Which Factors Affect Online Video Views and Subscriptions? Reference-Dependent Consumer Preferences in the Social Media Market

Abstract

In the attention-driven environment of online video platforms, understanding the factors that influence content selection and channel subscriptions is crucial for creators, marketers, and platform managers. This study investigates how thumbnails, view counts, video length, genre, and the number of advertisements affect user decision-making on YouTube. Grounded in random utility theory and reference-dependent preference theory, this study conducted a choice experiment with 525 respondents and employed a combined model of rank-ordered and binary logit methods to analyze viewing and subscription behaviors. The results indicate a significant preference for thumbnails with subtitles and shorter videos. Notably, we found evidence of reference-dependent effects, whereby a higher-than-expected number of ads decreased viewing probability, while a lower-than-expected number significantly increased subscription probability. This study advances our understanding of the factors that influence user behavior on social media, specifically in terms of viewing and subscribing, and empirically supports prospect theory in the online advertising market. Our findings offer both theoretical and practical insights into optimizing video content and monetization strategies in competitive social media markets.

1. Introduction

The rapid rise in social media and digital platforms has fundamentally transformed the media industry, shifting content consumption from traditional one-way broadcasts to interactive, user-driven media [1]. YouTube, launched in 2005, exemplifies this transformation as the world’s largest user-generated video platform. It allows users to freely upload and share videos, and the site is projected to reach over 2.7 billion users worldwide in 2025 [2]. An enormous supply of content matches this massive reach—with over 500 h of video uploaded to YouTube every minute—which underscores both the explosive growth of the platform and the intense competition for viewer attention [3].

Understanding consumer viewing preferences is critical for content creators’ success on such platforms. In a highly competitive creator economy, individuals strive to increase video views by growing their subscriber base and by designing engaging thumbnails and titles that capture audience interest. View counts translate directly into advertising revenue, while subscriber numbers signal a creator’s influence and reach. Thus, creators have a strong incentive to discern and cater to what attracts viewers [4]. By quantitatively identifying which content attributes audiences prefer, the present study can guide creators in developing more effective strategies—for example, optimizing thumbnail imagery or video length—to attract and retain viewers.

From the perspective of video platform providers, maintaining user engagement is equally crucial. These platforms rely on advertising as a primary revenue stream, so they must display ads in ways that do not alienate users. The online video advertising market continues to expand alongside the growth of video platforms, but user aversion to disruptive ads (e.g., excessive or unskippable ads) remains a significant challenge [5,6]. In response, platforms employ recommendation algorithms to tailor content to users’ interests and keep them watching [7]. If recommended video features or ad loads misalign with viewer preferences, however, users may skip content or abandon the platform. By identifying the elements that attract viewers versus those that trigger ad avoidance, this study can inform platform strategies to enhance user satisfaction while maximizing advertising effectiveness.

Previous studies have examined the factors driving video viewership and popularity on YouTube and similar platforms. For instance, YouTube’s recommendation system and video presentation features (such as thumbnails) have strongly influenced viewing behavior [8,9]. However, much of the existing research is either descriptive or focused on individual factors in isolation. The relative importance of different video attributes in viewers’ decision-making—and how these factors trade off against one another—remains insufficiently understood. In particular, there is a need for empirical research that quantifies consumer preferences across multiple video characteristics in a controlled setting. This study addresses that gap by employing a discrete choice experiment to measure how key attributes of videos (e.g., thumbnail imagery, view count, and advertisement quantity) impact a viewer’s likelihood to select a video or subscribe to a channel.

Our analysis is grounded in well-established consumer choice theories. We draw on random utility theory, which models individuals as choosing the option that maximizes their personal utility from a set of alternatives [10]. In addition, we apply a reference-dependent preference perspective derived from prospect theory [11,12]. The reference-dependent framework supposes that people’s evaluations of outcomes are relative to a reference point (such as their current expectations or status quo), leading to asymmetric responses to gains and losses. Incorporating this perspective allows us to account for how certain video attributes (for example, an increase or decrease in the number of ads relative to a viewer’s expectation) can differently affect viewing decisions.

This study focuses on the context of South Korea’s online video market. South Korea offers an ideal setting due to its advanced digital infrastructure and high social media usage. The country leads the world in smartphone penetration at roughly 95% [13] and was one of the first to launch commercial 5G mobile services [14], ensuring widespread high-speed connectivity. The YouTube usage rate among Koreans is expected to reach 84.9% as of 2024 [15]. With ubiquitous access and a strong appetite for online video content, Korean consumers provide a pertinent sample for examining reference-dependent preferences in video choice behavior. Accordingly, we implement our discrete choice experiment with a sample of Korean YouTube users, analyzing how recommended video features influence their video selection and channel subscription decisions.

The remainder of this paper is organized as follows: Section 2 reviews the literature on online video consumer behavior. Section 3 describes the methodology, while Section 4 presents the data and estimation results. Finally, Section 5 discusses the conclusions and implications.

2. Literature Review

Online video platforms such as YouTube have fundamentally changed how content is consumed, drawing billions of users and countless hours of viewing daily [2]. The Attention Economy Theory further contextualizes these individual choices at the platform level. In today’s digital environment, attention is a scarce resource. YouTube’s success depends on its ability to capture and retain viewer attention amid a flood of competing content [16,17]. Thus, even small differences in video attributes—such as an engaging thumbnail or the optimal balance of ad exposure—can significantly influence a viewer’s decision, ultimately affecting both content popularity and advertising revenue. This paradigm shift in the digital content consumption environment raises questions about how viewers make consumption decisions. Contrary to conventional media, which typically employs a one-way broadcast format for content delivery, digital platforms proffer users a plethora of options. Consequently, there is a necessity for an understanding of the trade-offs between video attributes. In this section, we summarize the random utility theory (RUT) theory and reference-dependent preference, which form the foundations of our analytical research. We then review related studies on the consumption behavior of digital content.

2.1. Random Utility Theory and Reference-Dependent Preference

A key theory underpinning our analysis is RUT, which has been widely applied to explain consumer choice behavior. RUT assumes that individuals faced with discrete alternatives will choose the option that maximizes their utility, which consists of a systematic component based on the observed attributes of the choices and a random error component [18]. In practice, digital content choices (e.g., which video to watch or which service to use) can be modeled as utility-maximizing decisions where each alternative (video, ad, subscription plan, etc.) offers a bundle of attributes contributing to the user’s utility. RUT provides the foundation for most consumer choice models, including those for online platforms and social media [19,20]. McKenzie et al. [20] employed the mixed logit model to estimate how video content features, quality, advertising load, and price affect user choice in the context of YouTube or streaming platforms. In this way, the choice model allows researchers to calculate the probability of choice and willingness to pay for specific features. Therefore, RUT is an effective theoretical tool for analyzing consumer behavior on digital platforms. Based on RUT, this study posits that a viewer’s choice is based on maximizing utility composed of both observable attributes (such as thumbnail quality, video length, and ad frequency) and unobservable random factors [18].

Reference-dependent preferences explain that consumers evaluate outcomes relative to a reference point—where losses (e.g., more ads than expected) weigh heavier than equivalent gains [21]. It originated from Kahneman and Tversky’s prospect theory [11], which argues that consumers evaluate outcomes relative to a reference point rather than in absolute terms. This leads to phenomena like loss aversion, where losses loom larger than equivalent gains [21]. In other words, the change in utility depends on whether it is viewed as a gain or a loss relative to the consumer’s status quo or expectations. Specifically, the value function in classical prospect theory is steeper in the loss region than in the gain region, meaning that consumers experience a greater decrease in utility from a loss than an increase in utility from a gain of the same size. This theory has influenced digital markets to explain behaviors like aversion to new fees or ads perceived as losses. For example, Kőszegi and Rabin [21] developed a formal model of reference-dependent preferences that has been applied in consumer choice modeling [21]. They found that consumers have reference points such as past experience or expected standards, and their utility from new consumption varies with these reference points. In online media, a reference point might be ad-free viewing or a free usage level, and introducing ads or payments could be experienced as losses, potentially causing dissatisfaction. Some researchers have extended this theory to digital contexts in case users adopt ads on a website as a reference-dependent process [22]. They found that users adapt to past levels of inconvenience over time, so strategic adjustments should account for both existing users’ adaptation processes and new users’ reference points to maximize long-term revenue. In the case of YouTube platforms, viewers’ expectations on the frequency of ads, subscription prices, or content quality can be the reference point. Therefore, we adopt these theoretical perspectives and consider a reference-dependent framework for analyzing how specific video attributes affect consumer choices.

2.2. Related Works on Consumer Behavior of Digital Contents

As mentioned above, it is important to understand what makes viewers choose a video and to quantify the impact of each factor. Previous studies have analyzed the factors that drive viewers to watch videos [23]. Influential factors include whether a video platform recommends a video [24,25], external interactions [26], view counts [27], the number of likes [9], the number of subscribers [9,28], thumbnails [9,29], and content credibility [30]. Chatzopoulou et al. [8] found that among the various elements of YouTube systems, the number of video views, comments, ratings, and likes were highly correlated with the popularity of a video.

Thumbnails are also one of the factors that influence video viewing. Thumbnails and titles provide information about the video while attracting potential viewers and persuading them to watch it [31]. Thumbnails are rated as the most crucial element of marketing strategies [29,32]. Wang et al. [33] analyzed the impact of positive and negative thumbnail emotions on view counts. The results showed that negative thumbnails increased the number of views; positive thumbnails also increased the number of views, but not significantly. Koh and Cui [9] analyzed the effects of thumbnails on video selection using a refined contingency model. They analyzed 3743 vehicle videos and used a refined elaboration likelihood model with celebrity endorsement, colorfulness, brightness, thumbnail image quality, and the year as variables. The analysis showed that celebrity endorsements, a higher number of subscriptions and likes, and a higher number of thumbnails are associated with higher views. The number of elements and colorfulness showed a U-shaped relationship with views, whereas object complexity and image quality showed an inverted U-shaped relationship. Sahu et al. [29] analyzed how thumbnails significantly influence content preferences, focusing on over-the-top (OTT) content. Using survey results and applying a one-way analysis of variance and simple regression, they found that thumbnail changes were a factor in video selection.

In terms of advertising, while traditional advertising is presented to an unspecified number of people, online video advertisements can identify and target users [34]. However, online video advertising also faces the challenge of overcoming user ad avoidance behaviors and poses some risks. Since consumers are more goal-oriented on the internet [35], online video advertisements can be perceived as much more intrusive than other media advertisements [36]. This perceived intrusiveness can lead to negative attitudes toward and the avoidance of advertised products [5]. Therefore, most studies have been conducted on online video ad avoidance behavior. Previous studies have focused on the internal factors of advertisements. For instance, Hussain and Lasage [5] analyzed survey data. They found that the meaninglessness of ad content, lack of ad authenticity, and lack of interactivity caused online video ad avoidance behavior similar to the TV viewing context. Using survey data, Li and Lo [37] analyzed the effects of ad length, ad location in videos, and ad context matching on brand awareness. They found that embedding an ad in a video distracts attention and improves brand recognition. Belanche et al. [38] analyzed the effects of ad arousal stimuli, contextual congruence, and product involvement on ad effectiveness. Jeon et al. [39] found that advertisements that announce the remaining time of an ad can reduce advertisement stimuli compared with advertisements that do not; consequently, reducing time uncertainty can lead to lower skip rates.

On the other hand, studies have also focused on factors external to advertising. Several studies have focused on individual characteristics in response to advertisements. Campbell et al. [6] attempted to explain skipping behavior through emotions, cognitive processing, and persuasion. Other studies analyzed the interactions with ads in the context of videos associated with advertisements. Li et al. [40] reported that the peripheral elements of a video clip affect pre-roll advertisements. When analyzed with machine learning-based attribution techniques, they found that a negative value of the thumbnail image decreased pre-roll ad effectiveness. In contrast, a negative value of the title text increased ad effectiveness. Kim et al. [41] analyzed the effects of video features on ad avoidance behavior based on viewing logs from a Korean video platform company. The results showed that the longer the content length, the higher the popularity, and the more informative the genre, the more likely users were to watch the ad; the longer the ad length, the less likely they were to watch it. Meanwhile, content freshness demonstrated no effect.

Both RUT and reference-dependent theories have been applied in empirical studies of user behavior on YouTube and similar digital platforms. However, to our knowledge, only a few studies have analyzed video selection factors based on thumbnails and advertisements. Some studies were limited to specific genres, and those that analyzed the impact of the online video context on ads were limited to pre-roll advertisements. Despite the monetary value of advertising on video platforms, there is no attribution of the number of ads, which makes it difficult to determine the appropriate amount of advertising from an economic perspective. In addition, most studies have only focused on subscription decisions [42], video selection [43], and interaction between creators and consumers [44,45], but few have analyzed whether viewing leads to subscriptions. This study aimed to fill these gaps in the literature and present the results.

3. Methodology

This study analyzes consumer behavior on an online video platform—specifically YouTube—focusing on two linked behaviors: video viewing and channel subscriptions. Our approach is grounded in RUT, which posits that when individuals are faced with a finite set of alternatives, they choose the option that maximizes their utility. In this context, a viewer’s utility for a given video or subscription choice is assumed to be a function of observed attributes (e.g., video features) and an unobserved stochastic term. Based on this framework, we employ discrete choice models to examine the factors influencing video selection and to explore how viewing choices relate to subsequent channel subscriptions. Discrete choice models are well suited to analyzing decisions from a limited set of alternatives because they are based on clear stochastic assumptions regarding the unobserved portion of utility. In a general discrete choice model, the utility $U_{njt}$ that respondent n choosing alternative j from choice set t is expressed by Equation (1):

$$U_{njt}=V_{njt}+{\epsilon }_{njt}=\beta \prime x_{njt}+{\epsilon }_{njt}$$

(1)

In Equation (1), we show the deterministic term $V_{njt}$ and a stochastic term ${\epsilon }_{njt}$. $V_{njt}$ consists of $x_{njt}$ representing a vector of each alternative $j$ as faced by respondent n in choice set $t$ and ${\beta }^{\prime }$ representing the coefficient vector of attributes. Here, we examined whether an alternative can be expressed in the binomial logit form. The choice probability equation is expressed as follows:

$$P_{nj}={\displaystyle \frac{\exp \left(V_{nj}\right)}{1+\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{nj}\right)}}$$

(2)

In this model, we assumed that ${\epsilon }_{njt}$ follows a type I extreme value distribution. This is the same as assuming that the deterministic term of the no-choice alternative is zero. The probability equation of the multinomial logit can be expressed as an extension of the binomial logit. The probability equation of the rank-ordered logit can be represented in the form of a multinomial logit. Furthermore, a rank-ordered logit model uses rank data that provide a preference sequence; thus, the probability that the respondent $n$ chooses the rank of alternatives such as ${\mathrm{r}}_1\mathrm{,}{\mathrm{r}}_2\mathrm{,}\dots \mathrm{,}{\mathrm{r}}_{\mathrm{J}}$ from the choice set $t$ is expressed as Equation (3), which is used to determine viewing rankings. In addition, whether or not to subscribe is presented in the form of a binary logit and is expressed as Equation (4), where $d_J$ = 1 means subscribing to alternative $J$.

$$P_n\left(r_1\mathrm{,} r_2\mathrm{,}\cdots \mathrm{,} r_J\right)=P\left[U_{{nr}_1}>U_{{nr}_2}>\dots >U_{{nr}_J}\right]=\prod _{h=1}^{J-1}{\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{n{r}_h}\right)}{{\Sigma }_{j=h}^J\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{n{r}_j}\right)}}$$

(3)

$$P_n\left(d_J\right)={\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{nj}^{subscription}\right)}{1+\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{nj}^{subscription}\right)}}$$

(4)

We jointly estimate the rank-ordered and binary logit models to capture the sequential decision-making process: selecting a video and then deciding on a subscription. The combined probability equation can be expressed as follows:

$$\begin{gathered} P_n\left(r_1\mathrm{,} r_2\mathrm{,} \dots \mathrm{,}{r}_J\mathrm{,} d_1\mathrm{,} d_2\mathrm{,} \dots \mathrm{,} d_J\right)=P\left[U_{{nr}_1}>U_{{\mathrm{n}\mathrm{r}}_2}>\cdots >{\mathrm{U}}_{{\mathrm{n}\mathrm{r}}_{\mathrm{J}}}\right]p\left[{\mathrm{d}}_1\mathrm{,} {\mathrm{d}}_2\mathrm{,} \dots \mathrm{,} {\mathrm{d}}_{\mathrm{J}}\right]\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \\ =\prod _{j=1}^{J-1}{\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{nj}^{viewing}\right)}{{\Sigma }_{i=j}^{\mathrm{J}}\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{ni}^{viewing}\right)}}\prod _{\mathrm{j}=1}^J{\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{nj}^{subscription}\right)}{1+\mathrm{e}\mathrm{x}\mathrm{p}\left(V_{nj}^{subscription}\right)}} \end{gathered}$$

(5)

By jointly estimating these models via maximum likelihood estimation (MLE), we capture both stages of user decision-making—video selection and subsequent subscription. The MLE has many desirable properties, including consistency, efficiency, and asymptotic normality, making it a popular method for parameter estimation in a wide range of statistical models [10].

The estimated $\beta$ coefficient can be utilized for calculating the amount of marginal willingness-to-pay (MWTP) and relative importance (RI). MWTP quantifies the change in an individual willingness to pay as a certain attribute grows or decreases by one unit. This measure is derived from the compensation value, which is analyzed from a microeconomic standpoint. The concept of RI is quantified by the ratio of the part worth associated with each factor, which indicates the extent to which each attribute influences consumer decision-making. In this particular scenario, the calculation of $part-wort{h}_x$ involves the multiplication of a parameter by the disparity between the highest and lowest attribute levels. The calculations for MWTP and RI are performed using Equations (6) and (7).

$${\mathrm{M}\mathrm{W}\mathrm{T}\mathrm{P}}_k=-{\displaystyle \frac{{\beta }_k}{{\beta }_{price}}}$$

(6)

$${\mathrm{R}\mathrm{I}}_k={\displaystyle \frac{part-wort{h}_k}{\sum _{x}part-wort{h}_k}}={\displaystyle \frac{{\beta }_k(x_{k\mathrm{,}max}-x_{k\mathrm{,}\mathrm{m}\mathrm{i}\mathrm{n}})}{\sum _k{\beta }_k(x_{k\mathrm{,}max}-x_{k\mathrm{,}\mathrm{m}\mathrm{i}\mathrm{n}})}}\times 100$$

(7)

4. Empirical Analysis

4.1. Survey Design and Data

This study collected stated preference data using a choice experiment. A choice experiment is a method used to collect data by providing hypothetical circumstances in which subjects choose among alternatives from a choice set [46]. Respondents chose or ranked the preferred alternative to maximize their utility in the given assumed choice situation. Before the main survey, we conducted an online pilot survey with 300 participants. The results were used to refine and modify the design of the main survey. Gallup Korea, a specialized survey firm, conducted the main choice experiment survey. We conducted an online survey with 525 participants over 11 days, from 12 May to 22 May 2020. Stratified sampling based on the demographic characteristics of South Korea was used to collect data.

Table 1. Demographic characteristics of the sample.

Category	Characteristic	Respondents (n)	Percentage (%)
Total		525	100.0
Gender	Male	271	51.6
	Female	254	48.4
Age	20–29	103	19.6
	30–39	117	22.3
	40–49	118	22.5
	50–59	114	21.7
	60–69	73	13.9
Education Level	High school or less	81	15.4
	University/college	376	71.6
	Over graduate school	68	13.0
Income (unit: 10,000 KRW)	~100	23	4.4
	100~149	11	2.1
	150~199	25	4.8
	200~249	35	6.7
	250~299	40	7.6
	300~399	86	16.4
	400~499	108	20.6
	500~699	111	21.1
	700~999	65	12.4
	1000~	21	4.0

shows the respondents’ demographic characteristics.

The survey consisted of three sections: The first section investigated the general characteristics of the respondents’ decisions on online video viewing and channel subscriptions. The results show that 88.8% of participants had a viewing experience on YouTube. Additionally, their weekly average viewing time was 361.7 min, meaning they viewed YouTube for an average of 50 min or more per day. Regarding the prime time, the responses were in the order of 21:00–24:00 (36.19%) and 18:00–21:00 (24.19%). This indicates that more than 60 percent of total viewing occurred after work and before bedtime. The most common place to watch YouTube was at home, with 401 respondents, and the average number of channels subscribed to per person was 13.5. Consumers experience an average of 3.4 advertisements per 10 min of YouTube videos, and 17.38% of respondents pay for premium services to avoid advertisements.

Table 2. Sample statistics for YouTube viewing and subscription behaviors.

Type		Description
		Respondents (n)	Ratio (%)
YouTube viewing experience		466	88.76
Weekly average YouTube watch time		361.7 min
YouTube viewing time	0:00–3:00	19	4.08
	3:00–6:00	2	0.43
	6:00–9:00	9	1.93
	9:00–12:00	33	7.08
	12:00–15:00	44	9.44
	15:00–18:00	42	9.01
	18:00–21:00	127	27.25
	21:00–24:00	190	40.77
Viewing place (duplicate responses)	Home	401	86.05
	School or work	91	19.53
	On the go	115	24.68
	Other indoors	63	13.25
	Other outdoors	26	5.58
Number of subscribed channels		13.5
Advertisements per 10 min on average (self-reported by respondents)		2.4
Subscribing to YouTube Premium Service		81	17.38

presents the sample statistics for YouTube usage behavior.

The second section delved into the diverse genres of YouTube video content. Genre, as we discovered, is a key factor influencing the view count of YouTube videos. In our study, we categorized the genres based on various sources, including the Korean Ministry of Culture, Sports, and Tourism [47], Statista’s YouTube classification [48], the Korea Communications Commission [49], the Korean live streaming service, Afreeca TV’s broadcasting jockey classification, and YOUSTAR’s statistical classification. We then asked respondents to share their preferred genres. The results, summarized in

Table 3. Categories and preferences of video genres.

Genre	Preference		Genre	Preference
	1st	2nd		1st	2nd
Entertainment	45	35	Life/V-log	20	24
Food	44	47	Education/Knowledge	20	20
Politics/Current events	41	30	Web series	19	17
Music	39	50	TV series	13	20
Games	39	17	ASMR	12	4
News	34	26	Technology	9	15
Hobbies	31	17	Fashion/Beauty	8	16
Economy	27	31	Kids	8	10
Movies	26	27	Religion	8	3
Sports	23	36	Live streaming	7	8
Animals	23	28	Global	5	3
Health	20	33	Cars	4	8

, revealed that entertainment emerged as the most popular category, piquing the interest of most viewers, followed by food, politics, and current events.

The last section used a choice experiment to examine the respondents’ preferences for video viewing and channel subscriptions on an online video platform. The thumbnails, titles, channel names, video lengths, view counts, and upload dates are displayed on YouTube before recommended videos are played. Similarly, our choice experiment considered thumbnails, view counts, video length, and genre attributes. Thumbnails were categorized based on whether they contained the corresponding creators or subtitles, which are the two most common elements found in YouTube thumbnails. While thumbnails may also include other elements, such as products or background images, text and creator images are the dominant visual cues that guide viewer attention. To ensure interpretability and design efficiency in the discrete choice experiment, we selected only two representative elements. For example, if the thumbnail only displays subtitles, we refer to it as an informative thumbnail, and if the thumbnail only shows the creator without subtitles, we call it a creator-focused thumbnail. We set the range between 1000 and 10,000,000 for the video views to accommodate a wide range. However, recognizing the potential for a non-linear effect, we presented these view counts in categories: 1000–10,000, 10,001–100,000, 100,001–1,000,000, and 1,000,001–10,000,000. This allows us to capture the diminishing marginal effects of higher view counts. We set the video lengths to 3, 6, and 15 min because the video bounce rate significantly increases at these durations [50]. Fishman [50] shows a sharp increase in bounce rates after 3 min, a further increase around 6 min, and a substantial drop-off in viewership beyond 15 min. For genre, we intended to expose the respondents’ decisions in the recommended video, as shown in Table 3. Furthermore, we included the number of advertisements inserted during video playback as an attribute from a commercial perspective.

Table 4. Attributes and attribute levels in choice experiment.

Type	Attribute	Level	Description
Recommended video	Thumbnail	Only subtitle	Elements in thumbnails
		Only creator
		Both
	Views	1000	Default units set in the YouTube user interface
		10,000
		100,000
		1,000,000
		10,000,000
	Length	3 min	The biggest bounce rates occur at 3, 6, and 15 min [50]
		6 min
		15 min
		30 min
	Number of advertisements	0	Number of advertisements present in the video
		1
		2
		3
	Genre	1st preference	Use of preferred genres as stated by respondents, as shown in Table 3
		2nd preference
		Others
YouTube channel	Number of satisfactory videos	1	Determined based on preliminary survey results
		5
		10
		20
	Number of subscribers	1000	Default units set in the YouTube user interface
		10,000
		100,000
		1,000,000

presents a detailed explanation of the attributes and attribute levels used in the choice experiment.

Combining the attributes and levels shown in Table 4, we utilized the visual choice experiment (VCA) proposed by Dahan and Srinivasan [51]. A traditional choice experiment describes the alternatives on a card using only text, whereas a VCA represents a set of alternatives using both text and images. Consumers tend to respond better to visual descriptions of products compared with written descriptions [52]. The participants in this experiment were asked to rank the alternatives constructed through the VCA, as shown in Figure 1.

Following the choice experiment for watching YouTube videos, we conducted an additional choice experiment on channel subscriptions. For subscriptions, we selected the number of satisfactory videos and the number of subscribers as attributes. After adding these attributes, we asked respondents whether they would subscribe to each YouTube channel. However, since all attributes and levels resulted in 11,520 card combinations, investigating all combinations would be burdensome for respondents. Therefore, 32 combinations were created using a fractional factorial design using IBM SPSS Statistics (Version 27.0). Eight sets were prepared, with four as one set. Each respondent completed four sets of choice tasks. As shown in Table 1, the survey was administered to 525 respondents, resulting in a total sample size of 8400 observations.

4.2. Estimation Results

This study aims to investigate online video platform users’ video viewing and channel subscription behaviors. To achieve this goal, we employ a combined rank-ordered logit and binomial logit model, where each behavior is modeled with its own utility function.

4.2.1. Estimation Results for Viewing Behavior

First, we examine consumer preferences for video viewing behavior using the rank-ordered logit model. We consider five core attributes of recommended videos on the online video platform and estimate the consumer utility function for video viewing. This utility is formally expressed in Equation (8).

$$\begin{gathered} U_{nj}^{viewing}={\beta }_1{D}_{subtitle}+{\beta }_2{D}_{creator}+{\beta }_3\log \left(x_{view}\right)+{\beta }_4{x}_{length}+{\beta }_5{D}_{genre1} \\ +{ \beta }_6{D}_{genre2}+{\beta }_7{x}_{ad-up}+{\beta }_8{x}_{ad-down} \end{gathered}$$

(8)

In Equation (8), $D_{subtitle}$ and $D_{creator}$ are dummy variables for the thumbnail types. $x_{view}$ is defined as a video’s view count, and the model uses the logarithm of the actual view count for scaling. We used $x_{length}$ as a linear variable for video length. $D_{genre1}$ and $D_{genre2}$ are dummy variables for viewers’ first and second favorite genres, respectively. To analyze how the discrepancy between the expected and actual ad counts affects user preference, we established each respondent’s expected number of ads as the ‘reference point.’ When the actual number of ads exceeded this reference point, the difference was defined as $x_{ad-up}$, a variable representing negative deviation. Conversely, when the actual number was below the reference point, the difference was defined as $x_{ad-down}$, representing positive deviation. For instance, if the reference point is two ads, an actual count of four yields a value of two for $x_{ad-up}$, while an actual count of zero yields a value of two for $x_{ad-down}$. This separation of variables allows for the independent estimation of the disutility from an increase in ads and the utility from a decrease. This asymmetry in estimated effects aligns with prospect theory, which suggests that losses loom larger than gains [12].

Table 5. Estimation results for video viewing.

Type	Attribute		Coef.		Std. Err	RI	MWTP (Up)	MWTP (Down)
Viewing	Thumbnail	Subtitle	0.516	***	0.037	14.6%	0.22	0.04
		Creator	0.177	***	0.035	5.0%	0.64	0.12
	log (Views of video)		0.089	***	0.013	10.1%	1.26	0.23
	Length of video (unit: min)		−0.010	***	0.003	7.5%	−11.39	−2.09
	Genre	Most preferred	0.971	***	0.038	27.5%	0.12	0.02
		Second preferred	0.603	***	0.037		0.19	0.03
	Advertisements	More than expected	−0.113	***	0.026	18.3%
		Less than expected	0.021	*	0.012
log-likelihood			−6191.37
Number of observations			8400

Note. * p < 0.05, ** p < 0.01, and *** p < 0.001.

presents the estimation results for video viewing behavior.

Table 5 shows that viewers significantly prefer “subtitle” and “creator” on a thumbnail. Subtitles were about three times more influential than creator images. They also preferred videos with a large number of views. This observation implies the presence of the bandwagon effect in video selection. Therefore, the frontrunner is more advantageous in the online video creator market. However, the result shows a negative preference for video length. This suggests that viewers prefer shorter videos, which explains the recent phenomenal growth in TikTok, YouTube Shorts, and Instagram Reels. Recently, many influencers have used short-form platforms to raise awareness and attract viewers. These can be interpreted as meaningful. Regarding genre, we found that preferences for both the best and second-best genres were statistically significant. This indicates that our survey design effectively showed the respondents’ preferred genre as the attribute level of alternative cards. Among all the coefficients, genre was the most significant factor influencing online video viewing behavior. Specifically, viewers were 1.5 times more likely to watch their favorite genre than the second-best genre. Similar to Jang et al. [38], whose research shows that fandom content positively affects VOD consumption, the preferred genre may have a higher impact on viewing behavior. Regarding the number of advertisements, we found that more advertisements than expected led to more dislikes, while fewer advertisements than expected led to more likes, as is common sense. Interestingly, for the absolute value, the negative preference for many advertisements was approximately five times higher than the positive preference for a small number of ads. This confirms prospect theory in online video viewing behavior, in which people are more sensitive to losses than gains [50].

The relative importance was in the order of genre, advertisement, thumbnail, view, and video length. Within the MWTP framework, “MWTP (up)” signifies a higher number of ads than initially anticipated, whereas “MWTP (down)” denotes a lower number of advertisements than initially anticipated. This analysis quantifies the extent to which viewers dislike additional ads. According to our model, the number of views should be 18.33 times (101.26) more to add one more advertisement if there are more advertisements than the viewer expected, and 1.70 times (100.23) more if there are fewer advertisements than the viewer expected. This demonstrates that viewers are highly punitive toward ad overload. Similarly, to offset that same annoyance, the video’s length would need to be shortened by about 11 min, showing how extra ads severely diminish the perceived value of a viewer’s time investment. This lopsided reaction, where losses (too many ads) are felt far more intensely than equivalent gains (fewer ads), strongly supports prospect theory.

Among these attributes, creators can adjust the thumbnail size, video lengths, and number of advertisements. Advertisements play a crucial role in generating cash for content creators. However, it is essential to note that commercials can potentially reduce the number of views. Creators should develop practical approaches to enhance the number of views using these results.

4.2.2. Estimation Results for Subscription Behavior

Second, we investigate channel subscription behavior using the joint model combining a rank-ordered and binary logit method. The consumer utility functions for subscription can be represented as follows in Equation (9):

$$\begin{gathered} U_{nj}^{subscription}={\beta }_0+{\beta }_1{D}_{subtitle}+{\beta }_2{D}_{creator}+{\beta }_3\log \left(x_{view}\right)+{\beta }_4{x}_{length}+{\beta }_5{D}_{genre1} \\ \hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}+{\beta }_6{D}_{genre2}+{\beta }_7{x}_{ad-up}+{\beta }_8{x}_{ad-down}+{\beta }_9{x}_{satis}+{\beta }_{10}\log \left(x_{subscriber}\right) \end{gathered}$$

(9)

For viewers to decide to subscribe to a channel, it is necessary that they first view the channel’s videos. Consequently, we consider the following variables in addition to explaining subscription decisions: We consider the number of satisfactory videos and subscribers as the influencing factors. $x_{satis}$ is a continuous variable representing the number of satisfactory videos, and $x_{subscriber}$ is a continuous variable representing the number of subscribers. We use a logarithmic function for $x_{subscriber}$ as with the number of views.

Table 6. Estimation results for subscription.

Type	Attribute		Coef.		Std. Err	RI	MWTP (Up)	MWTP (Down)
Subscription	Thumbnail	Subtitle	0.302	***	0.056	6.6%	0.15	0.25
		Creator	0.248	***	0.056	5.4%	0.18	0.30
	log (Views of video)		0.048	**	0.019	4.1%	0.94	1.57
	Length of video (unit: min)		−0.020	***	0.003	11.9%	−2.20	−3.69
	Genre	Most preferred	1.010	***	0.056	22.0%	0.04	0.07
		Second preferred	0.395	***	0.055		0.11	0.19
	Advertisements	More than expected	−0.044		0.038	27.2%
		Less than expected	0.074	***	0.012
	Number of satisfactory videos		0.021	***	0.003	8.7%	2.12	3.54
	log (Number of subscribers)		0.085	***	0.020	5.5%	0.52	0.88
	Intercept		−1.724	***	0.160
log-likelihood			−11,686.13
Number of observations			8400

Note. * p < 0.05, ** p < 0.01, and *** p < 0.001.

presents the results of our model used to explain channel subscription behaviors on an online video platform.

The interface of the YouTube platform allows users to choose a recommended video with a thumbnail, and users can subscribe to the corresponding channel. In this circumstance, the creator and subtitles in thumbnails had a positive effect on subscription behavior as well as viewing behavior, but a change in their importance was observed. Although subtitles on thumbnails retained their primary importance, the presence of a creator held nearly equal relevance. This is a notable difference, as subtitles were approximately three times more critical for viewing. It appears that the presence of a content creator in the thumbnails enhances a channel’s credibility. The coefficients of view counts also showed a positive number, albeit with less importance than the viewing behavior. Conversely, there was a comparatively greater aversion toward video lengths. Similar to viewing behavior, the preference for a particular genre was also determined, with the highest level of preference being assigned to the most favored and the second highest level of preference being assigned to the second most favored genre. However, the weight for the most preferred genre increased, which can be interpreted as online video platform users simply viewing content and not subscribing to the channel in the case of the second-best genre.

The preferences for the number of advertisements showed interesting differences between subscription and viewing behavior. In the model of viewing behavior, the absolute value of ${\beta }_{ad-up}$ was about five times greater than the absolute value of ${\beta }_{ad-\mathrm{d}own}$. This means that viewers feel a much stronger dislike for an increase in advertisements than their preference for a decrease. However, in the model of subscription behavior, ${\beta }_{ad-up}$ was not statistically significant. This suggests that subscribers do not show as clear a negative reaction to an increase in advertisements as viewers. However, a reduction in advertising had a significant positive impact on subscriptions. This means that subscribers react positively to decreased advertisements, likely leading to subscriptions. Because subscribing is an expression of the will to watch the channel’s content consistently, fewer advertisements lead to a more positive experience and help maintain subscriptions. This implies that decreasing the number of ads to ensure a higher number of views is ineffective; however, reducing the number of ads to ensure more subscribers can yield positive results. Consequently, video creators should consider these findings and develop advertising tactics accordingly. We also found that attributes such as the number of satisfactory videos and subscribers positively influenced channel subscriptions.

Therefore, online video creators who want to grow their subscribers should develop a strategy focusing on topics their potential subscribers are most interested in. Interestingly, an increasing number of advertisements was not related to the subscription behaviors of online video platform users. However, a reduction in advertising had a significant positive impact on subscriptions.

5. Conclusions

5.1. Contributions and Implications

This study quantifies the impact of video attributes on YouTube viewing and subscription decisions, employing a discrete choice model that incorporates reference-dependent preferences. The results reveal a distinct asymmetry concerning the number of advertisements, consistent with prospect theory. Specifically, the decrease in utility from facing more ads than expected was substantially greater than the increase in utility from facing fewer ads. The findings also confirm the significant impact of strong genre preferences and the importance of thumbnail captions in driving viewer engagement. These results contribute to a deeper theoretical understanding of video content consumption patterns and offer practical implications for content creators and platform managers.

This study makes several contributions to the social media literature. Firstly, we empirically investigate how multiple factors related to digital media content collectively affect consumer decision-making, whereas previous studies have typically examined single aspects, such as advertisement avoidance or thumbnail effects, in isolation. Our choice experiment instead considered these elements simultaneously in a controlled setting. Secondly, unlike most research that has examined either viewing or subscriptions separately, our work integrates these two stages of user engagement. We investigate the relationship between initial viewing preferences and longer-term subscription behavior, thereby extending prior findings to capture the conversion from a one-time view to sustained channel followership. Thirdly, while previous studies have mainly focused on pre-roll ads, the present study provides a more comprehensive view of the impact of ads on viewer choice by considering the mid-roll ad experience. Fourthly, we extend the literature by explicitly modeling reference-dependent preferences to capture how deviations from viewers’ ad load expectations create asymmetric changes in utility. This addition highlights loss aversion in the digital video context and broadens the theoretical discourse on advertising effectiveness beyond traditional utility-based frameworks.

Content creators, marketers, and platform administrators can apply our results to craft evidence-based strategies. If growing the subscriber base is the priority, a creator might limit the number of advertisements, even at the cost of some immediate ad revenue, and produce relatively short videos to foster a positive user experience. Conversely, if maximizing views for short-term gain is more critical, leveraging attention-grabbing thumbnails and the bandwagon effect of high view counts would be advantageous. At the platform level, recommendation algorithms might incorporate these insights by suggesting videos aligned with a user’s preferred genre and optimizing suggestions based on video length and view counts to sustain engagement. Considering viewers’ awareness of algorithmic curation and content sourcing [50], such data-driven refinements in recommendations and advertisement placements can help attract and retain viewers. Lastly, this study provides a foundation for balancing monetization with user experience by quantifying the trade-offs between advertisement load and user satisfaction to ensure sustainable growth. This balance is heightened on emerging short-video platforms: the majority of viewers drop off within the first quarter of a TikTok video advertisement [22], and perceived advertising clutter has been identified as a primary driver of ad avoidance on such platforms [53]. Given the multitasking habits and shorter attention spans of Gen Z users in these environments [53], excessive or intrusive advertising can rapidly alienate audiences, underscoring the need for ad strategies that are both engaging and non-intrusive.

5.2. Limitations and Future Research

While our findings reveal some interesting insights into consumer behavior toward online video media, we acknowledge that certain limitations may suggest directions for future research. Firstly, this study focused on interstitial advertisements, a format that has not been covered in previous research. Future work should examine the interactions of various advertisement formats (e.g., pre-roll, mid-roll, pop-up, and personalized ads) to provide a more comprehensive view of the effects of advertising on viewer behavior. Secondly, we relied on a choice experiment with stated preference data in a hypothetical situation; however, it would be valuable to obtain revealed preference data from actual viewing and subscription behaviors. Combining stated and revealed data could improve the external validity of the results and enhance the forecasting of future social media trends. Thirdly, we limited our scope to certain readily quantifiable factors encountered by viewers. Future research may include additional factors, such as thumbnail quality and text length, to further enhance the understanding of what drives video selection and channel subscription on platforms like YouTube.

Finally, although genre preference was considered as a control variable, user choices may be heterogeneous across content genres. In particular, the reference point for ad volume, a core concept in this study, is also likely to vary significantly with genre characteristics. For instance, a music listener, whose immersive experience is easily disrupted, could have a far lower tolerance for ads than someone watching an educational video. Future research should segment viewers by genre, examine how these reference points differ, and test how such differences moderate the effect of advertising on viewer decisions.