A Mimetic Approach to Social Influence on Instagram

We combine philosophical theories with quantitative analyses of online data to propose a sophisticated approach to social media influencers. Identifying influencers as communication systems emerging from a dialectic interactional process between content creators and in-development audiences, we define them mainly using the composition of their audience and the type of publications they use to communicate. To examine these two parameters, we analyse the audiences of 619 Instagram accounts of French, English, and American influencers and 2,400 of their publications in light of Girard’s mimetic theory and McLuhan’s media theory. We observe meaningful differences in influencers’ profiles, typical audiences, and content type across influencers’ classes, supporting the claim that such communication systems are articulated around ‘reading contracts’ upon which influencers’ image is based and from which their influence derives. While the upkeep of their influence relies on them sticking to this contract, we observe that successful influencers shift their content type when growing their audiences and explain the strategies they implement to address this double bind. Different types of contract breaches then lead to distinct outcomes, which we identify by analysing various types of followers’ feedback. In mediating social interactions, digital platforms reshape society in various ways; this interdisciplinary study helps understand how the digitalisation of social influencers affects reciprocity and mimetic behaviours.

All data analysed in the scope of this study are included in this published article within the associated dataset.xlsx file. To comply with existing regulations on users’ privacy, the IDs of Instagram users present in S0 were anonymized, and the images associated with their publications in S1 were removed.

1. The engagement metrics associated with contests in Fig. 8 for classes 1 and 2 are not relevant because they refer to only one publication each.

Authors and Affiliations

1. Ecole Normale Supérieure, Department of Philosophy, Paris, France

   Hubert Etienne

2. Laboratory of Computer Sciences (LIP6), Sorbonne Université, Paris, France

   Hubert Etienne

3. Facebook AI Research, Paris, France

   Hubert Etienne & François Charton

Contributions

Both authors, HE and FC, contributed to the same extent.

Corresponding author

Correspondence to Hubert Etienne.

Ethics Approval

The research was approved by Meta’s internal review board and complies with existing regulations. It should also be noted that all the identified data (page, posts, age, gender, activity) presented is publicly accessible online, with one exception: the volumes of user reports. The authors decided to release them because they are not considered as “sensitive personal data”, have very little chances to be detrimental to the mentioned celebrities, and because hiding the names of the celebrities would have made it significantly harder for the reader to follow the authors’ reasoning, since these reports levels are analyzed in the light of celebrities’ real life events (e.g., the Astroworld scandal). Yet, the number of celebrities’ names has been limited to a minimum and the number of repots in question are only released two years later.

Consent for Publication

All relevant parties, HE and FC, provided consent for publication.

Consent to Participate

N/A.

Funding & Conflicts of Interest/Competing Interests

Both authors are Meta’s employees and acknowledge that this affiliation presents a conflict of interests considering the topic of this research. However, the research was conducted with the highest integrity standards and the authors engage their probity to attest that their work did not suffer from any influence.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Annexes

1.1 Annotation Methodology

1.1.1 The Classification of Influencers per Fame

The classification of influencers per fame (CIF) contains four verticals. Influencers’ origin and type of fame were determined based on their activity. For instance, it was agreed by convention that reality TV personalities had a more person-based fame than fashion models and TV presenters. In contrast, the degree of content personification was estimated, as a general trend, from the systematic review of influencers’ accounts (S0). Creators’ activities were collected via internet research, and when an influencer cumulated different activities along their career (e.g., a sportsman becomes a TV presenter), the activity selected is that which truly made them famous to the general public. Among class 3 influencers represented in S0, sportsmen include football players, American football players, basketball players, cricketers, Formula 1 runners, boxers, skateboarders, pole vaulters, bull riders, dancers and ice dancers. The type of fame is associated with the individual’s activity: it is activity-based when the individual’s performance can be dissociated from their personhood (e.g., one may enjoy Booba’s music without appreciating the singer as a person and consume their music without being interested in their private life) or person-based when such a dissociation tends to become impossible.

Unlike external influencers, organic influencers grew their fame as content creators on social media. Their activity is, therefore, closely related to the kind of content they produce on Instagram, YouTube or Tiktok, although they may have another professional activity irrelevant to their influencer status and often difficult to identify – especially for class 6, whose creators have an almost anonymous and mediatic presence often exclusively limited to Instagram. As a result, while the degree of content personification appearing in the CIF results from the observations presented in Section 2, relevant to external influencers, it is more linked to the degree of content personification for organic influencers per the design of the classification. Class 4 contains pure social media players (e.g., bloggers born on Instagram or Tiktok) who mainly showcase their bodies. Class 5 contains pure social media players (e.g., family blogs, travel blogs, pages of YouTubers or Tiktokers who publish different types of content on their primary social media), coaches (sports, food, etc.) and passion broadcasters (wine, hunting, cars, etc.). Class 6 contains artists (tattoo, interior design, clothes, jewellery, flowers, etc.) and passion broadcasters (fishing, memes, motorcycles, cars, dogs, etc.)

1.1.2 The Typology of Influencers’ Posts

The typology of influencers’ posts (TIP) has three labelling levels.

The first level describes the publication’s pictorial content. It labels who is present in the post: ‘Me’ (the influencer alone, or eventually with their baby or young children), ‘Me & others’ (the influencer surrounded by other people), ‘others’ (only other people) or ‘nobody’ (the content only contains an object, a landscape, a screenshot of another post, etc.). It also distinguishes between personal content (1, 2, 3), associated with the influencer’s private life, and professional content (4, 5, 6), associated with the influencer’s activity. Hence, the ‘other’ people tend to be friends and family members for posts labelled 2 or 3, but colleagues, clients and professional contacts for posts labelled 5 or 6. No distinction was made for label 7 between personal and professional life, and posts only presenting influencers’ children or pets were rated 2 as we consider them extensions of their parents/owners.

The second level captures the holistic and general emotions that emanate from the post by combining both images and captions. It intends to account for different ways of presenting a given message, thus informing about how an influencer wants to be perceived when communicating. Our 14 categories include H for showy (influencers strike a serious pose without smiling to look inaccessible), E for excitement (usually related to an announcement such as a pregnancy or the release of an album, a movie, an article, etc.), P for pride (related to achievements), S for self-derision (posts where influencers are not presented at their best, mocking themselves or expressing weaknesses), J for humour, L for love (expressing the influencer’s attachment of goodwill vis-à-vis their family, friends or colleagues), M for motivational (spiritual or motivational claims to provide advice, support and encourage followers), G for gratefulness (influencers expresses their gratitude to specific people or, in general, for what they have), O for altruism (celebrating another person for their work or friendship), C for charity (explicit supporting a charity cause, often encouraging donations), X for sadness (expressing compassion, sadness or bereavement), D for disgusted (positioned statements when influencers express anger or disgust regarding a situation), A for authenticity (influencers being open and genuine about their failure and life difficulties) and N for neutral. As several of the categories can apply to a unique post, TIP₂ can combine several letters (e.g. ‘AGL’).

The third level captures the type of communication influencers set up with their posts. Every post is viewed as a message from an influencer to their community – otherwise, one would wish their friends a happy birthday in private rather than through public posts – but the way this message is structured suggests different kinds of responses. The communication can be entirely closed on the influencers’ self: type 1 is for posts where influencers only discuss their lives as if they were talking to themselves. It can also be relatively close: type 2 refers to posts where influencers are talking to a specific person as they would do in a private message (‘I am so blessed to have you X, as a friend’) or discuss something largely unrelated to themselves (e.g. a picture of a landscape with no caption). The communication can also be unilateral, which is the case when influencers directly address their community without expecting a response: type 3R for recommendations (‘I suggest you check out this artist’ page,’ ‘I recommend this product,’ ‘go check out my last show’), 3W for wishes (‘I wish you a great day,’ ‘a Happy New Year’, ‘a Merry Christmas’) and 3T for thanks (‘thank you so much for your support,’ ‘Your love and comments encourage me to continue’). Lastly, the communication can be open, calling for followers’ engagement: type 4 for direct questions (‘What do you think about my new look?’, ‘tag a friend if you agree with this,’ ‘reach out by direct messages if you want to buy this piece of art’), type 5 for contests (‘I am partnering with brand X to offer you this trip to Italy. Like and comment the post with your name to participate and the winner will be drawn at random’) or type 6 (replies to followers’ comments, live Q&A, meeting with followers at events such as book signings).

The labelling was performed by the researchers with the utmost seriousness, including multiple rounds of reviews. Each content items present in S1 was reviewed by both of the two authors and ratings result from a consensus which did not present any challenge for labels 1 and 3. With regards to label 2, which is relatively more subjective, ratings from the two annotators were combined (e.g. “E,L” + “L,G” = “E,L,G”). The fact that the annotation task was performed by the authors themselves should not be considered a limitation of the methodology for at least two reasons. First, because while outsourcing the labelling task to external annotators is relevant to label voluminous databases with simplistic labels (e.g. is there a plane on this picture?), it becomes irrelevant to analyse social interactions with more sophisticated labels. Second, because the typology was designed in such a way that the task can be replicated. Labels 1 and 3 are based on the structure of the content – the components of the pictorial content for the former, the semantic structure of the latter – which drastically limits the weight of annotators’ subjective appreciation in the annotation task. Label 2 is arguably more subjective, which is why the researchers mainly relied on keywords – e.g. almost all G-rated items’ captions include words such as ‘thanks,’ ‘cheers,’ ‘grateful,’ ‘honoured,’ ‘blessed.’

1.2 Presentation of the Influencers’ Accounts Dataset (S0)

S0 is composed of a random selection of N = 619 public Instagram accounts balanced across four regional groups: France, the US, the UK and ‘world_stars.’ This latter group gathers all accounts from S0 whose audience exceeds 10 M followers, whether they are owned by French, American, or British influencers. Table 5 presents the resulting composition of S0 per influencer class and region. All data was extracted from Instagram internal servers on the 31st of October 2021 and is available under an anonymised format in the file dataset.xlsx.

To ensure our analysis was not narrowed to ‘mega-influencers,’ we also aimed to select a critical number of accounts from different audience ranges and filtered both private accounts (for privacy reasons and because influencers’ accounts are typically public) and non-personified accounts held by organisations (e.g. National Geographic) or brands (e.g. Dior). Table 6 presents the resulting composition of S0 per audience size and region and estimates S0’s coverage vis-à-vis the population relevant to our study – e.g., the 34 British accounts with a 10-50 M follower base in S0 account for 77.5% of all UK public and personified influencers accounts followed by 10 K to 50 K followers on October 31st, 2021.

As shown, S0 covers a significant part of all accounts followed by more than 10 M people in the three countries, and 36 out of 40 of the influencers followed by more than 50 M people while still including micro- and nano-influencers. Therefore, if one argues that the world_stars category does not include enough instances to allow for a rigorous analysis (with the exception of class 3), we shall answer that it contains almost all the mega-influencers, representing dozens of millions of followers (Table 7).

1.3 Ages and Genders

Users’ ages and genders are those provided by Instagram users and accessed from the platform’s internal databases. To confirm the robustness of our results on gender analysis, we observe that there was no significant change in the distributions of genders at the influencer and country levels in a two-month period (from 31/10/2021 to 01/04/2022). At the influencer level: Avg(delta) = 0.28%, Med(delta) = 0.17% and Max(delta) = 2.68%. At the region level: Delta(World) = 0.11%, Delta(US) = 0.86%, Delta(UK) = 0.04%, Delta(FR) = 0.03%. In addition, while all users do not provide their gender, resulting in gaps in gender data, two signals allow us to assume that the gender ratios remain stable between known and unknown genders. First, the proportion of known genders largely outnumbers that of unknown genders (70–78% across countries). Second, these ratios remain stable when distributed across all ages for each of the four regional categories. Finally, while Instagram allows users to select between ‘male,’ ‘female’ and ‘custom’ when choosing a gender, we neglect this latter category, which only represents 0.3% of S0’s audiences (Fig. 5, Tables 8, 9, and 10).

Distributions of Instagram users’ genders per age in the US, UK and France

Full size image

1.4 Presentation of the Influencers’ Posts Dataset (S1)

S1 is composed of 2,404 content pieces selected among all posts from 204 influencers of S0 between the 31st of October 2021 and the 4th of January 2022. A content piece is composed of pictorial content (one or several images and videos) and semantic content (textual caption). The sampling was randomly stratified, and we tried to have in S1 as many different influencers from S0 as possible, with a target number of 10 posts per influencer, while keeping a balanced aggregate between the six influencers classes of CIF. We also selected influencers who had the most local audience (except for world_stars). The annotation is based on the TIP, which aggregates three levels of labels. Leaving aside sponsored posts, we are left with 2,142 items balanced between regions (N_{world_stars} = 551, N_US = 560, N_UK = 518, N_FR = 513) and classes (N₁ = 378, N₂ = 307, N₃ = 407, N₄ = 345, N₅ = 389, N₆ = 316).

Table 11 presents the distribution of non-promotional content per influencer class and content type for the three labels of the TIP (N), together with their weighted average numbers of likes (likes) and comments (com.), adjusted by influencers’ average audience size over the period (/K followers).

1.5 User Reports against Travis Scott and DaBaby (Fig. 6 and 7)

Travis Scott and DaBaby’s daily following evolution

Full size image

Travis Scott’s distributions of reporters and followers per age and gender

Full size image

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions