Advancing netnography – hybdrid ways of coding social media content Open Access

Since the advent of online communication 3 decades ago, research interest in social and business phenomena in this realm has kept growing. One major reason being that traditional phenomena such as consumer behaviors, communities or culture have emerged in the web realm (Tavakoli and Wijesinghe, 2019). The other reason is that a wealth of data is being generated on various digital platforms or through digital interactions (Kozinets and Gretzel, 2024), opening ways to generate insights both on established phenomena, such as interactions among supply chain members (Rynarzewska and Giunipero, 2024), and emerging new phenomena such as digitally immersive experiences in metaverse (Kozinets, 2023). The rapidly growing interest in online phenomena fosters the development of novel methodologies suitable for exploring and explaining phenomena and concepts typical for the online context and data that depart from traditional survey research or quantitative modelling (Rynarzewska and Giunipero, 2024). In this vein netnography, i.e. “qualitative research approach for gaining cultural understanding that involves the systematic, immersive, and multimodal use of observations, digital traces, and/or elicitations” (Kozinets and Gretzel, 2024), is gaining prominence with more than 320 studies identified in the period 1997 to 2017 (Heinonen and Medberg, 2018) and more than 720 in the period 2001 to 2023 (Bansal et al., 2024).

Netnography was originally designed by Kozinets (1997) to study consumer behavior, but since then has spread across multiple disciplines (Bartl et al., 2016). Reviews of this literature indicate that thematic analysis, coding and discourse analysis are the most popular analytical techniques (Tavakoli and Wijesinghe, 2019). However, despite the growth in popularity of this method and the vast opportunities associated with using it to understand online phenomena, there is a surprising lack of methodological guidance on how to collect and analyze data collected online. And yet challenges include the vast amount of data collected on social media, decontextualization, filtering by the platforms on which the digital traces are left by users, ethics and technological embeddedness of online interactions (Kozinets and Gretzel, 2023; Asname and Berrada, 2025). As a result, data analysis is equally challenging as data collection, and likely to depart from the traditional guidelines with some scholars advocating a prudent use of AI tools (Kozinets and Gretzel, 2024) while others use conventional coding (Madden et al., 2024).

Our study addresses the gap relative to guidelines, recommendations and practices useful in coding social media data. Coding is central to qualitative data analysis as it aims at exposing the “meaning, ideas and thought” (Khandkar, 2009), opening ways for structuring information, generating interpretations and formulating insights. The aim of this study is to develop a framework and guidelines useful for researchers when coding data collected on social media. We first contextualize social media data for business research. Next, we outline conventional coding approaches in order to provide a contrasting framework for social media data coding. Third, we develop our framework and illustrate how it helps address two key concerns, i.e. ambiguity of social media content reflected in low inter-coder agreement rates, and complexity reflected in multiple codes applying to the same piece of text. We conclude with avenues for non-textual data analysis and a further research agenda.

The rise of social media has profoundly influenced the organizations management, transforming communication strategies, stakeholder relationships and identity construction (Etter et al., 2019). As organizations follow digital transformation, social media emerges as a critical instrument in shaping corporate strategies and fostering stakeholder engagement (Jedynak et al., 2021). Social media has facilitated new forms of communication, resource sharing and collaboration, reshaping managerial practices and organizational structures (Weber et al., 2016). The shift from traditional, controlled messaging to interactive, decentralized communication has necessitated an evaluation of organizational identity (Fdhila et al.., 2021).

The integration of social media into organizational processes has become imperative, as it is linked with potential competitive advantage (Song et al., 2022), which makes this area of practice and research truly strategic. For instance, traditional top-down approaches to management are being supplemented by more agile, interactive models that prioritize digital engagement and responsiveness (Kasperiuniene and Zydziunaite, 2019). Identity formation is no longer a static process but rather an ongoing, dynamic exchange shaped by digital transformation. Social platforms serve as key analytical tools, providing real-time insights into user's behavior, market trends, and competitor strategies (Theodorakopoulos and Theodoropoulou, 2024). Organizations leverage social media for innovation and creation, enabling stakeholders to contribute to brand identity. This shift underscores the need for organizations to develop adaptive strategies that align with the evolving digital landscape (Essamri et al., 2019).

Despite its advantages, the reliance on social media poses substantial challenges (Czakon et al., 2024). The decentralized nature of third-order communication increases the risk of reputational crises, misinformation and stakeholder conflicts (Kietzmann et al., 2011). The growing influence of users in shaping corporate narratives can lead to organizational identity chaos, where brand perception is influenced by multiple, often unpredictable, external actors (Marsen, 2020). As a result, organizations must implement proactive communication strategies and reputation management frameworks.

Social media interactions are often perceived as voluntary (Meikle, 2016). The extent to which users engage with organizational content is also influenced by algorithmic mechanisms and promotional strategies (Cotter, 2019). While individuals may choose to comment or share content based on personal interest, their exposure to corporate messaging is frequently shaped by paid promotions and targeted advertising, which expand the reach of specific posts. This dynamic blurs the boundary between organic and incentivized engagement, raising questions about the authenticity of stakeholder interactions and the extent to which organizations can strategically shape their perceived identity through social media discourse (Oestreicher-Singer and Zalmanson, 2013).

While organizations maintain control over the initial stage of communication, such as crafting the content of their original posts, the subsequent discourse is largely shaped by audience interactions and external engagement (Liu et al., 2020). Once published, social media content becomes subject to reinterpretation, discussion and even distortion through user-generated comments, shares and reactions. This decentralized nature of communication means that organizations cannot fully control how their message is received, reshaped, or repurposed by others (Vergne, 2020). Moreover, individuals engaging with such content may do so with varying intentions, ranging from genuine interest and brand advocacy to criticism, or misinformation. This unpredictability underscores the challenges of managing organizational identity in the digital sphere, where stakeholder participation plays a crucial role in shaping public perception (Kietzmann et al., 2011).

All in all, tracing interactions on social media is likely to demand a distinctive set of techniques from those established in qualitative methods. The purpose of this study is to outline a hybrid way of coding social media data.

Coding as a method of data analysis in qualitative research has become a key technique for systematically organizing and interpreting qualitative data (Glinka and Czakon, 2021). One important aspect of coding is its role in increasing transparency and rigor in qualitative research, particularly in generating new concepts and theories (Gioia et al., 2013). Coding allows for the organization of the collected data and reduces its complexity, enabling further analysis (Miyaoka et al., 2023).

Three basic coding procedures are commonly used: (1) deductive coding, (2) inductive coding and (3) hybrid coding (Gibbs, 2015). Each of these methods serves different purposes and is characterized by a different approach to data analysis.

Deductive coding is a top-down approach in which researchers begin with predefined codes based on existing theories or frameworks (Fereday and Muir-Cochrane, 2006). This method involves applying these codes to data to test specific hypotheses or categorize information according to established concepts (Azungah, 2018).

Inductive coding is a bottom-up approach in which researchers generate codes directly from the data (Chandra and Shang, 2019). This exploratory method allows for new themes and concepts (Gioia et al., 2013) that could not have been anticipated before data collection. Inductive coding is beneficial in qualitative research because it captures the richness and complexity of participants' perspectives, enabling researchers to understand the phenomena under study better (Rivas, 2012).

Hybrid coding combines deductive and inductive approaches, allowing researchers to leverage the strengths of each method (Xu and Zammit, 2020). This approach begins with predefined codes based on existing literature or theoretical frameworks while remaining open to the emergence of new codes from the data (Fereday and Muir-Cochrane, 2006). This approach is beneficial in qualitative research because it enables comprehensive data analysis while maintaining the flexibility of the coding process. Using hybrid coding can also improve the reliability and validity of qualitative research findings. Integrating predefined codes with codes generated from data causes the analysis to be based on both theoretical frameworks and researchers' experiences (Proudfoot, 2023). When there are several researchers, there is a need to standardize the codes or re-code them, which leads to results that are more grounded in the data and increases credibility through triangulation of researchers (Baumgartner and Schneider, 2010). This approach is an iterative process in which researchers repeatedly return to the data and modify their codes to interpret the results more accurately (Braun and Clarke, 2006). Hybrid coding can facilitate the use of qualitative data analysis software, which can streamline the coding process and increase the efficiency of data management (Jiang et al., 2021; Nelson et al., 2021).

Different from conventional coding of interview-based data, the procedure for coding data collected from social media has not so far been standardized. Current practices depend on factors, such as the goals of the analysis (Blakeman et al., 2025), the specificity of the platform (Celik et al., 2022; Licen and Cermelj, 2024) or the availability of resources (Batrinca and Treleaven, 2015). Another challenge is the vast amount of user-generated content, which makes researchers look for opportunities to automate the coding process (Pulido Rodriguez et al., 2021). Batrinca and Treleaven (2015) emphasize that the ease of access to public data from sites such as Twitter and Facebook has accelerated the development of a wide range of analytical tools to process this information. It is important to consider that social media data is oftentimes unstructured, noisy and dynamic, which poses unique challenges for data mining efforts (Ahmadi et al., 2022). Furthermore, the rapid evolution of social media platforms often requires users to quickly adapt to new features, which can impose cognitive burdens that impact content creation (Ahmadi et al., 2022).

The process of coding content in social media is also challenging by user language dynamics, which also involves graphic content or symbols (Sutrisno and Ariesta, 2019). The majority of social media posts done by firms use links, i.e. active URLs that allow users to be redirected to a selected page or website with a single click, without having to manually enter its details (Chen et al., 2020). Some of those links redirect to a particular company's website, others may encourage users to sign up for in-person events, webinars, training courses, conferences, job fairs, etc. Moreover, some links provide access to additional educational content, such as articles, podcasts or recordings posted on YouTube channels. Furthermore, the posted links were practically always accompanied by direct, short and directive phrases such as “Join us!,” “Sign up today!” or “Don't miss this event!” These are designed to pique the interest of observers and stimulate them to take the action expected by the company (Molina et al., 2020).

A distinct category of elements oftentimes included in posts are hashtags, i.e. phrases prefixed by the # symbol, which are employed in online communication to label content (e.g. posts, photos, videos) so that it may be easily searched for (Habibi et al., 2021). They are generally placed at the conclusion of a post, although they are occasionally visible elsewhere within the post. Thoughtful use of hashtags directly influences users' attention and generates more interaction with the content (Cheng et al., 2020; Salomon, 2013). In relation to hashtags dual functionality may act both as an index for content retrieval and a tool for engagement, enhancing their efficacy as marketing instruments (Kumar et al., 2022). A detailed analysis of hashtags may reveal diverse usage patterns (Burikova and Ovchinnikova, 2021; Habibi et al., 2021). For instance, hashtags can be used to directly promote the name of a company, a series, a slogan or a project created by the company (Lin et al., 2023). Also, hashtags can refer to highly specialized industry jargon. Thirdly, hashtags can refer to common topics and words. Finally, mentions can refer to specific people or organizations (usually posted as a kind of variation on a hashtag – the so-called Facebook tag, e.g. @JanJohn Smith) (Cheng et al., 2020; Salomon, 2013).

Finally, emoticons and emojis are typically used in social media content. These are as small images employed in online communication to convey emotions through facial expressions, as well as to enhance text statements with images of places, objects, animals, plants, beverages, sports, international flags and many more. Their use poses a significant challenge in the digital content coding due to the variety of functions, which them particularly challenging to interpret unambiguously (Brito et al., 2020). Both their selection and subsequent interpretation by the audience are highly subjective and intuitive (Wall et al., 2016). On the other hand, the appropriate selection of emoticons and emoji often makes it possible to claim professional creation and management of one's online content (Khaksar et al., 2023). All in all, social media content poses distinctive challenges when it comes to coding.

The aim of this article is to propose a coding procedure suitable for social media content in netnography. In this article, we focus on the textual content of posts, leaving other elements beyond the scope of our analysis. Notwithstanding, the changing narrative style of social media users requires the creation of complex coding procedures, even if this process is automated capable of processing non-standard language (Adawiah et al., 2023). Since firms usually strive to maintain consistency in social media, this means adopting a strategy of indirectly influencing the created content to be consistent with corporate goals (Grafström and Falkman, 2017; Huotari et al., 2015). Moderation of discussions in social media can make it challenging to identify the actual reactions of recipients and, thus, the credibility of the coded content (Elsayed, 2017).

Our proposed coding guideline is illustrated on a study of social media digital organizational identity construction (Figure 1). We take this illustration to ground our structured recommendations in a real-life study and thus exemplify challenges in a particular context. We particularly emphasize the challenges relative to language ambiguity and complexity, which are consequential for the number of iterations necessary to develop a useful codebook. At the same time, the number ambiguity and complexity impact the inter-rater agreement ratios that can be expected in this process.

The process presented in the diagram provides a detailed and systematic model for coding qualitative data in content research, based on an iterative approach to developing and testing codes. This research contribution can be used by other researchers because: (1) the figure demonstrates the sequential stages of developing a coding scheme, from coding single words to coding full posts. Other researchers can adapt this structure as a framework for analyzing content from social media, organizational documents, or other forms of text; (2) iterations include consistency checks, which emphasize the importance of inter-coder reliability; researchers can adopt similar thresholds or tests (e.g. % agreement) to ensure the reliability and repeatability of results; (3) the process involves moving from preliminary codes to a more complex structure. This can serve as an example of a hybrid coding approach (both deductive and inductive), useful for exploratory research. (4) The final iteration includes validation of the entire coding process, which can be adopted as best practice in reporting qualitative research findings.

A content analysis of Facebook posts created by the group of firms selected for the study of their digital organizational identity reveals several key elements that collectively constitute their content. These elements embrace the textual content of the post itself, in addition to links, emoticons, emojis, hashtags, and tagging (Wang, 2021). Textual content of a post, encompassing its informative, encouraging or cautionary nature, is focal in this study.

We followed grounded theory because there are currently no guidelines in the literature on how to study/code social media material, and according to the principles of this theory, data should be collected systematically (data was collected over a 3-month period), analyzed and iterated (we conducted 8 iterations), samples should be taken theoretically, i.e. in order to understand the problem, not to obtain a representative sample. The study was based on 768 posts collected over a three-month period from 25 of the largest law firms, which were selected as examples of an industry where social media communication is particularly challenging due to the formality and specific language used. Above all, no preliminary assumptions regarding the theory should be made, because it should emerge only from the collected empirical material (Hensel and Glinka, 2018).

A primary challenge in analyzing post content stems from the use of foreign languages or the combination of multiple languages within a single post (Wolny, 2017). The ambiguity of the posts appeared as relevant, as the majority of the collected social media content addressed multiple topics and themes simultaneously within a single post, whilst mixing different temporal dimensions (Kreutzer and Rueede, 2019). For instance, some firms invited their observers to upcoming events by advertising past events, whilst formulating these messages in the present tense.

In accordance with Charmaz (2013), we undertook the coding process to categorize the data using short titles (codes) that were to summarize our data and show how we sorted them, in order to explain them analytically in the next step. Thanks to this, we did not have to refer to the content of all the collected data in the subsequent stages of our analysis and interpretation.

The coding unit can be word by word, line by line or event by event (Charmaz, 2013). We first contemplated focusing on individual words and their immediate contexts was also contemplated. This may allow rankings of the most frequently occurring words in a given firm's posts. In the first iteration, we established the coding process at the level of individual words, words were selected according to their frequency of occurrence, and the word limit was set at 250 for each company. However, adopting single words as pieces of information appeared as not viable due to a number of difficulties in the analysis, including different word endings affecting the ranking, occurring in different personal forms and/or depending on the tense form adopted in the sentence, as well as due to the semantic ambiguity of the words studied and the impossibility of taking into account the overall context of the message.

As a result, it turned out that, unlike in conventional coding, the data collected on social media should be treated as one post = multiple pieces of information. This is because the post describes only one event, but more than one code can be assigned to one post, and the codes can overlap. This conclusion was drawn in connection with the course of the process described below. For the coding process, four main codes were adopted in accordance with the dimensions described in Czakon et al. (2024): Awareness, Control, Orientation and Values. For each of the four dimensions, definitions were developed. It was assumed that during the analysis, we were looking for potential new codes that could be added. We began coding according to the scheme: (1) independent coding, (2) discussion of results in established pairs and recoding; coding at this stage was done in previously prepared Excel files.

During this process, it became clear that in the case of social media material, single words did not convey information, so we abandoned the concept of coding single words in favor of coding full sentences. At this point, we also decided to use computer-assisted qualitative research software MAXQDA Analytics Pro (24.6.0) for coding. Thanks to this tool, qualitative data is transformed into quantitative data by quantification (Kuckartz, 2017), among others, in order to enable the assessment of the strength of the phenomenon (O'Connell and Skevington, 2005).

At the beginning of this iteration, we also coded according to the dimensions described in Czakon et al. (2024): Awareness, Control, Orientation and Values, but while working on the collected material, it turned out that this approach was not sufficient. So we used a hybrid method to search for codes (Glinka and Czakon, 2021) - in addition to preconceptualized codes, we adopted an open approach to coding so that the list of codes could be expanded as the content of posts was analyzed. An additional justification for the hybrid approach was the fact that social media data sets are huge and it is better to use mixed methods to search for codes, so that we do not lose sight of the widest possible spectrum of issues. This approach was also used due to the fact that several researchers participated in the work, and it is difficult to imagine their “disciplining” based on one theoretical concept–each of the coders therefore had the right to propose new codes, but could not enter them into the code tree independently without consultation. As part of the open coding, a list of codes was created, which was unified, and the material was re-coded in order to obtain more established data. Ultimately, five research categories were selected as main codes: Awareness, Stakeholder Orientation, Time Orientation, Topic and Values (and therefore a modified form in relation to the initially assumed one).

In the third iteration, this method was also not sufficient – the sentences did not reflect the contextual meaning of the posts. It turned out that in order to find valuable statements that allow for understanding the phenomenon under study, the coding unit should be entire posts, regardless of their length. During the research, we stuck to the 5 main codes that were created earlier: Awareness, Stakeholder Orientation, Time Orientation, Topic and Values, while within the open coding, we identified 19 subcodes, which were assigned to individual main codes (see Table 1).

In order to ensure the rigor of the coding procedure, we proceeded in three steps: first, individual code development, then cross-validation of codes by two pairs of researchers, followed by a final validation by the whole research team. By discussing the similarities and differences in the codes and the approach to coding itself, we achieved greater reliability and consistency of coding (Baumgartner and Schneider, 2010). The format of the codes we adopted is to define them in the form of themes (concepts) (Charmaz, 2013), which allowed them to be related to the entire group of activities and described situations that were visible in the posts. During four subsequent iterations, not only was a list of codes created, but also a code book, in which each of the codes was provided with a reference definition and examples.

Despite efforts to match the codes to the data as best as possible, each of the researchers (coders) was faced with a typical challenge of interpreting the posts. Additionally, the large volume of data and the interpretation problem described above forced frequent meetings, discussions and, consequently, more iterations. This was reflected in the coding agreement results. The initial coding similarity coefficient was as low as 55%. Therefore, we turned to examining the differences between each document, codes, and the way they were interpreted to avoid obtaining further incorrect results and began recoding. We tested the coding agreement based on the similarity coefficient created in accordance with the Kuckartz and Rädiker (2019) zeta analysis method. Only after obtaining a satisfactory level of agreement – 78.85% on a pilot sample of 5 firms from the database, did the coders begin coding data from the remaining dataset. The final similarity coefficient for all coded posts was 75.58%. It is difficult to define what should be considered a low or high percentage of agreement, because the percentage of agreement between coders depends not only on the number of coded segments, but also on other factors, such as the number and variability of different subcategories, as well as the difficulty of the coding process itself (Kuckartz and Rädiker, 2019).

Our study aimed at developing a framework and guidelines useful for researchers when coding data collected on social media. While the general guidelines relative to conventional coding practices appear as a useful starting point, we recognize multiple challenges relative to the nature of social media content. They include a vast amount of specifically structured, complex communication between organizations and social media users. This results in increased challenges when researchers are confronted with coding such data. We believe that while automated coding may be effective once a useful code book is developed, human coding is difficult to replace at this initial stage of code book development. Our findings offer substantial contributions to digital content qualitative analysis methodology by focusing on a key challenge to rigorous netnography, that is, coding.

We recognize that the challenges typical of social media data sets impact the coding process in five distinct ways. Firstly, posts are likely to be pieces of information. Not single words or sentences, but whole posts. Secondly, individual posts are likely to be coded with multiple codes at the same time. This is different from conventional coding, aimed at unveiling a structure of meanings by isolating distinct pieces of information. Social media posts typically convey multiple information, and are therefore likely to match multiple codes. Thirdly, the process of developing codes in such complex datasets requires multiple iterations. Our exemplification involves up to eight iterations before developing an acceptable code book. Fourthly, inter-rater agreement ratios are likely to be lower than conventional data coding. We pave ways for a critical debate on acceptable levels by indicating that a rate of 75% required multiple rounds of discussion and vast amounts of time. Fifthly, we believe that involving multiple raters, more than two, adds to the rigor of the code book development process. Once agreed upon, the code book may serve to automate coding of the remaining dataset.

We are aware of some limitations of our study. Qualitative research faces distinct challenges that require consensus around acceptable practices. We believe that our study may spark debates necessary to advance towards such a consensus. One limitation of our recommendations is connected with the data set, which focuses on interactions between firms and their stakeholders. Interactions among firms, or among individuals, may pose different challenges. Additionally, we use an exemplary dataset, which while recent and manageable in terms of size, might not be directly transferable to much larger datasets. Hence, we encourage further debates on how to select appropriate samples from large datasets, useful in developing codebooks. Thirdly, we are aware that a critical review of practices across multiple studies might add to a literature-driven methodology framework development. As available reviews indicate, the number of netnography studies using coding is relatively low to date, and authors seldom reveal details of their coding procedures. Therefore, we encourage authors to report in detail how they developed their codebooks in netnographic studies, and reviewers to explicitly ask for increased transparency in this respect. Finally, we left beyond the scope of our framework nontextual data. We are aware that they interact with textual content to some extent, and that they may be content on their own as well. Frameworks that may include those elements of social media content may be more comprehensive and thus offer more complex findings.