Use the right words : research on linguistic style dimensions influencer marketing and its effect on influencer success

(1)

Use the Right Words

Research on Linguistic Style Dimensions in Influencer Marketing

and its Effect on Influencer Success

Name: Fleur Griffioen Student number: 10176160 Date of submission: 22/06/2017 Version: Final Word Count: 16.120 MSc. in Business Administration – Digital Business University of Amsterdam Thesis Supervisor: Frederik Situmeang

(2)

Statement of Originality

This document is written by Fleur Griffioen who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

Abstract

Marketers are increasingly relying on social media as a channel for marketing communications. However, instead of brand-initiated content, marketers let digital influencers spread branded content on these channels, as customers perceive these influencers as more trustworthy. This research aims to predict the success of digital influencers, by looking into the linguistic style dimensions they use in their

expressions. Following the Communication Accommodation Theory, it is

hypothesized that influencers’ linguistic style alignment with their followers linguistic style, affects the followers’ evaluation of the influencer (i.e., liking the page,

increasing the follower base of the influencer) and their engagement behavior (i.e., liking, commenting on or sharing the post). Linguistic style is based upon function words such as articles, prepositions, and pronouns. Additionally, more in-depth, this research tries to predict influencer success based upon personal pronoun use in posts. The research uses text mining to extract the linguistic style properties of Facebook posts and comments on posts. In the research 841 posts and their comments were analyzed. Five hierarchical regression models and several multilevel models were applied. The results show that a greater synchronicity in linguistic style negatively relates with influencer success in terms of follower growth and engagement. Furthermore, the results show that personal pronoun use doesn’t affect influencer success. Contradicting the existing literature on linguistic style match, this research opens up a new discussion regarding the use of textual attributes in marketing communications spread by a digital influencer. There is argued that when an influencer doesn’t match the linguistic style of the follower, the influencer diverges from the norm by showing a higher level of autonomy. Consecutively, the perceived coolness of this person is increased. This research argues that perceived coolness

(5)

plays a key role in establishing influencer success. Therefore, there is suggested that marketers search for these cool influencers when they are setting up influencer marketing campaigns, instructing them to be different than their followers in their textual expressions, in order to ensure effective campaigns.

(6)

“Language is a social art.

In acquiring it we have to depend entirely on intersubjectively available cues as to what to say and when.”

(7)

1. Introduction

The emergence of social media - from Facebook to Twitter, blogs, LinkedIn, and YouTube - has facilitated new ways of social interaction for both customers and companies (Fischer & Reuber, 2010; Hall, 2016), offering them the tools to join a worldwide conversation every moment (SAS HBR, 2010). Like never before, millions of people have the opportunity to talk to each other and exchange information;

similarly, companies now have the opportunity to talk with millions of their customers (SAS HBR, 2010). Social media provide marketers interactive

communication environments with opportunities to build and foster relationships with customers (Chung & Austria, 2010) and increasingly, organizations use them to do so (De Vries, Gensler & Leeflang, 2012). Recognized as the most potentially powerful medium in business practice (Chung & Austria, 2010) and given their revolutionary reach, firms are increasingly relying on social media as a channel for marketing communication (Kumar, Bezawada, Rishika, Janakiraman & Kannan, 2016). According to research by eMarketer (2015b), advertisers worldwide have spend $23.68 billion on social media ads – including search, display and video ads – to reach consumers in 2015 (an increase of 33.5% compared to 2014). By 2017, social network ad spending will reach an expenditure of $35.98 billion, representing 16% of all digital ad spending globally (eMarketer, 2015b). However, customers don’t always trust marketer-initiated branded content anymore. Research shows that online

shoppers put 12 times more trust in a peers’ opinion (i.e., user generated content) than in marketer-initiated sources (eMarketer, 2010). Accordingly, brands have started to use a peer influencer type, to spread their branded content, also referred to as

influencer marketing. Influencer marketing shows its potential, as influencers can have a far-reaching impact on their followers and there is a potential for viral growth

(8)

(De Veirman, Cauberghe & Hudders, 2016). Instead of using a celebrity to represent and promote a brand, the trend is shifting towards the use of a digital influencer to do so. Research shows that 92% of the consumers seems to trust a digital influencer more than an organizations’ advertisement or traditional celebrity endorsement (Weinswig, 2016). These digital influencers use their social platforms to share stories and

interpretation and have the ability to influence (brand) attitudes and opinions of others, as opinion leaders (Uzunoğlu & Kip, 2014). In May 2015, 84% of marketing and communication professionals did expect to launch at least one campaign

involving a digital influencer in the next 12 months (eMarketer, 2015a).

The credibility of this digital influencer may also rely on message characteristics used by this source (Pornpitakpan, 2004). Limited research has however yet focused on the content of the message spread by the influencer type, especially in the online context. According to Ludwig & De Ruyter (2016), the language used in the message is the key to decoding social media speak. Following the Speech Act Theory (SAT), language construction in speech or writing, i.e., using specific words, sentence structure and interactional exchanges, predicts the speakers’ identity, perceptions and the intended behavior of the speaker (Austin, 1962). People (i.e., writers) use words (and images) to convey substantial information about who they are, their relationship with their audience and their intentions (Pennebaker, Mehl & Niederhoffer, 2003). Following this theory Giles (2009) found out that, in any conversation, we adjust our communication style to the opposite speaker. This can be referred to as the

Communication Accommodation Theory (CAT). In computer-mediated contexts, such communication style accommodations are present through the assimilations in linguistic styles among interactants. A linguistic style match (LSM) symbolically

(9)

reflects social identification with the partner conversant in this scenario (Ludwig, De Ruyter, Mahr, Wetzels, Brüggen & De Ruck, 2014). Identification with a source influences message effects or, in other words, whether an audience can identify with the spokesperson determines the attitudinal and behavioral effects of a message. Research that has applied the CAT to the online context shows that in an online review setting, LSM increases conversion rates of the reviewed product, suggesting that managers should match the linguistic style of their own firm-initiated reviews to the style of the product interest group’s (Ludwig, De Ruyter, Friedman, Brüggen, Wetzels & Pfann, 2013). Moreover, Ludwig et al. (2014) investigated textual features in online user communities. They found that at a community level, a greater match in linguistic style increases members’ participation in the community. Synchronicity in communicative behavior adds to the explanation of individual members’ participation. From a marketing perspective, this may be beneficial in terms of engaging with the customer online.

Linguistic style matches are build upon function words that represent the style of a message. One function word category is the category of pronouns such as ‘I’, ‘you’ and ‘they’. Research shows that subtle changes in the use of closeness-implying terms affect customers’ perceptions of and attitudes towards brands (Sela, Wheeler & Sarial-Abi, 2012). Using first-person plural pronouns (‘we’) indicates closeness, a shared identity (Brown & Gilman, 1960), an outgrowth or partnership, and confidence in being able to solve problems together (Seider, Hirschberger, Nelson & Levenson, 2009). On the other hand, first-person singular pronouns (‘I’) indicate an individuated identity and self-focus (Pennebaker et al., 2003). Applied to marketing, there is found that the use of ‘we’ leads to a more positive product attitude compared to the use of

(10)

‘you and [the brand]’ in advertising (Sela et al., 2012). Moreover, Chen, Lin, Choi & Hamn (2015) find that the use of personal pronouns in a social media context

increases customer engagement. It is interesting to investigate how the different personal pronouns used in influencer marketing affect customers’ perceptions of the influencer and their engagement with the content, as this is not yet examined before in the literature.

Overall, the literature has paid little attention to investigate the textual attributes of marketing expressions online. As social media speech shows its marketing potential, this research aims to apply the Communication Accommodation Theory (CAT) to a social media context. The relatively minimal media richness available on social media means that textual communication is fundamental in how users construct, share and form views, perceptions, identities and define their relationship with the sender or receiver of the message (Ludwig & De Ruyter, 2016; Pennebaker et al., 2003). Explicitly, this research will focus on Facebook as a social medium. The research extends the literature by examining the linguistic style dimensions used in Facebook expressions spread by a digital influencer. Specifically, we will look into how these dimensions used affect the growth of - and engagement with the influencer, as indicators of influencer success. Following the CAT, this research tries to predict the influencers’ success by the use of linguistic style matching (LSM) and more in-depth, personal pronouns. The following research question will be answered:

RQ: What is the effect of linguistic style dimensions in influencer marketing posts on Facebook on the influencers’ success?

(11)

Theoretical & Managerial Relevance

The research contributes to the field in several ways. Firstly, the textual analysis will be applied to the marketing context, which researchers have yet paid little attention to. Text analytics is an emergent and growing field, showing its potential by allowing us to analyze vast amounts of data and quantifying data that is mostly qualitative

(Ludwig et al., 2013). Secondly, the research contributes to the yet unexplored field of textual analysis of social media content spread by a digital influencer type. This will deeper our understanding of – and build upon the framework of textual analytics in influencer-communications. Thirdly, LSM has been applied in several researches, but yet limited research has examined the effective use of personal pronouns in textual expressions. Decoding the content of social media communications in textual elements can further improve the efficiency of determining potential impacts and quality of social media messages (Ludwig & De Ruyter, 2016).

From a managerial perspective, this research taps into a current trend in the marketing field. More and more marketers use influencer marketing a part of their digital

marketing activities, but they face challenges in executing it effectively (eMarketer, 2015a). There is a lack of understanding how people perceive social media marketing messages (Chung & Austria, 2010). Identifying the right influencers, finding the right engagement tactics and measuring performance of the influencer campaigns are the main challenges (eMarketer, 2015a). This research will provide guidelines for this, for both marketers and influencers, as they will gain insights in which text properties could be used best in sponsored content, to make sure the post is effective in terms of post engagement and moreover, maintaining and increasing the follower base of the influencer. The research will help marketers to steer for the right influencers and steer

(12)

for the right, fitting marketing content for them to spread, to enhance both influencer success as well as the effectiveness of the organizational, branded campaign.

Applying the results of this study may help predict the performance of influencer marketing, enhancing marketing intelligence. For the influencer, the research outcomes might help them to adjust their own content in general to make sure they continuously connect and engage with their follower base in the most effective way. Doing so may give them an advantage over their fellow influencers, being more attractive for brands to cooperate with.

The research is structured as follows: We begin by further elaborating on the theory, defining the concept of digital influencers and reviewing and listing current literature on linguistic style dimensions. Following the literature, six hypotheses will be posited. Then we proceed to the Method chapter, in which there will be elaborated on the way this research is conducted. In Chapter 4 the results of the analyses conducted will be displayed. In Chapter 5 conclusions will be drawn based on the results. The

conclusions will be heavily discussed and also several managerial implications and future research directions will be presented in this last chapter.

2. Theoretical Framework

Social Media

Social media can be defined as “a group of Internet-based applications that build on the ideological and technological foundations of Web 2.0, and that allow the creation and exchange of User Generated Content” (Kaplan & Haenlein, 2010). Social media allow for both social and professional interactions (Trusov, Bucklin & Pauwels,

(13)

2009). Members of social media can become friends with other members, but they can also become fans of brands on brand pages. Brands increasingly create these brand pages on social network sites, as social media form excellent platforms to develop customer relationships (De Vries et al., 2012). On these brand pages, brands can place posts containing, among others, messages, videos, and quizzes. The

challenge for brands is to attract people’s attention with these posts and engage them with the provided content (De Vries et al., 2012). The source of this content spread is the brand itself. Research indicates the importance of using a credible source for advertising purposes. People seem not to trust ads anymore. They do however, increasingly trust people and their voice of authority (Newman, 2015). Therefore, over the years, different people that radiate a certain authority were used by marketers to spread their marketing message in both offline and online advertising, mostly celebrity endorsers (i.e., actors). Since engagement is becoming more important than reach only, brands are changing their perceptions of the type of influencer they want to work with. As building a relationship with the customers is what matters, also referred to as relational marketing, organizations recognize the value of working with influencers with a smaller reach and more niche, also called ‘micro’ influencers, rather than using celebrity endorsers in campaigns (Reynaud, 2016).

Digital Influencers

Applied to the online context, these influencers can be referred to as digital

influencers. These digital influencers take a similar approach as brands do with their brand pages online. Influencers create a personal fan page on social media, on which they spread their (sponsored) content and interact with their follower base. To define a digital influencer, Uzunoğlu & Kip (2014) refer to the two-step flow theory of Katz &

(14)

Lazarsfeld (1955). This theory states that messages disseminated by mass media are filtered by opinion leaders, instead of having a general and direct effect. Certain people are recognized as opinion leaders, acting as intermediaries in the web of social interactions; they decode the messages and mediate the transmission of information, as they interpret the media information they receive and pass it on to others (Katz & Lazarsfeld, 1995). These opinion leaders can be defined as “the individuals who are likely to influence other persons in their immediate environment” (Katz & Lazarsfeld, 1995, p.3). They have a wide set of personal connections, play a key influential role (Weimann, 1994) and are considered as both source and guide (McQuail & Windahl, 1993). The theory states that interpersonal communication is more powerful

compared to mass media in affecting attitudes of individuals (Weimann, 1994). The majority of studies in the last decade were dominated by the necessity of face-to-face and personal interaction for the presence of opinion leadership. However, with the rise of Web 2.0, face-to-face-communication is not the only form of personal interaction anymore; rather than in geographical space, people are connected through the Internet in online communities (Bagozzi, Bergami, Marzocchi & Morandin, 2012). Therefore, the existence of online opinion leaders is increasingly acknowledged (Uzunoğlu & Kip, 2014). Applied to the online context, opinion leaders can be referred to as digital influencers (Kapitan & Silvera, 2015). The stated characteristics of traditional opinion leaders such as openness to messages, taking the role of the discussant, and being valued and influential among social contacts (Uzunoğlu & Kip, 2014), can be observed in these digital influencers. As an illustration, Kapitan and Silvera (2015) found that digital influencers have a strong influence on their reading audience.

Moreover, they found that individuals consider digital influencers as more trustworthy than celebrities. Pornpitakpan (2004) indicates the superiority of a high-credibility

(15)

source over a low-credibility source, resulting in more persuasion in both attitude and behavioral measures. More specifically, 84% of the millennials say that user

generated content, such as generated by influencers affects their purchase decision (Kapitan & Silvera, 2015). Additionally, the content digital influencers distribute is perceived as more powerful than branded content shared by organizations (Hall, 2010; Goh, Heng & Lin, 2013). This can be explained by the fact that electronic word of mouth (eWom) is perceived as more trustworthy by consumers and less likely to be manipulated compared to advertisements, empowering communication among peers on social media (Uzunoğlu & Kip, 2014).

Organizations use word-of-mouth marketing to reach out to a wide set of potential consumers and attract their attention via social interactions. Via word-of-mouth, information can be distributed more quickly and easily among social networks. In a social network, people are affected by the decisions of their peers and therefore, word-of-mouth is extremely powerful (Li, Lai & Chen, 2011). Organizations gradually discovered the far-reaching impact and viral growth potential of

approaching digital influencers too (De Veirman et al., 2016). According to Captive8, a company that connects influencers with brands, brands are spending more than $255 million on influencer marketing on Instagram every month (Weinswig, 2016). An example of brand that has been using an influencer to spread branded communications is Pantene, a company selling hair products. As illustrated in Figure 1, Pantene and Blogilates, the influencer in this example, have a partnership. In this post Blogilates introduces a new shampoo and conditioner. She expresses her positive thoughts regarding the brand and being their hair model. Followers could easily recognize the post is a sponsored post, meaning the influencer received a (monetary) compensation for posting this information. In this example this is indicated by ‘Paid’, mentioned

(16)

next to the date the post was published. In the post itself, the influencer refers to it by using the hashtag #ad, short for advertisement. Also #sponsored or #sp, which is short for sponsored, are often used to indicate a post is an advertisement, paid for by brands (Schwab, 2016).

(17)

This is a perfect example of how influencer marketing could look like. However, marketers have no idea if this post will be effective. The indicators of success are unknown for marketers (eMarketer, 2015a). This ‘uncertainty’ problem in marketing results in the risk that resources will be wasted on inefficient marketing. Guidelines could be helpful, also in terms of cost savings (Li et al., 2011).

Predicting Influence

For organizations, to carry out a successful influencer marketing strategy, marketers need to know what determines a successful influencer, and be able to identify and target them. According to Li et al. (2011), an influencer must be influential and powerful towards others in online social networks. An influencer needs to be

considered popular and their posts need to allow for engagement with their follower base. According to research by Launchmetrics (2017), a successful influencer is based on both quantity and quality of the audience and (levels of) engagement. Several researchers have addressed the question whether influence can be predicted. Research on Instagram influencers shows that influencers with a high number of followers are considered more likeable, mostly because they are considered to be more popular (De Veirman et al., 2016). Also, the number of post comments on the authors’ post (Li et al., 2011) and the amount of likes for the authors’ post are indicators of influence (Chen, Tang, Wu & Jheng, 2014).

Additionally, post-related (content-based) characteristics could be taken into account, when predicting the potential influence of an influencer (Li et al., 2011). Social networks consist of both text and multimedia options and this way, allow for various kinds of content in posts (Aggarwal & Wang, 2011). De Vries et al. (2012) have

(18)

investigated the possible drivers for brand post popularity, which is reflected in liking and commenting on the posts. They found out that vivid and interactive brand post characteristics enhance the number of likes of a post. Furthermore, an interactive brand post characteristic, like a question, can enhance the number of comments. In contradiction to their expectations, the researchers didn’t find any support for the factors of a post being either informative or entertaining, affecting the post popularity. Fans don’t consider an informative or entertaining post more attractive compared to a non-informative or non-entertaining brand post (De Vries et al., 2012). Berger & Milkman (2012) try to explain why people share online news content, which results in content going viral. They find that positive content is more viral compared to negative content, and both lead to more shares than neutral content. Also people tend to share surprising, interesting or practically useful content more. Another important message element is whether the message elicits an emotion: to go viral, content needs to include a surprising aspect, in combination with another emotion like joy, anger, or fear (Dobele, Lindgreen, Beverland, Vanhamme & Van Wijk, 2007). Li et al. (2011) take several other post-related characteristics into account and identify length of a (blog) post and subjectiveness of a post as predictors for the influential value of a (bloggers’) post.

Language

Not only is it important which message elements or content affects how a consumer interacts with the sender of the message; it is also essential to examine how this content is communicated (Labraque & Swani, 2017). According to Ludwig and De Ruyter (2016), language is the key to decoding social media speak. They refer to the most influential linguistic theory to study the use of language: the Speech Act Theory

(19)

(SAT) as introduced by J.L. Austin in 1962. SAT refers to “how word categories and sentence constructions, apparent in people’s everyday language use, give insights into their intentions, perceptions, and identities” (Ludwig & De Ruyter, 2016, p.125). Meaning, words can be used not only to present information, but also to carry out actions. Applied to the online context, speech acts in social media communications reflect the writers’ intentions and behaviors and define the accessibility and

diagnosticity of social media messages for their audience (Ludwig & De Ruyter, 2016). Emerging research on text-based communications suggest that both linguistic content and style elements of social media speak are relevant inputs that help

determine the accessibility and diagnosticity of the message (Huffaker, Swaab & Diermeier, 2011). The content of communication is expressed by content words, which are generally nouns, regular verbs, and many adjectives and adverbs. These words express what people are saying, i.e., the sentential meaning of the message. Content can’t be communicated without style words, often referred to as function words. Function words are made up of pronouns (such as I, you, they), prepositions (to, of, for), articles (a, an, the), conjunctions (and, but, if), auxiliary verbs (is, am, have), and a few other esoteric word categories. These words express how people communicate, i.e., the sentential style of the message (Pennebaker, 2011; Tausczik & Pennebaker, 2010; Ludwig et al., 2013). Consider the following example, as presented by Chung and Pennebaker (2007), which shows the way in which three different people summarize how they feel about drinking a juice:

Person A: “The experience of drinking this bottle of juice, is certainly quite satisfactory.”

Person B: “I’d have to say that I like this juice.” Person C: “Yummy. Good stuff.”

(20)

All three say essentially the same thing, but they have different ways of expressing themselves. Person A is very formal and stiff, person B is a bit hesitant. Person C is more extrovert and easy-going (Chung & Pennebaker, 2007). Specific linguistic styles reveal aspects of the authors’ personality (Ludwig et al., 2013). All three people use different pronouns, have different usage of large versus small words, verbs, and different other style dimensions to express themselves. Detecting linguistic style may start by identifying ‘junk words’ - words that don’t convey much real content. These junk words serve as the cement that holds content words together, and can be referred to as function words (Chung & Pennebaker, 2007). Prior research has demonstrated the importance of function words for determining conversational outcomes (Huffaker et al., 2011). In the English language, around 500 words are function words. This is only 0.05% of the total of 100.000 English words in our vocabulary. However, these function words reflect 55% of our daily word usage - the words we speak, hear and read (Tausczik & Pennebaker, 2010).

LSM and Identification

The Communication Accommodation Theory, developed by Howard Giles in the 1970s, suggests that in interaction we adjust and accommodate our communication style to others (Giles, 2009). Meaning, we adjust our function words use to one another. “People consciously or subconsciously accommodate their dialogue partners to develop closer relationships and signal empathy, credibility, and a common social identity.” (Ludwig et al., 2014, p.1203). Greater degrees of synchronization in communication styles in conversations lead participants to perceive a common social identity, decrease their perceptions of social distance and elicit more approval and trust (Pickering & Garrod, 2004). For example, for couples on a first date, a greater

(21)

match in their conversation style predicts their subsequent relationship initiation and stability, as founded by Ireland and Pennebaker (2010). A linguistic style match (LSM) between conversants - that is, the use of similar function words - represents a form of psychological synchrony (Ireland & Pennebaker, 2010; Pickering & Garrod, 2004). LSM signals social identification, increases shared perceptions and credibility (Ireland & Pennebaker, 2010) and decreases perceptions of social distance (Chung & Pennebaker, 2007), in both verbal and written (text-based) communication.

Applied to the online context, Huffaker et al. (2011) find that in online text-based negotiations, a greater match in function word usage increases interpersonal rapport and agreement among conversants. Also in an online review setting a greater LSM score helps readers to establish rapport with the source (Ludwig et al., 2013). As perceived rapport allows for readily accessible diagnostic information, the receivers’ judgments and behaviors are affected. Readers are stimulated to rely on the source cues to form attitudes, perhaps even being stimulated towards the exclusion of the message content (Pornpitakpan, 2004; Ludwig et al., 2013). Linguistic style establishes source perceptions and elicits a positive bias (Ludwig et al., 2013).

Following the CAT, we argue that accommodations in linguistic style online represent symbolic acts. LSM signals social identification with the conversation partner

(Ludwig et al., 2014). We argue that function words serve as a subtle, subconscious way to construct shared meaning in the online context. Identification mediates message effects (Basil, 1996; Jin & Phua, 2014); a spokesperson which whom the audience identifies, is most likely to achieve lasting attitude or behavior change (Basil, 1996). Regarding behavior, Ludwig et al. (2013) find that in an online review setting, increasing congruence with the product interest group’s typical linguistic style increases conversion rates directly. Online reviews influence customer behavior to a

(22)

greater extend, when they match the linguistic style of the target audience. In this setting, LSM serves as a good predictor of attitudes and behavior (Huffaker et al., 2011; Ireland & Pennebaker, 2010; Ludwig et al., 2013). These results suggest that a message with a high LSM score, spread by an influencer, signals and increases identification with the influencer, which results in a positive attitude towards the influencer and the behavioral action of liking the influencer. If overall, a great LSM in influencer posts is present, this results in a bigger follower base of the influencer, i.e., a more successful influencer in terms of quantity. Consequently, we posit the

following:

H1: A linguistic style match between the influencers’ and their followers’ linguistic style is positively related with the number of followers of the influencer.

Ludwig et al. (2014) have investigated textual features and synchronicity in

communicative behavior in online user communities. They seek to explain how text-based communication may drive participation efforts, also taking on a language-as-action perspective, diverging from the transitional language-as-product view (Brennan & Clark, 1996; Ludwig et al., 2014). In the research they find that, at a community level, members’ community identification critically influences their participation efforts. A greater LSM fosters individual members’ participation behavior in both quantity and quality within the community. Participation quantity refers, in this case, to the amount of posts. The participation quality refers to contributing posts with more developed arguments, leading to better group discussion outcomes (Ludwig et al., 2014). Participation efforts in a social media setting refer to engagement activities at the message level: liking and sharing of - and commenting on posts. This may also be

(23)

referred to as post popularity (De Vries et al., 2012; Labrecque & Swani, 2017). In this specific research this variable will be referred to as post engagement. Engagement on social media sites is important, since the success of social media pages depends on consumers engaging with the communications that help the page owner (Berger & Milkman, 2012). We believe the results of the study of Ludwig et al. (2014) can be applied to a social media setting, i.e., that LSM increases post engagement in this setting. Consequently, we posit the following:

H2a: The greater the convergence (divergence) in LSM between the influencers’ post and their followers’ linguistic style, the higher (lower) the post likes.

H2b: The greater the convergence (divergence) in LSM between the influencers’ post and their followers’ linguistic style, the higher (lower) the post comments.

H2c: The greater the convergence (divergence) in LSM between the influencers’ post and their followers’ linguistic style, the higher (lower) the post shares.

What about Pronouns?

Taking a closer look at the function word categories, indicating the linguistic style of a message, several research has pointed the importance of pronoun use in language as an indicator of message evaluations. Pronouns reveal information about the sender of the message and have the power to directly involve the message receiver (or not) and therefore, are important as part of speech (Noguti, 2016). Furthermore, pronouns provide an important means for both communications as reinforcing perceptions about relationships (Gordon, Grosz & Gilliom, 1993; Sela et al., 2012). As personal

pronouns implicitly claim relationships between brands and their audiences (Pollach, 2005), brands are increasingly using them in their online expressions, in order to

(24)

emphasize continuing relationships (Kwon & Sung, 2013) and to reduce the impersonality of mass communication (Fairclough, 1989).

First-person singular pronouns (e.g., ‘I’, ‘me’) act as an indicator for ‘self’, whereas first-person plural (e.g., ‘we’), second (e.g., ‘you’), and third-person (e.g., ‘he’, ‘she’) pronouns emphasize the ‘other’ (Zimmermann, Wolf, Bock, Peham & Benecke, 2013; Pennebaker, 2011). First-person plural pronouns (‘we’) indicate closeness and shared identity (Brown & Gilman, 1960). For example, couples that use first-person plural pronouns such as ‘we, ‘our’, and ‘us’, behave more positively to one another than couples who emphasize their separateness by using first-person singular and second-person pronouns such as ‘I’, ‘me’, and ‘you’, who have proven to be less satisfied in their relationship (Seider et al., 2009). Furthermore, people rate both their own and others’ interpersonal relationships as closer and of higher quality when they are described using the pronoun ‘we’, rather than ‘the other person and I’ (Fitzsimmons & Kay, 2004). This could also be applied to marketing communications, as subtle

changes in the use of closeness-implying terms have proven to affect customers’ perceptions of - and attitudes towards brands (Sela et al., 2012). The use of closeness-implying pronouns (‘we’ vs. ‘you and [the brand]’) in advertising creates more positive product attitudes. Customers expecting a close relationship with the brand evaluated a brand more favorable after reading a message in which the brand referred to the brand and the customer using the pronoun ‘we’ (Sela et al., 2012). Based on this research we posit that the use of first-person plural pronouns in influencer

marketing will result in a more positive evaluation of the influencer as we assume, the receiver of the message expects (to develop) a close relationship with the sender of the message, i.e., the influencer. A positive evaluation of the influencer is reflected in the number of followers of this influencer. Accordingly, the following hypothesis is

(25)

composed:

H3: The use of first-person plural pronouns in posts relates positively with the number of followers of the influencer.

Moreover, Chen, Lin, Choi & Hamn (2015) find that the use of personal pronouns in general, influences customers’ engagement including likes, shares, and comments. In their research context (Facebook), they found second-person pronouns to be most common personification strategy used in brand posts. Danescu-Niculescu-Mizil, Gamon, and Dumais (2011) confirm the hypothesis of linguistic style accommodation in a real life social media setting in their research. In Twitter conversations they identify the existence of synchronization in style dimensions. However, they also find differences in style categories. For first-person plural pronouns (‘we’) for example, they find that even the least accommodating users still match the style of the most accommodating participants, which they refer to as symmetric accommodation. For second-person pronouns (‘you’) they find that only one participant in the conversation accommodates (asymmetric accommodation) (Danescu-Niculescu-Mizil et al., 2011). This makes sense, given the fact that a word like ‘you’ has a different meaning for the two participants in a conversation. In a marketing context, this could mean that brands are talking to their customers and use second-person pronouns to directly address them (e.g., by asking them a question) (Chen et al., 2015). Brands would expect a reaction to this appeal, expressed in any engagement level. Using second-person pronouns typically invites readers into the conversation and can be regarded as an invitation to directly engage (Pollach, 2005).

(26)

entity from the audience (Halmari, 2005), but at the same time first-person singular pronouns reveal an authorial voice through which one another can portray him- or herself as an expert (Dueñas, 2007). The use of these pronouns may be related to better perceptions of message credibility, given the potential association with a more competent or experienced source (Noguti, 2016). Noguti (2016) finds that, when experience is relevant, the use of first-person singular pronouns increases engagement with a post (post ranking and comments) in online content communities. Based on both the results of the study of Pollach (2005) and the study of Noguti (2016) we posit that both the use of first-person singular pronouns and the use of second-person pronouns individually can increase post engagement.

H4a: The use of first-person singular pronouns in posts relates positively with post likes.

H4b: The use of first-person singular pronouns in posts relates positively with post comments.

H4c: The use of first-person singular pronouns in posts relates positively with post shares.

H5a: The use of second-person pronouns in posts relates positively with post likes. H5b: The use of second-person pronouns in posts relates positively with post comments.

H5c: The use of second-person pronouns in posts relates positively with post shares.

The use of first-person singular pronouns (‘I’) shows a great symmetric

(27)

relationships, meaning that, in a relationship, a person may feel obligated to reveal as much as information as he has had from another in order to maintain equality in the exchange (Derlega, Harris & Chaikin, 1973). In order to maintain a ‘norm or

reciprocity’, the value of behaviors exchanged between individuals is expected to be of comparable value (Gouldner, 1960). Applied to the online context it may be that, the more the influencer discloses about his or her personal life, the more followers are willing to share about their experiences online as well, in order to maintain the ‘norm of reciprocity’. Sharing personal information (“I have [..]”, “I am [..]” etc.) or

individual attitudes (“I think [..]”, “My opinion [..]” etc.) can be indicated by the use of first-person singular pronouns. Accordingly, we posit the following:

H6: The use of first-person singular pronouns in influencer posts leads to an increase in the use of first-singular pronouns in comments.

3. Method

To answer the research question a content analysis was conducted. A content analysis is “a research technique for the objective, systematic, and quantitative description of the manifest content of communication” (Bryman, 2001, p. 178). Specifically, a data mining method was used, in a social media context. This method was chosen as it allows for gathering a large dataset to analyze, examining a large pre-existing

database in order to generate new information. Among social media, Facebook is the largest platform with 1.871 million active users (Statista, January 2017). According to US influencers, Facebook was the best and most popular channel to use for influencer marketing in 2016 (eMarketer, 2016). Moreover, Facebook has not restricted the data

(28)

access via the API and had not announced to do so in the period of data collection. Therefore, as Facebook gives the access opportunity to acquire a large set of social data and the fact that it allows for gathering accurate data of active users (both influencers and followers), Facebook was chosen as the social media platform to investigate in this research.

Influencer Characteristics

As elaborated on in Chapter 2, a digital influencer can be defined as the opinion leader “who is likely to influence other persons in their immediate environment” (Katz & Lazarsfeld, 1995, p.3). Katz (1957) defines the dimensions of opinion leaders based on “the personification of certain values (who one is), competence (what one knows), and strategic social location (whom one knows)” (p.73). Applied to the online context, opinion leaders’ influence is carried out online via social media like Twitter, Facebook, or YouTube (e.g., via vlogs), or websites (e.g., in blogs) (Chaffey, 2015). Digital influencers can be various types of people. It can be a celebrity, with a big audience size and wide reach. But it can also be an expert, who has a high

expertise level in his/her focus area and is focused on contextual fit and engagement with the audience mainly, instead of reach. It can also be a blogger, who exhales his/her personal brand, or a journalist (Chaffey, 2015).

In practice, the definition of Brian Solis is used often. According to him, an influencer is “someone of notable status and focus within a community, who possess the ability to cause effect or change behavior among those to whom they’re connected” (Solis, 2012). Mainly marketers group influencers by dividing them in mega, macro and micro lists; mega influencers referring to the influencers with the most reach, that is, more than a million followers, macro influence referring to the ones with an audience

(29)

size of 10.000 to a million follower, and micro influencers referring to influencers with between 500 and 10.000 followers (Insightpool, 2017). However, there is no such a specific operationalization of a digital influencer verified yet in academic research. Besides, engagement is becoming increasingly important, compared to reach only (Chaffey, 2015). The focus of marketers shifts from quantity to quality when dealing with social influencers. Several brands have experienced successful partnerships with micro influencers as their followers seem to be more engaged (Weinswig, 2016). Moreover, in this scenario, “a brand receives intangible benefits like authenticity, a unique point of view, deeper storytelling, and the potential of reaching a more tailored audience”, according to Rebecca Suhrawardi (2016), fashion journalist and Forbes contributor. So not only reach, but also resonance (engagement with the audience) and relevance (contextual fit) should be criteria when defining in influencer (Chaffey, 2015).

Combining the various influencer dimensions mentioned in the previous paragraphs, we’ve defined seven criteria to identify the influencers for this research. Also

practical issues played a role in defining the criteria, as the text-mining program used only allowed for analyzing English text. Moreover, we needed to assure to gather enough data points (both post and comments) for each influencer. The criteria are listed below.

The influencer:

i) had to post a post at least once in two weeks;

ii) had to promote a brand in their posts at least once a month; iii) had to post in English all the time;

(30)

specific field of interest, expressing that clearly on their page and in their posts;

v) had to have a considerable amount of followers (> 500 page likes); vi) had to have more than one like and one English comment on each post,

and

vii) had to be an individual human being or a group of individuals.

In total 34 influencers were selected by applying a convenience sampling method. A list of the selected influencers can be found in Appendix 1. Of these 34 influencers, 19 specified their gender as ‘female’ and five specified their gender as ‘male’. The remaining ten didn’t specify a gender. At the beginning of the data collection period, April 13th 2017, the number of page likes, also referred to as number of followers of a page, was gathered for all the individual pages. The influencers were selected in a way that variance in size of the follower base was represented in the selection. On average, influencers had 4.464.885,97 page likes (SD = 7.765.704,10), with a minimum of 5.642 and a maximum of 35.233.251 likes, indicating the sample was quite diverse in terms of follower base size of the influencers. Also we made sure to include various influencer types in different categories in the sample. The variable ‘category’ was manually categorized by a coder, as this value wasn’t consistently and accurately indicated on the individual Facebook pages. ‘Category’ indicates the specific field of interest of an influencer and/or the industry the influencer is active or works in. The 34 influencers were categorized as Celebrity (n = 5), Beauty (n = 4), Sports (n = 6), Fashion (n = 4), Food (n = 3), Lifestyle (n = 5), Music (n = 3), News (n =3), and Travel (n = 1). Moreover, the coder classified the variable ‘influencer type’. This variable indicated the influencers’ composition, i.e., whether the influencer was

(31)

an individual, a duo, or a group of people posting on behalf of a magazine or newspaper online. Five influencers were classified as duo; six were classified as a magazine or newspaper. The remaining 23 were classified as solo influencers.

Controlling for both gender, category and influencer type in the analysis allowed us to see whether there were differences in influencer characteristics, affecting influencer success.

Data Collection

Data was scraped from Facebook by communicating with the application-programming interface (API) of Facebook. The program R was used to build a scraper. This scraper scanned the Facebook pages of the selected influencers and collected the page likes at that point in time. Moreover, it scraped all the posts of the influencers’ public Facebook page, including the posts (id, post type and message), the amount of likes on the posts (indicated as likes count), the posts’ comments (comments count), and shares (shares count). Additional code was written and used to gather the comments on the posts as text input.

The scraper collected data from the public Facebook pages of the influencers. The engagement on this page, i.e., the likes, shares, comments (text) and number of page likes, was created by the visitors of their page, also known as the followers of the page. The data input of these participants has been collected anonymously, meaning there was no data available regarding their demographics, ensuring ethical standards. The data was gathered in a real-word setting. Participants weren’t aware that

anonymous data was collected about them, meaning the data was gathered in the ultimate natural environment. Data was collected during a period of 4 weeks, from April 13th until May 11th 2017. On the first and last day of this period the amount of

(32)

likes of the individual pages were collected. This panel setting allowed us to observe the influencers’ follower base over time, indicating a growth rate of followers. This growth rate was used to indicate the number of followers of the influencer as

presented in the first and third hypothesis. The actual number of followers couldn’t be compared fairly as a score and since there was no information about the date of page creation available at the moment this research was conducted, the growth rate over a period of 4 weeks acted as an accurate and fair means to indicate and compare the influencers’ follower bases. The posts and comments on the pages, published between April 13th and May 11th 2017, were gathered all at once on May 12th 2017.

In total 8.971 posts and 1.262.860 comments were retrieved. Of these posts, 8.429 were collected from magazine/newspaper pages and only 542 posts were collected from pages of duo and solo influencers. Therefore, for newspapers and magazines, a random sample of 75 posts was selected from all the gathered posts, in order to present equal categories for each subject in multilevel analysis later on. After this selection was applied, 992 posts and 205.850 comments were left in the database. Comments and posts without any textual description were deleted from the dataset. Also posts without any comments were deleted from the dataset. In the end, the dataset consisted of 841 posts that were analyzed.

Measurement Development - Independent Variables

After the data was gathered, a content analysis of the influencers’ posts content and the text comments on the posts was conducted. The Linguistic Inquiry and Word Count (LIWC) program was used to systematically and automatically analyze the texts. This program has been developed to analyze emotional writing. As the program is coded by software and human coders, the LIWC offers a strong, reliable

(33)

convergence between the dimensions they extract (Pennebaker, Booth & Francis, 2007). The validity of the program has also been confirmed by more than 100 studies that have applied this methodology to various texts, including analyzing emotions in online content like blogs or instant messaging (Cohn, Mehl, & Pennebaker, 2004; Slatcher & Pennebaker, 2006). LIWC has been used in marketing research

successfully and consistently over the last years (Agnihotri & Bhattacharya, 2016) and therefore, this present study uses the same program. Using word counts for a given text, the program calculates the proportion of words that match predefined dictionaries (Ludwig et al., 2013). In this research one dictionary from the LIWC program was used, called ‘function’.

In line with recent research on LSM (Gonzales, Hancock & Pennebaker, 2010; Ireland & Pennebaker, 2010;. Ludwig et al., 2014), LSM was operationalized as a measure of the degree to which two or more conversant produce similar usage intensities of function words. The linguistic style profile of each post was compared to the

linguistic style profiles of its comments, to determine the linguistic style match score between influencer and follower. More specifically, eight function words categories were identified separately within the dictionary of function words (Table 1). The different word categories can be referred to as dimensions. For the personal pronouns, five individual dimensions were identified (Ludwig et al., 2013; Pennebaker, Boyd, Jordan & Blackburn, 2015).

The automatic retrieval of the function words in a text produced a measure to create a linguistic style profile for each Facebook post spread by an influencer. For example, ‘a’ appears in the function word dictionary ‘articles’ and would be counted as 1 in the total amount of function words in a text. If the word ‘the’ was in the same post, which

(34)

Table 1: Function Word Categories Used to Calculate LSM and Pronoun Use

LIWC Dimension Example(s)

Personal pronouns First-person singular I First-person plural We

Second-person You

Third-person singular She, he Third-person plural They

Impersonal pronouns It, that, anything

Articles A, an, the

Prepositions In, under, about

Auxiliary verbs Shall, we, was

Common adverbs Very, rather, just

Conjunctions And, but, because

Negations No, not, never

Note: We conducted text mining using the 2015 LIWC Program (Pennebaker et al., 2015).

also appears in the dictionary (specifically, as a preposition), this word would be counted too and the total score of function words would be 2. At the end of the content analysis, LIWC calculates the total number of function words that appear in a text, divided by the total number of words of the text, to determine the percentage of the text that falls into the function word linguistic style category (Ludwig et al., 2013). In this research this percentage is indicated as function word usage intensity (FWC). The LIWC program was applied to determine the FWC for the posts (p) of each influencer for each function word category (FWC_p) and to determine the FWC for each individual comment (c) on a post for each function word category (FWC_c).

To indicate an average function word usage intensity of the followers of an influencer, all comments were grouped per post (id). An average FWC score was calculated by

(35)

taking the sum of FWC_c per group, i.e., for all comments with the same post id, divided by the number of data points in that group, i.e., the amount of comments for that post (id). That way average function word usage intensity for comments per post was calculated (FWC_ac) for each individual post. Then, the linguistic style match (LSM) between each post and its comments was derived using the following equation:

LSM = 1 −

, resulting in a score between 0 and 1.

In this equation, LSM is the ratio of overlap between the usage intensity by each individual influencer for each function word category and the cumulative average usage intensity of the same function word category by all comments that were posted by followers on the Facebook post. Higher numbers represent greater linguistic similarity between influencer and follower. In the denominator, we added .0001 to prevent empty sets. In this dataset the average LSM score was 0.64 (SD = 0.30) with a minimum score of 0.00002 and a maximum of 1.00 - indicating a 100% match.

Additionally, for each (sub) dimension in the category of personal pronouns (See Table 1) a score was calculated separately. This means an individual score for first-person singular pronoun, first-first-person plural pronoun, second-first-person pronoun, third-person singular pronoun and, third-third-person plural pronoun use was calculated to indicate the percentage of each personal pronoun type used in the text. This was calculated for both post and comments text. For comments, an average score was calculated for all comments per post, indicating the average use of several pronoun categories in posts’ comments. Calculation of this average score was executed

(36)

similarly to the calculation of FWC_ac. The usage intensity of first-person plural pronouns in posts was used as the independent variable for testing hypothesis 3. The usage intensity of first-person singular pronouns in posts was used as an independent variable for testing hypothesis 4; the usage intensity of second-person pronouns in posts was used as an independent variable to test hypothesis 5. To test hypothesis 6, first-person singular pronoun usage intensity in posts served as an independent variable. The average first-person singular pronoun usage intensity in comments served as the dependent variable in the model.

Measurement Development - Dependent Variables

With regard to the other dependent variables in this research, influencer success was operationalized as an indication of follower growth and post engagement. Follower growth was indicated by a growth rate, calculated by comparing the amount of page likes at the start of data collection (April 13th 2017) indicated by followers1, and the amount of page likes at the end date of data collection (May 11th_{2017), indicated by} followers2. The following formula was applied to calculate a growth rate for each influencer:

growth rate = followers2 – followers1 x 100% followers1

On average, the 34 influencers grew with 1.37% in follower base within 4 weeks (SD = 2.96) with a minimum of -0.29% and a maximum growth of 15.96%.

Post engagement was measured by separately taking the amount of post likes, comments, and shares, and running individual models on these post engagement

(37)

indicators. While collecting the data from Facebook, for each individual post a likes, comments, and shares count was included in the output. In order to compare the amount of likes, comments and shares for each post fairly, a ratio score was calculated by dividing the amount of likes, comments, and shares by the number of followers of the influencer page the post belonged to. The following calculations were applied:

likes ratio = likes count per post x 100%

followers1 of page post belongs to

comments ratio = comments count per post x 100% followers1 of page post belongs to

shares ratio = shares count per post x 100% followers1 of page post belongs to

In further analysis only these ratio scores where used to indicate the amount of likes, comments, and shares of a post. In the results section these variables are simply referred to as the likes, comments, and shares. Descriptives of these variables are indicated in Table 2. Notable is the relatively low engagement with the various posts; for example, on average only 0.14% of the followers likes the posts on the

influencers’ page.

Control Variables

In addition to communication style dimensions, several other influencer-related characteristics might influence the influencers’ success. Therefore, in this research, there was controlled for gender, category, and influencer type, as indicated before. Moreover, there is controlled for post regularity. Post regularity indicates the frequency in which an influencer has been posting during the period of data

(38)

collection. More specifically, post regularity was operationalized as the sum of posts with a textual description an influencer publishes on their individual Facebook page within the period of data collection. On average, the influencers in the sample posted 258 times (SD = 552.18) with a minimum frequency of 1 and a maximum of 1707. As elaborated on before, magazines have higher post regularity (M = 1386.17; SD = 394; MIN = 683; MAX = 1707) compared to duo or solo influencers, who have an average post regularity of 16 posts (SD = 18.46; MIN = 1; MAX = 79). As mentioned before, to control for this difference, a random sample of posts from magazine and newspaper pages was included in the final posts dataset. Post regularity doesn’t take into account this random sample applied for these subjects and indicates the post frequency before the random sampling was applied.

Lastly, as also post-related characteristics might be of influence, ‘post type’ was identified as a control variable in the research. Post type indicates whether the post spread by an influencer was a status update, link, photo, or video. Several examples of the different post types are added to the Appendix (Appendix 2). Per post, the R scraper collected the ‘post type value’ automatically. Of all 841 posts, 375 posts were links, 301 posts were photos, 160 posts were videos, and 5 posts were status updates.

The control variables gender, category, influencer type, and post type were transformed into dummy variables in order to include them in the model. For

example, for post type, four dummy variables were composed: ‘link’, ‘photo’, ‘video’, and ‘status’. For the dummy variable ‘link’ applies that a score of 1 indicates that the post is a link update; a score of 0 indicates the post is not a link, but belongs to any other category for post type; for ‘photo’ applies that a score of 1 indicates that the post is a photo update; a score of 0 indicates that the post is not a photo update, but

(39)

belongs to any other category for post type, etc. For gender, we distinguish three dummies: ‘male’, ‘female’ and ‘not specified’. This last dummy variable is used to indicate that an influencer didn’t specify a gender on the public Facebook page, i.e., a score of 0 for this dummy variable indicates that an influencer did specify a gender online, either being a male or female.

Data Analysis

After the dataset was cleaned as missing values were deleted and all variables were constructed in the way mentioned above, a correlation analysis was applied. This analysis gave a first indication of correlation between all variables in the study. After that, several hierarchical regression analyses were performed using SPSS to test the hypothesis in the research, assuming the gathered data points were each independent. This analysis was applied, as it is a technique that examines the linear relationship between one or more quantitative independent variables and one quantitative

dependent variable (Fox, 1984). This could be applied to the hypothesis in this study. However, to verify the results of the hierarchical regression models, additional analysis was required. The dataset asks for a multilevel analysis, as multiple data points were assigned to each influencer. The data points may not be individual

independent data points as assumed in the hierarchical modeling. Multilevel modeling was performed using MLwIN, software for advanced multilevel modeling. All

hypotheses were tested gain. After that the results from the hierarchical and multilevel modeling were compared. A conclusion was drawn about the hypothesis and accuracy of the different results. Finally, additional curve estimation was performed to further investigate the relationship between LSM and the success metrics.

(40)

4. Results

Correlation Analysis

The correlation analysis conducted indicates the relationship between the variables used in the study. Table 2 outlines the results, i.e., descriptive statistics and

correlations of the influencer-related, post-related, and comment-related variables. The correlation between LSM and growth rate indicates a negative relationship, against the expected positive direction. Similarly, LSM correlates negatively with both post likes and comments, against the expected positive correlation. LSM doesn’t significantly correlate with shares. The use of ‘we’ in posts doesn’t significantly correlate with growth, indicating the first proof of rejection of Hypothesis 3. The use of ‘I’ in posts only correlates positively with post engagement in likes, and not with comments or shares. Using ‘you’ in posts doesn’t show any correlation with post engagement metrics. And lastly, the use of ‘I’ in posts correlates with the occurrence of ‘I’ in comments, indicating first proof for support of Hypothesis 6.

Notable in the correlation analysis are the control variables post regularity and gender, indicating a significant strong correlation with each other and several medium to strong correlations with other variables, like likes and comments. Small correlations exist between gender, category, and post type. Growth rate correlates among others, with likes and comments, but these correlations are rather small effects. Further analysis of several regression models has to prove the causality of the indicated correlations.

(41)

(42)

Regression Analysis

To capture the influence of the explanatory variables on subsequent influencer success, firstly, five hierarchical linear models (HLMs) were specified assuming the gathered data points were each independent, and differences between the individual data points could be found. The modeling controls for the possibility that various communication spread by the same influencer (as multiple posts are nested in each influencer page) may differ, influencing the dependent variables in the research.

Firstly, we estimated how variance in LSM and the use of ‘we’ in posts can explain influencer success indicated by follower growth rate, in the first model specified; Hierarchal multiple regression was performed to investigate the ability of LSM and the use of first-person plural pronouns in posts to predict a growth in followers, after controlling for gender, category, influencer type, post regularity, and post type (Model 1). In the first step of hierarchical multiple regression, five predictors were entered: gender, category, influencer type, post regularity and post type. Gender, category, influencer type, and post type were added as dummy variables. The model was statistically significant, F (16, 824) = 12.46; p < .001, and explained 19.5% of the variance in follower growth. After entry of LSM and first-person plural pronoun use at Step 2, the total variance explained by the model as a whole was 20.2%, F (18, 822) = 11.55, p < .001. The introduction of LSM and first-person plural pronoun use

explained additionally 0.7% of the variance in follower growth, after controlling for gender, category, influencer type, post regularity, and post type (R2 Change = .007; F (2, 822) = 3.59; p < .05). In the final model four out of seven predictor variables were statistically significant, indicating five dummy variables for categories, post

(43)

-0.09 (β = -.09, p < .01), meaning if LSM increases for one, growth rate will decrease with 0.09. This result indicates a negative relationship, meaning that H1 is not

supported. The effect of first-person plural pronoun use is insignificant (p = .836), indicating that H3 is not supported either. As for categories, the category ‘news’ was used as a baseline variable to compare with. Results show that if an influencer is categorized as ‘beauty’ (β = .13, p < .01), ‘lifestyle’ (β = .36, p < .001), ‘sports’ (β = .23, p < .01) or ‘travel’ (β = .16, p < .01), this influencer is more likely to grow faster (compared to one in the category ‘news’). An influencer in the category ‘music’ (β = -.10, p < .05) is less likely to grow (relative to one in the category ‘news’). Lastly, the model shows a significant result for both the dummy variable female (β = -.56, p < .001) and male (β = -.44, p < .001), indicating that specifying a gender as an influencer results in a slower growth pace. Moreover, post regularity shows a

significant negative relationship, indicating that the more regular one posts, the lower the growth rate of the influencer (β = -.24, p < .05). The results are displayed in Table 3. In this table, first-person plural pronoun usage is indicated as ‘we’.

Secondly, a hierarchical multiple regression was performed to investigate the ability of LSM, the use of first-person singular pronouns, and the use of second-person pronouns in posts to predict post engagement in terms of likes, after controlling for gender, category, influencer type, post regularity, and post type (Model 2). In the first step of hierarchical multiple regression, five predictors were entered: gender,

category, influencer type, post regularity, and post type. Gender, category, influencer type and post type were again added as dummy variables. This model was statistically significant, F (16, 824) = 33.46, p < .001, and explained 39.4% of the variance in post likes. After entry of LSM and first-person singular pronoun use and second-person

(44)

pronoun use at Step 2, the total variance explained by the model as a whole was 40.5%, F (19, 821) = 29.41, p < .001. The introduction of LSM and first-person singular pronoun use and second-person pronoun use explained additionally 1.1% of the variance in likes, after controlling for gender, category, influencer type, post regularity, and post type (R2 Change = .011; F (3, 821) = 5.13; p < .01).

In the final model five out of eight predictor variables were statistically significant, indicating three dummy variables for categories, two dummy variables for influencer type, one dummy variable for post type, gender, and LSM influencing likes. LSM recorded a Beta value of -0.10 (p < .001), meaning if LSM increases for one, post likes will decrease with 0.10. This result indicates that H2a is not supported.

First-person singular pronoun use has an insignificant effect on post likes (p = .47). H4a is rejected. Also the effect of second-person pronoun use is insignificant (p = .11), indicating that H5a is not supported either. Again gender shows a negative effect, this time on post likes (male: β = -.73, p < .001; female: β = -1.12, p < .001). Influencers in the category ‘celebrity’ and ‘food’ have significantly higher chances on receiving post likes, compared to influencers in the category ‘news’. ‘Music’

influencers have less chance to get post likes (celebrity: β = .32, p < .001; food: β = .15, p < .01; music: β = -.12; p < .01). A status update receives slightly less likes compared to a link (β = -.07, p < .05) and both duo and solo influencers receive relatively more likes on posts compared to magazines (duo: β = .50, p < .001; solo: β = 1.36, p < .001). Results of the model are shown in Table 3. In this table first-person singular pronoun usage is indicated as ‘I’, second-person pronoun usage is indicated as ‘you’.

Use the right words : research on linguistic style dimensions influencer marketing and its effect on influencer success