You are on page 1of 10

Proceedings of the 56th Hawaii International Conference on System Sciences | 2023

Content Creator versus Brand Advertiser? The Effect of Inserting


Advertisements in Videos on Influencers Engagement
Tengteng Ma Yingda Lu Yuheng Hu
University of Illinois at Chicago University of Illinois at Chicago University of Illinois at Chicago
tma24@uic.edu yingdalu@uic.edu yuhenghu@uic.edu
Xi Chen Yuxin Chen
Zhejiang University NYU Shanghai
chen_xi@zju.edu.cn yc18@nyu.edu

Abstract social media, brands collaborate with influencers by


Influencer advertising has become an providing monetary incentives in exchange for
indispensable component of online marketing due to brand/product mention, review or endorsement in their
the exponential growth of social influencers and their videos (Stubb et al. 2019). Unlike the pop-up ads on
influence. Whereas the effectiveness of using YouTube, these inserted advertisements are usually
influencer endorsements is well studied from the brand designed by the influencer, thus can be related to the
or company perspective, how the commercial video content (e.g., a food blogger endorses a cheese
endorsements affect influencers themselves is an brand when making a cheesecake). Influencers also
important yet unrevealed question. We empirically decide the position and the length of the
examine the instantaneous (measured using live advertisement. This flexibility enables influencer
comment sentiment) and longer-term (measured using advertising to target users accurately and potentially
video feedback and follower number change) alleviate negative attitudes toward advertisements,
influence of inserting advertisements in videos on which further explain the prosperity of influencer
influencers’ reputation. We further investigate how marketing in recent years. In this paper, we define
this effect can be moderated when influencers “inserted advertisements in influencers’ videos” as
demonstrate stronger endorsement by showing their sponsored content and advertisements designed and
faces during advertisements. Our result suggests that inserted by the influencers in their videos. Note that
inserting advertisements have a negative impact on these advertisements are highly embedded in the video
both instantaneous and longer-term viewer content and usually cannot be skipped through ad-
engagement; advertisements with influencers’ face blockers.
showing moderate the negative effect of To date, existing studies in the influencer
advertisements on viewers’ instantaneous response, marketing literature primarily focus on how brands
while the different impact between advertisements can benefit from using influencer advertising and the
with/out influencers showing their faces is not corresponding strategy of prompting product exposure
significant in the longer term. (Lou and Yuan 2019; Ren 2021; Schouten et al. 2020).
Yet, there is a scarcity of studies on the impact of
Keywords: influencer marketing, video product endorsement on influencers themselves. In
advertisement, social media. fact, social media influencers face a trade-off between
endorsing products and maintaining popularity. On
1. Introduction one hand, sponsorship is a major source of income for
influencers. On the other hand, advertisements have a
negative effect on viewers’ experience (Evans et al.
Social media platforms such as YouTube,
2017; Kelly et al. 2010), which can be detrimental to
Instagram, and more recently TikTok, have reached
influencers’ popularity. Therefore, a comprehensive
unprecedented success in breeding social media stars
analysis of advertisement effect on viewer responses
who have a significant number of followers online.
from an omni-dimensional perspective is needed.
These social media stars, or micro-celebrities have
Specifically, we distinguish instantaneous feedback
built strong bonds with followers by sharing extensive
(measured using Live Comments) versus relatively
experience and insight into their followers (Hudders et
long-term feedback (measured using videos’ numbers
al. 2021). Given the emergence of social media
of likes, favorites, coins, shares and influencers’
influencers, ever-increasing brands are using social
follower change) in this study. Figure 1 illustrates an
media influencers to broadcast their products. On
example of the omni-dimensional audience feedback.

URI: https://hdl.handle.net/10125/103068
978-0-9981331-6-4 Page 3567
(CC BY-NC-ND 4.0)
The key difference between instantaneous and long- viewers to express their feedbacks (Figure 1),
term feedback is that, viewers can post live comments including: (1) posting live comments while watching
at the moment while watching the video content, so the video, i.e., a unique feature of Bilibili.com that
that live comments reflect the instantaneous attitude of enables viewers to post comments. Distinguished from
audiences towards advertisement; however, long-term traditional comments that are listed below the video
feedback usually happens when audiences finish together, live comments are recorded and rolling
watching the entire video (e.g., give a thumbs up after displayed in sync with video content; (2) leaving
watching a YouTube video), the evaluation of which feedback toward the video (e.g., like, favorite, coin,
is based on the overall content of the video or the share), (3) following or unfollowing the video author.
impression of the influencer. While traditional
marketing literatures attempt to examine the lasting
effect of advertising compared with the short-term
change in purchasing decisions (Wang et al 2009),
there is a dearth of literature examine the
instantaneous versus long term effect of advertising on
content generators’ reputation. To this definitive end,
it is requisite to answer the following question:
RQ1: How do inserted advertisements affect
different types of audience feedback (i.e., Figure 1. Example of audience feedback on
instantaneous feedback v.s. long-term feedback)? Bilibili.com
While researchers are overwhelmingly leveraging We collect live comments and video feedback
multiple methods to explore strategies to enhance the indicators of more than 344,130 videos that are
efficacy of influencer marketing in the light of product generated by the top 5,000 popular influencers on the
or brand broadcasting and promotion, we focus on a website from Jun.01 2020 to Jun.01 2021, as well as
rarely studied angle on the optimal advertising strategy the influencers’ daily follower number during the
for influencers: face showing. Influencers can control same time period. We leverage Propensity Score
the way of communicating with audiences by Matching and difference-in-differences (DID) models
manipulating the degree of self-disclosure and social to empirically examine the effect of inserting
presence by adopt various strategies to endorse advertisements on viewers’ preference toward
sponsored products or brand, such as changing voice influencers by exploiting instantaneous feedback (i.e.,
(Hwang et al. 2021) and mentioning the brand in a live comments) and long-term feedback (i.e., like,
particular time (Rajaram and Manchanda 2021). favorite, coin, share, follower change). Next, we
However, the role of face communication is a vital but employ deep learning based face recognition
understudied aspect in influencer marketing. Face techniques to identify whether the influencer shows up
showing has displayed its avails of prompting viewers’ during the advertisement, and further examine whether
engagement rate indicated by the number of likes and enhanced endorsement can moderate the effect of
comments (Bakhshi et al. 2014). Whether this result inserted advertisement on audience feedback. The key
can be generalized to influencer marketing on social result from our empirical estimation suggests that 1)
media platforms is yet not clear. On one hand, inserting advertisements have a negative impact on
advertising products with displaying faces might viewers’ instantaneous and long-term feedback; 2)
increase a sense of trust for viewers; on the other hand, Compared with advertisements without influencers’
identified as a group of people who are both organic faces, influencers’ face displaying moderate the
content creators and brand advertisers, influencers negative effect on instantaneous audience attitude in
should be extremely carefully in their way of product live comments. 3) however, the difference impact
endorsement as it is a reputation consumption. Thus between advertisements with/out influencers’ face
we propose the second research question: showing in long term is not significant.
RQ2: Whether influencers’ face showing This study contributes to prior literature in three
moderates the effect of advertisement on different ways. First, we are one of the first research that
types of audience feedback? examines the impact of advertising on influencers’
To answer the above research questions, we reputation. We emphasize the important question of
collect data from Bilibili.com, one of the major influencers’ dilemma on balancing between
Chinese video platforms. Bilibili.com is a Chinese maintaining popularity and earning income from
video sharing website similar to YouTube where users sponsored content, and we pioneer in exploring the
can submit, view and add overlaid commentary on effect of inserting advertisements in videos from the
videos. On Bilibili.com, there are several channels for perspective of influencers, rather than brands or

Page 3568
companies. In particular, we distinguish the positively affect followers’ trust in influencers’
instantaneous viewer feedback from long-term branded posts. Evans et al. (2017) reveal that
feedback using a unique dataset. Second, we are one advertising disclosure will lead to higher advertising
of the first studies that investigate the moderate effect recognition compared to no advertising disclosure, yet
of influencers’ face showing during advertisements. In no evidence on more negative attitudes or lower
addition, we also show that using deep learning purchase intention. Gerrath and Usrey (2020) examine
algorithms can extract important video features that how followers react to influencers’ incentivized
are critical for the managerial decisions of influencers review, and the results of one survey, three
and platforms. Lastly, our findings provide guidelines experiments, and one field study indicate that
for both influencers and social media platforms. influencer type (e.g., review vs. lifestyle) plays an
important role in determining followers’ perception of
2. Theoretical Background and Related influencer authenticity.
Literature However, most existing studies on audience
perception of sponsorship disclosure focus on brand or
product related outcomes, such as self-reported
2.1. Influencer Marketing purchase intention and brand awareness. Despite the
rapid growth of influencer marketing, the possible side
With the flourish of social media platforms such effect of such product endorsement content on
as Instagram and TikTok, brands are increasingly influencers themselves has been understudied in
favoring social media influencers for advertisement academia. One exception is Cheng and Zhang (2022),
compared to traditional celebrity. Influencer who investigate the impact of brand sponsorship on
endorsements now become a popular way for social influencers by examining sponsored videos on
marketers to promote their brands, products, and YouTube and corresponding audience engagement
services. Ever-increasing studies have evidenced the and sentiment toward influencers. Their result shows
advantages of influencer marketing. A recent study that the negative effect of brand sponsorship on
(Schouten et al. 2020) investigates the impact of reputation involves less engagement with sponsored
celebrity vs. influencer endorsements on advertising videos and less trust in the video/influencer. However,
effectiveness and finds that consumers identify more they focus on relatively longer-term viewer feedbacks
with influencers than celebrities, feel closer to (i.e., like action or comment sentiment), hence missing
influencers than celebrities, and trust influencers more a comprehensive reveal of the overall effect at
than celebrities. Due to the extensive experience with different levels. Different from traditional comments
and insight into their followers, influencers build that are usually posted by audiences after watching the
strong bonds with followers by creating a sense of video and listed below the video, live comments
shared interests in a specific area (Campbell and provide a function that allows audiences to post
Farrell 2020). A wealth of research shows that comments while watching video play synchronously.
increasing influencer-product congruence has a This unique feature enables us to capture
positive effect on followers’ product attitude, brand instantaneous feedbacks.
engagement, brand awareness and purchase intentions
(Argyris et al. 2020; Kim and Kim 2021; Lou and
2.3. Advertising Strategy and Enhanced Face
Yuan 2019). Parasocial relationship between
influencers and audience has also received substantial communication
attention in influencer marketing domain in the past
years. Explicit evidence is provided to show that a The optimal strategy of influencer marketing has
strong parasocial relationship increases the perceived been increasingly discussed in recent literature.
credibility of the influencer (Reinikainen et al. 2020; Hwang et al. (2021) examine the voice change during
Shan et al. 2020; Yuan and Lou 2020). influencers’ sponsored videos and find that influencers
choose a soft selling strategy and tone down their
voice loudness significantly in sponsored videos. This
2.2. Audience Perception of Sponsorship
voice modulation strategy can help influencers win
Disclosure back favorable consumer sentiment. Ren (2021)
proposes that high real-time interactivity should be
Some studies explore the audience perception or adopted to prompt brand loyalty during live stream
feedback on sponsorship disclosure in influencer shopping. Rajaram and Manchanda (2021) find that
marketing. By conducting an online survey, Lou and brand mentions in the first 30 seconds of a video are
Yuan (2019) find that informative value of influencer- on average associated with a significant increase in
generated content and the influencer’s trustworthiness

Page 3569
attention to the brand but a significant decrease in information of the influencers during the same period
sentiment expressed towards the video. mentioned above. Our dataset consists of 344,130
While the voice and time related strategies are videos generated by 5000 influencers, including video
explored in existing studies, the visual factors are still information, live comment information, author
understudied. In the context of video sharing platforms information and their follower number history during
such as Bilibili.com, face showing during the collection period.
advertisement can be regarded as an indicator of Compared with popup advertisement, it is much
enhanced endorsement. To understand whether face more challenging in identifying in-video
communication influences the relationship between advertisements as such advertisements are generally
inserted advertisements and audience feedback, we not labeled by influencers. To address this challenge,
turn to social presence theory, which refers to the we utilize a unique comment culture among viewers to
degree of sociable, warm and personal of the identify advertisements inserted into videos. On
communication perceived by people (Kim 2021). In a Bilibili.com, viewers usually post specific slangs in
computer-mediated communication environment, live comment channel to indicate their recognition of
more anthropomorphic (human-like) computer advertisements. We take advantage of this convention
representations elicit more social responses from to identify the presence and location of endorsement
people (Gong 2008). Theory of self-disclosure content through investigating related keywords. Given
indicates that communication through sharing intimate the large amount of viewership for each video, even if
personal information about themselves enhances a small percentage of viewers post these slangs, we can
feelings of connection (Utz 2015) and perceived still observe a reasonable number of slangs to help
closeness (Lin and Utz 2017). Leite and Baptista identify the existence and location of product
(2021) find that self-disclosure is positively related to endorsement within video. To further validate this
parasocial relationship intensity, in turn casts a strategy, we recruit a renowned data labeling company
positive effect on consumers’ purchase intention. to manually watch and identify the videos that might
Therefore, face showing is a typical format of social have inserted advertisements. Specifically, we
presence and self-disclosure in social media, which distinguish the sponsored videos 1 and those regular
might reduce audiences’ negative feedback. This line videos with inserted advertisements and label the
of reasoning is further supported by a wealth of corresponding starting and ending position of the
research showing that showing face can improve inserted advertisements in the latter case. The labeling
viewer engagement (Bakhshi et al. 2014; Zhao et al. result is further validated by two professional research
2021). assistants, suggesting an above 95% accuracy. After
data cleaning and removing outliers (e.g., extremely
3. Research Methodology long/short videos), we obtain 10,100 videos with
inserted advertisements. We further exploit propensity
To investigate the abovementioned research score matching to match videos with inserted
questions, we collect data from Bilibili.com, a leading advertisements using the remaining videos without
Chinese video sharing website that has over 233 advertisements, and regard them as a control group for
million active users. Through Bilibili API, we collect inserted advertisement videos. We use both influencer
information of top 5000 popular influencers (ranked information (number of followers, account created
by follower numbers), including their id, name, time, and total number of videos) and video
number of followers and account created time. Then, information (video length, type id, upload time, play
we collect the video information posted by these count) in our propensity score matching. The well-
influencers between Jun.01 2020 and Jun.01 2021, aligned distribution illustrates the balanced matching
including video id, length, description, type id, created performance (Figure 2). To further ensure no ads are
time, play count, audience feedback indicators (like inserted in videos in control group, we again recruit
count, favorite count, coin count, share count), and the the data labeling company to manually watch the
corresponding live comments that consist of text videos to confirm. Finally, we are able to match the
content, timestamp and user id of each comment. 10,100 videos and build balanced treated and control
Furthermore, we also gather the number of follower groups.

1 new products to influencers for free and ask them to produce a


“Sponsored videos” refer to videos that are purely commercial
product review/evaluation video to achieve the effect of
collaboration between influencers and brands where the entire
broadcasting new products. This study does not focus on how the
video content is product promotion. This is in comparison with
purely sponsored content affect influencers’ reputation. Instead, we
advertisement inserted into videos where only a small proportion
examine the effectiveness of inserting advertisement in
of videos are advertisements. For instance, some brands send their
influencers’ regular videos on influencers reputation.

Page 3570
Follo The follower number
wersh on day 𝑑 minus the -
Follower_ 1027
ip follower number on 973 3543 5650
change 793
day 𝑑 − 1 for 82
influencer 𝑖.

4. Empirical Model and Results

4.1. Instantaneous feedback


(a) Distribution of propensity score
We use live comments to measure viewers’
Figure 2. Distribution of propensity score and
variable before and after propensity score instantaneous feedback. To investigate such
matching instantaneous change of audience engagement and
To address our research questions, we conduct sentiment influenced by advertisements, we examine
face recognition using deep learning method to detect the live comments posted between a specific time
whether the influencer shows his or her face during window. Specifically, we set the starting position of
advertisements. First, we collect the 10,100 treated the inserted advertisement as time 𝑡0 , and label 1-
videos and capture frames every 5 seconds. Next, we minute sessions before and after 𝑡0 respectively, that
perform face recognition using a deep learning model is [𝑡0 − 1 𝑚𝑖𝑛, 𝑡0 + 1 𝑚𝑖𝑛].
named “face_recognition” to detect whether a face 𝐸𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡𝑣𝑖𝑡 = 𝛽0 + 𝛽1 ∙ 𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 ∙
shows in each frame. If a human face is detected, we 𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑡 + 𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 + 𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑡 + 𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒𝑟𝑖 +
identify whether the face is from the influencer. We 𝜖𝑣𝑖𝑡 (1-1)
extract each influencer’s face manually and compare Here 𝑡 represents the index of 5-second time slots,
them with the faces captured in the videos. As a result, ranging from -12 (1 minute before 𝑡0 ) to 12 (1 minute
we find 3,858 videos have the influencer’s face during after 𝑡0 ). The dependent variable 𝐸𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡𝑣𝑖𝑡
advertisement period. captures instantaneous feedback at an aggregated
We propose nine measurements to level, including the number of comments,
comprehensively depict audience reaction to inserted commenters, comments per commenter, and average
advertisements, which reveal the effect on influencers’ comment sentiment at time slot 𝑡 in video 𝑣 that are
reputation. Specifically, we consider live comment posted by influencer 𝑖 . The treatment variable
level, video feedback level and followership level 𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 equals 1 if the video 𝑣 belongs to treated
variables to measure the omni-dimensional effect. The group (i.e., videos with inserted advertisements), and
detailed descriptions of measurements are listed in zero otherwise. The time variable 𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑡 equals 1 if
Table 1. time slot 𝑡 is greater or equal to 𝑡0 , and 0 otherwise.
Table 1. Descriptive statistics of key variables We also control unobserved heterogeneity across
Variable Description Mean SD Min Max influencers by introducing fixed effects of
Live Number of comments
com
Comment_
posted for video 𝑣 11.16 15.62 1 1181 𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒𝑟𝑖 . 𝜖𝑣𝑖𝑡 is the error term.
count
ment within session 𝑡. Table 2 shows the key results. We observe a
Number of viewers significant increase in the number of live comments
Commenter who post comment for
_count video 𝑣 within session
10.95 15.27 1 1054 (column (1), 3.249, p<0.001), number of live
𝑡. commenters (column (2), 3.228, p<0.001). However,
Comment_p
The average comment the comment frequency indicated by
number for each
er_comment
viewer for video 𝑣
1.02 0.12 1 18 comment/commenter decreases after advertisements
er (column (3), -0.0032, p<0.1). In terms of average
within session 𝑡.
The averaged sentiment of live comments, we see a significant drop
sentiment score of off (column (4), -0.027, p<0.001). Apparently, after
Sentiment 0.59 0.15 0 1
comments for video 𝑣
within session 𝑡. seeing inserted advertisements, more viewers tend to
Video
Like
The number of likes of
24211 56854 65
1307 post live comments, and they are more likely to post
feedb the video 𝑣. 926 negative live comments. Therefore, inserting
ack The number of
Favorite favorites of the
4200.
14409 0
4731 advertisements leads to negative instantaneous
85 80 feedback in live comments.
video 𝑣.
Coin
The number of coins
7972 33899 1
1347 Table 2. Estimation result for live comment (model
of the video 𝑣. 057 1-1)
The number of shares 1745 (1) (2) (3) (4)
Share 1175 4520 0
of the video 𝑣. 88 Dep. Comment Comment Comment_per_ Average
_count er_count commenter sentiment
After* 3.249*** 3.228*** -0.003* -0.027***

Page 3571
Treat (0.250) (0.244) (0.002) (0.002) hypothesis of these two coefficients at 1% level for
Treat -0.681*** -0.667*** 0.002 -0.002
(0.238) (0.233) (0.002) (0.002) comment count, commenter count and average
After -0.797*** -0.793*** 0.00165 0.004* sentiment, 10% for comment per commenter, which
(0.239) (0.233) (0.00174) (0.002) means that the coefficients 𝛽1 and 𝛽2 are
Const_ 10.970*** 10.76*** 1.017*** 0.589***
ant (0.225) (0.219) (0.00164) (0.002) statistically different from each other, indicating
FE Y Y Y Y that advertisements without influencers’ faces
N 188,103 188,103 188,103 188,103 prompt audience engagement in terms of number of
R^2 0.165 0.167 0.153 0.107
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 comment and commenters, yet leading to worse
To further examine the moderate role of face sentiment. The result reveals that the negative
showing, we stratify the videos with inserted instantaneous feedback posted by audiences can be
advertisements into two groups according to moderated if influencers show faces during
influencers’ face showing during the advertisement. advertisements.
The treatment variable 𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 equals 1 if the Table 4 Wald test result of linear hypotheses
video 𝑣 belongs to treated group and has the estimation for live comments (𝑯𝟎 : 𝜷𝟏 = 𝜷𝟐 )
influencer’s face showing during advertisement time (1) (2) (3) (4)
Dep. Commen Commen Comment_per Average_se
as well. The other treatment variable, t_count ter_count _commenter ntiment
𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 equals 1 if the video 𝑣 belongs to Prob > F 0*** 0*** 0.0713* 0***
treated group but with no face showing of the
influencer during the advertisement time. For the 4.2. Long-term feedback
control group, videos have no advertisement, thus two
treatment variables are zero. We introduce the We categorize long-term feedback into two types:
estimation model (1-2). 1) video level feedback based on video content and 2)
𝐸𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡𝑣𝑖𝑡 = 𝛽0 + 𝛽1 ∙ 𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 ∙ influencer level feedback based on followership
𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑡 + 𝛽2 ∙ 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 ∙ change.
𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑡 + 𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 + 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 + On Bilibili.com, there are four types of video
𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑡 + 𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒𝑟𝑖 + 𝜖𝑣𝑖𝑡 (1-2) feedback actions through which viewers can express
As shown in Table 3, both coefficients of their preference towards the video. Viewers have the
𝐴𝑓𝑡𝑒𝑟 ∗ 𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡 and 𝐴𝑓𝑡𝑒𝑟 ∗ 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡 on option to “like”, “favorite”, “give coin” or “share” the
most of the columns are significant, except the video they like. Note that the effort required by
coefficient of 𝐴𝑓𝑡𝑒𝑟 ∗ 𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡 for different endorsement actions varies among each other
𝐶𝑜𝑚𝑚𝑒𝑛𝑡_𝑝𝑒𝑟_𝑐𝑜𝑚𝑚𝑒𝑛𝑡𝑒𝑟 . The estimation results (Table 5). For example, compared with a “like” action
are consistent with the average main effect in (1-1). that requires no additional effort, “Give coin” action
Table 3. Estimation result for live comment with costs a virtual coin of the viewer/donator (note that
grouped treatment (model 1-2) most users on Bilibili.com only get one free coin
(1) (2) (3) (4) daily); “favorite” action adds the video to a private list
Dep. Comment Commente Comment_ Average
_count r_count per_ sentiment (invisible to public) of the viewer for future reference,
commenter while “share” action adds the video to the viewer’s
After* 2.305*** 2.292*** -0.002 -0.022*** Bilibli.com personal page that are visible to the public.
FaceTreat (0.261) (0.255) (0.002) (0.003) Table 5. Video feedback description
After*No 4.155*** 4.127*** -0.004** -0.032*** Action Description
FaceTreat (0.260) (0.254) (0.002) (0.003) Like Click a button to express love to the video.
FaceTreat -0.256 -0.249 0.001 -0.005** Favorite Add the video to favorite lists.
(0.252) (0.246) (0.002) (0.002) Coin Donate a coin (a virtual currency use on the platform)
NoFace -1.081*** -1.060*** 0.004** -0.000 to the influencer.
Treat (0.250) (0.244) (0.002) (0.002)
Share Share the video on personal page that can be viewed
After -0.793*** -0.788*** 0.002 0.004*
by visitors.
(0.239) (0.233) (0.002) (0.002)
Constant 10.970*** 10.770*** 1.017*** 0.589*** We examine the treatment effect of inserting
(0.225) (0.219) (0.002) (0.002) advertisement on video level feedback indicated by the
FE Y Y Y Y four endorsement actions mentioned above separately.
N 188,103 188,103 188,103 188,103 Furthermore, we investigate whether influencers’ face
R^2 0.166 0.168 0.153 0.107
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
showing during advertisements moderates the effect
Next, we employ Wald test to investigate using a two-way interaction:
𝐸𝑛𝑑𝑜𝑟𝑠𝑒𝑚𝑒𝑛𝑡𝑣𝑖 = 𝛽0 + 𝛽1 ∙ 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙
whether the coefficients before two treatments
𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖 + 𝛽2 ∙ 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙ 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖 +
differ from each other. The result is reported in
𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑉𝑎𝑟𝑖𝑎𝑏𝑙𝑒𝑠 + 𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒𝑟𝑖 + 𝜖𝑣𝑖 (2)
Table 4. We find that we can reject the equality

Page 3572
The dependent variable, 𝐸𝑛𝑑𝑜𝑟𝑠𝑒𝑚𝑒𝑛𝑡𝑣𝑖 FE Y Y Y Y
represents the number of 𝑙𝑖𝑘𝑒 , 𝑓𝑎𝑣𝑜𝑟𝑖𝑡𝑒 , 𝑐𝑜𝑖𝑛 N 14,161 14,161 14,161 14,161
R^2 0.663 0.402 0.411 0.374
or 𝑠ℎ𝑎𝑟𝑒 of video 𝑣 posted by influencer 𝑖 . The Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
treatment variable 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 equals 1 if the video 𝑣 has In the next step, we aim to examine the moderate
inserted ads, and 0 otherwise. The treatment variable role of face showing on video feedback level. Wald
𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 equals 1 if the video 𝑣 belongs to test is adopted to reveal whether the showing faces
treated group and has the influencer’s face showing moderate the negative effect. The result is shown in
during advertisement time. The other treatment Table 7. Interestingly, the result indicates that there is
variable, 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑡 equals 1 if the video 𝑣 no significant difference between the coefficients of
belongs to treated group but with no face showing of 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙ 𝐹𝑎𝑐𝑒𝑣𝑖 and 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙ 𝑁𝑜𝐹𝑎𝑐𝑒𝑣𝑖 , and we
the influencer during advertisement time. The cannot reject the null hypothesis that 𝛽1 and 𝛽2 are
interaction terms 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙ 𝐹𝑎𝑐𝑒𝑣𝑖 and 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙ equal. Therefore, whether showing faces during
𝑁𝑜𝐹𝑎𝑐𝑒𝑣𝑖 aim to capture if the effect differs among advertisements does not affect the relationship
groups with/out influencers’ face showing during between inserting advertisement and audience video
advertisements. We add an influencer-level fixed level feedback.
effect to control the time-invariant heterogeneities Table 7. Wald test result of linear hypotheses
among videos. In addition, we introduce two control estimation for video level feedback (𝑯𝟎 : 𝜷𝟏 = 𝜷𝟐 )
variables, the number of video endorsement of the (1) (2) (3) (4)
most recent video posted by the same influencer Dep. Like_ Favorite Coin_ Share_
𝑃𝑟𝑒𝑣𝑖𝑜𝑢𝑠_𝑒𝑛𝑑𝑜𝑟𝑠𝑒𝑚𝑒𝑛𝑡𝑣𝑖 and the influencers’ count _count coint count
number of followers 𝐹𝑜𝑙𝑙𝑜𝑤𝑒𝑟𝑖 on the day of video Prob > F 0.1727 0.1087 0.2488 0.1437
post, to eliminate the time-variant heterogeneities Moreover, we capture the effect of inserting
among videos. Here 𝜖𝑣𝑖𝑡 is the error term. advertisement on follower number change through the
As shown in Table 6, the coefficients of both following model. We set the day of video posting as
𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙ 𝐹𝑎𝑐𝑒𝑣𝑖 and 𝑇𝑟𝑒𝑎𝑡𝑣𝑖 ∙ 𝑁𝑜𝐹𝑎𝑐𝑒𝑣𝑖 in all 𝑑0 , and define the following model:
columns are significantly negative. Specifically, 𝐹𝑜𝑙𝑙𝑜𝑤𝑒𝑟_𝑐ℎ𝑎𝑛𝑔𝑒𝑣𝑖𝑑 = 𝛽0 + 𝛽1 ∙ 𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 ∙ 𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑑 +
posting a video with inserted advertisements lead to + 𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 + 𝐴𝑓𝑡𝑒𝑟𝑑 + 𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒𝑟𝑖 + 𝜖𝑣𝑖𝑑 (3-1)
8,833 (column (1), p<0.001) decrease of like count, Where 𝐹𝑜𝑙𝑙𝑜𝑤𝑒𝑟_𝑐ℎ𝑎𝑛𝑔𝑒𝑣𝑖𝑑 is the follower
2,382 (column (2), p<0.001) decrease of favorite count, number change between adjacent two days which can
5,080 decrease of coin count (column (3), p<0.001) be positive (follower increased) or negative (follower
and 658.8 decrease of share count (column (4), decreased). The treatment variable 𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 equals 1
p<0.001) if influencers show their faces during if the video 𝑣 belongs to treated group, otherwise 0.
advertisements, whereas a video without influencers’ The time variable 𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑑 equals 1 if day 𝑑 is greater
face showing tend to reduce 7,149 (column (1), or equal to 𝑑0 , and 0 otherwise. Fixed effects of
p<0.001) likes, 1,749 (column (2), p<0.001) favorites, 𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒𝑟𝑖 is also employed in the model. 𝜖𝑣𝑖𝑑
3,975 (column (3), p<0.001) coins, 476 (column (4), represents the error term.
p<0.001) shares. To explore whether there is a time decay trend of
Table 6 Estimation result for video level feedback the inserting advertisement effect of followership
(model 2) change, we run the same model with a different time
(1) (2) (3) (4) range, that are 7 days, 5 days, 3 days, 2 days and 1 day.
Dep. Like_ Favorite_ Coin_coun Share_co Result presented in Table 8 shows that within the next
count count t unt following 3 days of video posting, there is a significant
Treat*Fac -8,833*** -2,382*** -5,080*** -658.8*** decrease in followership change. Specifically, on
eTreat (1,059) (338.6) (821.6) (107.2) average, the number of followers decreases by 133 in
Treat*No -7,149*** -1,749*** -3,975*** -476.0***
FaceTreat (996.5) (318.7) (773.3) (100.9) the next day (column (5), p<0.001). And this effect is
Follower -0.016*** -0.003*** -0.005*** -0.002*** 161.3 (column (4), p<0.001) and 130.7 (column (3),
(0.002) (0.001) (0.002) (0.000) p<0.001) for day 2 and day 3 respectively. However,
Previous 0.158*** the negative effect becomes weaker on day 5, and
_like (0.009)
unsignificant on day 7. Note that the influencer might
Previous 0.023**
_faveorite (0.009) upload new videos within the 7-day time window
Previous 0.084*** (usually after day 5). To sum up, we observe that there
_coin (0.009) is a negative effect of posting videos with
Previous 0.003 advertisements inserted on followership change,
_share (0.008)
Constant 35,574*** 7,493*** 12,854*** 2,992***
whereas the negative effect is gradually vanishing in a
(1,532) (486.7) (1,179) (154.1) relative long term.

Page 3573
Table 8. Estimation result for followership level Constant 804.7*** 924.4*** 880.8*** 869.0*** 860.4***
feedback (model 3-1) (15.34) (20.01) (17.19) (21.01) (29.33)
(1) (2) (3) (4) (5) FE Y Y Y Y Y
Dep. 7 days 5 days 3 days 2 days 1 day N 218,180 159,917 101,631 72,505 43,396
follower follower follower follower follower R^2 0.182 0.186 0.331 0.339 0.349
change change change change change Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
After* -26.07 -64.61* -130.7*** -161.3*** -133.0*** Table 10. Wald test result of linear hypotheses
Treat (26.48) (34.24) (28.87) (34.60) (46.43) estimation for follower change (𝑯𝟎 : 𝜷𝟏 = 𝜷𝟐 )
Treat -39.66* -27.60 -18.21 -23.29 -46.09
(1) (2) (3) (4) (5)
(23.17) (30.19) (25.84) (31.44) (43.45)
After 168.6*** 95.87*** 206.4*** 251.7*** 226.4*** Dep. 7 days 5 days 3 days 2 days 1 day
(18.98) (24.55) (20.71) (24.83) (33.34) follower follower follower follower follower
Cons- 804.4*** 924.1*** 880.4*** 868.5*** 859.5*** change change change change change
tant (15.33) (20.01) (17.19) (21.00) (29.33) Prob > F 0.9140 0.6481 0.7702 0.9114 0.6990
FE Y Y Y Y Y
N 218,180 159,917 101,631 72,505 43,396 Combining the results above, we believe that face
R^2 0.182 0.186 0.331 0.339 0.349 showing does have a moderating role in the negative
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 effect of inserting advertisements on audience
Similar to the examination method adopted in live feedback, whereas this only holds in the case of
comment, here the treatment variable 𝑇𝑟𝑒𝑎𝑡𝑣 equals 1 instantaneous feedbacks, i.e., live comments.
if the video 𝑣 has inserted advertisements, and zero if
not (control group). The treatment dummy variable 4.3. Robustness check
𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 equals 1 or 0 depending on whether the
video 𝑣 has the influencer’s face showing during We validate the parallel trends assumption using
advertisement time. The other treatment variable, the dynamic difference-in-differences associated with
𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 equals 1 if the video 𝑣 belongs to the event of inserting advertisement at time 𝑡 in model
treated group but with no face showing of the (1-1) and day 𝑑 in model (3-1). This analysis allows
influencer during advertisement time. For the control us to detect the pre-advertising effect on the audience
group, both two treatment variables are zero. feedback outcome variables. Taking live comment
𝐹𝑜𝑙𝑙𝑜𝑤𝑒𝑟_𝑐ℎ𝑎𝑛𝑔𝑒𝑣𝑖𝑑 = 𝛽0 + 𝛽1 ∙ 𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 ∙ sentiment and follower number change as two
𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑑 + 𝛽2 ∙ 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 ∙ example outcomes, we observe that the test of
𝐴𝑓𝑡𝑒𝑟𝑣𝑖𝑑 + 𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 + 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡𝑣𝑖𝑑 + common trend assumption is passed (Figure 3).
𝐴𝑓𝑡𝑒𝑟𝑑 + 𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒𝑟𝑖 + 𝜖𝑣𝑖𝑑 (3-2)
The results of model regression and Wald test are
demonstrated in Table 9 and Table 10 respectively.
Result presented in Table 9 shows that the average
negative effect of inserting advertisements on
followership growth is similar between group
𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡 and 𝑁𝑜𝐹𝑎𝑐𝑒𝑇𝑟𝑒𝑎𝑡, that is significant in
(a)Instantaneous feedback, (b)Long-term feedback,
the first three days, and decays gradually. According outcome variable: sentiment; outcome variable: follower change;
to the Walt test result shown in Table 10, we can pre=4, post=4; F(4, 93698) = 1.52; pre=3, post=3; F(3,106658) = 1.81;
neither reject the null hypothesis that 𝛽1 and 𝛽2 are Prob > F = 0.1928; Prob > F = 0.1430;
equal, which indicates that there is no difference Parallel-trend passed Parallel-trend passed
between whether influencers show their faces during Figure 3. Dynamic difference-in-differences
advertisements regarding to the negative effect of While results thus far have shown remarkable
advertisements on followership growth. consistency, it remains plausible that the results are
Table 9. Estimation result for live comment with falsely significant due to serial correlation in the
grouped treatment (model 3-2) dependent variable or spurious effects. We implement
(1) (2) (3) (4) (5) a random treatment test to exclude this possibility. We
Dep. 7 days 5 days 3 days 2 days 1 day first randomly shuffle the treatment variables (i.e.
follower follower follower follower follower
change change change change change 𝑇𝑟𝑒𝑎𝑡), then run the Difference-in-differences models
After*Fa -23.95 -53.11 -124.5*** -164.2*** -119.8** using the shuffled treatment variables. As expected, all
ceTreat (32.87) (42.51) (35.85) (42.96) (57.64) coefficients of 𝐴𝑓𝑡𝑒𝑟 ∗ 𝑇𝑟𝑒𝑎𝑡 term are insignificant.
After*No -27.95 -74.93* -136.3*** -158.8*** -145*** Table 11. Estimation result for live comments after
FaceTreat (31.73) (41.03) (34.59) (41.44) (55.58)
FaceTreat -56.15* -49.09 -42.64 -52.02 -106.6*
treatment shuffle (Model 1-1)
(29.57) (38.52) (32.94) (40.02) (55.13) (1) (2) (3) (4)
NoFace -25.87 -9.316 2.304 0.438 4.681 Dep. Comment Commenter Comment Average
Treat (27.98) (36.45) (31.18) (37.91) (52.30) _count _count _per_com sentiment
After 168.6*** 95.86*** 206.4*** 251.7*** 226.4*** menter
(18.98) (24.55) (20.71) (24.83) (33.34) After* -0.183 -0.180 0.001 -0.000

Page 3574
Treat (0.188) (0.184) (0.001) (0.002) side for brands and companies, our finding shows that
Treat 1.432*** 1.432*** 0.001 -0.020*** product endorsement in videos negatively affects
(0.174) (0.170) (0.001) (0.002)
After 2.330*** 2.314*** -0.002 -0.020*** audience feedback, including instantaneous feedback
(0.186) (0.181) (0.001) (0.002) via live comments, video feedback via feedback
Cons- 8.983*** 8.793*** 1.019*** 0.605*** buttons (e.g., like, favorite, coin, share), as well as
tant (0.165) (0.161) (0.001) (0.002) follower number change. This finding further
FE Y Y Y Y
highlights the importance of this study, and point out
N 188,103 188,103 188,103 188,103
R^2 0.164 0.166 0.153 0.106 the dilemma of influencers, that is, how to take
Standard errors in parentheses advertising without affecting popularity. By
*** p<0.01, ** p<0.05, * p<0.1 decomposing the empirical analysis into different
Table 12. Estimation result for followership groups according to influencers’ face showing, we
change after treatment shuffle (Model 3-1) propose two meaningful suggestions for influencers
(1) (2) (3) (4) (5)
Dep. 7 days 5 days 3 days 2 days 1 day
that can be used as guidelines in advertisement design.
follower follower follower follower follower First, influencers should pay attention to the frequency
change change change change change of inserting advertisements in videos as it hurts their
After* -8.630 0.352 -13.68 -12.80 -0.135 reputation and popularity in an omni-dimensional
Treat (15.49) (19.52) (18.47) (21.77) (27.24) manner. Second, showing faces during advertisement
Treat -53.57*** -62.80*** -92.86*** -120.1*** -134.7***
(11.68) (14.82) (14.20) (16.95) (21.69) can mitigate negative emotions from audience to some
After 167.0*** 62.77*** 138.4*** 164.3*** 149.1*** extent.
(10.96) (13.88) (13.31) (15.94) (20.84) Our work makes theoretical contributions to the
Cons- 848.0*** 984.9*** 967.4*** 971.0*** 957.9*** influencer marketing literature. To best of our
tant (9.278) (11.84) (11.53) (14.03) (18.93)
knowledge, we are one of the first studies that
FE Y Y Y Y Y
N 368,117 269,844 171,525 122,347 73,255 examines the product endorsement effect on
R^2 0.221 0.229 0.334 0.339 0.348 influencers empirically. Moreover, we introduce
Standard errors in parentheses omni-dimensional measurement for social media
*** p<0.01, ** p<0.05, * p<0.1 video platforms, including live comment level, video
feedback level and follower number level. This allows
5. Conclusion and Discussion us to evaluate the instantaneous and long-term
feedback when advertisings were inserted into videos.
With the prevalence of social media platforms, This study also demonstrates the potential of using
many social media users have gained online fame and deep learning algorithms to extract important
become social media influencers who have a strong managerial information.
online identity as well as a significant number of Besides, our study has several implications. Our
followers. This phenomenon flourishes influencer research demonstrates the overall effect of commercial
marketing in which influencers endorse products in endorsement on influencer advertising from the
their generated video clips. However, it is unclear how perspective of influencers and offers meaningful
commercial endorsement can affect influencers’ strategies for influencers to guide advertisement
popularity, and what is the optimal strategy that design. One of the strategies to alleviate these negative
influencers should adopt when designing endorsement impacts is to enhance social present by showing faces
content. during the advertisement, which can make viewers
To tackle the abovementioned problems, we feel connected and increase the perceived closeness
collect data from Bilibili.com and conduct an (Utz 2015). Social media platforms can also gain
empirical examination. We find a negative effect of insights about how to establish rules for advertisement
product endorsement on viewers’ live comment regularization to maintain user volume.
sentiment, video feedback and follower number Our research has limitations, which also open
change. Particularly, this negative effect on opportunities for future research. In the next step, we
instantaneous feedback represented by live comments would like to introduce more control variables and
is stronger when the influencer does not show his/her conduct extensive robustness check experiments to
face during the inserted advertisement, whereas this further validate the results. While we examined that
moderating role is insignificant for long term the inserted advertisements cast a negative effect on
feedback. the influencers’ reputation and popularity, it is unclear
Our findings augment existing literature on whether the impact varies among different types of
influencer advertising, one of the latest forms of influencers. Additionally, the content of product
influencer marketing. Whereas existing research on endorsement is another important aspect of
influencer advertising mainly focus on the positive investigating advertisement effect, thus leveraging

Page 3575
such information is another topic we would like to Kim, H. 2021. “Keeping up with Influencers: Exploring the Impact
of Social Presence and Parasocial Interactions on Instagram,”
explore for future research. Moreover, we would like
International Journal of Advertising, Taylor and Francis Ltd.
to examine the underlying reasons behind the negative (https://doi.org/10.1080/02650487.2021.1886477).
comments further using Natural Language Processing Leite, F. P., and Baptista, P. de P. 2021. “The Effects of Social
techniques. Media Influencers’ Self-Disclosure on Behavioral Intentions:
The Role of Source Credibility, Parasocial Relationships, and
Brand Trust,” Journal of Marketing Theory and Practice,
References Routledge.
(https://doi.org/10.1080/10696679.2021.1935275).
Argyris, Y. A., Wang, Z., Kim, Y., and Yin, Z. 2020. “The Effects Lin, R., and Utz, S. 2017. “Self-Disclosure on SNS: Do Disclosure
of Visual Congruence on Increasing Consumers’ Brand Intimacy and Narrativity Influence Interpersonal Closeness
Engagement: An Empirical Investigation of Influencer and Social Attraction?,” Computers in Human Behavior (70),
Marketing on Instagram Using Deep-Learning Algorithms Elsevier Ltd, pp. 426–436.
for Automatic Image Classification,” Computers in Human (https://doi.org/10.1016/j.chb.2017.01.012).
Behavior (112), Elsevier Ltd. Lou, C., and Yuan, S. 2019. “Influencer Marketing: How Message
(https://doi.org/10.1016/j.chb.2020.106443). Value and Credibility Affect Consumer Trust of Branded
Bakhshi, S., Shamma, D. A., and Gilbert, E. 2014. “Faces Engage Content on Social Media,” Journal of Interactive Advertising
Us: Photos with Faces Attract More Likes and Comments on (19:1), Routledge, pp. 58–73.
Instagram,” in Conference on Human Factors in Computing (https://doi.org/10.1080/15252019.2018.1533501).
Systems - Proceedings, Association for Computing Rajaram, P., and Manchanda, P. 2021. “Video Influencers:
Machinery, pp. 965–974. Unboxing the Mystique,” SSRN Electronic Journal (June).
(https://doi.org/10.1145/2556288.2557403). (https://doi.org/10.2139/ssrn.3752107).
Campbell, C., and Farrell, J. R. 2020. “More than Meets the Eye: Reinikainen, H., Munnukka, J., Maity, D., and Luoma-aho, V.
The Functional Components Underlying Influencer 2020. “‘You Really Are a Great Big Sister’ – Parasocial
Marketing,” Business Horizons (63:4), Elsevier Ltd, pp. 469– Relationships, Credibility, and the Moderating Role of
479. (https://doi.org/10.1016/j.bushor.2020.03.003). Audience Comments in Influencer Marketing,” Journal of
Cheng, M., School, H. B., and Zhang, S. (n.d.). “Reputation Marketing Management (36:3–4), pp. 279–298.
Burning: Analyzing the Impact of Brand Sponsorship on (https://doi.org/10.1080/0267257X.2019.1708781).
Social Influencers.” Ren, L. 2021. How Live Stream Shopping Influences Brand
(https://papers.ssrn.com/sol3/papers.cfm?abstract_id=407118 Loyalty on Taobao Live: A Perspective of Parasocial
8#:~:text=The%20reputation%2Dburning%20effect%20is,s Interaction: The Effects of Affective Commitment, Real-
maller)%20audiences.). Time Interactivity, Visual Complexity, Live Streamer Type.
Evans, N. J., Phua, J., Lim, J., and Jun, H. 2017. “Disclosing Schouten, A. P., Janssen, L., and Verspaget, M. 2020. “Celebrity
Instagram Influencer Advertising: The Effects of Disclosure vs. Influencer Endorsements in Advertising: The Role of
Language on Advertising Recognition, Attitudes, and Identification, Credibility, and Product-Endorser Fit,”
Behavioral Intent,” Journal of Interactive Advertising (17:2), International Journal of Advertising (39:2), Routledge, pp.
Informa UK Limited, pp. 138–149. 258–281. (https://doi.org/10.1080/02650487.2019.1634898).
(https://doi.org/10.1080/15252019.2017.1366885). Shan, Y., Chen, K.-J., and Lin, J.-S. (Elaine). 2020. “When Social
Gerrath, M. H. E. E., and Usrey, B. 2020. “The Impact of Media Influencers Endorse Brands: The Effects of Self-
Influencer Motives and Commonness Perceptions on Influencer Congruence, Parasocial Identification, and
Follower Reactions toward Incentivized Reviews,” Perceived Endorser Motive,” International Journal of
International Journal of Research in Marketing, Elsevier B.V. Advertising (39:5), pp. 590–610.
(https://doi.org/10.1016/j.ijresmar.2020.09.010). (https://doi.org/10.1080/02650487.2019.1678322).
Gong, L. 2008. “How Social Is Social Responses to Computers? Stubb, C., Nyström, A. G., and Colliander, J. 2019. “Influencer
The Function of the Degree of Anthropomorphism in Marketing: The Impact of Disclosing Sponsorship
Computer Representations,” Computers in Human Behavior Compensation Justification on Sponsored Content
(24:4), pp. 1494–1509. Effectiveness,” Journal of Communication Management
(https://doi.org/10.1016/j.chb.2007.05.007). (23:2), Emerald Group Holdings Ltd., pp. 109–122.
Hudders, L., de Jans, S., and de Veirman, M. 2021. “The (https://doi.org/10.1108/JCOM-11-2018-0119).
Commercialization of Social Media Stars: A Literature Utz, S. 2015. “The Function of Self-Disclosure on Social Network
Review and Conceptual Framework on the Strategic Use of Sites: Not Only Intimate, but Also Positive and Entertaining
Social Media Influencers,” International Journal of Self-Disclosures Increase the Feeling of Connection,”
Advertising (40:3), Taylor and Francis Ltd., pp. 327–375. Computers in Human Behavior (45), Elsevier Ltd, pp. 1–10.
(https://doi.org/10.1080/02650487.2020.1836925). (https://doi.org/10.1016/j.chb.2014.11.076).
Hwang, S., Liu, X., and Srinivasan, K. 2021. “Voice Analytics of Yuan, S., and Lou, C. 2020. “How Social Media Influencers Foster
Online Influencers—Soft Selling in Branded Videos,” SSRN Relationships with Followers: The Roles of Source
Electronic Journal, Elsevier BV. Credibility and Fairness in Parasocial Relationship and
(https://doi.org/10.2139/ssrn.3773825). Product Interest,” Journal of Interactive Advertising (20:2),
Kelly, L., Kerr, G., and Drennan, J. 2010. “Avoidance of pp. 133–147.
Advertising in Social Networking Sites,” Journal of (https://doi.org/10.1080/15252019.2020.1769514).
Interactive Advertising (10:2), pp. 16–27. Zhao, K., Hu, Y., Hong, Y., and Westland, J. C. 2021.
(https://doi.org/10.1080/15252019.2010.10722167). “Understanding Characteristics of Popular Streamers on Live
Kim, D. Y., and Kim, H. Y. 2021. “Influencer Advertising on Streaming Platforms: Evidence from Twitch.Tv,” Journal of
Social Media: The Multiple Inference Model on Influencer- the Association for Information Systems (22:4), pp. 1076–
Product Congruence and Sponsorship Disclosure,” Journal of 1098. (https://doi.org/10.17705/1jais.00689).
Business Research (130), Elsevier Inc., pp. 405–415.
(https://doi.org/10.1016/j.jbusres.2020.02.020).

Page 3576

You might also like