https://orcid.org/0000-0002-8199-585X
Figures
Abstract
Political influence on social media is not confined to explicitly political spaces. We examine how UK football fan communities on Twitter became conduits for political messaging in the post-Brexit Referendum period (2016–2017). Although political tweets represented fewer than 0.1% of 152 million football-related posts, they achieved substantial amplification through network position, identity alignment, and platform dynamics. Using a hybrid approach that integrates social network analysis with iterative content filtering, we isolate 94,846 political tweets and map how political narratives entered and propagated within fan networks. We propose an analytical framework that links recurring mechanisms of political influence to observable network and content signatures, enabling the systematic identification of candidate regions for closer investigation. We identify six distinct influence mechanisms: hashtag hijacking, suspected coordinated inauthentic amplification, hybrid activism embedded within fan groups, elite political “megaphones”, media provocation and inadvertent amplification. Five case studies collectively illustrate these six mechanisms. Each mechanism exhibits characteristic network signatures, including star-like amplification hubs, tightly reciprocal clusters, and dense ideologically aligned user communities that facilitate the diffusion of political narratives. Our findings show that political messaging was not randomly dispersed but concentrated within connective nodes of the football discourse network, enabling cross-community transmission. Propagation and amplification of political messages depended less on message volume than on structural integration. Beyond documenting a historical episode, we contribute a transferable analytical pipeline and a reproducible framework for identifying candidate regions of political influence through integrated network and content analysis, providing a foundation for future automated detection approaches.
Author summary
Political influence on social media is often studied in explicitly political spaces, yet much of everyday online life unfolds in communities organised around culture, sport, and shared identity. In this study, we examine how UK football fan communities on Twitter became sites of political engagement following the Brexit referendum. We combine network analysis and content analysis to identify how political messages entered and spread within a predominantly non-political environment. Rather than focusing only on individual actors or viral posts, we introduce a framework that combines network and content analysis to identify regions of online communities where recurring influence mechanisms are likely to operate. Through case studies, we show how political content was inserted into football conversations via hashtag hijacking, coordinated amplification, activist mobilisation, elite “megaphone” accounts, media-triggered provocation and inadvertent amplification. Our findings demonstrate that political influence in digital environments often operates indirectly, leveraging shared cultural identities and network structures rather than overt campaigning alone. Even low volumes of political content can achieve substantial amplification when aligned with structurally central nodes or cohesive communities. By outlining a reproducible workflow for identifying and interpreting these dynamics, we provide tools for understanding how online cultural spaces can become conduits for political influence and how such processes might be identified in other domains.
Citation: Paffard H, Pacheco D (2026) Kicking politics: Towards a framework for detecting political influence in online communities. PLOS Complex Syst 3(10): e0000141. https://doi.org/10.1371/journal.pcsy.0000141
Editor: Giulio Rossetti, Alessandro Faedo Institute of Science and Technology National Research Council: Consiglio Nazionale delle Ricerche Istituto di Scienza e Tecnologie dell’Informazione Alessandro Faedo, ITALY
Received: February 27, 2026; Accepted: October 5, 2026; Published: October 30, 2026
Copyright: © 2026 Paffard, Pacheco. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: To ensure reproducibility of this observational study, we share our data. We do not share the full raw data to comply with Twitter’s terms that prohibit sharing content obtained from the Twitter API with third parties. User and tweet IDs are anonymised to protect subject privacy. This allows other researchers to reproduce our results and/or compare their results. The link to the external repository is available at https://zenodo.org/records/21415527.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Political influence in networked online communities
The UK’s 2016 Brexit Referendum, in which 52% voted to leave the European Union, was a watershed moment in British politics, reshaping the political landscape and deepening societal divisions. The event triggered a distinct period of heightened political volatility and accelerated electoral realignment [1,2]. New polarised identities emerged that extended beyond traditional party affiliations, reflecting nationalist, globalist, and economic viewpoints [3]. Similar dynamics were evident in the subsequent 2016 US presidential election, prompting scholars to draw parallels between the two events [4].
In this shifting landscape, online social media platforms such as Twitter (now X) have become central arenas for political influence, shaping opinion formation and election outcomes. Social media has reshaped political communication by enabling direct interaction between politicians and the public, with political candidates using platforms to boost votes [5]. At the same time, it has empowered non-traditional influencers to shape debate at an unprecedented scale [6]. While some prior works have emphasised benefits for mobilisation and democratic participation [7], others highlight risks related to agenda-setting [8], selective amplification, and weakened editorial control [9].
In computational social science, political influence in online environments has been studied primarily through Social Network Analysis (SNA) and content-based approaches. SNA has been widely applied to electoral events, from inferring political affiliations [10] and identifying influencers [7], to tracking misinformation [11]. At a system level, the network propaganda model [12] highlights how network structures can facilitate political manipulation, while within networks, SNA has exposed coordinated tactics, including bots, sockpuppets, and astroturfing [13,14]. In parallel, content analysis has provided thematic insights into political discourse to capture issue priorities, measure polarisation, and approximate public opinion [3,15].
Political influence on social media is not confined to explicitly political communities, though; it also operates in broader cultural spaces. Influence in these settings is often not apparent at the level of individual messages, but can be observed instead through collective interaction patterns and discursive alignment. Analyses relying either on network structure or content in isolation risk overlooking more subtle forms of influence. Hybrid approaches that integrate SNA and content analysis partially address this limitation by combining structural and thematic perspectives, for example, in Brexit studies [7] and Italian election networks [16]. However, prior studies have focused predominantly on election-centred networks or identification of coordinated inauthentic behaviour, rather than the broader spectrum of political influence mechanisms we may experience in everyday online interactions.
In this study, we adopt an operational definition of political influence as the structured propagation and amplification of political content within a networked environment. This definition focuses on exposure and diffusion dynamics rather than direct measurement of belief or behavioural change, which are not observable within the available data. Accordingly, our analysis examines how network structure, actor roles, and discourse patterns enable political messages to spread within non-political communities.
There remains a lack of integrated frameworks for detecting mechanisms of political influence that operate within ostensibly non-political communities. This study addresses this gap by proposing a framework and examining whether these mechanisms can be identified using observable structural and content signals. The framework is intended as an analytical taxonomy that generates empirically testable hypotheses regarding the network and content signatures associated with different mechanisms of political influence, rather than as a fully automated classification system.
Football fandom as a site of political influence
Online spaces organised around shared cultural affiliations can function as sites of political influence, with political content introduced into the flow of everyday conversation. Football fandom, long considered a manifestation of British identity [17], represents one such space. Online fan communities increasingly express identities that extend beyond sport, and their strong collective identities and tribal behaviours make them attractive targets for those seeking to shape public opinion. Political actors have recognised these communities as vehicles for influence, whether mobilising activism, promoting nationalist ideologies, or managing international reputations [18].
Social media platforms have further transformed football communication. Clubs use these platforms to engage fans and broader stakeholders [19,20], while fans themselves act as content creators and broadcasters, also interacting directly with players. These interactions extend to social and political issues [21], bypassing traditional media intermediaries. At the same time, fan networks have been shown to enable toxic behaviours such as hate speech [22]. Together, these factors create an environment where political narratives can circulate within, and potentially be obscured by, broader cultural discourse.
The politically turbulent period surrounding the UK’s Brexit Referendum and subsequent general election (2016–2017) provides a compelling context for examining these dynamics. Prior research demonstrated that political content was present but relatively sparse within online UK football communities during this period, embedded within a substantial volume of football-related and other discourse [23]. Influence operated through structural patterns not visible by message volume alone. The present study extends that work by exploring the underlying influence mechanisms, motivating the use of a hybrid network– and content-based approach to analyse the propagation and amplification of political narratives.
While Guzmán et al. [19] identified sub-communities oriented towards politics and current affairs within Manchester United’s Twitter network, how political influence is enacted within online football fandom - and other cultural communities - remains under-explored. Computational studies of political influence in other sport-related networks are also limited. Existing work is focused primarily on cultural and commercial dimensions, with prior research exploring global fan sentiment during the 2014 World Cup using language similarity [24], club rivalries in the UK and Brazil via attention dynamics [25], and supporter diversity as a proxy for social disorganisation in the UK [26].
Research question and contributions
Addressing these gaps, this study builds on and extends our previous large-scale network analysis of political content within UK football-related Twitter interactions during 2016–2017 [23]. We advance and explore the applicability of a new analytical pipeline and framework for detecting political influence mechanisms in online cultural communities, demonstrating how such mechanisms can be identified systematically through a hybrid network- and content-based analysis of UK online football fan communities.
The primary research question is: How did political campaigns engage with and influence UK online football fan communities on Twitter during the post-Brexit Referendum period (2016–2017)? We examine patterns of interaction with political content, the role of influential actors, and the ways in which political narratives entered, circulated within, and gained traction in these communities. We show how political and football discourse became closely intertwined, with political parties, media outlets, activist groups, and pseudonymous influencers leveraging fan identities to mobilise support and provoke reactions.
Our previous study introduced three mechanisms through case studies: hashtag hijacking, embedded activism, and political “megaphones”. We extend these case studies to illustrate three additional mechanisms: suspected coordinated inauthentic activity, media provocation and inadvertent amplification – also providing expanded analysis of the previous cases. Together, these cases illustrate how distinct influence mechanisms can be identified and characterised through the proposed framework.
Acknowledging the limits of generalisation inherent in case-based analysis, our aim is to provide a reproducible pathway progressing from identifying political content in large-scale social media data to isolating and characterising the mechanisms through which political influence is enacted. The study makes three main contributions:
- A new framework for detecting political influence mechanisms in cultural online communities, linking influence mechanisms to observable network and content features.
- Five empirically grounded case studies illustrating six distinct influence mechanisms operating within these networks: hashtag hijacking, suspected coordinated inauthentic amplification, embedded “hybrid” activism, political “megaphones”, media provocation and inadvertent amplification, demonstrating how these can be identified systematically.
- A multi-step process for isolating political discourse from a predominantly non-political Twitter dataset, providing further methodological detail on our previous work for the purpose of supporting replication in other cultural domains.
Materials and methods
Dataset
The analysis drew on a UK football Twitter dataset comprising 152.3M UK football-related tweets, collected via the Twitter Streaming API between July 2016 and October 2017. Tweets were collected using accounts and hashtags associated with 44 UK Premier League and English Football League Championship clubs (Table A in S1 Appendix). From this corpus, 94,846 tweets containing political content were identified through an iterative filtering process (described below) for subsequent analysis. Although political tweets constituted fewer than 0.1% of the initial dataset, prior work demonstrates that low-volume political content can achieve substantial amplification through platform dynamics and agenda-setting effects [27]. To estimate the potential audience capacity embedded within the network, an Aggregate Follower Capacity (AFC) was calculated, defined as the sum of follower counts across all accounts represented in the political dataset. The same individual following multiple accounts was therefore counted multiple times. Accordingly, AFC provided an aggregate measure of potential audience capacity but did not estimate unique user reach or observed exposure, as follower overlap and tweet-level exposure could not be directly measured with the data available.
Analytical pipeline
To assess whether political influence can be detected through network and content analysis, we applied a staged analysis pipeline (Fig 1) to the large-scale UK football Twitter dataset. While developed for this study, the pipeline is conceptually transferable to other cultural contexts with domain-specific adaptations.
Political discourse was identified through iterative keyword and hashtag filtering, supported by a hashtag co-occurrence network. Influence structures were mapped using interaction-based networks and both user and hashtag similarity networks. A detection framework was applied to identify candidate influence cases through structural filtering, followed by behavioural, content and identity analysis to characterise influence mechanisms. Selected case studies were prioritised for qualitative validation.
First, political discourse was identified through an iterative filtering procedure combining political keyword and hashtag filtering with conversational context, further augmented by a hashtag co-occurrence network used to surface topics and iteratively expand the set of political markers.
Second, influence structures were mapped through two complementary network-based analyses. Interaction-based networks captured direct engagement patterns between users, supporting the identification of influential actors and conversational sub-groups. User and hashtag similarity networks captured latent discourse affiliations, enabling the identification of politically oriented user communities, independent of direct interaction. All networks were visualised using Gephi [28].
Third, a detection framework was applied to these analyses to identify cases of interest. Structural filtering reduced the search space by isolating network communities where concentrated influence dynamics were more likely. Behavioural, content, and identity characteristics were analysed to characterise influence mechanisms.
Fourth, communities meeting multiple diagnostic indicators were prioritised for further qualitative validation as case studies, with detailed analysis of broader evidence.
Identifying political content
Political content was identified through an iterative filtering process, the robustness of which was independently assessed through post hoc validation of both the extracted tweet set and the manually expanded political marker list. An initial set of political tweets was extracted using a curated list of hashtags and keywords derived from prior studies [7,15,29]. To preserve conversational context, original tweets quoted or replied to by these baseline political tweets were also retained, where present in the football dataset. By contrast, downstream replies that did not themselves contain political markers were excluded, as they typically diverged from the initial topic and were fragmentary in the football-based sampling frame. This approach prioritised salient political interactions while limiting the inclusion of off-topic content.
To expand the set of political markers, we constructed an undirected hashtag co-occurrence network in which nodes represented hashtags and edges connected hashtags appearing in the same tweet, weighted by the co-occurrence frequency of the hashtag pair. Hashtags were manually annotated as political, football, location, or other based on visual inspection. Community detection using the Louvain algorithm [30] was used to identify clusters of co-occurring hashtags corresponding to distinct discourse topics. Political hashtags emerging within clearly political clusters that were absent from the initial list were added to the filter, and the extraction process was repeated.
Throughout this process, curation balanced the inclusion of neutral political terms (e.g., #election) with explicitly partisan markers (e.g., #VoteLeave), while excluding overly broad terms (e.g., vote) to limit false positives.
Validation of the political filtering procedure
To assess the robustness of the political content extraction process, we conducted a manual validation of a random sample of tweets from the extracted dataset. This provided an empirical check on the precision of the iterative filtering procedure and helped identify the main sources of false positives. Recall cannot be directly estimated without a complete ground truth for the full corpus; however, this post hoc validation supports the reliability of the extraction approach. The validation was conducted after the filtering pipeline had been finalised and was not used to modify or refine the extraction procedure.
Mapping influence structures
Interaction-based influence.
To capture direct engagement patterns, we constructed directed user interaction networks in which nodes represented users, and edges represented retweet, reply, quote, or mention interactions. For retweets, replies and quotes, an edge from user u to user v indicated engagement with content produced by v; for mentions, an edge indicated that u mentioned v in their tweet. Edge weights corresponded to interaction frequency. Community detection using the Louvain algorithm supported the identification of conversational sub-groups, within which users engaged primarily with each other.
Influential actors were identified using multiple centrality measures - in- and out-degree, betweenness, and PageRank - to capture complementary dimensions of influence. Degree-based centralities distinguish between visibility and activity (in-degree versus out-degree), betweenness highlights brokerage or gatekeeping positions, and PageRank captures structural influence according to the influence of neighbouring nodes. The combination enabled differentiation of actor roles in the network, whether high-visibility broadcasters, key amplifiers, or central intermediaries.
Finally, to further characterise influencers, the top 20 users per centrality metric and interaction type were manually annotated by actor category (for example, politician, media outlet, football club), based on account profile descriptions and verification status, and other publicly available information.
User community affiliations.
Similarity networks were used to capture indirect, latent associations between users and discourse topics that were not observable through direct interactions, enabling the identification of politically aligned user communities. These networks were derived as one-mode projections of a user–hashtag bipartite matrix, yielding two undirected graphs: a user similarity network, capturing affinities between users based on shared hashtag usage, and a hashtag similarity network, capturing topical associations between hashtags based on shared user engagement.
The user similarity network was constructed from hashtag usage patterns, reflecting topical alignment rather than semantic similarity, even in the absence of direct interaction. A user–hashtag matrix was defined (where rows are users, columns are hashtags) with values corresponding to a user’s posting frequency for each hashtag. Low-activity users and infrequently used hashtags were removed to reduce noise and avoid spurious associations, enhancing community detection. The matrix was projected onto user–user space, with edge weights given by the cosine similarity of users’ hashtag usage vectors. To assess the robustness of the observed similarities, they were compared with a null distribution generated by randomly permuting the user–hashtag matrix and recomputing similarities. Edges were retained only where similarity values exceeded a threshold beyond which all edges were statistically significant, ensuring that they reflected only significant non-random alignment in hashtag usage.
An analogous projection was used to construct the hashtag similarity network, with edges weighted by the cosine similarity of hashtag pairs’ user engagement vectors. Again, only statistically significant edges relative to the null model were retained, ensuring that connections captured non-random patterns of user co-engagement.
The two similarity networks served as the foundation for a higher-level analysis linking user and hashtag communities to broader discourse themes. We characterised user communities by their engagement with hashtag communities, with engagement strength defined as the number of user–hashtag co-occurrences between members of a given user community and hashtags belonging to a given hashtag community, aggregated from the user–hashtag matrix. This yielded weighted associations allowing user communities to be characterised according to their participation in distinct discourse themes.
To capture broader thematic structure, hashtag communities were described by their proportional hashtag composition and clustered into overarching discourse themes (football, political, UK location, other) using hierarchical clustering with Ward linkage [31]. The number of clusters was selected via the elbow method (Fig A in S1 Appendix). User community–hashtag community engagement patterns were visualised with polar plots, which mapped the distribution of engagements and overlaid the higher-level discourse themes identified through clustering. This integrated analysis provides a compact representation of user community discourse affiliation, enabling the identification of politically orientated user communities.
Detecting influence mechanisms
A three-stage detection framework was applied to the outputs of these network analyses to identify cases exhibiting observable patterns of political influence. (Fig 2).
First, structural filtering isolated communities most likely to contain concentrated influence dynamics. Second, community structures were classified as centralised or distributed multi-hub systems. Third, behavioural, content, and identity indicators were analysed to detect and characterise influence mechanisms, following different analytical procedures for each community structure.
Detection stage 1 – isolating influence structures.
The first stage reduced the search space by isolating candidate conversational communities of interest within each interaction network (retweets, replies, quotes and mentions). Users were ranked according to each centrality metric, and the resulting rank-ordered centrality distributions were inspected. Influence was treated as a continuous structural property, rather than a binary classification. Key influencers were identified as high-centrality nodes above a distribution-informed threshold, defined by the elbow separating the high-centrality region from the long-tail distribution of lower-centrality users. This method calibrated the threshold to the dataset and network characteristics, as well as to the sensitivity of the search required. Accounts representing football clubs, players and institutions (e.g., the Premier League) were excluded, as they were considered moderators of discussion rather than potential originators of political influence mechanisms.
Louvain conversational communities were retained only if they contained at least one key influencer, on the basis that these communities were more likely to exhibit identifiable influence mechanisms. The retained communities were further prioritised using the user community-hashtag community engagement analysis described previously. Only communities containing key influencers affiliated with politically-aligned user communities were taken forward for detailed investigation.
Detection stage 2 – determining influence structure topology.
The retained communities were classified according to their internal influence topology. Communities containing a single dominant key influencer were classified as centralised, whereas those containing multiple key influencers were classified as distributed multi-hub. For distributed multi-hub communities, relative centrality differences were examined to determine whether influence was genuinely distributed or concentrated around one dominant account. Communities where one key influencer substantially dominated the others were reclassified as centralised. This classification determined the analytical pathway applied in stage 3.
Detection stage 3 – inferring influence mechanisms.
The third stage characterised candidate communities according to six recurring political influence mechanisms identified in the literature on political communication theory, networked publics and computational propaganda [32,33]: hashtag hijack, coordinated inauthentic amplification, hybrid activism, political megaphone, media provocation, and inadvertent amplification. Each influence mechanism represents a distinct pathway through which political narratives may be introduced, amplified or normalised within cultural communities. Analysis did not seek to infer political intent; instead, it inferred the mechanism from observable behavioural, content and identity characteristics. Communities exhibiting multiple indicators were prioritised for detailed qualitative case study analysis for final validation.
The influence mechanisms encompass both active mechanisms, characterised by deliberate political intervention, and passive or ambiguous mechanisms, where political influence arises indirectly, incidentally or with unclear intent. This distinction reflects that political influence may emerge not only through coordinated posting but also through routine interaction, platform affordances and properties implicit in social networks themselves.
Because these mechanisms manifest differently across community structures, separate analytical procedures for detection were developed for centralised and distributed multi-hub communities. The influence mechanism taxonomy and corresponding detection procedures are described below.
Operationalising influence mechanisms
Influence mechanism taxonomy.
Six mechanisms were operationalised as a set of network, content-based and behavioural features (Table 1), providing the target categories against which observable features were evaluated. This provided a basis for systematic identification and a foundation for future automation. Mechanisms were not assumed to be mutually exclusive, but could co-occur or operate sequentially. For example, a hashtag hijack followed by coordinated amplification.
These observable network and content signatures are intended as empirically grounded indicators rather than deterministic decision rules. Individual mechanisms should therefore be interpreted using the combined pattern of evidence across structural and content features, rather than any single characteristic in isolation.
- Hashtag hijack. Insertion of politically-oriented content into popular and highly visible non-political cultural hashtags, to gain attention from additional audiences [34]. This mechanism reflects agenda-setting dynamics, exploiting hashtag connectivity to introduce political narratives into unrelated conversations.
- Coordinated inauthentic amplification. Coordinated systematic dissemination of political content across multiple accounts, simulating organic support and consensus. This mechanism is grounded in media manipulation and online influence campaign literature [13,16].
- Hybrid activism. Organic amplification of political messaging by activist users embedded within cultural communities, often linking online participation with offline mobilisation. This mechanism reflects decentralised, identity-driven political participation within networked publics [32], rather than influence driven by formal organisational control.
- Political megaphone. One-way broadcasting of partisan political content by elite or institutional actors holding formal political roles, such as politicians and political parties. Consistent with classic elite-driven agenda-setting and the network propaganda model [12], influence operates through asymmetric amplification, with elite accounts receiving substantial attention despite limited direct engagement with others.
- Media provocation. Dissemination of politically neutral or ambiguous content framed as news, with content eliciting a political reaction in its audience. This mechanism typically reflects media-driven agenda-setting effects [35] rather than explicit political intent.
- Inadvertent amplification. Automatic redistribution of political content by culturally focused bot-automated accounts [36]. Unlike coordinated inauthentic amplification, this mechanism is reactive, triggered by predefined reposting criteria such as hashtag presence rather than coordinated interaction. Although associated with non-political content, political messages may be incidentally captured, rendering inadvertent amplification a passive mechanism through which platform automation drives message propagation.
The observable network and content signatures described above are intended to identify candidate regions of the interaction network that warrant closer investigation, rather than to provide deterministic labels for influence mechanisms. Mechanism identification is therefore based on the combined interpretation of network structure and content within these regions, rather than on any single structural criterion.
Low-intensity political discussion also occurs in online communities, embedded within everyday conversations through humour, commentary and other routine interaction. Although such discourse may contribute cumulatively to opinion formation, perceived consensus and the normalisation of political viewpoints, and to silencing dynamics associated with the ‘spiral of silence’, effects do not produce distinct structural patterns so were not operationalised.
Detection in centralised communities.
In centralised communities, the influence mechanism was inferred primarily from the identity and behaviour of the key influencer. First, to identify elite media or political affiliation, publicly verifiable identity information, including verified status, profile description, external organisational links and other metadata, was assessed.
Where elite affiliation was present, interaction patterns were examined to determine whether the account functioned as an institutional broadcaster, rather than as a participant in reciprocal discussion. Broadcast behaviour was operationalised through calculating the difference in influencer in-degree and out-degree. High degree asymmetry, where in-degree substantially exceeded out-degree, indicated one-way broadcasting. Elite political actors exhibiting this pattern were classified as candidates for the political megaphone mechanism, while equivalent elite media actors were classified as candidates for media provocation.
Where elite affiliation was absent, or broadcast behaviour was not observed, alternative amplification indicators were considered. Hashtag hijack candidates were inferred manually from tweet-level semantic misalignment between hashtag theme and accompanying post content. Inadvertent amplification was identified through the presence of redistributive behaviour with minimal original content production. The number of retweets as a proportion of the key influencer’s total tweets was calculated and redistributive behaviour confirmed where retweets comprised the majority of an account’s activity (greater than 90%, as a guide).
Detection in distributed multi-hub communities.
In distributed multi-hub communities, influence mechanisms were inferred from relationships between multiple key influencers, rather than the behaviour of a single account. Evidence of coordination between influencers was assessed before examining broader patterns of amplification.
Coordination was evaluated using three complementary indicators: interaction strength, content overlap and temporal synchronisation. Interaction strength was measured as the total interactions exchanged between pairs of key influencers (the sum of edge weights in both directions). Content overlap measured the extent to which key influencers amplified one another’s posts through retweets or quote tweets, with a high number of shared posts originating from one influencer as a proportion of another influencer’s content (e.g., more than 70% of posts) treated as an indicator of possible coordinated dissemination.
Temporal synchronisation was assessed from the distribution of reposting time lags between influencers, with repeated reposting behaviour within a particular time period treated as an indicator of coordinated amplification. Rather than applying a fixed threshold, the full distribution of lag times was examined to detect varying coordination tempos. A comparatively structured lag distribution, with an identifiable peak in the distribution around a given time window, was indicative of coordination. A flatter and more variable lag distribution was indicative of organic, less synchronised activity.
Where multiple coordination indicators were present, influencer identity was evaluated using the same profile-level indicators applied to centralised cases. Cases exhibiting coordination combined with ambiguous identity signals were classified as candidates for coordinated inauthentic amplification.
Where coordination was not observed, analysis shifted to broader amplification patterns within retweet communities, reflecting retweets’ role in organic narrative propagation. For each community, an influencer amplification concentration metric was calculated as the proportion of retweets directed towards key influencers relative to total community retweet activity. Higher values indicated greater dependence on a small number of key influencers for information diffusion. These influencers’ profiles and post content were examined for evidence of political or activist affiliation. Communities characterised by politically engaged influencers, elevated amplification concentration and an absence of coordination indicators were classified as candidates for hybrid activism.
Validating mechanisms with case studies
Candidate cases underwent final detailed qualitative validation using complementary content, actor, and network-based evidence. Manual inspection of posts confirmed dominant narratives and partisan framing, while influencer profiles and publicly available metadata provided further context on the actors involved. Media reports were used to relate online activity to relevant offline events, and previous network analysis findings were examined to verify the scale and pattern of information diffusion in the context of the identified user community affiliations.
Results
Political content and topical structure
Political content filtering retrieved 94,846 tweets, with the final filter criteria comprising 163 hashtags and 8 keywords (Table B in S1 Appendix). The political dataset is summarised in Table 2. The robustness of the political filtering procedure was assessed through independent validation of both the extracted tweet sample and the manually expanded political marker list (Section 2 in S1 Appendix). Independent annotation demonstrated high agreement for both tweet classification and political marker identification, supporting the reliability of the iterative extraction procedure.
Table 3 reports structural metrics for each network constructed in this study. The hashtag co-occurrence network (15,926 hashtags) revealed dense connectivity, with a giant component covering 98% of nodes and high average degree, density and clustering coefficients. These metrics indicate a cohesive space with heavily overlapping themes, in which most hashtags are directly or indirectly connected.
In the hashtag co-occurrence network visualisation (Fig 3, k-core > 25), restricted to political and football-related hashtag nodes only, 264 tightly connected hashtags formed thematic clusters. Community detection identified seven communities representing discourse topics, summarised with their dominant political and football hashtags in Table 4. Discourse topics were predominantly oriented around political themes, including partisan politics, electoral events and cross-cutting issues such as Brexit - interwoven with football hashtags. Central political hashtags included #Brexit, #Trump, #UKIP, and #EDL (English Defence League, a far-right anti-Muslim extremist movement.). Football club hashtags were distributed across all communities, with club hashtags such as #MUFC (Manchester United), #LFC (Liverpool), and #Arsenal appearing to bridge across both UK and US political contexts in hashtag co-occurrence.
Football and politics-related hashtags with > 25 co-occurring hashtags. Node colour represents discourse topic community, nodes sized by weighted degree. Zoom is recommended for detailed inspection.
Scottish club hashtags #Celtic and #RFC (Rangers) clustered with Scottish independence topics (#SNP, referring to the Scottish National Party, and #IndyRef2, referring to the proposed second Scottish Referendum on leaving the UK). Other clubs (for example, #NFFC - Nottingham Forest, #EFC - Everton, #CFC - Chelsea, #THFC - Tottenham Hotspur) aligned with far-right and nationalist discourse. Some peripheral synonymous hashtags appeared across clusters, suggesting distinct narrative use. Lower degree hashtags populate the periphery, forming a sparser background of contextually-related but more niche terms.
Although political content represented fewer than 0.1% of tweets in the broader football dataset, the accounts participating in this network represented substantial potential audience capacity. The Aggregate Follower Capacity (AFC) of accounts in the dataset was 511.4 million. Audience capacity was highly concentrated, with the most highly followed 1% of accounts accounting for 83% of total AFC. Within this top 1%, football-related accounts – including clubs, governing bodies, players, sports media and fan sites – accounted for 230.6 million AFC, approximately 54% of the AFC held by this group and 45.1% of total AFC. News media accounts accounted for 29% of the AFC held by the top 1%.
Interaction-based influence structures
Interaction networks and conversation groups.
In contrast to the densely connected discourse structure observed in the hashtag co-occurrence network, user interaction networks (retweets, quotes, replies, and mentions) were fragmented, with low average degree, density, and clustering (Table 3). Community detection identified many small conversational groups rather than a single cohesive structure, consistent with prior studies of social media networks [37].
Across all interaction types, attention was concentrated around a number of highly visible accounts, forming hub-and-spoke structures in the network visualisations (Fig 4). In the retweet network (Fig 4A), these took the form of star-like retweet cascades centred around influential accounts, indicating amplification of political content posted by these users. Some hubs were surrounded by secondary layers of interaction, suggesting diffusion beyond the initial cascade. However, because platform data attribute all retweets to the original author, these star-like structures may partially reflect platform-imposed diffusion assumptions rather than the true pathways of information flow, potentially obscuring intermediate amplifiers [38].
Zoom is recommended for detailed inspection. Node size represents PageRank, colour indicates communities. A: Retweets. B: Quotes. C: Replies. D: Mentions.
Beyond these central hubs, community detection revealed additional sub-structures of interest. A large conversational sub-group community containing overlapping retweet cascades was identified in the retweet network (in red, Fig 4A), suggesting that political messages circulated widely within a shared conversational space. Additionally, a clique of three accounts with frequent repeat interactions was observed in both the retweet and quote networks (in mid green, Fig 4A and B). These structures are examined further in our case studies.
The mention network differed from other interaction types. It was more cohesive, with fewer connected components (838 compared with over 2,000 for replies, retweets, and quotes) and a giant component containing 87% of nodes. At the same time, it exhibited the longest average path length (4.04 vs. 1.36-1.54 for other interaction types) and largest diameter (10 vs. 6), indicating more extended chains of interaction. Mentions therefore connected otherwise separate users, while replies, retweets, and quotes tended to form smaller, more fragmented conversational clusters.
Key influencers.
Actor prominence varied by interaction type, with centrality measures identifying distinct influence roles. Activist groups and unverified individuals dominated retweet and quote networks, reflecting the use of these interaction modalities for message amplification. In contrast, reply and mention networks were led by verified institutional accounts, including media outlets, football clubs, and politicians. These actors functioned as conversational anchors rather than active participants, attracting a high volume of replies and mentions (high in-degree) yet rarely engaging in return (low out-degree). Table 5 shows in-degree rankings by actor type across interaction sub-networks.
Betweenness centrality revealed contrasting bridging functions across networks. In the mention network, institutional football accounts (e.g., @premierleague, @ManUtd, @Carabao_Cup) dominated, accounting for 19 of the top 20 accounts by betweenness. These actors served as shared cultural reference points linking otherwise weakly connected communities. In retweets, though, political individuals and activist groups comprised half of the top 20 by betweenness, indicating a stronger role as discourse bridges transmitting political content across the broader football discussion space.
Taken together, these patterns indicate multiple co-existing influence mechanisms at play, including one-to-many institutional broadcasting, multi-hub amplification by political and activist actors, and more reciprocal forms of engagement.
User community discourse affiliations
Discourse themes.
As in the hashtag co-occurrence network, the hashtag similarity network (Fig 5) represented a largely cohesive discourse space, with most hashtags directly or indirectly connected – consistent with the global network metrics reported in Table 3. Community detection identified 27 hashtag communities, each corresponding to a group of hashtags tending to be used by similar users. Fig 5A shows these communities, excluding four small, uniformly non-political hashtag communities containing fewer than 10 unique hashtags that were excluded from further analysis.
Detected hashtag communities and assignment to overarching discourse themes based on proportional hashtag category composition. A: Nodes coloured by hashtag community, representing distinct discourse topic clusters. B: Nodes coloured by overarching discourse theme (Political, Football, UK, General). Hashtags occurring fewer than 20 times and users contributing only a single tweet were removed to enhance community detection. Only statistically significant edges with similarity > 0.1 were retained (p < 0.05).
To discern higher-level discourse structure, hierarchical clustering condensed the hashtag communities into overarching discourse themes based on their proportional composition by hashtag content category (Fig 5B). Dendrogram inspection and the elbow method indicated four clusters, defining the discourse themes: Political, Football, UK, and General. The Political theme is dominated by political hashtags with minimal football content. The Football theme is primarily football-related, with limited political crossover. The UK theme is characterised by a high proportion of UK location-based hashtags, combined with mixed topical content. The General theme shows low thematic concentration, containing a broad mix of hashtags, with the majority classified as ‘Other’. The proportional hashtag composition by hashtag community and associated dendrogram are given in Table C and Fig A in S1 Appendix, respectively.
Political and football-related themes formed distinguishable but interlinked overarching discourse clusters. The presence of intermingled regions reflects conversations with meaningful thematic convergence alongside separate political and football conversations, suggesting that political and football conversations frequently overlapped, rather than remaining separate.
From user communities to discourse affiliations.
The user similarity network (Fig 6) identified 25 user communities, excluding those with fewer than 10 users, which were omitted from further analysis. 11 communities accounted for 90% of the active users.
Nodes coloured by user community. The network was constructed as a one-mode projection of a user–hashtag bipartite matrix, capturing topical alignment among users based on hashtags. Hashtags occurring fewer than 20 times and users contributing only a single tweet were removed to enhance community detection. Only statistically significant edges with similarity > 0.45 were retained (p < 0.05).
The discourse affiliation of each user community was mapped based on users’ participation in hashtag communities and those hashtag communities’ relationships with the overarching discourse themes. This analysis revealed constituent user communities with strong political discourse affiliations. The five largest of these communities, selected on the basis of total users, are shown in the polar plots of Fig 7. Plots for the other communities within the top 11 are given in Fig B in S1 Appendix.
Polar plots for the largest politically-focused user communities display the distribution of a community’s tweets against hashtag communities and discourse themes. A & B: Communities 21 and 17, centred on Labour Party politics and Jeremy Corbyn. C: Community 4, linked to Brexit and Manchester United. D: Community 8, engaged with US politics. E: Community 11, representing broader UK political conversation.
Influence mechanism case detection
The detection framework identified 80 unique key influencers, representing the top 20 users per centrality metric across retweet, quote, reply and mention networks. These influencers were associated with 65 conversational communities. Filtering to influence affiliated with politically aligned user communities reduced this to 19 communities (Table 6), of which 13 exhibited diagnostic indicators consistent with the proposed influence mechanisms. Five representative cases were subsequently selected from these for detailed validation as case studies.
The political alignment filter proved most effective for the quote and retweet networks, where politically aligned influencers were readily identified. By contrast, only one politically aligned influencer was identified among the shortlisted influencers in the mention and reply networks, consistent with the conversational nature of reply and mention interactions, rather than information diffusion. Because reply and mention networks contained few politically aligned influencers, the highest-ranked user by in-degree in each network was also retained for subsequent analysis.
Influence topology classification identified 14 centralised communities and five distributed multi-hub communities. Among centralised communities, three were characterised as hashtag hijacking, five as media provocation, and two as political megaphone behaviour. The key influencer in one additional centralised mention community was intertwined with activity in a multi-hub retweet community, and the case was classified there. Among distributed multi-hub communities, one was classified as hybrid activism and one as coordinated inauthentic amplification. The remaining communities were ambiguous and were set aside for further manual review.
Case studies of political influence mechanisms
The following five case studies illustrate the six distinct mechanisms of political influence, with strong indicators found through the detection framework. Cases were chosen to reflect influential actors embedded in prominent interaction structures and politically oriented user communities. Case 1 demonstrates hashtag hijacking, alongside inadvertent amplification as a secondary influence mechanism, while Cases 2–5 illustrate the remaining four mechanisms. Associated tweet examples are shown in Table 7.
Case 1: hashtag hijacking and inadvertent amplification.
This case documents the politically motivated appropriation of the football hashtag #MUFC to disseminate partisan content. A pro-Trump influencer of ambiguous identity, ranked second and third by PageRank in the retweet and quote networks respectively, hijacked #MUFC alongside political hashtags #tcot (Top Conservatives on Twitter), #MAGA (Make America Great Again), and the ironic #LiberalsUnite to disseminate partisan content in a meme mocking Democrat Barack Obama: “Obama’s greatest accomplishment is that he built the strongest Republican Party. Thank you, Obama. #LiberalsUnite #MUFC #UNSC #MAGA #tcot”. The tweet generated 1,413 retweets and 205 quotes, forming a distinct hub-and-spoke structure in both interaction networks.
92% of retweeters belonged to US-politics-orientated user communities (communities 22 and 25, shown in Fig B in S1 Appendix). Manual inspection of retweeter profile descriptions further revealed the majority of US Conservative orientations, and keyword-based profile analysis found negligible evidence of football affiliation among these users: < 1% of retweeter profiles contained football-related terms, compared with a mean of 22% among typical #MUFC users. These findings indicate that amplification was driven primarily by politically aligned users.
Direct engagement from football fans was therefore limited. However, a football news bot retweeted the post 11 times, highlighting how the inclusion of a high-visibility cultural hashtag can help political messages infiltrate fan spaces via automated accounts. This example reveals two distinct suspicious dynamics: first, hashtag hijacking, whereby a political message strategically appropriates a football-related tag to penetrate an unrelated audience space; and second, bot-driven engagement manipulation through repetitive retweet activity that exploits platform visibility mechanisms. Prior research identifies interaction-toggling strategies as a means of producing transient engagement signals that can influence platform ranking mechanisms [39]. Although that study examines like/unlike dynamics rather than retweet activity, it demonstrates how engagement cycling can function as a manipulation vector. In the present case, repeated retweeting by an automated account operates as an analogous mechanism of amplification. We therefore find that this case demonstrates the hashtag hijack mechanism alongside a secondary passive influence mechanism of inadvertent amplification.
Case 2: coordinated inauthentic amplification.
This case documents coordinated posting consistent with inauthentic amplification of partisan political content to simulate organic political support. Most original posts in this case originated from a pseudonymous account with an ambiguous identity (profile description: “Dedicated to ensuring every echo chamber is fitted with acoustic panels”), ranked fourth by PageRank in the retweet network, and seventh in the quote network.
Closer inspection of the interaction networks revealed this user was part of a clique of three accounts from a UK politics-oriented multi-hub user community (Fig 7E). These accounts exhibited high reciprocity in the retweet and quote networks - excerpt in Fig 8). The accounts posted partisan content combining political and football club hashtags, focusing primarily on Scottish independence, Brexit, and the UK general election.
Excerpt from quote network showing reciprocal quoting among a clique of three core accounts. Two less active bots visible to the left of the visualisation were excluded from analysis, but display similar behaviour. Nodes sized by PageRank.
The original posts were subsequently amplified via retweets or quote tweets by one or more of the accounts within minutes to hours. Analysis of amplification time lags (Fig 9) showed sharp peaks for two accounts, consistent with automated behaviour. The third account, which entered the cluster in early 2017, displayed a flatter time-lag distribution, suggestive of manual posting. This could indicate the combination of automated amplification with human-like timing to evade detection, consistent with “cyborg account” amplification strategies that blend human and automated behaviours to increase apparent authenticity.
Time lag distributions between original tweets and their replies or quotes, for a sample of 211 original tweets. Logarithmic scale, labels represent absolute values.
Taken together, the dense reciprocal interaction network structure, pseudonymous accounts and temporal signatures strongly suggest the presence of the coordinated inauthentic amplification influence mechanism.
Case 3: hybrid activism.
This case illustrates hybrid activism, in which openly political actors embedded within a fan community exploit the network to amplify partisan messaging through grassroots engagement. A prominent and dense multi-hub retweet network community comprising 1,379 users (Fig 10) was observed in interaction analysis, with users primarily drawn from the Labour-aligned user communities shown in Fig 7A and B. The largest retweet cascades were driven by two prominent activist accounts: one self-described as advocating “ethical socialism”, ranked sixth by in-degree and eleventh by PageRank in the retweet network, and another positioning itself as a disseminator of Labour campaign messaging, (profile description “helping to get Labour’s General Election messages out and Jeremy Corbyn into No 10”).
Excerpt from retweet user interaction network showing the 1,379 users amplifying the Corbyn banner event.
Content analysis linked this retweet community to an offline event on 7 May 2017 when, during a Liverpool football match, activists unveiled a banner supporting Labour leader Jeremy Corbyn alongside a campaign slogan signalling electoral support (https://bit.ly/hybrid_activism). Using the keywords Corbyn, banner, and the campaign slogan, 619 tweets referencing the event were identified. The Labour-aligned community exhibited significantly higher per-user engagement (Fig 11) than the rest of the network, as well as retweet rates (97% vs. 65%).
Time series of tweets per user among the activist cluster (red) vs. the rest of the network (blue). A sharp activity peak on the event date is observed in the left-leaning cluster.
Amplification was driven by explicitly political accounts and bottom-up engagement among ideologically sympathetic users present in the fan community, distinguishing the hybrid activism mechanism from coordinated inauthentic amplification that simulates grassroots support through concealed coordination. The case demonstrates how dense, ideologically aligned sub-networks propagate political messages, providing conduits for political mobilisation when political and cultural identities co-exist.
Case 4: political megaphone.
Verified political actors accounted for 5 of the top 20 users by in-degree across all user interactions. Jeremy Corbyn and the Labour Party (@UKLabour) both ranked highly as influencers across quote, reply, and mention sub-networks by in-degree and PageRank, despite minimal football-related posting: Corbyn tweeted once on Arsenal’s FA Cup performance, and @UKLabour twice on grassroots football investment. Their prominence stemmed instead from positional authority via frequent mentions (Corbyn 948, Labour 581).
Mentions came from both supporters and opponents, clustering along partisan lines. Classifying mentions manually based on tweet content and campaign hashtag usage into five categories (Supportive, Opposed, Neutral, Irrelevant, or Humorous), we found that supportive mentions concentrated within Labour-affiliated user communities, while oppositional mentions aggregated in broader UK politics user communities. These accounts functioned as political megaphones, attracting distributed amplification despite limited direct engagement. Supporters and opponents contributed to the onwards propagation of political messages within the sampled fan communities.
Case 5: media provocation.
This case illustrates a provocation-based influence mechanism in which media actors stimulate political engagement through ambiguous or neutral commentary, generating polarised responses without sustained interaction. Influencer analysis identified football pundit Gary Lineker as the 4th most replied-to user in the reply network by in-degree, as well as 3rd and 7th by PageRank and betweenness, respectively, despite limited posting activity. Content analysis found that Lineker had quoted a Daily Mirror journalist’s tweet unveiling UKIP’s proposed logo, adding a brief sarcastic remark that mocked the party’s branding and its resemblance to the Premier League emblem.
This single post triggered a burst of polarised responses. Among 49 replies and quotes captured within the dataset, around 50% opposed UKIP, 20% supported UKIP, while the remainder criticised Lineker or responded humorously. The incident was amplified further through the mainstream media when the Mirror reported on the online reaction (https://bit.ly/mediaprovocation). Lineker did not engage further, yet centrality metrics indicate his account’s structurally influential position in routing attention and responses.
An additional 263 tweets referencing the UKIP logo episode were identified in the dataset, using keywords ‘UKIP’ and ‘logo’. The most retweeted of these was a neutral summary post by Sky News, whose account ranked 15th by in-degree in the reply network. The post linked to its online news article (https://bit.ly/mediaprovocation_2) and generated a comparable level of engagement (49 replies and quotes), again with a partisan split: 31% supported UKIP, 22% opposed it, and 12% criticised the media framing itself.
Across both instances, media actors acted as highly followed broadcaster accounts, engaging in a one-way interaction modality associated with polarised responses. These dynamics reflect the media provocation mechanism, with media actors introducing political discourse into fan communities without reciprocal interaction.
Discussion
Key findings and implications
Our findings identify multiple, co-existing political influence mechanisms operating within UK football fan communities on Twitter during 2016–2017. Case studies were selected by applying the detection framework, integrating four analytical components arising from observable structural properties: interaction network structures, influential actor analysis, community discourse affiliations and content analysis.
Across the dataset, political content constituted a small proportion of overall football-related discourse, yet individual political messages were capable of generating substantial amplification. Interaction networks were fragmented at the level of direct user-to-user exchanges, but remained connected through overlapping conversational structures, shared hashtags, and thematic linkages. Across both interaction-based and similarity-based network representations, user communities exhibited strong internal cohesion and thematic alignment, indicating stable patterns of affiliation and discourse practice. This produced a cohesive discourse space even without direct user-to-user interaction. Diverse actors embedded political messages into everyday football discussion, frequently leveraging football-specific contexts to frame broader issues. Evidence of hashtag hijacking, suspected coordinated political posting, and embedded activism indicates that online football fandom provided a setting in which political mobilisation and amplification occurred.
Political content was not uniformly distributed across communities but was concentrated around particular actors, hashtags, and conversational communities that bridged otherwise non-political football discussion. This pattern is consistent with instances of targeted embedding, with political messaging introduced through structurally central or connective nodes rather than broadly dispersed across fan communities. Mechanisms such as hashtag hijacking, coordinated inauthentic amplification, hybrid activism, and political megaphone dynamics reflect distinct pathways through which political content entered and propagated within cultural discourse spaces.
The dominant political narratives identified through thematic analysis of the hashtag networks were consistent with prior studies of the period. These spanned the political spectrum, including Brexit, the 2017 General Election, Corbyn-era Labour politics, Scottish independence, nationalist and far-right discourse, and pro-Trump/MAGA messaging. The presence of transatlantic linkages between UK and US populist narratives reflects broader political currents of the time. These narratives were not isolated intrusions but were interwoven with everyday football discussion, circulating both deliberately and organically within user communities. In some cases, discourse originating within football-tagged conversations crossed into mainstream media coverage.
While high-profile accounts such as politicians and media commentators played predictable agenda-setting roles, non-traditional actors played a substantial role in message propagation. Activist groups, hyper-partisan influencers and likely bot or cyborg accounts frequently acted as local amplifiers within tightly clustered communities. These findings align with network propaganda theory, which emphasises decentralised and mutually reinforcing influence dynamics over top-down institutional broadcasting models. The propagation and amplification of political messaging was shaped less by message content or posting frequency than by structural position, underlining the importance of combining network and content analysis when examining influence strategies.
We found that a distinctive structural role was played by major football clubs and organisations – most notably Manchester United, Arsenal, Liverpool, the Premier League and the Carabao Cup – which functioned as central cultural reference points connecting otherwise disparate groups. Their accounts and hashtags occupied structurally significant positions in interaction networks through vast global fan bases and visibility within football discourse. Political content appearing alongside club-related hashtags was able to traverse community boundaries beyond overtly political groups. In this way, these neutral cultural anchors unintentionally became structural enablers of political influence. The concentration of audience capacity among highly connected football-related accounts, exceeding that of media accounts, further reinforces their role as structural connectors within cultural communities, where politically relevant content can be disseminated beyond conventional political spaces.
Although political content constituted only around 0.1% of tweets in the broader football dataset, this modest prevalence did not imply limited potential for dissemination. Instead, political messaging operated within online communities characterised by substantial audience capacity. Its embedding within highly connected football communities created opportunities for political narratives to diffuse through existing cultural networks rather than solely through explicitly political channels. Network position acted as an enabling structural condition that moderated the transmission and potential visibility of political messages. This interpretation is consistent with research demonstrating how football institutions can serve as vehicles for soft power and reputation management [40,41]. Community structure analysis indicated some degree of fan tribalism among larger clubs, but this was less pronounced than anticipated, suggesting that shared cultural affiliation outweighed club-level segmentation in enabling the diffusion of political messaging.
The findings have several implications for how political influence in online communities is conceptualised and detected. Network analysis identified not only influential actors but also the structural conditions under which political narratives gained traction. For example, the large Corbyn-supporting activist cluster exhibited substantial message propagation alongside strong structural integration within football fan communities. In contrast, isolated hashtag hijacking attempts that lacked such integration showed limited penetration into football discourse. Collectively, these findings demonstrate that political influence in online communities operates as a complex system shaped by network topology. Properties emerging from networked discourse systems provide the infrastructure for political transmission. Detection requires hybrid analytical approaches that integrate content analysis with structural network features. Content-centric or actor-centric approaches risk missing influence signals that are structurally latent but socially consequential.
Limitations
This study focuses on English-language Twitter content related to Premier League and Championship football clubs, offering a broad but necessarily partial view of UK online football fandom. It excludes platforms such as Facebook and Instagram, limiting demographic representativeness. Twitter’s user base skews towards politically engaged demographics [5], so findings may reflect primarily politically active segments of football fandom rather than the general public.
Further limitations include: (1) reliance on hashtags and keywords, although independent validation indicates high precision, which may still overlook indirect forms of political expression; (2) focus on political content within football discourse, excluding users’ wider political activity beyond this context; (3) incomplete data coverage, during key periods (June 2016 EU Referendum; partial June 2017 UK General Election); and (4) challenges in interpreting tone or intent, particularly distinguishing between genuine political discourse, sarcasm, and humour, introducing unavoidable subjectivity, even with manual review.
Finally, while the hybrid content–network approach enables robust identification of structural and thematic patterns, the study remains observational in nature and cannot establish actor intent nor definitively confirm coordination strategies. The case studies allow for fine-grained analysis but limit generalisability and the scope of inference. Nevertheless, these limitations do not detract from our core contributions, isolating political discourse and demonstrating how hybrid network and content analysis can be applied to detect political influence in online communities.
Future research directions
Our findings provide a foundation for a framework to detect and classify political influence in football and other cultural online spaces, informed by both network structures and content patterns. Although developed in the context of UK football fan communities, the proposed framework is not specific to this domain. The influence mechanisms are defined in terms of general network and content patterns, suggesting applicability to other culturally grounded online communities where shared identities structure interaction. Future work should test this framework across different domains and platforms and assess its generalisability. An interpretable approach integrating the observable network and content signatures identified in this study could support scalable detection for academic and policy purposes, while also underpinning supervised classifiers for systematic identification and risk assessment. Achieving this will require formalising these signatures into explicit quantitative decision rules that can be evaluated across different datasets.
Conclusion
Using a multi-step political content extraction process and a mixed-methods approach combining SNA and content analysis, this study examined how political narratives penetrated non-political online football fan communities during 2016–2017. It identified distinct mechanisms of political influence by combining network structure and content analysis while highlighting the key actors, communities and structural conditions that enabled political messaging to circulate within everyday football discourse. Rather than relying on isolated structural indicators, the proposed framework integrates network and content signatures to identify candidate regions for closer investigation, enabling mechanisms of political influence to be interpreted within their social context.
Findings confirm that political influence was not incidental but embedded within online discourse on Twitter during the observation period. However, the contribution of this study lies beyond documenting a historical episode; it lays the groundwork for a detection and classification framework that could be extended into AI-based systems for identifying influence strategies. Such work would require formalising network signatures, but could inform governance frameworks and protective measures for cultural communities online. At the same time, any such detection efforts must remain sensitive to the distinction between political manipulation and legitimate political expression. Safeguards for freedom of expression must remain central: excessive monitoring risks reinforcing dominant narratives and prompting self-censorship, echoing the ‘spiral of silence’.
This research remains highly relevant, as social media continues to shape political and sporting discourse, with X a key platform. The findings advance understanding of how political actors leverage pre-existing social structures, including fan networks and club affiliations, for mobilisation. State actors also increasingly use football for soft power projection, as seen in high-profile ‘sportswashing’ examples such as the 2022 Qatar World Cup [40,41]. In an era of geopolitical instability, misinformation, and persistent polarisation [11], understanding how divisive viewpoints spread is vital. By showing that football fan communities can act as conduits for political influence, this study contributes to wider efforts to build societal resilience against digital manipulation, protecting both fans and democratic discourse.
By combining interpretable network and content signatures with qualitative contextual analysis, the proposed framework provides a reproducible workflow for identifying candidate regions of political influence in large online communities and interpreting the mechanisms operating within them. In doing so, it establishes a foundation for future automated detection approaches while preserving the contextual interpretation needed to distinguish political manipulation from legitimate political expression.
Supporting information
S1 Appendix. Supporting figures, tables, and validation analyses.
The appendix contains Fig A–B and Tables A–D, including the hierarchical clustering dendrogram, football club accounts and hashtags used for data collection, political hashtags and keywords used for content filtering, hashtag community composition, additional user-community discourse affiliations, validation of the political filtering procedure, and validation of the manually expanded political marker list.
https://doi.org/10.1371/journal.pcsy.0000141.s001
(PDF)
Acknowledgments
For the purpose of open access, the authors have applied a Creative Commons Attribution (CC BY) licence to any accepted manuscript version arising from this submission.
References
- 1.
Blick A. Democratic turbulence in the United Kingdom. First edition ed. De Gruyter. 2024. https://doi.org/10.1515/9783110735925
- 2. Griffiths JD, Perrett S, Fieldhouse E, Prosser C, Green J, Mellon J, et al. The Brexit realignment amid electoral volatility: The role of party blocs in the 2024 General Election. Parliamentary Affairs. 2025;79(1):154–89.
- 3. Bastos M, Mercea D. Parametrizing Brexit: mapping Twitter political space to parliamentary constituencies. Information, Communication & Society. 2018;21(7):921–39.
- 4. Ramswell PQ. derision, division–decision: parallels between Brexit and the 2016 US presidential election. Eur Polit Sci. 2017;16(2):217–32.
- 5. Kartsounidou E, Papaxanthi D, Andreadis I. Measuring the impact of candidates’ tweets on their electoral results. Journal of Information Technology & Politics. 2022;20(2):154–68.
- 6. Riedl MJ, Lukito J, Woolley SC. Political Influencers on Social Media: An Introduction. Social Media Society. 2023;9(2):20563051231177938.
- 7. Lynn T, Rosati P, Nair B. Calculated vs. Ad Hoc Publics in the #Brexit Discourse on Twitter and the Role of Business Actors. Information. 2020;11(9):435.
- 8. McCombs ME, Shaw DL. The Agenda-Setting Function of Mass Media. Public Opinion Quarterly. 1972;36(2):176.
- 9. Walsh M, Singer JB. Platform speech: Journalists and political campaigners reflect on Facebook and disintermediation in three UK general elections. Journalism. 2024;26(10):2087–104.
- 10. Mora-Cantallops M, Sánchez-Alonso S, Visvizi A. The influence of external political events on social networks: the case of the Brexit Twitter Network. J Ambient Intell Human Comput. 2019;12(4):4363–75.
- 11. Bovet A, Makse HA. Influence of fake news in Twitter during the 2016 US presidential election. Nat Commun. 2019;10(1):7. pmid:30602729
- 12.
Benkler Y, Farris R, Roberts H. Network propaganda: manipulation, disinformation, and radicalization in American politics. Oxford University Press. 2018. https://doi.org/10.1093/oso/9780190923624.001.0001
- 13. Pacheco D, Hui P-M, Torres-Lugo C, Truong BT, Flammini A, Menczer F. Uncovering Coordinated Networks on Social Media: Methods and Case Studies. ICWSM. 2021;15:455–66.
- 14. Sela A, Neter O, Lohr V, Cihelka P, Wang F, Zwilling M, et al. Signals of propaganda-Detecting and estimating political influences in information spread in social networks. PLoS One. 2025;20(1):e0309688. pmid:39883667
- 15.
Cram L, Llewellyn C, Hill R, Magdy W. UK general election 2017: A Twitter analysis. In: 2017. https://doi.org/arXiv:170602271
- 16. Giglietto F, Righetti N, Rossi L, Marino G. It takes a village to manipulate the media: coordinated link sharing behavior during 2018 and 2019 Italian elections. Information, Communication & Society. 2020;23(6):867–91.
- 17. Poulton E. Mediated patriot games: The construction and representation of national identities in the British television production of Euro ‘96. International Review for the Sociology of Sport. 2004;39(4):437–55.
- 18. Moreau N, Roy M, Wilson A, Atlani Duault L. “Life is more important than football”: Comparative analysis of Tweets and Facebook comments regarding the cancellation of the 2015 African Cup of Nations in Morocco. International Review for the Sociology of Sport. 2020;56(2):252–75.
- 19. Guzmán EM, Zhang Z, Ahmed W. Towards understanding a football club’s social media network: an exploratory case study of Manchester United. IDD. 2021;49(1):71–83.
- 20. Romero-Jara E, Solanellas F, López-Carril S, Kolyperas D, Anagnostopoulos C. The more we post, the better? A comparative analysis of fan engagement on social media profiles of football leagues. International Journal of Sports Marketing and Sponsorship. 2024 Jul;25(3):578–99. Available from: https://www.emerald.com/insight/content/doi/10.1108/IJSMS-12-2023-0252/full/html
- 21.
Billings AC, Anderson J. Social media, digital technology, and national identity in sport. In: Sanderson J. Research in the sociology of sport. Emerald Publishing Limited. 2022. p. 107–25. https://doi.org/10.1108/S1476-285420220000015013
- 22. Seijbel J, van Sterkenburg J, Oonk G. Expressing rivalry online: antisemitic rhetoric among Dutch football supporters on Twitter. Soccer & Society. 2022;23(8):834–48.
- 23.
Paffard H, Pacheco D. Kicking Politics: How Football Fan Communities Became Arenas for Political Influence. Complex Networks & Their Applications XIV. Springer Nature Switzerland; 2026. p. 182–94. https://doi.org/10.1007/978-3-032-16723-1_16
- 24.
Pacheco D, de Lima NFB, Moyano L, Menezes R. Football conversations: what Twitter reveals about the 2014 World Cup. In: Brazilian Workshop on Social Network Analysis and Mining (BraSNAM). SBC; 2015.
- 25.
Pacheco DF, Pinheiro D, De L-NFB, Ribeiro E, Menezes R. Characterization of Football Supporters from Twitter Conversations. In: 2016 IEEE/WIC/ACM International Conference on Web Intelligence (WI). IEEE Computer Society; 2016. p. 169–76.
- 26.
Pacheco DF, Oliveira M, Menezes R. Using social media to assess neighborhood social disorganization: A case study in the United Kingdom. In: 30th International Florida Artificial Intelligence Research Society Conference, FLAIRS 2017, 2017. 341–6.
- 27. Huszár F, Ktena SI, O’Brien C, Belli L, Schlaikjer A, Hardt M. Algorithmic amplification of politics on Twitter. Proc Natl Acad Sci U S A. 2022;119(1):e2025334119. pmid:34934011
- 28. Bastian M, Heymann S, Jacomy M. Gephi: An Open Source Software for Exploring and Manipulating Networks. ICWSM. 2009;3(1):361–2.
- 29.
Lee HW. Using Twitter Hashtags to Gauge Real-Time Changes in Public Opinion: An Examination of the 2016 US Presidential Election. In: Social Informatics. ISSN: 1611-3349. Springer, Cham; 2018. p. 168–75.
- 30. Blondel VD, Guillaume J-L, Lambiotte R, Lefebvre E. Fast unfolding of communities in large networks. J Stat Mech. 2008;2008(10):P10008.
- 31. Ward JH Jr. Hierarchical Grouping to Optimize an Objective Function. Journal of the American Statistical Association. 1963;58(301):236–44.
- 32.
Boyd D. Social network sites as networked publics: affordances, dynamics, and implications. In: Papacharissi Z. A networked self. Routledge. 2011. p. 47–66. https://doi.org/10.4324/9780203876527-8
- 33.
Woolley SC, Howard PN. Political communication, computational propaganda, and autonomous agents. 2016. https://comprop.oii.ox.ac.uk/research/working-papers/political-communication-computational-propaganda-and-autonomous-agents/
- 34. Xanthopoulos P, Panagopoulos OP, Bakamitsos GA, Freudmann E. Hashtag hijacking: What it is, why it happens and how to avoid it. JDSMM. 2016;3(4):353.
- 35.
McCombs M. Setting the agenda: mass media and public opinion. Cambridge: Polity Press. 2014.
- 36. Ferrara E, Varol O, Davis C, Menczer F, Flammini A. The rise of social bots. Commun ACM. 2016;59(7):96–104.
- 37. Himelboim I, Smith MA, Rainie L, Shneiderman B, Espina C. Classifying Twitter Topic-Networks Using Social Network Analysis. Social Media + Society. 2017;3(1).
- 38.
DeVerna MR, Pierri F, Aiyappa R, Pacheco D, Bryden J, Menczer F. Information diffusion assumptions can distort our understanding of social network dynamics. 2024. https://arxiv.org/abs/241021554
- 39.
Torres-Lugo C, Pote M, Nwala AC, Menczer F. In: Proceedings of the International AAAI Conference on Web and Social Media, 2022. 1029–39.
- 40. Ganji SK. The Rise of Sportswashing. Journal of Democracy. 2023;34(2):62–76.
- 41. Hassan AAM, Wang J. The Qatar World Cup and Twitter sentiment: Unraveling the interplay of soft power, public opinion, and media scrutiny. International Review for the Sociology of Sport. 2023;59(5):679–704.