that these applications present little information may also explain why the discoverability of movies still also happens via the web.²⁹

In contrast, catalogs’ websites tend to present eye-catching landing pages, original thumbnail stacking (in mosaic of larger tiles on BFI Player or by evoking frames in a film stock on LaCinetek), and dedicated or thematic sections, each shortly introduced and using specific layouts. However useful this method of capturing and archiving scrolling screenshots or ice carrots may be, it captures only the visual content of the page. It doesn’t extract the links, HTML tags, or code behind the displayed elements. Studying those requires making short exploratory searches, by following the different hyperlinks and manually scanning the source code.

How, then, can one consider all the visual forms previously identified and archived, while maintaining a manual and heuristic approach to these materials? To achieve this, I used a mnemonic method of location and referencing modestly inspired by the work of Austrian formalist filmmaker Gustav Deutsch.³⁰ In addition to noting the aesthetic aspects of the pages, I draw a quick sketch of their composition. Deutsch developed a method for finding a series of recurring motifs in large corpuses of films, starting with home movies and then archival materials. His process was a very slow one where he would watch each movie, take notes, and draw quick sketches of outstanding shots. The set of motifs thus produced, which Deutsch then edited to make parallels or highlight contrasts, is less a posteriori synthesis than something that emerged as the research progressed, throughout the consultation of the material found in the archives. The study of SVoD interfaces can take inspiration from this technique to better describe the different ways in which multimodal elements are arranged on the pages and how their layout follows and enacts cultural programming and distribution strategies.

Stitching a Patchwork of Data Processing Tools

In addition to the interest in the shape of SVoD catalogs, understanding the main trends in title selections is a key area of study in the current research on streaming. This is especially relevant for cinephile catalogs. An important part of the marketing discourse they produce is meant to distinguish them from other more generalist SVoD services by showcasing their programming efforts and the fact they offer niche or under-seen movies. When conducting my research, I was confronted with the lack of reliable data on VoD services.³¹ In Quebec, those statistics still do not exist.³² In France, the catalogs of my corpus disappeared from the CNC graphs about VoD consumption after 2020, due to their small subscriber base. Multiple methods have been devised to try to describe the specificities of cinephile SVoD programming. Research projects classified services by specialty, studied Mubi’s scheduled programming on a given date, followed a sampling method, or took a broader approach, analyzing the overall classification of subscription video by genre.³³ Catalog analysis also benefited from the data collected by third-party data providers, such as Ampere Analysis, which specialized in gathering movies selection and metadata across different platforms and countries.³⁴ But accessing this commercial information is costly and often beyond the financial means of small academic teams and individual scholars.

These constraints made essential the need to draw up a complete picture of the titles offered by cinephile SVoD services at a given time. Working with data samples makes it difficult to make comparisons between catalogs, and accessing commercial databases is too costly, as just stated. I have then sought to overcome these limitations by gathering information on the movies selected (their title, date, director, and country of production) through web-scraping. I used two different methods for this purpose: the importHTML or importXML function of the Google drive software and small capture scripts that I developed in Python. The two methods emulate human browsing to collect webpage data. The first looks into the HTML or XML source code to find tags that indicate content structured as lists. For the second, I used the Anaconda development tools, a programming environment, and the Beautiful Soup library, a series of prewritten tools and functions, to send requests to the catalogs and capture the data sent back by their servers. Each catalog used slightly different systems to dynamically display content and metadata to subscribers, and those scripts had to be adapted each time following a trial-and-error method.³⁵

The movie information collected is limited to data already available on other platforms (e.g., from the Internet Movie Database (IMDb) to Rotten Tomatoes, via library catalogs) and does not include information produced by the catalogs themselves. Only the title selections specific to each catalog were the subject of my analysis.³⁶ The creation of this capability hinged on designing a tailored set of procedures for data collection, processing, and analysis. These procedures were built up through a series of back-and-forth processes, from collection to data cleaning. I worked to harmonize country names, to find missing information for certain titles by reconciling data tables with authority file systems (Virtual International Authority File (VIAF) and Wikidata), and to aggregate the various data sources together. OpenRefine, a free and open source software supported by a lively community, proved to be invaluable to do so. I used it to join different datasets together and to start cleaning and structuring them, while always keeping an eye on the data by previewing tables and using simple transform functions. The same applies to Tableau, a data visualization software with a general user-oriented interface that also provides direct access to the underlying tables. Through OpenRefine and Tableau, I was able to create exploratory graphs of the data and then come back to the tables to adjust parameters and correct the remaining errors. I thus recreated, on a very modest scale, some of the interactive and iterative data cleaning methods used in much larger projects.³⁷

In the end, each data gathering method needed to be adjusted to every catalog. The selection of titles presented in a unique catalog could differ between countries, following programming logic–titles chosen with a local audience in mind–or because of restrictions related to the user’s geographical location. The latter stems from the negotiation of digital rights by territory, sometimes understood as broad regions–such as North America or Europe–or as individual countries. Beyond these general considerations, some catalogs posed specific challenges. Mubi’s programming required adjustments due to its daily update and the small selection of titles available at any given time. I annualized the sample using the Wayback Machine archive to reconstitute a selection of 325 titles.

The data collection method created for this project was thus made of a very diverse set of tools cobbled together to achieve the specific goal of analyzing catalogs’ programming. To accomplish this, researchers need to be able to adapt these tools to the unique characteristics of the web objects they wish to study (API, database structure, scraping methods, etc.). This leads to two observations. First, the selected tools must be easy to access and flexible enough to accommodate this adaptation. Second, the methodology has to be flexible enough to accept borrowings from other fields of research. The use of the Wayback Machine archive, a tool often queried to study the history of the web, provides a good example of this. Despite the difficulty in navigating the site layouts due to image storage and display problems, the archive.org database keeps track of title selections. This allowed me to annualize Mubi’s programming and add it into my dataset.

Data Visualization to Locate Where to Dig Further

Beyond cobbling together tools, the cobbling metaphor also applies to the need for combining methods. The aim is to better follow and highlight the relation among, in this case, title selection, interface design, and larger cultural trends. After gathering the data, I created a set of exploratory graphs to identify patterns in title choices. Far from constituting full-fledged results, these graphs enabled me to highlight certain aspects of catalog programming, showing me where to take my research further.

At the time of data collection, the catalogs provided access to the following numbers of films: Fandor (3386), Criterion (1900), LaCinetek (1034), BFI Player (697), Mubi (325), and Sundance Now (244). Figure 2 shows that 40 percent of titles come from the so-called classic period, roughly between 1950 and 1980. If we exclude Fandor’s and its big group of recent titles, it represents the majority of movies offered (54 percent). These figures, influenced by the more voluminous offerings in our corpus, nevertheless conceal other relevant clues.

Figure 2

Breakdown of Selected Titles By Decade and Catalog.

Sundance Now offers a majority of films set in the first two decades of the 21st century. This is also the case for Mubi, albeit in less marked ways. On the LaCinetek side, the selection takes the opposite form, with very few recent titles and many works made between 1907 and 2011. A closer look at the distribution of titles for each year reveals the strategies of the cinephile catalogs. For example, Fandor seems to maintain a small sampling of titles (between fifteen and thirty) for each year, separating the oldest film, Monkey Shines, No. 1 (William K.L. Dickson and William Heise, 1889) from the most recent, You’re Welcome (Rebecca Panian, 2017). It is also worth noting the presence of many early cinema titles in Fandor. It’s an unprecedented presence in the subscription video landscape, and one that contributed to Fandor’s reputation in its heyday.

There is one exception to Fandor’s coverage of different periods. The 1930s and early 1940s are much less represented. Criterion seems to do better in the same period, presenting a selection mostly occupied by films from French Poetic Realism and the Hollywood Golden Age. In the case of these two movements, their absence at Fandor is perhaps explained by the difficulty in acquiring licenses for them. The former, including films by Jean Renoir, Jean Duvivier and Marcel Carné, is part of the French pantheon. The latter belongs to the major studios and is part of already-established distribution strategies. Their presence in the Criterion catalog undoubtedly helps “validate” their canonical aspect, while their broadcasting rights are often acquired by larger platforms (e.g., Netflix, Amazon, and Hulu) or “in-house” services (e.g., FilmStruck then Max for Warner, Crackle for Sony, or Peacock for Universal).³⁸

This second graph (Figure 3) illustrates the concentration of titles in three main areas: North America, Central Europe, and Japan. Those are by far the main sources of titles offered in the catalogs.³⁹ Perhaps less expected, Brazil, India, and China account for less than 2 percent of all titles. The Fandor catalog stands out for its presence in many countries. Leaving aside the Latin American countries (a few dozen titles) and the countries of the Near and Middle East renowned for their major film productions: Israel, Lebanon, and Iran (between ten and twenty titles each), Fandor’s strategy was one of sampling. Several countries are represented by just a handful of titles: Ukraine (2), Bulgaria (1), Cambodia (2), Vietnam (2), Algeria (6), Senegal (6), and Nigeria (1), to name but a few.

Figure 3

Breakdown of Selected Titles By Country and Catalog.

Visualizing data in charts helps highlight these major geographical trends. Here the aim, however, is not to draw macro-level typologies or to study regional or global distribution strategies. Larger projects, both state-run and independent, do that well today and provide a useful context to the more focused research described here. The goal is rather to shed light on programming trends and decision-making by taking a snapshot of title selections across a very specific category of SVoD services. What differentiates this snapshot from other visualization methods is that it is very close to the source where data have been collected. Having created the data collection and processing workflow by themselves, researchers can answer specific questions about interface design or programming logics by coming back to each catalog under analysis. They would also have a better understanding of the trade-off and decisions made during data capture and processing: how movies produced in countries that no longer exist, such as the German Democratic Republic or Yugoslavia, were coded, for instance. What defines a title? Were feature-length films the only ones considered? Moreover, as stated earlier in this article, each data gathering method had to be adjusted to each catalog functioning. This caveat, when it is time to bring everything together and to try to draw comparisons between catalogs, could then turn into a series of blind spots if not known by researchers. Not being able to check the quality of data supplied by third-party companies is, in fact, a problem raised by other research teams. Amanda Lotz et al. mention that caveat while conducting their large-scale research on streaming services.⁴⁰

Visualizing large trends and quickly accessing data sources are also conducive to another key aspect of this methodology. Charts can also help researchers identify new areas of interest, allowing them to focus their efforts on particular aspects. Quantitative analysis is not intended to provide final conclusions but rather to guide further analysis. The distinctive trends observed in the data, such as the global distribution of Fandor’s chosen titles, should be regarded as indicators and incentives for further investigations. These inquiries must go beyond the raw data to examine underlying strategies and the way those strategies take forms in the websites of the catalogs. Small complementary enquiries can then profit from the analysis done on the design and formal aspects of the catalogs to provide context and help explain data patterns and trends, thus bridging two or more of the methodological pieces previously cobbled together.

To demonstrate how different methodological threads can inform each other and shed light on the relations between aspects of the cinephile catalog programming and curating work, I now describe how the trend identified earlier in Fandor can be explained by the catalog’s design choices. Fandor devoted a section of its website entitled International to the sorting and browsing of movies according to their region of origin. For each part of a world map divided into six regions, from North America to Eurasia to Oceania, users were invited to discover the countries covered by the catalog. These were displayed in overlay when the mouse was moved over the region name. When clicked, they linked to groups of films from these countries. Having studied the graph, I went back to the site’s interactive filter. I then realized that each group of movies (even a single title) made their country of origin appear on the interactive list of the website. Selecting a very small number of films from a multitude of countries meant that this interactive menu would display a large number of entries, even if many of them referred to just one or two titles. This strategy that blended title selection and web design then made Fandor’s international reach look stronger.

Conclusion

Assembling a patchwork of tools to study streaming services allows to create adapted sets of tools in order to grasp the multifaceted aspects of the SVoD offering. Since I conducted the field work of this research, a steady stream of new studies was published on SVoD interfaces and programming.⁴¹ They tend to work on large datasets, provide longitudinal or comparative analysis, and focus on big national and international brands. This article aims to contribute to the development of suitable methodologies by promoting a particular type of research designed to address small-scale and dedicated streaming services. Partly because of their hybrid nature and partly due to their limited means, cinephile catalogs use less dynamic layouts than the constantly rearticulated interface of Netflix and are, in many ways, far away from AI’s big data shuffling, described, for instance, in Somaini’s Grey Room article.⁴² They nonetheless contribute to cinema online distribution by constantly working to find new ways to shape the flows of subscribers and movies circulating through their websites. A multifaceted methodology can then help researchers gain a plurality of views on their cultural role. This may involve reorienting certain methods to pursue new objectives. Analyzing interface design and title selection, a method usually used for the measurement of content prominence and discoverability, can also serve the study of the movie-watching experience and how catalogs experiment around it. But the cobbling metaphor furthermore points to open and accessible research. Studying the streaming phenomenon calls not only for big infrastructure-led projects but also for the development of a multiplicity of focused research, maybe conducted with more proximity to their subject and certainly leveraging free or affordable tools and developing tailor-made and accessible data gathering strategies. Contributing to this still emerging field of research, this article makes the case that studying SVoD catalogs requires appropriate methods, both small-scale and pieced together, to capture their marginal and nonetheless essential place in cinema online distribution.