To verify the specific role, aim of Instagram in tourism discourse and marketing, the nature of imagery in three popular tourist boards' Instagram and website pages was analyzed with social semiotics and empirical methods. This was done by looking at visual features, or materialities (Bateman, 2018), their frequency and clustering across digital channels and by testing the validity of the main framework available in multimodality research for image analysis, i.e., the linguistics-based Grammar of Visual Design (Kress and van Leeuwen, 2006). Such qualitative framework, indeed, is grounded in a general and presumed symbolic, socially attached value of images that may lack empirical testing and potentially objectivity (Bateman, 2019a, p. 533–535; Bateman, 2019b, p. 90–92). By studying larger amounts of data with bottom-up approaches, conversely, theories on the presence of tourist driving forces, or desires—such as the quest for the uncontaminated, unknown, remote—may be validated (Mattei, 2023c).

This procedure was guided by two research questions:

Through the development of a tagging system and the conduction of bottom-up statistical analysis for the detection of data clustering and dimensions of variation, it becomes visible how Instagram imagery varies from website narratives, and seems to be systematically constructed to capture the attention and foster both a perceptual pre-consumption of the travel experience and emotionally charged reactions by representing visually such conditions. Indeed, Instagram mainly features aerial views of pristine, aesthetically pleasant settings, often complemented with rear views of solitary individuals performing static processes of peaceful contemplation of natural wonders. IG photographs are shown specifically to be taken from either long shots and high angles in natural settings, often without human footprint, or in close contact with wild animals. Thus, IG is shown to play a particular role in the marketing funnel of persuasion of tourists, slowly becoming “a historically stabilised site for the deployment and distribution of some selection of semiotic modes for the achievement of varied communicative purposes” that may depend on the channel's audience (Bateman et al., 2017, p. 123). The findings were obtained by analyzing the presence of particular participants, their actions and gaze in specific settings, as well as the shots, angles, and salience or positioning in the picture.

Website photography, on the other hand, is demonstrated to feature mostly active and human participants photographed in collective and social moments, also from close distance, and involved in a variety of gastronomic and cultural events in artificial settings like cities. The importance of providing information and agency is confirmed by the predominance of humans involved in activities or in the fruition of services.

For the project, three popular national tourist boards located in English-speaking countries were selected, especially for their daily sharing of high-quality images combined with long captions. Data consisting of photographs and corresponding texts were gathered from two main communication platforms:

The corpora were collected in 2019 by accessing Instagram's application programming interface (API) to retrieve posts for each tourist board in a timespan of 6 months of the same year, and by scraping websites once permission was granted. Table 1 presents a breakdown of the visual sub-corpora, categorized by tourist board and their respective channels.

The statistical analysis of visual strategies in tourism photography was supported by a tree tagging system based on the Grammar of Visual Design but adapted to tourism photography (Mai et al., 2011). The model enabled the quantification of frequencies and correlations between objective features, like the type of settings, entities, actions or reactions, shots, gazes, photography techniques that are of interest to multimodalists (Kress and van Leeuwen, 2006). This empirical process enabled to cautiously move from theory to evidence-based, sociological interpretations of the data.

This model thus tested an adapted version of a general model of visual knowledge by categorizing and measuring occurrences, variations of data-informed visual features after close data inspection and (inter)annotation. The tags were then grouped by metafunction (Halliday and Matthiessen, 2014).

Specifically, the model examines the decisions made by discourse specialists regarding the selection of participants (subjects), the presence and type of processes (activities), and settings (macro-category Representation of Reality); the selection of camera shots and angles (macro-category Relationship with the Audience), and both the position, size and visual weight of elements within the photograph (macro-category Composition).

For preliminary analyses, the macro-category Tourism Strategies was included to quantify the presence of tourists' driving forces or recurring themes (Maci, 2020) based on the identification of particular features in each image³. For example, the search for what is unknown, remote or pristine was encoded as strangerhood (trope) when detecting uncontaminated environments, like exotic beaches or clear waters. Or again, the representation of meals, swimwear or alcohol was encoded as indexical reference, an object symbolizing postmodern desires and taboos embodied by represented participants and transferred to tourists (Dann, 1996).

The macro-categories linked to the three linguistic metafunctions are summarized in Table 2 and encompass a range of sub-categories of tagging possibilities (or macro-variables) that group “choices” in meaning construction (main variables).

In the tagging system, the macro-category Representation of Reality relates to the experiential metafunction and classifies both narrative and conceptual images and related circumstances (Kress and van Leeuwen, 2006, p. 48–62)⁴, i.e., settings, here identified and coded by the author as natural, artificial, historical, cultural, analytical and gastronomic. In particular, narrative images are detected by coding action and reaction (observation) processes⁵, together with the subject and nature of the process, i.e., whether it is transactional or non-transactional and therefore includes or not the goal of the action or the object of contemplation. Conceptual images, conversely, are articulated into analytical (maps, icons, logos) or symbolic suggestive, coded implicitly through the absence of human beings.

The macro-category Relationship with the Audience explores choices concerning the relationships between the producer of the semiotic artifact, the represented participants, and the interpreter who makes sense of the sign in the social context, i.e., the viewer. The producer communicates with the viewer through three main tools, which correspond to the categories included: (a) the gaze, if present, directed toward the viewer or a represented participant; a close, medium, or long, very long shot, which progressively provide a bigger picture with a less focus on details, as close shots, for example, capture only the portions of an entity; and an angle, which may be objective (top-down or frontal, neutralizing any distortions) or subjective. The adoption of a subjective perspective includes choices regarding vertical (high, eye-level, low) and horizontal angles, the latter being oblique when the frontal plane of the photographer does not run parallel to the one where the main represented participants are.

The hierarchical tagging system was created using Statistically Reliable Image Tagging⁶, a software designed for complex manual tagging procedures on large visual corpora⁷. SRI Tagging was developed to provide a user-friendly interface for categorizing lexicogrammatical units in multimodal studies⁸.

To ensure the objectivity and consistency of the tagging process and rule out the possibility of chance agreement, inter-coder reliability measures such as Cohen's kappa and Krippendorff's alpha were implemented. These were further elaborated in R to consider tag dependencies and both mutually and non-exclusive child nodes (Bateman et al., 2017, p. 198–204). In this project, 18% of the images were independently tagged by another coder, instructed on tag meanings through a reading scheme. Most variables in this study report substantial reliability values (>0.61), ensuring the robustness of the tagging procedure⁹.

This section presents the results derived from the manual tagging of the visual corpora through Principal Component Analysis (PCA) in R. In particular, this section offers an overview of data patterns and clustering, including the contribution of each variable to statistical variation. PCA reduces the complexity of the data under investigation and explores “to what extent the annotations allow the data to be grouped into clusters” (Bateman and Hiippala, 2021, p. 7) by means of fictitious dimensions of variance. Due to its bottom-up nature, this test enables the generation of data-driven hypotheses of variance, correlation by mapping data onto a multidimensional space. Eventually, this analysis helped understand whether tags (i.e., variables) grouped according to channel, preparing the ground for inferential statistical analyses such as One-Way ANOVA in Jamovi, chi-square and Correspondence Analysis (CA) in R within and across macro-categories¹⁰. These tests helped identify cluster patterning and correlations, i.e., variations in the use of specific tags across different channels (Field et al., 2012). This was possible after grouping of each image's tag occurrences of each variable according to channel. To conduct ANOVA tests, both Shapiro-Wilk test of normality of the distribution of the data and Levene test were performed, reporting non-significant values and confirming the assumption of normality and homogeneity of variance ¹¹.

The input data used for the calculation of variation across dimensions was a table with a list of all images and the presence or absence of each feature (tag) for each image. All metadata pertaining to the type of sub-corpus was excluded and mapped afterwards. The first round of exploratory PCA includes all variables, without any filtering based on reliability values¹². This was done to provide an overview of potential clustering patterns, and due to space constraints, considering the complex procedure of implementation of reliability tests on hierarchical systems. Analyses excluding non-reliable and redundant features have been extensively reported in Mattei (2023b) and partially in Mattei (2023a) and Mattei (forthcoming) and describe the implementation of Cohen's kappa and Krippendorff's alpha. To facilitate comparison with more reliable results, Supplementary material report a PCA (alpha > 0.8) and CA (Kappa and alpha > 0.6) of reliable features only.

An overview of the percentage of explained variance is provided in Figure 1, in the graph on the left side. As can be seen from the scree plot, the first ten dimensions account for 45.9% of the variance detected by the Principal Component Analysis. Each one of the remaining 191 dimensions detected accounts for too little variation of the data; this phenomenon might be investigated further by annotating a larger body of images and providing more annotations for each one of the (sub)categories of the tagging system, or by reducing the number of variables (tags) in the taxonomy. For reasons of space, the current section focuses only on the exploration of the first two dimensions of variance, which contribute the most to the explained variance.

The graph on the right side shows the mapping of each tagged image onto a multidimensional space. The PCA allowed for a preliminary observation of potential regularities across the dataset with respect to their use of visual elements according to channels. In particular, each image was plotted according to the presence or absence of the features (variables) of the tagging system, and images displaying a similar pattern of occurrences and co-occurrences of features clustered together. The specific weight, or contribution, of each variable to the first two dimensions of variance calculated allowed for a particular placement of each image by taking into account its sum of features and related scores.

As can be noticed from the graph, two distinct clusters may be distinguished, occurring in opposing quadrants along the first and second dimension. The presence of two different clusters shows how selected classes of data items group together. Clusters are represented graphically as ellipses referring to the mean of a set of values with a 95% confidence level (Bateman et al., 2021, p. 16). In particular, the spread of the ellipses along the dimensions indicates the deviation of the data samples from the mean values. In the current graph, the narrow ellipses suggest that there is not considerable variation within the samples, i.e., website and Instagram images.

This finding is significant if accompanied by the interpretation of the position of the two distinguished clusters along dimensions 1 and 2. Indeed, website images are grouped together in the top right quadrant, showing a positive contribution to both dimensions. Conversely, Instagram images cluster in the bottom left quadrant, which indicates a negative contribution to the same dimensions. This distinction suggests that digital, visual communication in tourism discourse is constructed differently depending on the channel of dissemination of the promotional message.

In terms of variation across tourism boards, the PCA showed a less distinct and significant difference between the communicative, visual strategies; indeed, Tourism Ireland¹³,¹⁴ and Tourism Western Australia's¹⁵,¹⁶ ellipses overlap in the same quadrant, whereas only Destination Canada's¹⁷,¹⁸ images cluster in the opposite quadrant, but within closer distance, compared to the graph reported (see Supplementary material).

The characterization of the first dimensions of variance, which account for the largest contribution to variation in the corpora under investigation, allowed to discuss the potential nature of variance in the corpus. Particularly, the calculation of the weightings of each variable sharpens an understanding of the variables behind variance in the data items, i.e., the images.

Table 3 shows the values of the variables which contribute the most to the largest dimensions of variance that were calculated by the PCA. The features with values > 0.1 contribute positively to a placement along the corresponding dimension; conversely, features with weightings < −0.1 contribute negatively to the dimension, indicating the presence of an inverse correlation between the dimension and the variables. For ease of reading, results report only variables which proved to be reliable in subsequent analyses.

As can be seen from the table, the variance in the first dimension is mostly due to positive contributions, whereas the second is mainly characterized by negative ones. As website images cluster along the top-right quadrant, showing positive values of significant variation in both dimensions, their nature may be said to be defined by dimension 1. The latter includes the variables Direct Presence of Living Being, which encompasses animal and human life, Process of Action performed by Humans (A), the sub-category of gaze Toward the Represented Participant(s) combined with Medium (shots), Eye-level (angle), and Represented Living Participant (a sub-category of Visual Weight).

Conversely, Instagram imagery, showing negative values of variation in the graph, seems to be defined by dimension 2, and in particular by the variables Transactional Reaction, Oblique (Angle) and Vertical Angle, Converging Lines, Landscape Element (a sub-category of Visual Weight), and Direct Presence of Living Being. These are combined with Process of Reaction and Humans (R), Long (Shot), Rule of Thirds and correlated Water, Mountains, Land, which encode large and definite portions of pictures representing landscapes. If we consider the variables' presence, weight, and positive or negative value, we may formulate the following hypotheses, confirmed by the chi-square and CA:

On the one hand, results show website communication focuses on narrative representations that visually shape a multifaceted tourist experience, in which humans are the main characters, occupying a good portion of the photograph, and are involved in various activities. These include sports, collective gastronomic experiences, and tours with guides or in more exclusive conditions (e.g., by car, on a cruise). Also, the presence of gazes and eye-level angles shapes a close and equal relationship with viewers (Messaris, 1997), possibly mirroring instances of everyday life or mass tourism (Urry and Larsen, 2011). The representation of activities in various settings, including content like maps, signals an attempt to show what can be practically done and where when visiting a new place. This is because websites are visited mostly by individuals who have developed informative or purchase intents related to a brand or service, increasingly influenced by social media's awareness- and community-building activities through content sharing and interaction with a broad audience (Leung et al., 2013; Jamil et al., 2022)¹⁹.

Conversely, Instagram communication is defined by conceptual and static narrative representations in which a distant, commodifying view of the tourist destination is constructed, often combined with professional compositional techniques like converging lines and Rule of Thirds that make the picture attractive to the human eye (Mai et al., 2011). This is line with computational studies showing how posting aesthetically appealing photography is a shared practice in the IG community (Manovich, 2017). Specifically, the combination of long shots of vast, natural views and rear views of reaction processes with vertical angles establishes an impersonal relationship with dehumanized subjects who, when present, rarely engage with viewers through gazes; rather, they perform solitary, static processes of perceiving natural wonders, as also reported by qualitative and marketing studies (Garrod, 2009; Smith, 2021). Through such semiotic acts, represented tourists become themselves commodities, symbolizing consumption and fostering identification (Debord, 1977; Francesconi, 2014). The use of vertical angles—often high, as reported by absolute counts and CA—might signal the intention to build an unequal relationship between represented participants and viewers, as the latter view the destination from a superior position denoting control, ownership and power (Kress and van Leeuwen, 2006). Compared to website imagery, thus, Instagram imagery prompts imagination through scapes and a sense of pristine, remote experience, defined extra-ordinary by tourism sociologists (Urry and Larsen, 2011), i.e., different from everyday routines and mass-targeted, collective experiences in industrialized, polluted areas (Dann, 1996).