Taste is key to understanding algorithms as cultural interlopers as it is both the input and output of algorithmic processes. Taste is also a crucial aspect of the browsing process because it directly influences how users interact with digital platforms and make choices about content consumption. It is clear that exercising cultural taste influences the browsing process on Netflix, largely because movies and similar media are hedonic products: products and services that offer more experiential and emotional value compared to utilitarian products, which provide functional value (Dhar and Wertenbroch 2000). Bourdieu proposed that taste preferences are not individual or innate but deeply influenced by one’s position in social space, structured by various forms of capital and embodied through the habitus. Classification, from Bourdieu’s perspective, is a manifestation of these tastes, reflecting and reproducing social distinctions and hierarchies; hence, “taste classifies, and it classifies the classifier” (Bourdieu 1984 6). However, some scholars have argued that taste is complex and idiosyncratic in ways that are not easily mapped onto class divisions. Peterson (1992) challenged the assumption that taste is a strictly hierarchical manifestation of class through his model of cultural omnivorousness, where individuals increasingly engage in diverse taste practices that transcend their social position. While Bourdieu emphasized the exclusivity of highbrow culture as a source of cultural capital, Peterson suggested that the ability to appreciate and engage with a wide range of cultural forms may also serve as a form of capital, signaling cosmopolitanism and higher social status. Lahire (2006) also critiqued the notion of a unified habitus, arguing that individuals have dissonant tastes, exercising contradictory preferences across different cultural fields. For decades now, sociologists have examined how taste is socialy constructed, how it is used to enact boundries and assert social status. However, what this research has failed to grapple with is how taste manifests in the culturally laden exercise of content selection.

Classification, the invisible infrastructure that provides structure to our social lives, involves the organization of cultural objects into distinct genres, types, or categories. The classification of cultural objects contributes to the social construction of taste by defining what is considered valuable, worthy, or prestigious. Bowker and Star (2000) define a classification system as “a set of boxes (real or metaphorical) into which things can be put to then do some kind of work.” So then, what work do classifications do? Classifications create and reinforce social hierarchies and norms. They are not merely reflections of pre-existing social orders but actively construct them by categorizing the world in specific ways. Classification is also a form of boundary work, where people create separations between themselves and what they deem to be above or below them by determining what is included or excluded from categories. By analyzing data from the 1993 General Social Survey about musical dislikes, Bryson (1996) found that people use cultural taste to reinforce symbolic boundaries between themselves and categories of people they dislike. Contrary to Bourdieu’s hypothesis of musical exclusiveness increasing with education, Bryson, like Peterson, found that higher education correlates with broader musical tastes, suggesting a form of cultural tolerance. However, this tolerance is not indiscriminate; genres associated with lower education levels (e.g., gospel, country, rap, and heavy metal) are more likely to be rejected by the musically tolerant, revealing a pattern of exclusion based on class rather than mere taste diversity.

Despite paying a great deal of attention to cultural taste and its consequences, until quite recently, sociological research has handled generic taste coarsely. Surveys ask whether people like rock or blues music, horror films or romantic comedies. This generalization not only does a disservice to the incredibly granular ways in which individuals classify cultural objects, it ignores the rich array of knowledge people use to do so. These studies, while acknowledging the complexity of taste, fall short of capturing this complexity empirically. Examining the act of digital browsing on Netflix requires a more nuanced and delicate approach to generic taste because the NRS already sorts content into categories finer than typical generic categories. Furthermore, the very act of browsing is additional filtering beyond mere genre. This shortcoming in sociological literature motivates a closer analysis into the ways that subtle cultural cues influence browsing decisions.

To grasp how users employ cultural knowledge while browsing, it is necessary to first explore research that investigates the impact of culture on people’s actions. There is a rich sociological debate surrounding whether cultural knowledge serves as a motivation or justification for people’s behavior. Vaisey (2009) sought to bridge these two perspectives with a dual-process model of culture in action, distinguishing between discursive (conscious, justificatory) and practical (unconscious, motivational) modes of culture and cognition. He leveraged panel data from the National Study of Youth and Religion to illustrate how individuals may not articulate clear principles of moral judgment (aligning with the justification camp) yet display behaviors strongly predicted by their moral and cultural scripts (aligning with the motivation camp). Discursive culture pertains to the conscious, articulated aspects of culture that individuals use to rationalize their actions. Practical culture, however, refers to the unconscious, taken-for-granted knowledge and dispositions that guide behavior in a more automatic and intuitive manner; similar to what Lizardo (2017) calls non-declarative, that is, culture that is “phenomenologically opaque and not open to linguistic articulation.”

Vaisey’s model has gained significant traction, contributing to the rise of a new subfield, the sociology of culture and cognition. Although Vaisey provides an abstract model to understand cultural knowledge in action, we do not know how these processes are enacted and operationalized while browsing. In a recent paper, Vaisey urged scholars siloed within this subfield to engage more directly with other disciplines (Vaisey 2021). I respond to this call to action by engaging with scholarship and methodological tactics from HCI.

Most empirical studies of digital browsing behavior have been conducted by researchers within the fields of HCI and Library and Information sciences. Some recent scholarship has focused on how people navigate social media platforms and how through their interactions, users are able or unable to evade algorithmic surveillance and influence outcomes. Yeung (2017) proposed that algorithms influence user decisions in the form of a hypernudge, “a particular form of choice architecture that alters people’s behaviour in a predictable way without forbidding any options or significantly changing their economic incentives.” Witzenberger (2018) introduced the term hyperdodge to describe a strategy where users intentionally feed manipulated or misleading data to algorithms. This practice can disrupt the predictive capabilities of these algorithms, challenging their objectives and potentially affecting their accuracy. Witzenberger’s interviews revealed that while users may not fully understand the technical workings of algorithms, they are aware of their presence and influence. This awareness leads to the development of resistance practices that range from simple avoidance tactics to more sophisticated methods of data manipulation. Similarly, Ellison and a team of researchers explored the motivations and implications behind the “non-click,” instances where users deliberately decide not to click on digital content they have viewed or engaged with. Interviewing and eye-tracking Facebook users revealed three audience-related concerns, namely participants not wanting the poster, their network, and the platform itself perceiving their engagement with a particular post (Ellison et al. 2020). These results seem to suggest that social consequences influence online interactions.

Early studies of browsing behavior, especially those in Library and Information sciences, examined browsing as an information-seeking activity. Branch (2000) investigated how 12-15-year-old students sought out information using Microsoft Encarta 98, a digital encyclopedia originally sold as a CD. She found that while all participants were able to answer the research questions that prompted their search, they used a variety of search terms, categories, and strategies to do so. These tactics change when users struggle to find what they are looking for. Researchers at Google found that when having difficulty finding information, users formulate more diverse queries, use advanced operators more frequently, and spend more time on the search results page compared to the successful tasks (Aula, Khan, and Guan 2010). These studies demonstrate that fundamentally, browsing is a quest for information. However, further research has shown that cultural differences influence the types of information people are looking for.

A limited body of HCI research has examined how cultural differences can influence browsing behavior and interface acceptance. Chau and a team of researchers examined how students from the U.S. and Hong Kong differed in their online consumer behavior while browsing automobile manufacturers’ websites. They found that American and Chinese participants had different motivations for using the internet and that simply translating websites into a different language is insufficient without considering the deeper cultural nuances that influence user interaction and perception (Chau et al. 2002). Evers and Day (1997) researched the influence of culturally specific design preferences and their impact on attitudes and behaviors towards interface acceptance. The study confirms that design preferences rooted in cultural background significantly affect user acceptance of interfaces. For instance, Chinese users’ interface acceptance is influenced more by system usefulness driven by their design preferences, whereas Indonesian users place more emphasis on the system’s ease of use.

While HCI scholarship has considered the importance of cultural adaptivity in software design and the influence of cultural background on online behavior, overall, the field is in need of more robust theories and models of culture. Most papers lack a model of culture; they merely use the word, and sometimes incorrectly by equating it with nationality (Clemmensen and Roese 2010). As digital interfaces and algorithmic systems increasingly govern social and cultural life, there should be more collaboration between sociologists and technologists. By turning to sociology’s extensive literature on culture, HCI research can gain a deeper understanding of the nuanced ways culture influences human interaction with technology, offering a more complex and effective framework for designing and studying culturally adaptive systems.

Scholars studying taste with surveys have been limited by the small number of items that can be feasibly covered by a survey, yet a growing scholarship of Netflix itself may shine more light on how decision-making processes play out in a digital space with hundreds of items for evaluation. The evolution of the NRS is well documented in a wide breadth of scholarship that includes Netflix’s proprietary research documents and the work of external academics. Some scholars have explored the personalization strategies of Netflix by examining thumbnail optimization. Thumbnails are excellent sites for research as they are tangible embodiments of algorithmic processes in action, “providing a face to something that is often lost in the world of big data” (Eklund 2022 738). Pajkovic (2022) set up three Netflix user profiles, each with distinct taste personas, in an attempt to reverse-engineer the operational logics of the NRS. Over two weeks, Pajkovic interacted with the NRS daily by selecting one new film or television show for each fictional user to watch based on their predetermined tastes. He observed and recorded changes in each profile’s homepage as they became increasingly personalized. By the end of the experiment, “when identical titles were offered to each profile, they almost always had different artwork images” (Pajkovic 2022). Pajkovic’s results suggest that the NRS’s circular logic tends to reinforce users’ existing preferences i.e., if a user watches certain genres or titles, the NRS is more likely to recommend similar content, potentially creating an echo chamber of content that mirrors the user’s historical data. Other qualitative studies of the perceptions and impacts of the NRS adopted indirect methods for collecting user feedback. Gaw (2022) used a commercial tool connected to the Twitter API to collect and analyze tweets containing keywords directly related to the NRS. After filtering, tweets were coded and examined alongside themes identified from previously analyzed proprietary press releases, corporate documentation, and media discourse about the NRS. This method aimed to uncover recurring patterns and highlight the tensions, inconsistencies, and contradictions in the accounts of the NRS as constructed by its engineers, depicted in the media, and experienced by users. Gaw proposes the concept of “algorithmic logics” to define the assumptions, processes, and mechanisms that govern the construction of taste within the platform. Gaw argues that on Netflix, taste-making is constituted by four algorithmic logics—datafication, reconfiguration, interpellation, and reproduction—and that these logics reappropriate traditional mechanisms of social control and have created new ways to engineer cultural processes.

Researchers have also conducted studies of the NRS that involve real users of the platform. Schaffner and a team of researchers interviewed 20 regular Netflix users, asking about their experiences and perceptions of features in the Netflix platform design that may affect their sense of agency. Participants expressed that even while using a variety of strategies to manage their time on Netflix, design features such as autoplay and a lack of stopping cues often influenced them into watching more than they had planned to. Furthermore, the design of the homepage and its emphasis on recommendations influenced what the participants chose to watch. Most participants were not thrilled about this, preferring “more human” recommendations from friends and family over the algorithmically generated recommendations pushed by the NRS. Overall, the study concluded that “a user’s sense of agency was at odds with the platform’s design,” especially when it came to content selection and time spent on the platform (Schaffner et al. 2023).

While Netflix’s technical papers and blog posts illustrate how their algorithms work, human interactions with these algorithms reveal how they are not working as intended. Within cultural sociology, there has been a great deal of research about taste, classification, and how cultural knowledge influences people’s behavior; however, this research fails to consider how these processes are enacted through browsing content recommendation systems like the NRS. Within HCI, there have been empirical studies investigating browsing behavior and user agency in algorithmic systems, but the role of cultural knowledge in content selection has been largely unexplored. Scholars of the NRS have presented a somewhat dystopian vision of the effects of personalization algorithms, where users passively accept recommendations that reproduce echo chambers and cultural hegemony, but there have been few user-centric studies conducted by external academics, none of which have explicitly examined the browsing processes of users in real time. By borrowing methodological techniques from the field of HCI and Vaisey’s dual model of culture in action, this study aims to present a novel user-centric approach that scrutinizes real-time browsing behaviors to investigate how cultural knowledge and algorithmic mediation interact within the browsing process of Netflix users.

Regular Netflix users in Chicago, aged 21-34, were selected as the population for this study. Regular Netflix users (defined here as those who use the platform at least once a week) were chosen because of their familiarity with Netflix’s user interface and their increased likelihood of having developed browsing strategies and opinions about the platform. Young adults residing in Chicago were chosen due to their relative technical literacy and accessibility to the researcher. This study implemented a multi-faceted recruitment strategy beginning with the distribution of informative flyers at the David Rubenstein Forum and various public spaces around Hyde Park, leveraging the researcher’s workplace networks and the area’s high foot traffic for diverse participant outreach. Six of the participants were University of Chicago graduate and undergraduate students. All participants were compensated with $30 gift cards.

Figure 1.1: Participant Demographic Information

Participant ID	Age	Sex	Student Status
U1	21	M	Student
U2	22	F	Student
U3	22	F	Student
U4	21	F	Student
U5	23	M	Non-Student
U6	34	M	Non-Student
U7	31	M	Student
U8	21	F	Student
U9	26	F	Non-Student
U10	28	M	Non-Student

This study’s primary site of investigation is the Netflix homepage, the user interface of the NRS. This focus allows for an examination of the interaction between users and the NRS during the browsing process. A study of browsing could have taken on many forms and focused on any number of platforms; the decision to concentrate on Netflix users’ search for content was intentional. Because of their length, picking a movie or show to watch is a more consequential choice than selecting a YouTube video or TikTok, often requiring deeper consideration and more research before the user feels ready to commit. Netflix was selected because, compared to other streaming platforms, it offers a virtually unparalleled level of personalization integrated into its user experience and recommendations. Additionally, the popularity and ubiquity of Netflix content provides an opportunity to observe users’ symbolic references and inquire about content they have “heard of” but have not seen.

The semi-structured interview was designed to probe how Netflix users employ cultural knowledge to facilitate their browsing process. The interview procedure consisted of four parts: introductory questions, personalized and impersonalized browsing tasks, browsing task review, and concluding questions. At the start of each interview, I asked the participant about their tastes, viewing habits, feature usage, and external sources of recommendation. After that, the participant logged in to their Netflix account and completed the first two browsing tasks on their personalized profile. The second set of browsing tasks took place on a “default” account with no viewing history to emulate an impersonalized browsing environment. I tested these two conditions to compare the potential differences in browsing tactics across personalized and impersonalized browsing environments.

While the interview guide provided foundational structure, context-specific questions were necessary to probe implicit judgments (e.g. “What makes it look dumb to you?”). As the participant browsed, I would ask follow-up questions based on what they said or did. After completing the browsing tasks, the participant and I reviewed the screen-recorded footage together in an attempt to explicate their browsing process. During the browsing task review, I asked more impromptu questions that required the participant to explain their rationale behind certain actions (e.g. “Why did you click on Whiplash?”), entertain hypotheticals (“How about Love is Blind?”), and reflect on a string of decisions that led them somewhere on the page (“How did you get to ‘Movies Based on Real Life?’”). At the end of the interview I asked the participant to reflect on their experience browsing Netflix, the efficacy of its personalized recommendations, and theorize about the NRS’ mechanics and innerworkings.

Browsing tasks were designed to encourage participants to engage with the Netflix interface in ways that would reveal their browsing strategies and decision-making processes. To avoid participants choosing titles they already watched, they were instructed not to pick things they had seen before. Additionally, these browsing tasks were ambivalent to genre and format to encourage participants to consider as many titles as possible. Two types of browsing tasks were used:

1. Open-Ended: Participants were asked to find something they were in the mood to watch at the time of the interview.
2. Situational: Based on information provided by participants earlier in the interview, a situational task was created. For example, if a participant mentioned they often watch Netflix while doing chores, I would ask them to find something to watch while folding laundry. This task was designed to simulate a real-life scenario that influences their content selection process.

For each participant, identical prompts were used in both the personalized and impersonalized environments. While these artificial browsing tasks motivated the participant’s search, they primarily served to initiate the browsing process. The goal of the browsing tasks was to examine the sequence of choices, distinctions, and navigational behaviors that led participants to their final content selection, not to highlight what was ultimately chosen.

While participants browsed the Netflix homepage, I asked them to think aloud, verbalizing their thought processes, intuitive judgements, and confusions. Originally hailing from cognitive science, the think aloud method is used widely in HCI research to understand how people work with digital products and interfaces. With the most common approach, concurrent think aloud (CTA), users perform a task and verbalize it at the same time. However, researchers have raised concerns about reactivity, the way in which participants alter their behavior in response to being observed or having to articulate their thought processes aloud. These concerns about CTA have led to the emergence of retrospective think aloud (RTA), a method where participants describe their thought processes and decision-making after completing a task, rather than during the task itself. Retrospective think aloud has been used to examine why Facebook users choose not to interact with certain content on their feed (Ellison et al. 2020) and how students seek out information using a digital encyclopedia (Branch 2000). I decided to use a combined CTA-RTA approach because I wanted to gauge how participants made decisions in real-time while also hearing post-hoc reflections on their navigational process.  

Using an inductive approach, interview transcripts were qualitatively coded to identify emergent themes and patterns in how users evaluate content. This process began with locating when and where participants were engaging in classification. Following that, I generated specific subcodes based on the knowledge or information participants used to make these classifications (see fig. 2.1 for codebook). Additionally, screen recordings of the browsing sessions were utilized to complement the interview transcripts, providing necessary nuance and context.

The NRS automatically presents recommendations to users based on generic categories. These categories are often already highly specific, like “Revenge Action Thrillers” or “TV Dramedies Featuring a Strong Female Lead,” but rarely are these genres fine-grained enough to identify movies or shows that the user is actually interested in. To go beyond the initial filtering of the NRS, users employ a filtering system of their own by classifying content along a variety of valences to digest the information presented to them on the homepage. These techniques are numerous, but I describe five that were particularly prevalent in my interviews: (1) identifying tropes, (2) making associations, (3) gauging situational suitability, (4) ascribing content to perceived audiences, and (5) considering its social and moral acceptability.

Figure 2.1: Five Content Classification Schemas

Code	Definition	Guiding Questions	Code Instances
Trope	Participant identified themes, motifs, and clichés embedded in a title’s plot.	What is this about? What’s the deal? What’s the shtick?	28
Association	Participant directly linked a title with other content due to their perceived similarities.	What is this like that I’ve already seen? What does this remind me of?	42
Perceived Audience	Participant ascribed content to particular social categories based on who they thought would watch it.	Is this for me? If not me, who is this for?	11
Situational Suitability	Participant classified title based on where or when they deemed it appropriate to watch.	Is this the appropriate time to watch this? What would be an appropriate context to watch this?	34
Social and Moral Acceptability	Participant classified title based on its social or moral acceptability.	Is it okay to watch this? Does this align with my values? Does this align with my community’s values?	6

Tropes

Netflix users I spoke to classified content using their knowledge of tropes: themes, motifs, and clichés embedded in a show or movie’s plot. This process involved the participant identifying a piece of information from a thumbnail, preview, or description and recognizing a trope based on that. Genres serve as broad categorizations of media, characterized by shared conventions, settings, themes, and narrative structures. Tropes, on the other hand, become recognizable through their repeated use across various media. While genres categorize the overarching framework of narratives, tropes delve into specifics, detailing how stories within these broad categories differentiate themselves. While browsing, tropes functioned as a narrative shorthand, enabling participants to infer a film’s plot without actually watching it. By recognizing Gerard Butler from the thumbnail, U6 quickly pegged Hunter Killer as one of those movies where a “retired CIA guy gets called back to action.”

I know this guy. This seems kind of cool. Top ten. I might watch. Seems kind of a little bit, um, bro-y, action, whatever. So it might be a little too basic … classic, like, missile military guy, action, uh, save the world … Gerard Butler, like that’s classic him … You know, retired CIA guy gets called back to action because he’s the only one who can do it!  … I guarantee you there’s a dozen movies with him where he’s the military guy saving the day and it’s like, okay, been there done that.

This classification led U6 to deem the movie as maybe something he would watch “if I’m doing laundry or something” but not compelling enough to dedicate his full attention due to its perceived lack of originality and narrative depth. U6’s discernment—bored by the redundancy of the trope yet still considering the movie for casual watching—was made possible by combining the information he was picking up on from the thumbnail with his knowledge of other films where a “retired CIA guy gets called back to action.”

Participants also used tropes to distinguish between what they did and did not like within a specific subgenre. U2 liked romance reality, so I asked, “what about Love Island?” She said, “the ones where they’re wearing bikinis and have like a fanny pack with the team-building exercises are just not my thing.” However, she did like romance reality that used some sort of technology to mediate the interactions between the contestants—things like The Circle and Too Hot to Handle—and ended up selecting Deep Fake Love. Because U2 has a nuanced understanding of the thematic and aesthetic elements of the romance reality genre, she was able to pick up on subtle variations between shows to make a more informed decision.

Users’ tendency to seek out tropes to evaluate content extended to episode-level information. U4 cared a lot about what happens during episodes of Cold Case Files and would cherry-pick them based on their plot.

I am very particular about the kind of unsolved mystery or cold case they’re dealing with … The ones where it’s like, you can kinda tell it’s a suicide and the family can’t accept it is too depressing for me.

She even recognized language within an episode description that suggested a resolution, saying “this one implies that there is an ending to it to it because ‘unraveling a dark family secret.’”

U4’s browsing behavior—parsing through descriptions of episodes looking for tropes—is in part driven by her knowledge of these tropes and her preferences for some over others. This user is not merely interested in murder mysteries, she specifically wants the murder mysteries where it is obvious there was a murder, and it will be resolved. Because she has seen so many murder mysteries, she knows what she likes, what information to look for, and where to find it.

But tropes and the narrative elements they suggested were not always as crucial to participants. U1, who also enjoys crime television, seemed to care a lot less about the plot of these shows and would put them on as background noise. In fact, he experienced a formulaic plot as an asset, citing his relationship with Criminal Minds.

I don’t really care about, you know, the greater plot, if fucking Spencer Reed is gonna get his girlfriend or not. I’m just there for the skit, the routine. What is it? It opens up. Some guy’s getting fucking murdered. Where is he? Who is he? The team flies out. I like that repetitiveness, that’s why my number one genre on Netflix is probably crime.

U4’s meticulous selection process, driven by her desire for particular plot outcomes and her ability to discern narrative cues from episode descriptions, highlights how specific tropes can significantly impact viewer choice. Conversely, U1’s appreciation for the predictability and routine of shows like Criminal Minds suggests that the familiarity and repetition of tropes can also enhance viewing experience. However, they can also detract from a film’s attractiveness, as seen with U6’s evaluation of Hunter Killer. This suggests that the significance of tropes varies based on individual viewer investment and preference. Recognizing a trope requires first knowing it exists; therefore trope recognition was limited to genres that the participants knew well. Tropes are one way that users go beyond the coarse genre labels to make fine-grained classifications necessary to identify what they really like.

Associations

Browsing the Netflix homepage while instructed to think aloud placed participants in a state akin to free association, a spontaneous, non-linear process of making connections. This process revealed how participants’ minds naturally navigated through their own web of references and experiences. As a result, another significant method emerged for how participants classified content: by making associations, which involved directly linking titles with other titles or groups of titles due to their perceived similarities. While both tropes and associations serve as mechanisms for content classification, they operate on different levels of narrative understanding and viewer engagement. Tropes are about identifying specific narrative patterns, whereas associations involve drawing broader connections based on thematic or stylistic parallels to specific titles the user knows. Participants made associations based on information provided to them by the NRS and their own database of references. These associations would naturally come up during the browsing tasks. For example, U8 noticed The Kissing Booth, which she had seen. When I asked what she thought, she said “it was fun … similar to Never Have I Ever, like, same genre … something of a teenage person’s life … getting into university, both of them same, and then navigating love lives, things like that.”

Associations happened most often when participants were looking at thumbnails and previews. I asked U3 why she hovered on The Family Business. “That one did catch my eye. It interests me because I feel like it could be a Succession kind of thing.”

This hunch is unsurprising given the thumbnail’s visual and stylistic cues. The depiction of two adults in suits, engaged in a confrontational stance, coupled with the use of a shallow depth of field and a cool-toned color palette, are distinctive visual elements characteristic of Succession, suggesting a thematic or aesthetic parallel between the two shows.

While considering Young Sheldon, U5 remarked “I guess it brought me in because it gave like The Office;” perhaps this was because the title was written in a typewriter font. It is unclear if Netflix is intentionally mimicking the graphic design and stylistic elements of other shows; regardless, these subtle nods were perceptible and effective.

Other times participants would associate a title with a broader style of television rather than a particular show. The preview of Hack My Home reminded U3 of HGTV; “This could be good. I love HGTV. This is a Netflix series, so it’s not on HGTV but it seems like that it’s that kind of style.” Associating Hack My Home with a style of television that she already knew and liked enabled U3 to confidently settle on this as her final choice.

Associations allow users to make sense of unfamiliar content by connecting it to what they already know or enjoy, illustrating the importance of cultural knowledge in browsing decision-making. Much like tropes, participants were only able to make associations to content they have seen or heard of. By grouping unfamiliar content with familiar categories based on their perceived similarities, participants were able to make sense of new material.

Perceived Audiences

The Netflix users I spoke to also classified content based on who they thought would watch it. In this case, people would ascribe content to particular social categories, namely, age, race, and gender. Take this example of U6 explaining why he was uninterested in some of the titles in the “TV Comedies” row while browsing the impersonalized homepage:

Gilmore Girls, not my genre, probably for women.

Good Girls, not my genre, probably for women.

Um, Tom and Jerry, literally a cartoon for kids.

[Scooby Doo] Cartoon for kids.

[Nicky Ricky Dicky & Dawn] Um, maybe something for like pre-teens or something, this looks like children, Disney show.

Um, Cobra Kai, I know … I’ve watched this before, it just didn’t hook me.

Jane the Virgin. Probably for women, or kids, or having kids, having babies. Not interested.

U6 did not just say these titles were not for him, he suggested who they might be for. His intuitive judgments draw upon stereotypes related to both the shows themselves and the social categories he assigns to them. For U6 to classify Gilmore Girls, Good Girls, and Jane the Virgin as “probably for women,” he must tap into a broader understanding of which themes and narratives are traditionally marketed to or associated with female audiences. This implies awareness of the societal norms that dictate certain genres or storylines as being of particular interest to women.

Sometimes participants made these classifications based on their personal experience with particular demographic groups. U10 liked history documentaries, so I asked him about Age of Tanks, a documentary about WWI. After hovering over it to watch the preview, he concluded that “this is for dads, like, middle-aged white guys that are a little too invested in World War II.” This classification was rooted in personal experience, given that U10 said Age of Tanks was “something my stepdad would watch” and “maybe I’d watch it with him if it was already on and I was bored.” By ascribing certain shows and genres to specific demographic categories, users like U6 and U10 navigate their content consumption choices through the lens of social identity, potentially limiting their exposure to content they deem to belong to another social group.

Being recommended children’s content in the impersonalized environment surprised a few of the participants who did not typically see this on their own homepages. U5 was shocked to see Coco Melon in the trending row, “I’ve never seen that before … because that’s always in the kids one.” During one impersonalized browsing task, U3 remarked, “not interested. All this stuff looks like it’s for kids.” U3’s discernment of this content as being “for kids” showcases the application of cultural knowledge in the assessment of age appropriateness. This ability to instantly categorize shows within age-specific domains indicates a sophisticated understanding of the visual and thematic cues that signal content’s intended demographic. Rather than merely reflecting on personal preferences, such judgments rely on an extensive backdrop of cultural contexts and industry norms that dictate content marketing for different age groups.

Participants classifying content based on who they deemed it was for demonstrates how Netflix users utilize their cultural insights to navigate the vast media landscape, guiding their selections towards what aligns with their identity or away from what is perceived as outside their demographic relevance.

Situational Suitability

Beyond simply identifying whether a program is something that they might like, participants also decided whether the program fit their current situation or “mood.” Participants often encountered content they liked but did not suit the situational prompt I gave them or what they felt like watching at the time. In this case, evaluating content was not a question of “is this for me?” rather, “is this for now?” When this happened, participants classified titles based on where or when they deemed it appropriate to watch them. For example, U7 rejected Money Heist because he was not in the mood to get riled up by an action thriller at the end of a workday.

Like I said, I want to relax, right? I don’t want to keep stuck on my toes because of an accident and stuff. So I don’t want to watch that. I’d rather go for one that makes me relax and smile…

Because of his previous exposure to action thrillers, U7 knew that watching Money Heist would be a high-energy viewing experience. His discernment highlights the fact that users engage with Netflix not merely for entertainment but as a tool for managing their emotional states, seeking content that aligns with their immediate psychological needs. Some participants expressed that they did not like watching particular forms of content alone. U2 said “I don’t really like watching reality TV by myself” and U1 said “movies I can only ever watch with other people.”

Most participants expressed that sometimes they liked to watch things passively while folding laundry, unloading the dishwasher, or doing other tasks. When I asked U3 whether she liked to multitask while watching TV, she said it depends on the show.

For something like Suits, yes. For something like Fall of the House of Usher, no … because that’s one where I’m watching for the story … I actually care about what's happening … It’s really about whether or not I feel like I have to pay attention.

Attention and time were often cited as constraints while choosing content. Nyad had been on U3’s list of movies to watch since her girlfriend recommended it. However, for the first browsing task, she was asked to find something she was in the mood for, and she did not want to watch a movie. “Too long … otherwise I would have chosen that.”

There were numerous cases of participants desiring to watch a title outside the context of their browsing task. U5 was prompted to find something to put on in the background. He initially shrugged off Pablo Escobar, but after seriously considering it for a minute—explaining how he believed Pablo Escobar is one of the lesser discussed narcos and the possible reasons why—he admitted, “I feel like I’m too into it now. I feel like I wouldn’t play that in the background … Like, I’d want to watch it, watch it.” The distinction between wanting to “watch it watch it” versus a more passive form of consumption was something alluded to by other participants. U4 would watch Scarface, but not casually. “It is something I would watch on a plane though.” For her, finding something to watch passively was difficult because it forced her to sacrifice on standards.

The hard part is, I need a show that is like a B rating in my mind because I can tune out, and it won’t matter. But it’s so hard because I have pretty extreme opinions.

Of course, determining what a B rating is requires cinematic taste, which encompasses the user’s understanding of genre conventions, narrative complexities, and thematic nuances. This balancing act requires selecting content that is sufficiently engaging to hold interest without demanding full cognitive engagement, allowing viewers to enjoy it without the need for intense focus.

The context-dependence of viewing preferences poses an obvious challenge for the NRS: recommending the correct program not only requires knowing the user’s preferences, but also their mood at that moment. This is one reason why the user must collaborate with the NRS to identify the right show for the occasion. Users often select shows or movies aligned with their immediate feelings or the specific context of their viewing, adopting a discerning method for choosing content that suits the present moment rather than just their general preferences. Viewing habits are influenced by the desire for passive consumption, indicating that not all viewing is active or requires full attention. And there is a conscious negotiation between wanting to be fully engaged with content versus selecting something that allows for multitasking, reflecting a context-specific balancing act while browsing. Participants classifying content based on their ideal viewing situation demonstrated their awareness that sometimes the best choice in theory was not always optimal in practice.

Social and Moral Acceptability

Some choices are more contentious than others. Within a particular cultural context some shows are deemed acceptable or appropriate to watch while others are not. The social and moral acceptability of a film depends on its viewing context and the background and values of the audience. U7 explained that some of his selections are based on his culture experience as a Nigerian immigrant “and how I negotiate that contestation between where I’m from and where I am now and what is socially acceptable here.” He spotted Woman King, a movie he already watched and liked, which led him to explain some of the considerations he makes while selecting content.

In Nigerian movies, before now, they moderate certain contents because of its social acceptability, so um, gender kind of movies were not really made in Nigeria because, of course, it's a highly patriarchal society. So, you won’t really find movies like Woman King trending in Nigeria. But because I have that interest in international norms … movies like this would appeal to me, regardless of the fact that I’m coming from a patriarchal society where those things are not. These are some of the, um, should I say, silent conversations I have within me while I’m thinking of what movies to watch … but again, I don’t want to watch movies that are extreme because I also want to be able to have conversations about them and feel comfortable having those conversations with my friends, especially with my family and my wife.

Media consumption is a site of cultural contestation, and this is particularly pronounced for individuals in diaspora. Being able to comfortably talk about what he watches with his family was important to U7, demonstrating that movies and TV shows are not just personal entertainment, they are also mediums through which social relationships are maintained and cultural values are negotiated and transmitted.

Other users were conscientious of the political impact of their consumption habits. U4 noticed SWAT by recognizing Shemar Moore but decided not to watch it because she did not want to glorify the police.

Wait, is this the guy who’s in Criminal Minds? … I love that man, so much … But the hard part is, too, I want to be, at least semi culturally conscious when I'm picking police-themed TV shows, because I’m like, am I glorifying the profession by watching it?

While she deemed SWAT socially unacceptable because “it felt too pro police,” other users were unbothered or did not consider this altogether. Two other participants chose SWAT as their final selection for their browsing tasks. This discrepancy suggests that while some content is fine to particular social groups, others may deem it unacceptable based on their political beliefs.

There were also instances where content did not align with users’ morals. While browsing romance reality, I asked U9 what she thought of Down for Love, a Netflix original series that follows adolescents with Down syndrome while trying to navigate the complicated world of dating.

That one I’m not so sure about. I’ve seen the other one, Love on the Spectrum, and it made me feel kinda weird.

[Interviewer: Why?]

I don’t know. Something about it isn’t right. It treads this odd line where it’s like, on one hand, the show will be very heartwarming and hopeful and like, human and kind. But then it will also try to be funny and entertaining and will make fun of their autism. It feels predatory … I have very complicated feelings towards that show, and this feels very similar to that, so no.

U9’s discomfort, characterized by the perceived exploitation of the show’s participants for entertainment, reveals the moral ambiguity of certain content choices. This instance diverges from social acceptability in that it is not just about societal standards but requires the participant to take moral inventory and judge whether content aligns with their personal values.

Observing this phenomenon was difficult given it is a socially undesirable case. People are uncomfortable talking about why they or others deem certain content taboo. As U7 said, these are the “silent conversations” he has inside his head. However, the testimonies of U4, U7, and U9 suggest that social and moral acceptability is a consideration Netflix users make while determining what to watch. Classifying content based on whether it aligns with personal and community values highlights how viewers’ choices are deeply influenced by their cultural backgrounds, personal beliefs, and the broader societal and political implications of their media consumption. These considerations, similar to the audience-related concerns raised by the non-clickers from Ellison’s study, affirm that media selection is a complex negotiation, reflecting the interplay between individual viewer identities and the cultural and social contexts in which they exist.

Having explored how participants navigate the Netflix homepage, employing cultural knowledge to classify content, our investigation naturally progresses from the individual’s thought processes to their interactions with the mechanisms that shape and guide these experiences. The browsing experience on Netflix is a dynamic, non-linear process described by iterative funneling, a theoretical model I propose that categorizes user interaction into four distinct phases: search, discovery, evaluation, and selection. With this model in mind, I discuss some of the browsing strategies used by the participants I interviewed; most of these tactics either narrowed search criteria or were information-seeking behaviors aimed at helping the user evaluate content. Some users’ strategic navigation and evaluation of content demonstrate an active negotiation with algorithmic suggestions, revealing a complex interplay between user agency, algorithmic mediation, and cultural knowledge in the content selection process.

Iterative Funneling: A Speculative Model of the Browsing Process

This section will propose a model to describe and delineate the browsing process as it is experienced by users, who are channeled by the platform into particular paths of action. This model arose out of patterns in my data and observations. However, the claims I suggest about the cognition involved in browsing cannot be proven with interview data—as Vaisey notes, the semi-structured interview “puts us in direct contact with discursive consciousness but gives us little leverage on unconscious cognitive processes” (Vaisey 2009 1688)—so these claims are purely speculative. I present them to suggest that researchers within psychology and neuroscience could continue investigating the types of thinking involved in browsing.

Picture a funnel: on one end lies every option (in our case, the entire Netflix library), and on the other, one final choice. As the user moves through the funnel, they make dozens of tiny decisions—where to look, which rows to scroll through, how many, whether to hover on a title to watch its preview or not—often subconsciously, often half-heartedly, often backtracking and changing their mind.

Every browsing session begins with a search. Most of the time, participants started by glossing through the homepage, but in a few cases, participants immediately narrowed the search criteria by looking for specific titles, genres, or formats. At this point, participants wandered, somewhat aimlessly, waiting for something to hook them. During this search phase, participants made snap judgments about content based on their thumbnails (e.g. U6 quickly dismissed titles as “not my genre”), often without conscious deliberation or reasoning. The images the user encounters in the first phase are not enough information to decide whether they want to watch it, but they are enough to determine whether they want to pursue it further. When something eventually caught their eye, participants would “discover” it.

A multitude of factors led to discovery: previous exposure (e.g. U4 recognized Shemar Moore in SWAT), visual cues embedded in thumbnail (U3 was drawn to The Family Business because it reminded her of Succession), and its relative position on the homepage. Despite being instructed to think aloud, participants did much of this early navigational work silently, instinctually, automatically, and without direction. This observation leads me to suggest that during search and discovery—the shallow mode of browsing—participants embodied a more practical consciousness. Since they could not realistically examine every option presented to them with a fine-toothed comb, they relied on “quick and dirty” inferences to make decisions efficiently.

When I tried probing decisions made during these phases, participants were often frustrated or at a loss for words, since I was essentially asking them to verbalize internalized schemas that seemed obvious to them. For instance, I asked U10 why he ignored 3 Body Problem, which was the marquee image on his homepage. He said “I don’t know. Seems boring.” When I asked why, he flipped the question back on me. “I don’t know, why does anything seem boring? It just does. I don’t have a reason.” Oftentimes when I asked participants these sorts of questions, they would provide rationale. I asked U4 why she clicked on the thumbnail for Blue Zones. She said, “the reason I was drawn to it was because they had a very bright color and I like their different fonts.” However, “discursive consciousness is incredibly good at offering reasons that may not be at all related to the real motives behind a person’s behavior” (Vaisey 2009 1688).

While participants began by doing a rapid pass over a wide swath of options, they switched techniques as they narrowed in on individual titles. During the early phases of browsing, participants were guided by interest and intrigue, but once they had singled-out a particular title, their actions became deliberate information-seeking behaviors guided by specific questions like, “how many seasons does this have?” (U2), “is this in German?” (U1), and “when was this made?” (U10). To seek out the additional information necessary to evaluate content, participants watched previews, read synopses, and examined the evidence provided to them by the NRS with the aim of finding something that would persuade or dissuade them from watching.

When evaluating a single title, participants were able to point to specific details that turned them on or off. For instance, U1 was originally drawn to Manifest because he likes planes. Judging from the thumbnail, he thought “it was gonna be like, plane mystery investigations, or maybe a docufiction about this hijacking, or maybe even something about the history of the commercial aviation industry.”

However, he was disappointed when he read the show’s description and learned this was not the case.

That whole gimmick, “we’ve disappeared,” even the Avengers has done that. And also, the fuck? The plot? That makes no sense from the get-go? A plane is flying mysteriously for years and then lands and everybody’s lives messed up because they’ve been gone for years… You're asking a lot from your audience within the first sentence of your bio. Fuck physics, fuck time … Yeah, I was just there for the planes.

Many cases of misalignment like this occurred, where there was a discrepancy between participants’ first impressions and what the content was actually about. When this happened, participants had specific rationale and information to justify their choices. Because these considerations were explicit and articulable, I would suggest that during the latter phases of browsing, participants embodied a more discursive consciousness. After completing what the participant deemed to be an exhaustive search, they would come to a final selection.

A single browsing session can be broken down into these four distinct phases: search, discovery, evaluation, and selection. As this iterative process unfolds, the user makes increasingly granular and deliberate judgements, oscillating between the realm of practical to discursive consciousness. The five classification systems I outlined in the first section do not directly correspond to the four phases of iterative funneling, they were used by participants throughout the entire browsing process as filters to understand unfamiliar content and disqualify potential candidates. For instance, U6 identified a trope while looking at the thumbnail of Hunter Killer (discovery phase) and U4 did so while scrutinizing episode descriptions of Cold Case Files (evaluation phase).  

The iterative funneling model is useful because it delineates the process of browsing into concrete phases while accounting for back-and-forth navigational behaviors. The model also provides a framework for comparing user behaviors across different digital platforms or within different browsing contexts. By applying the model to various scenarios, researchers can explore how platform affordances, content types, and user goals influence browsing behavior. Following Vaisey’s dual model of culture in action, I would suggest that in the early phases of browsing (search and discovery), cultural knowledge primarily operates through practical consciousness as a motivational force guiding users implicitly; at the latter end (evaluation and selection), it shifts towards discursive consciousness, serving as a justification for the choices made.

Settling

Many participants experienced browsing as a process of settling—reaching the point at which they decide to stop searching and opt to watch a show or movie despite it not their first choice. This primarily served as a method to achieve satisfactory content choices under limited time constraints. Most settlers experienced something in line with a phenomenon described by U1: “I see something I half recognized. I wouldn’t mind watching it. Let’s see. But I’m not totally convinced by it, so I keep scrolling, and then everything else disappoints me.” The most common tactic employed by settlers was having backup options. The way this worked was while browsing, participants would flag items to potentially return to. At the top of the page, U7 was recommended The Flash, which had been on his radar for a while.

What I do when I see a recommendation like this is to see if there’s any other one that is more interesting than it. If not, then I come back to it.

After browsing for a few minutes, U7 ended up selecting The Flash. As U3 explained, “it’s nice to have the fallback because if it’s been like a minute and I haven’t found something else, I’m just gonna be like, ugh, fine. It’s a lot of that.” Some users alluded to there being a hierarchy of backup options. When I asked about Leave the World Behind, U6 said “that would be like a back pocket.” He moused over Bad Surgeon. “This is like back back pocket.” The strategy of settling demonstrates an awareness that rarely users are going to find the perfect movie or show, most of the time they are looking for the next best thing.

Proactive Engagement: Platform Affordances and Browsing Strategies

The architecture of the Netflix homepage is central to understanding how the NRS mediates the browsing process and how participants use features of the platform to facilitate their search, discovery, evaluation, and selection of content. While most participants relied on the NRS to provide them with adequate recommendations on their homepage, more savvy Netflix users will leverage the platform’s affordances to extract the information they deem most useful to evaluate content. Most of the browsing strategies I observed were aimed at narrowing search criteria and sourcing additional information to aid evaluation.

Narrowing Search Criteria

While the majority of participants’ search was contained within the homepage, some participants employed browsing strategies—namely using the “similar to” search feature and filtering by microtags and generic categories—aimed at narrowing the scope of their search. Users effectively reduced the amount of work they would have to do during the initial search phase of browsing. The majority of participants stayed exclusively on the homepage while they browsed. Of these users, some explored only the first few rows on the page and would usually decide on something to watch within two to five minutes. For these users, the ranking and sorting of the NRS significantly impacted their navigation through and selection of content because naturally, users would gravitate towards what they saw first. U4 explained how the NRS’ ranking and sorting guided her to navigate to somewhat arbitrary places out of convenience.

The earlier up it is in these rows, the more I’m drawn to it … like, “Familiar TV Favorites.” I don’t have a reason to really be picking at this selection… I went there because it was the first thing.

In a similar vein, U3 assumed content depreciated in quality as the search went on.

I also always find that whatever they’re showing you first, is always better than whatever comes next … the first 6 clips I always think are better than whatever the next 24 are … So it is pretty rare that I’ll really scroll through the categories.

Here, she is making implicit assumptions about quality based on the show’s positioning on the page, which is determined by the ranking and sorting algorithms of the NRS. These assumptions are not unfounded. Netflix’s behavioral research indicated that users are more likely to scan vertically than horizontally (Alvino and Basilico 2015), therefore, content the NRS deems most promising is placed towards the top left of the page.

To go beyond homepage recommendations, participants employed filtering tactics to narrow the criteria of their search. One of the most common strategies was participants searching for titles that were not on the platform to see algorithmically generated recommendations of similar content. As U1 explained, “I’ll put a title of a movie that I know is not on Netflix, but they do the thing where they find similar things to it.” Four participants said they have done this before and three of them did it during the interview. U4 explained that in order to use this feature to find shows like Criminal Minds, she would start by looking at all the shows Netflix “thinks” are similar and “find the one that like, has the right qualifications for why I liked Criminal Minds.” Her approach reveals an essential layer of human judgment required within the interaction with the NRS. While the algorithm provides a list of ostensibly similar shows, the crucial task of discerning which of these recommendations aligns with the specific reasons she appreciated Criminal Minds falls squarely on her. Searching for “similar to” titles enabled participants to leverage the NRS’ advanced categorization capabilities—as of 2014, Netflix had 76,897 altgenres (Madrigal 2014)—while controlling the input keywords to generate more tailored results. This tactic allowed participants to exercise agency within a system that does not grant much. However, the “similar to” search feature was not without its limitations. U4 explained that sometimes she felt like “the Netflix selection in ‘shows like this’ isn’t that good.” She was confused why they listed some shows as like the one she was looking for; “like maybe it’s just the same director but it’s something completely different.” When the NRS’s “similar to” recommendations failed to align with the reasons why she liked a particular show, she would turn to Google.

When participants decided to browse by genetic category or microtag, it was because they had something specific in mind, like if they wanted to watch a particular genre, in a particular language, or a TV show instead of a movie. Two participants with more technical literacy and familiarity with the Netflix platform used microtags to help them browse. U2 employed the most intricate and sophisticated browsing strategies of any user I spoke to. During two of her browsing tasks, she utilized micro tags to help her browse, looking through tags entitled “Romantic,” “Reality TV,” and “Columbian.”

She explained that she was looking at the Columbian tag “because I don't watch enough stuff in Spanish and I prefer Colombian Spanish, obviously.” But she had to parse through these options since “so many of them are about narco trafficking, which I’m not into.”

Some participants said time constraints limited their ability to browse as thoroughly as they would like. For instance, U8 has browsed using microtags in the past, saying, “I used to be very picky, I used to take a lot of time to find what I wanted to watch,” but now she is under stricter time constraints while pursuing her master’s degree so “I just put on something because, you know, I have to eat and then get stuff done later, so.” Deeper browsing requires time and persistence, because of this many participants relied heavily on the NRS and experienced browsing as a process of settling.

Sourcing Additional Information

Sourcing additional information beyond the thumbnail was necessary for participants to make selections. Participants explored further details by watching previews, reading descriptions, and reviewing the various pieces of evidence provided by the NRS. None of the users I spoke to were able to make a selection without doing some research. While the extent of this research varied, it was always a crucial step necessary to evaluate content. Most of this research happened within the Netflix platform with participants relying on information provided by the NRS. Most of these techniques have been detailed or alluded to in previous sections of this paper, such as hovering over a thumbnail to watch a preview, clicking on it to expand its details, and reading descriptive adjectives and synopses. With each successive action, participants had access to more information (see fig. 3.2).

In rare cases, some participants would conduct research outside of Netflix on the internet. For instance, U2 was interested in Surviving Summer, an Australian drama about a group of teenagers who surf. She watched the trailer on Netflix but needed to know more, so she googled “surviving summer” and clicked on its IMDB page.

During the browsing task review, she explained what she was looking for.

I wanted to know who Ari is. Because there was a guy with a girlfriend that she seemed kind of mad that he had a girlfriend, like the main character. So I was wondering if that was the guy that she was like, interested in, surfing-wise, and I think it was.

In this scenario, U2 was using the internet to learn more about specific character dynamics. Other times, participants used the internet to find spoilers. U4 was considering Hellbound, a Korean drama about a cult-like entity founded on the idea of religious justice. She said she might watch it but “would have to honestly look online and also see the reviews” first. She googled “hellbound” and clicked on its Wikipedia page. She scanned the first two paragraphs and scrolled towards the bottom of the page. I asked what information she was looking for. She said, “I’m looking to make sure that the religious group has, like, a horrible downfall.” Once she confirmed this, U4 selected Hellbound.

Participants invariably sought out more information beyond the initial presentation of content on Netflix to make informed decisions about what to watch. They did this by utilizing features within Netflix, such as hovering over a thumbnail or clicking on it to see the expanded details to access more information about content. This reliance on the platform’s built-in tools underscores the importance of these features in guiding users’ decision-making processes. This suggests that the initial algorithmic recommendations and visual cues provided by thumbnails are starting points rather than endpoints in the content selection process. Some users, particularly those with specific interests or questions, went beyond Netflix to conduct external research. This was done to gather detailed information about character dynamics, plot specifics, or to find reviews and spoilers that are not available within the Netflix platform.

Users engage with the system, leveraging its recommendations to refine their search, but also contribute actively by classifying content, employing browsing strategies, and sometimes trying to “game” the NRS for more relevant suggestions. Many participants spoke about browsing as being a collaborative effort between them and the NRS. Some users liked to do this work more than others. U3 found it tedious. “I really don’t like to browse. I want to just start watching the show. This is like work, and I don’t want to do work.” This made her more comfortable with settling. On the other hand, users like U2 and U4 conducted extremely thorough searches by using micro-tags, the “similar to” search feature, and by close-reading episode descriptions. U2 believed she puts in more work than the NRS, and moreover, that she games the system:

I think I work it more than it works me … I think I take what it gives me and maximize it, rather than them laying it all out for me … It shows me a lot of options of similar content that I’ve watched. But like, I think I do the hard work of going through all of that to see something that is actually similar to what I watched and not like, different. Well, I think it does very basic level curation, but I don’t think it actually can predict what I like.

The perspective offered by U2 suggests a dissonance between the system’s intended functionality of delivering highly personalized content and the reality that for some especially picky users, it fails to deliver fully on this promise. This sentiment, “I think I work it more than it works me,” reveals strategic engagement with the platform, where the user assumes a more proactive role in navigating the options presented by the NRS, rather than relying on the system’s automated curation. Despite her characterization of browsing as labor, U3 recognized the necessity of her participation in the search process, saying “I guess I just don’t really expect it to be giving me the good stuff. I expect to be finding it.” Participants recognized that relying on algorithmically generated recommendations could bring them close, but only so far.

At the end of the interview, I asked participants to compare their experience browsing the personalized and impersonalized environments. Most participants found it easier to browse their own homepage and found personalization to be generally helpful. U3 said “I think mine kind of jumps into genres pretty quickly. And here you get genres, but they’re way less specific.” U8 noticed in the default profile “there are no Indian movies,” compared to her homepage which had many. It is clear that the NRS does not merely curate content; it approximates and adapts to the habitus of its users. When asked to describe the process of browsing Netflix, U6 compared it to finding a culturally specific grocery store:

It’s like going to a grocery store in your neighborhood even like arguably to your culture. So let’s say like for me if I was to go to a Mexican grocery store. It kind of knows what you want to a certain extent, right? This is foods in your culture. It’s gonna have things you like. It’s gonna piece them together a little bit. Whereas if I was making a Mexican meal and I went to Whole Foods, it’s gonna be a lot harder, right? Whole Foods is gonna be a little bit more generic. Whole Foods is gonna have the end-all be-all for everyone. And like, you’ll get close enough, right? You’ll find something, of course. It’ll be good … But um, as you navigate the system, you start to find your grocery store.

This analogy was apt, I would only ask, are people really “finding their grocery store” or is Whole Foods changing the items on the shelves? In this metaphor, U6 is recognizing that the NRS intuits information about his cultural taste and caters to those tastes. By using data to intuit user’s backgrounds, habits, and cultural context, the NRS attempts to approximate the habitus of its users.

Users engage with the NRS in varied ways, from passive acceptance of recommendations to active searching and manipulation of the system to find content that aligns more closely with their preferences. This suggests a spectrum of reliance on the NRS, where some users’ interactions are characterized by proactive engagement, while others would rather leave the burden of choice to the NRS. Discussions around personalized and impersonalized browsing experiences reveals the importance of cultural specificity in content recommendation. Users like U8 noted a stark difference in the availability of culturally relevant content between personalized and default profiles, underscoring the NRS’s capacity to adapt to users’ cultural backgrounds and tastes.