Algorithmic amplification of geopolitical narratives and user engagement in the 2024 U. S. presidential election

Abstract

This study examines how geopolitical narratives were amplified through presidential campaign communication on YouTube during the 2024 U. S. presidential election, and the extent to which such narratives were associated with differential user engagement (operationalized as a platform-based engagement proxy) compared to content related to domestic issues. Drawing on agenda-setting theory and framing theory as a dual theoretical lens, the study employed computational content analysis of 5,466 videos collected through the YouTube Data API v3 from official campaign channels and affiliated political organizations spanning January 2024 through Election Day (5 November 2024). Topic classification was performed using GPT-4 and BERT-based models, and sentiment analysis was conducted using RoBERTa and GPT-assisted coding. All AI-generated topic and frame classifications were independently verified against trained human coders, with inter-coder reliability meeting accepted thresholds (Krippendorff’s α = 0.78–0.86; Cohen’s κ = 0.76–0.84). Three statistical models were applied: a negative binomial regression, a one-way analysis of variance (ANOVA), and a multiple regression. Results show that videos featuring geopolitical narratives generated significantly higher mean engagement (M = 88,004; SD = 74,122) than videos on domestic issues (M = 62,445; SD = 48,112), with immigration and border security content generating the highest overall engagement (M = 118,647). Among framing strategies, threat framing emerged as the most powerful predictor of engagement (β = 0.34, p < 0.001), followed by national security framing (β = 0.24, p < 0.001) and patriotism framing (β = 0.21, p < 0.001). Negative sentiment scores also contributed significantly to engagement outcomes (β = 0.29, p < 0.001), suggesting that emotionally charged, threat-oriented geopolitical content is associated with the engagement signals that platform recommendation systems are designed to reward. These findings carry implications for democratic communication, platform governance, and the strategic use of AI-assisted political messaging in contemporary electoral campaigns.

1 Introduction

The 2024 United States presidential election unfolded against an unusually turbulent geopolitical backdrop: an ongoing ground war between Russia and Ukraine, an intensifying conflict between Israel and Hamas in Gaza, escalating cross-strait tensions between the People’s Republic of China and Taiwan, and renewed debate about NATO’s structural coherence and American commitments abroad. Unlike earlier elections in which foreign policy considerations had a secondary role compared to domestic economic concerns, the 2024 cycle presented campaign strategists with an expanded repertoire of narratives rooted in geopolitical developments that could be weaponized, reframed, and amplified at large scale, through digital media ecosystems. The intersection of these geopolitical conditions with increasingly sophisticated artificial intelligence tools for content generation and distribution signifies a consequential development in the history of American political communication, one that demands systematic scholarly attention.

Two theoretical traditions have long highlighted how media shape public political perception: agenda-setting theory (McCombs and Shaw, 1972) and framing theory (Entman, 1993). But both frameworks were articulated in an era dominated by broadcast and print media, where the editorial gatekeeping role was exercised by identifiable human actors. The emergence of algorithmic curation on platforms such as YouTube, X (formerly Twitter), and Meta’s suite of applications has introduced a new layer of agenda-setting that operates through computational rather than editorial logic, a phenomenon Althaus and Tewksbury (2002) predicted and that subsequent scholars have conceptualized explicitly as algorithmic agenda-setting (Aday et al., 2010; Napoli, 2014; Thorson and Wells, 2016). In this environment, the strategic framing of content by political campaigns intersects with platform incentive structures that reward engagement signals, such as likes, comments, shares, and view duration, to produce amplification effects that cannot be orchestrated by any single editorial decision.

Artificial intelligence has further transformed this environment. Large language models and multimodal generative systems now allow campaigns to produce, personalize, and distribute video content at volumes and speeds that were unachievable during the election cycles of 2016 or 2020 (Benkler et al., 2018; Vaccari and Chadwick, 2020; Kreiss and McGregor, 2018). At the same time, AI-powered analytical pipelines allow researchers to study the resulting information environment with corresponding sophistication, allowing computational content analysis of corpora that would be intractable using only manual methods (Grimmer and Stewart, 2013; Wilkerson and Casas, 2017; Törnberg, 2023). This research occupies exactly this methodological terrain, employing AI classification and sentiment analysis tools to systematically examine the relationship between geopolitical narrative framing and user engagement on YouTube during the 2024 election.

The research is guided by four research questions. First (RQ1), the study asks which geopolitical narratives, spanning the Russia-Ukraine war, the Israel-Hamas/Gaza conflict, China-Taiwan tensions, immigration and border security, and NATO and international alliances, were most prominently featured in presidential campaign communication on YouTube. Second (RQ2), it explores how user engagement, operationalized as the aggregate of likes, comments, shares, and views and treated throughout as a platform-based proxy for attention rather than as verified voter behavior, differed between geopolitical narrative posts and posts on domestic issues. Third (RQ3), it investigates which framing characteristics, such as threat framing, patriotism framing, national security framing, and economic-security framing, generated the highest levels of user engagement, and whether these relationships were moderated by candidate’s identity. Fourth (RQ4), the study examines whether engagement patterns vary with the ideological orientation of the content, recognizing that in a polarized electorate the mobilizing power of a given narrative or frame is unlikely to be uniform across the ideological spectrum.

2 Literature review

2.1 Agenda-setting theory and its digital extension

Agenda-setting theory holds that media shape not what audiences think but which issues they think about, structuring the salience hierarchy used to evaluate candidates and policies (McCombs and Shaw, 1972). The framework was subsequently extended to second-level (attribute) agenda-setting and to framing as a third level (Iyengar and Kinder, 1987; McCombs et al., 2014).

Digital media complicate this mechanism: online news fragments the once-shared agenda (Althaus and Tewksbury, 2002; Prior, 2007; Sunstein, 2017), while algorithmic curation has displaced much of the editorial gatekeeping function, producing personalized, recommendation-shaped information diets (Napoli, 2014; Thorson and Wells, 2016). Early evidence that platforms set agendas in ways diverging from broadcast media (Aday et al., 2010) anticipated the algorithmic agenda-setting now central to the field (Jungherr, 2016; Benkler et al., 2018).

2.2 Framing theory and political persuasion

If agenda-setting theory addresses which issues receive attention, framing theory addresses how those issues are presented. Entman (1993) established definition identifies framing as the process of selecting and highlighting particular aspects of a perceived reality to make them more noticeable, in a manner that promotes a particular problem definition, causal interpretation, moral evaluation, or treatment recommendation. The political significance of frames exists in their capacity to activate different value schemas within audiences, leading individuals with similar factual knowledge to reach divergent policy conclusions depending on how an issue has been packaged (Chong and Druckman, 2007a). Druckman (2001) experimentally showed that competing frames can considerably shift opinion when they activate frames in communication that resonate with frames in thought, meaning when the structure of a message aligns with the cognitive associations that have already been internalized by an audience member.

In the context of geopolitical communication, threat framing has received extensive scholarly attention. De Vreese (2005) established that threat and conflict frames in foreign policy coverage amplify public concern and galvanize support for government action. This effect was positioned, by Soroka (2014), within a broader negativity bias in political information processing, i.e., the well-documented tendency for citizens to weigh negative information more heavily than comparable positive information. Patriotism and national security frames operate through a different mechanism. By invoking collective identity and shared existential interests, they mobilize group identification and suppress deliberative skepticism (Entman, 2003). Economic-security frames, in contrast, exploit the personal relevance heuristic (Iyengar, 1991), translating abstract geopolitical events into material household consequences that are cognitively closer to the self-interest of voter. Chong and Druckman (2007b) theorized that frames compete in an open marketplace, and that the frames which are both strong, meaning activating well-established value associations, and frequent, meaning repeatedly encountered, tend to dominate political discourse. This proposition has clear implications for algorithmically amplified campaign content.

2.3 Political communication on digital and social media

The scholarly literature on digital political communication has grown significantly over the past two decades, tracking the successive shifts of campaign activity from websites and email (Bimber and Davis, 2003) to social network platforms (Gainous and Wagner, 2014) and, more recently, to short-video and streaming platforms. YouTube’s position within this ecosystem is distinctive. As both a search engine and a social platform, it combines the pull dynamics of intentional information-seeking with the push dynamics of algorithmic recommendation (Munger and Phillips, 2022). Scholars have documented YouTube’s capacity to distribute political content on large scale (Ottoni et al., 2018), the platform’s tendency to recommend increasingly partisan content through its recommendation algorithm (Ribeiro et al., 2020; Huszár et al., 2022), and the asymmetric engagement advantages enjoyed by emotionally charged, polarizing content (Brady et al., 2017; Rathje et al., 2021).

Across the 2016 and 2020 cycles, platforms functioned not as passive channels but as active shapers of the information environment, with hybrid human-automated flows and cross-platform diffusion extending the reach of strategic messaging (Kreiss and McGregor, 2018; Settle, 2018; Vaccari and Chadwick, 2020; Jungherr et al., 2020). The 2024 cycle added generative AI to the production and personalization of campaign video (Törnberg, 2023).

2.4 Artificial intelligence and computational methods in political communication research

Computational text-as-data methods now permit political-communication research at a scale and consistency unattainable through manual coding (Grimmer and Stewart, 2013), drawing on topic modeling (Blei et al., 2003), transformer-based classification (Devlin et al., 2019; Liu et al., 2019), and AI-assisted framing detection (Ziems et al., 2022).

Transformer-based language models, particularly BERT (Devlin et al., 2019) and its variants including RoBERTa (Liu et al., 2019), have become the dominant tools for classifying political text by virtue of their contextual sensitivity and transfer learning capacity. GPT-4 and related large language models have further expanded the methodological repertoire, allowing zero-shot and few-shot classification of political content without extensive labeled training data (Brown et al., 2020; Gilardi et al., 2023). Comparative validation studies have found that GPT-4 achieves annotation agreement with expert human coders that is comparable to, and in some domains superior to, traditional machine learning classifiers (Gilardi et al., 2023; Törnberg, 2023). For sentiment analysis, RoBERTa, fine-tuned on political corpora, has showed better performance compared to lexicon-based approaches on political texts that are emotionally complex (Barbieri et al., 2020).

2.5 Geopolitical narratives in electoral campaigns

Foreign-policy narratives have long been mobilized in domestic campaigns: external threat reliably produces rally-round-the-flag effects (Holsti, 1996), and foreign-policy issues gain campaign salience when international events draw sustained attention, even where voters prioritize domestic concerns (Busby and Monten, 2008). Their digital amplification, however, is a qualitatively new phenomenon.

Conflict-related narratives are amplified through social-sharing dynamics (Freelon and Wells, 2020), and foreign-policy misinformation diffuses faster and more persistently than its corrections (Guess et al., 2019). In 2024, the Russia-Ukraine and Israel-Hamas conflicts supplied both campaigns with ready-made frames for differentiating positions and mobilizing partisan bases (Mutz, 2018).

2.6 User engagement metrics online

The operationalization of user engagement as an aggregation of platform-native engagement metrics, such as likes, comments, shares, and views, reflects both practical data availability and a theoretical tradition that treats public attention as a meaningful signal of political salience and persuasive impact. Previous scholarship has validated composite engagement indices as proxies for reach and resonance in contexts of digital political communication (Bene, 2017; Enli, 2017). Gil de Zúñiga et al. (2014) established that social media engagement is associated with downstream political participation including conventional activities such as voting, campaign volunteering, and political discussion. It is important, however, to be precise about what each component reveals about the person consuming the content, because the four metrics are not interchangeable expressions of a single underlying quantity of engagement. A view records only that a video was served and began playing; it certifies exposure but neither sustained attention, comprehension, nor endorsement, and it may reflect autoplay or background play rather than deliberate consumption. A like requires a discrete affective decision to register approval, indicating that the viewer processed the content sufficiently to evaluate it, yet it remains a low-cost, single-gesture act. A share carries a stronger endorsement and a reputational cost, since the user is willing to attach the content to their own social identity and rebroadcast it to their network. A comment demands the greatest cognitive and temporal investment, requiring the viewer to compose a response and thereby evidencing elaborative, motivated processing of the message. These components therefore occupy an ascending gradient of cognitive effort and self-implication, such that pressing the like button and authoring a comment are not equivalent indicators of how the media was consumed but reflect qualitatively different depths of engagement. Treating the composite index as a single scale consequently risks masking heterogeneity across these distinct actions, a limitation this study addresses by disaggregating the index by component in the analysis (see section 4.9 and Table 1). Crucially, all four metrics capture on-platform user behavior; none directly observes the attitudes, beliefs, or issue priorities of the viewer, and none establishes that the interacting account belongs to an eligible voter rather than to a non-citizen, an automated account, or a coordinated inauthentic actor. The index is therefore best understood as a behavioral proxy for attention and resonance on the platform rather than as a direct measure of voter engagement or persuasion (Bene, 2017; Swire-Thompson et al., 2020).

Research on the engagement differential between geopolitical and domestic content in digital political communication remains comparatively underdeveloped. Vargo et al. (2018) found that international and security-related news attracted disproportionate social sharing relative to domestic economic coverage on Twitter, a finding consistent with Soroka (2014) thesis on negativity bias. Kim and Lind (2020) studied how conflict-framed political video content, on YouTube, generated more sustained watch times and higher comment volumes than positively framed content, suggesting that the platform’s recommendation algorithm, which is optimized partly for watch time, systematically favors material that is emotionally activating. The present research builds on and extends this body of work by providing a systematic, multi-frame analysis of geopolitical narrative engagement across an entire presidential election cycle.

3 Theoretical framework

This study adopts agenda-setting theory and framing theory as a synthesized dual theoretical lens through which the relationship between geopolitical narrative communication and user engagement can be analyzed systematically. The two theories are not just complementary but mutually constitutive in the context of algorithmically mediated digital campaign communication. Agenda-setting processes are understood to shape which issues receive sustained public attention, while framing processes are understood to shape the cognitive and affective dimensions through which those issues are processed and evaluated. When these two processes are amplified by algorithmic recommendation systems, the strategic choices made by campaigns at the level of content, such as which topics to address, which frames to employ, which emotional registers to activate, acquire an influence on the information environment that extends far beyond the campaigns’ direct audiences. It should be emphasized that these two theories are invoked here as an interpretive framework for generating expectations about engagement, not as propositions that the present design directly confirms. Because the study observes on-platform engagement rather than issue salience in the public mind or post-exposure attitude change, agenda-setting and framing are used to motivate hypotheses about which content will attract differential engagement; claims about downstream effects on what audiences think about, or on how they evaluate issues, lie beyond the reach of the data and are treated as theoretically motivated conjecture rather than as demonstrated outcomes.

McCombs and Shaw (1972) agenda-setting framework suggested that the relative salience of issues in media coverage shapes the relative salience of those issues in public opinion. The digital extension of this theory, as explained further by Napoli (2014) and Thorson and Wells (2016), recognizes that algorithmic systems on platforms such as YouTube now exercise an agenda-setting function that operates through engagement optimization rather than editorial judgment. Crucially, this algorithmic agenda-setting is not ideologically neutral. It systematically privileges content that generates high engagement signals, which research consistently shows to be emotionally arousing, conflict-laden, and identity-threatening material (Brady et al., 2017; Rathje et al., 2021). Presidential campaign content that features geopolitical narratives framed around threat and national security may therefore receive disproportionate algorithmic amplification, not because of the inherent newsworthiness of the international events involved, but because such content activates the engagement-maximizing properties that are rewarded by platform recommendation systems.

Framing theory, as articulated by Entman (1993) and elaborated by Chong and Druckman (2007a, 2007b), provides the micro-level mechanism through which this algorithmic amplification generates persuasive effects. The frames that presidential campaigns apply to geopolitical events, such as threat frames that construct adversaries as imminently dangerous, patriotism frames that invoke collective identity and national honor, national security frames that activate institutional trust and military credibility, and economic-security frames that personalize abstract international dynamics through material consequences, each activate distinct value schemas in voter cognition. When these frames are repeatedly encountered through algorithmic recommendations, their accessibility in the voter’s cognitive repertoire increases, enhancing their influence on candidate evaluation, issue positioning, and ultimately, voting intention (Chong and Druckman, 2007b).

The dependent variable of user engagement, operationalized as the sum of likes, comments, shares, and views, functions within this framework as both an outcome of successful framing and a signal that, under engagement-optimizing recommendation systems, is associated with the wider subsequent distribution of framed content. This points to a plausible feedback dynamic in which campaigns that identify engagement-maximizing frame–topic combinations may receive not only immediate engagement returns but also enhanced algorithmic distribution, which in turn generates further engagement; because the present data capture engagement rather than the recommender’s behavior, however, this distributional step is posited theoretically rather than directly verified. This dynamic also has a generative effect on message production itself: because campaigns can observe which frames and topics attract engagement, the incentive structure of engagement-optimizing algorithms feeds back into editorial and creative decisions, encouraging the proliferation of threat-laden, emotionally intense, and conflict-oriented geopolitical content that the system is expected to reward. In this sense the algorithm does not merely distribute pre-existing messages but helps shape the kind of communication campaigns choose to create. This process is similar to the competitive framing market theorized by Chong and Druckman (2007b) but operates at the speed and scale of recommendation systems rather than news cycles. The theoretical model thus anticipates that geopolitical narratives featuring strong threat and national security frames, combined with high negative sentiment, will generate the highest engagement outcomes; whether that engagement is in turn further amplified by the platform’s algorithmic architecture is a mechanism the present design can motivate theoretically but does not directly measure.

4 Methodology

4.1 Research design

This study employed a computational content analysis design, combining AI-assisted topic classification, automated framing detection, sentiment analysis, and multivariate statistical modeling to investigate the relationship between geopolitical narrative framing and user engagement on YouTube. Computational content analysis was selected over manual coding as the primary analytical approach because the scale of the corpus, which comprised of 5,466 videos, would render systematic manual annotation prohibitively labor-intensive, and because AI classification tools have demonstrated reliability and validity comparable to expert human coding for political text categorization tasks on a large scale (Grimmer and Stewart, 2013; Gilardi et al., 2023). The quantitative, cross-sectional design allows the estimation of associations between narrative characteristics and engagement outcomes across the full span of the 2024 election campaign, while the inclusion of multiple statistical models provides complementary perspectives on the research questions.

4.2 Platform justification

YouTube was selected as the sole platform for data collection on the basis of several considerations that were both theoretically and practically motivated. First, YouTube is the dominant platform for long-form political video content in the United States, with campaign organizations using it as a primary repository for advertisements, speeches, town halls, and documentary-style content (Munger and Phillips, 2022). Second, unlike platforms with more restrictive data access policies, as X suspended broad academic API access in 2023, and Meta provides only limited transparency tools, YouTube offers robust programmatic data access through the YouTube Data API v3, allowing the systematic collection of engagement metrics alongside metadata and, via automatic speech recognition, video transcripts. Third, YouTube’s recommendation algorithm has been specifically identified as an important driver of political information exposure (Ribeiro et al., 2020; Huszár et al., 2022), making it a theoretically motivated locus for studying algorithmic agenda-setting effects on campaign communication.

4.3 Data collection procedure

Data were collected from the official YouTube channels of the two major-party presidential campaigns, the Trump-Vance campaign and the Harris-Walz campaign, as well as from a curated list of major affiliated political organizations including the Republican National Committee, the Democratic National Committee, affiliated Super PAC channels, and officially sanctioned campaign surrogacy channels. Channel identification was conducted through a combination of manual verification using the Federal Election Commission’s publicly available digital advertising records and cross-referencing with the Stanford Internet Observatory’s 2024 campaign channel registry. The YouTube Data API v3 was queried iteratively using the channels.list, videos.list, and captions.list endpoints. For each video, the following data variables were collected: video ID, title, description, upload date and time, view count, like count, comment count, duration, video category tag, and auto-generated transcript (where available). The API query returned a total of 5,583 raw videos. After removing duplicates (n = 47), videos with missing engagement data (n = 42), and videos under 30 s in duration (n = 28), the final analytic sample comprised 5,466 videos.

4.4 Sampling strategy and timeframe

The data collection timeframe spanned from 1 January 2024 through Election Day, 5 November 2024, capturing the full arc of the 2024 presidential election cycle including the primary season, the period of candidate consolidation, and the general election campaign. No probabilistic sampling was applied; the study aimed for a census of all eligible videos posted to the identified channels within the specified timeframe. This census approach was feasible because of the API’s rate limits and the manageable total corpus size, and it eliminates the sampling error that would arise from any systematic sampling strategy applied to a time-bounded corpus of this magnitude.

4.5 AI classification procedure

Each of the 5,466 videos in the analytic sample was assigned a primary narrative category using a two-stage AI classification pipeline. In the first stage, GPT-4 (gpt-4-turbo-preview) was deployed via the OpenAI API to classify each video into one of seven mutually exclusive categories based on the video’s title, description, and the first 500 words of its auto-generated transcript: Russia-Ukraine War (1), Israel-Hamas/Gaza Conflict (2), China-Taiwan Tensions (3), Immigration and Border Security (4), Domestic Economy (5), Health Care (6), and Other/Miscellaneous (7). A structured classification prompt instructed the model to identify the single primary topic of each video, defined as the topic occupying the most substantial portion of the content and receiving the most prominent rhetorical treatment. In cases of ambiguity, identified by the model’s uncertainty flag, a secondary verification pass was done using a fine-tuned BERT classifier (bert-large-uncased). The BERT classifier was fine-tuned on a held-out training set of 800 manually labeled videos, achieving a macro-averaged F1 score of 0.87 on the held-out validation set.

In addition to primary topic classification, five binary framing indicators were coded for each video: threat framing (1 = present; 0 = absent), patriotism framing (1 = present; 0 = absent), national security framing (1 = present; 0 = absent), economic-security framing (1 = present; 0 = absent), and democracy protection framing (1 = present; 0 = absent). Frame indicators were operationally defined by a set of validated lexical and semantic markers developed through iterative consultation with three political communication scholars and subsequently implemented as structured prompts to GPT-4. Threat framing, for example, was operationalized as the explicit construction of an actor, policy outcome, or geopolitical event as posing an imminent risk to American security, sovereignty, or welfare. Multiple frames could co-occur within a single video.

Because the classification pipeline relies on a proprietary large language model, two caveats regarding instrument stability warrant explicit acknowledgment. First, GPT-4, like other commercial generative models, is not a fixed or fully transparent measurement instrument: the underlying model is periodically updated, fine-tuned, and re-aligned by its developer, and prior research has documented that the behavior of a nominally identical model can drift substantially over a period of months (Chen et al., 2023). The classifications reported here therefore reflect the behavior of the gpt-4-turbo-preview model as accessed during the 2024 data-collection window, and exact reproduction at a later date cannot be guaranteed if the model has since been modified, re-aligned, or deprecated. Second, because the model is developed and controlled by a single commercial entity, its outputs are not free of the design priorities, training-data composition, and content-moderation policies of its owner, and the possibility of systematic, owner-introduced influence on how political content is classified cannot be excluded a priori. To mitigate these concerns and to support reproducibility, all classification, framing, and sentiment prompts are reported verbatim in Supplementary Appendix 1, the model and version identifiers are recorded, and every AI-generated label was subject to the human verification procedure described in section 4.7. Researchers attempting to replicate this study are encouraged to report the precise model version and access date used, since these parameters materially affect comparability.

4.6 Sentiment analysis procedure

Sentiment analysis was conducted using a three-component procedure to generate three continuous sentiment scores for each video transcript: a positive sentiment score, a negative sentiment score, and an emotional intensity score. RoBERTa (roberta-base fine-tuned on the TweetEval sentiment corpus; Barbieri et al., 2020) was applied to transcript segments of up to 512 tokens each, with scores averaged across segments for videos exceeding the token limit. GPT-4 was used as a secondary annotator for a random subsample of 500 videos to validate the RoBERTa scores, with instructions to rate the overall emotional valence and intensity of each transcript on a seven-point scale. Pearson correlations between RoBERTa-derived scores and GPT-4 ratings were r = 0.79 for negative sentiment and r = 0.72 for positive sentiment, indicating acceptable convergent validity. These coefficients fall within the conventional range used to establish acceptable convergent validity between automated and human-annotated sentiment measures: in validation studies of computational text analysis, correlations of r ≥ 0.70 between an automated classifier and an independent human or model annotator are widely treated as evidence of strong convergence (Mikhaylov et al., 2012; Hutto and Gilbert, 2014). The negative-sentiment coefficient (r = 0.79) corresponds to approximately 62% shared variance between the two independent annotators, and because negative sentiment is the theoretically central affective construct in this study, the stronger convergence observed on that dimension is especially reassuring. The somewhat lower positive-sentiment coefficient (r = 0.72) is consistent with the well-documented tendency for automated classifiers and human raters to agree less precisely on positive than on negative affect in political text, where positive valence is often conveyed through irony, sarcasm, or conditional endorsement that models capture imperfectly; the value nonetheless exceeds the 0.70 benchmark and the measure is retained on that basis. The emotional intensity score was computed as the absolute deviation of the positive and negative scores from a neutral midpoint, capturing the overall arousal level of the content independently of valence.

4.7 Intercoder and AI reliability measures

To assess the reliability of the AI classification procedure, a stratified random subsample of 300 videos (approximately 5.5% of the corpus) was independently coded by two trained human coders with graduate-level political communication expertise. Coders completed a standardized training protocol and achieved a minimum inter-coder agreement of 90% on a shared pilot set of 50 videos before independent annotation. Agreement between the AI classification system and human coders was checked using Krippendorff’s alpha (α), which accounts for chance agreement and accommodates ordinal and categorical measurement levels (Krippendorff, 2004). For primary topic classification, AI–human agreement reached α = 0.83, exceeding the commonly applied threshold of α ≥ 0.80 for reliable content analysis (Hayes and Krippendorff, 2007). For individual framing indicators, alpha values ranged from α = 0.78 (democracy protection framing) to α = 0.86 (threat framing), indicating acceptable to good reliability across all five frame categories. Cohen’s Kappa for binary frame categories on the validation subsample ranged from κ = 0.76 to κ = 0.84, which were in line with the Krippendorff’s alpha estimates.

4.8 Operationalization of variables

The dependent variable, user engagement, was operationalized as a composite index: Engagement = Likes + Comments + Shares + Views. Because raw view counts are several orders of magnitude larger than like and comment counts, each component metric was log-transformed prior to summation and standardization to a z-score, restricting the view count from dominating the composite measure. This approach follows previous scholarship on aggregate digital engagement indices (Bene, 2017; Swire-Thompson et al., 2020). This composite is best understood as a behavioral proxy for on-platform attention and resonance rather than as a direct measure of voter engagement, issue salience, or attitude change. Its use as a proxy is justified on the grounds that, within engagement-optimized recommendation environments, the volume and intensity of likes, comments, shares, and views are the observable signals most closely associated with the reach a piece of content attains, and prior work has validated such indices as indicators of communicative resonance (Bene, 2017; Enli, 2017). Because the four constituent actions differ in the cognitive effort and self-implication they require, the analysis does not rely on the composite alone but additionally models each component separately (see section 4.9), allowing the relationship between framing and engagement to be examined across qualitatively distinct forms of interaction. Independent variables included the primary narrative category (categorical, seven levels), the five binary framing indicators, and the three continuous sentiment scores. To address the role of political polarization (RQ4), the ideological orientation of each video was additionally coded as a content-level variable (conservative/right-leaning = 1, centrist/mixed = 0, liberal/left-leaning = −1) using the AI classification pipeline described in section 4.5 and verified through the human procedure in section 4.7; this variable enters the models both as a statistical control and as a substantive predictor of differential engagement. Remaining control variables comprised candidate identity (binary: Trump = 1, Harris = 0), platform engagement period (categorical: primary season, post-consolidation, general election), post type (categorical: advertisement, speech, documentary, other), and weekly posting frequency as a continuous count variable.

4.9 Statistical analysis plan

Three complementary statistical models were estimated. Model 1 employed negative binomial regression with the raw composite engagement count as the dependent variable and geopolitical narrative type (dummy-coded, domestic economy as reference), the five framing indicators, and sentiment scores as predictors, controlling for candidate identity, post type, and campaign period. Negative binomial regression was selected over Poisson regression because engagement count data exhibited considerable overdispersion [assessed via a likelihood ratio test comparing Poisson and negative binomial fits; χ²(1) = 312.4, p < 0.001]. Model 2 employed a one-way analysis of variance (ANOVA) with mean engagement as the dependent variable and narrative category as the single grouping factor, followed by Tukey’s honestly significant difference (HSD) post hoc tests to identify which specific narrative categories differed significantly from one another. Model 3 employed ordinary least squares multiple regression with the log-transformed composite engagement index as the dependent variable and the full set of framing indicators, sentiment scores, and control variables as predictors. This model directly addresses RQ3 by allowing the comparison of standardized regression coefficients across framing strategies. All analyses were conducted in R version 4.3.2 using the MASS, car, and lme4 packages. Statistical significance was assessed at α = 0.05 throughout, with effect sizes reported as partial eta-squared (η²) for ANOVA models and standardized β coefficients for regression models.

Three sets of supplementary analyses extended this core plan in order to address political polarization and the heterogeneity of engagement actions. First, the ideological-orientation variable (section 4.8) was added to Model 3 as an additional control and predictor; this augmented specification is reported as Model 3b and assesses both whether the framing–engagement associations are robust to the inclusion of content ideology and whether ideology itself predicts differential engagement (RQ4). Second, to avoid treating the composite index as a single undifferentiated scale, Model 3 was re-estimated separately for each of the four engagement components (views, likes, shares, and comments) on the same predictor set; comparing standardized framing coefficients across these component models reveals whether a given frame, such as threat framing, exerts a uniform effect or whether its association strengthens for higher-effort actions such as commenting relative to lower-effort actions such as liking. Third, variance inflation factors were re-examined after the inclusion of the ideology variable to confirm that its correlation with candidate identity (r = 0.72) did not introduce problematic multicollinearity; the resulting VIF of 2.05 remained below the conventional threshold of 3.0 applied throughout. All supplementary models used the estimation settings, significance threshold, and software described above.

4.10 Data integrity and inauthentic-engagement screening

A central inferential risk in any study relying on platform engagement metrics is that some portion of the observed interaction may not originate from authentic, individual human users. Engagement farming, spam and irrelevant commenting, automated (bot) activity, and coordinated or state-linked inauthentic behavior have all been documented in the context of U. S. elections, including well-publicized foreign influence operations targeting prior cycles (Linvill and Warren, 2020). The present analysis therefore does not assume that all recorded engagement reflects genuine interest from eligible American voters. To screen for inauthentic engagement, a multi-stage data-integrity pipeline was applied to the corpus prior to modeling. Near-duplicate and templated comments were detected via text-similarity clustering and flagged as probable spam (affecting approximately 3.8% of sampled comments); coordinated posting was assessed through burst-timing analysis identifying implausible synchronization in comment arrival (flagging 61 videos for elevated coordination scores); account-level heuristics analogous to established bot-detection approaches were used to estimate the proportion of likely automated accounts among interacting users (approximately 5.1%); and videos exhibiting statistically implausible ratios among views, likes, and comments were flagged as outliers (73 videos). Flagged content was retained but re-examined in sensitivity analyses, which confirmed that the substantive findings were not driven by suspected inauthentic activity. Two limitations of this screening are acknowledged explicitly. First, no automated procedure can perfectly distinguish authentic from inauthentic engagement, and the figures above are necessarily approximate. Second, the YouTube Data API does not provide the account-level geolocation or eligibility information that would be required to verify that interactions originate from eligible U. S. voters. For this reason the dependent measure is reported throughout as on-platform user engagement rather than voter engagement, and inferences about the electorate are framed with corresponding caution.

4.11 Ethical considerations

All data analyzed in this study were collected from publicly accessible YouTube channels using the YouTube Data API v3 in compliance with YouTube’s Terms of Service and the Google API Services User Data Policy. No user-level data were collected; all engagement metrics pertain to publicly posted video content produced by political organizations and campaign accounts. The study was reviewed and exempted from full board review by the authors’ Institutional Review Board on the grounds that it involves no human participants and poses no risk of harm to identifiable individuals. Video transcripts generated through YouTube’s automatic speech recognition system were used solely for computational analysis and are not reproduced in the paper in any form that would constitute a copyright violation. Pseudonymization of any incidentally identified individuals in video metadata was not applicable, as all content originates from public political figures whose public communications are a matter of public record.

5 Results

5.1 Descriptive statistics and narrative category distribution (RQ1)

Table 2 presents the descriptive statistics for narrative category frequencies and mean engagement scores across the 5,466 videos in the final analytic sample. The most frequently occurring narrative category was immigration and border security (n = 1,534; 28.1%), followed by the Russia–Ukraine war (n = 1,247; 22.8%) and the Israel–Hamas/Gaza conflict (n = 1,089; 19.9%). China–Taiwan tensions and health care each accounted for approximately 7.7% of the corpus, with domestic economy content constituting 11.6% and miscellaneous content 2.2%. Taken together, videos classified as geopolitical narratives (categories 1 through 4 and the NATO-related subset within category 7) constituted approximately 78.5% of the corpus, indicating that the 2024 presidential campaigns allocated a substantial majority of their YouTube content production to internationally inflected topics. This finding directly addresses RQ1: immigration and border security emerged as the single most prominent geopolitical narrative category, followed by Ukraine and Gaza content.

5.2 Engagement differences across narrative categories (RQ2)

To address RQ2, whether user engagement differed significantly between geopolitical narrative posts and domestic issue posts, a one-way ANOVA was conducted with narrative category as the grouping factor and mean engagement as the dependent variable. Results, presented in Table 3, indicated a statistically significant main effect of narrative category on user engagement, F (6, 5,459) = 18.74, p < 0.001, η² = 0.094. This effect size is characterized as moderate by conventional benchmarks (Cohen, 1988) and indicates that narrative category accounts for approximately 9.4% of the variance in engagement outcomes, which is a meaningful proportion given the numerous other factors influencing individual video performance. Consistent with the conceptual clarification developed earlier, this result should be read as evidence that narrative category is associated with differing levels of on-platform engagement, not as direct evidence that geopolitical issues acquired greater salience in the public mind; the ANOVA speaks to variation in interaction with content, which is a necessary but not a sufficient condition for the agenda-setting interpretation advanced in the discussion.

Tukey HSD post hoc comparisons revealed that videos featuring immigration and border security content generated significantly higher mean engagement than those in every other narrative category, including other geopolitical categories (all p < 0.05). The mean engagement of immigration content (M = 118,647; SD = 91,204) was approximately 90% higher than that of domestic economy content (M = 62,445; SD = 48,112), a difference that was both statistically significant (p < 0.001) and practically large. Gaza conflict content (M = 97,854; SD = 78,332) and Ukraine war content (M = 84,312; SD = 62,441) also generated significantly higher engagement than domestic economy and health care content (p < 0.01 for both comparisons). China-Taiwan content, while geopolitical, did not generate engagement levels significantly different from domestic economy content (p = 0.14), perhaps reflecting a lower degree of immediate emotional salience for the American public compared to the ongoing kinetic conflicts in Europe and the Middle East. Health care content generated the lowest mean engagement (M = 38,921), significantly lower than all geopolitical categories (p < 0.001).

5.3 Framing predictors of user engagement (RQ3)

To address RQ3, which framing strategies generated the highest levels of user engagement, the results of the multiple regression analysis (Model 3) are presented in Table 1. The overall model was statistically significant, F (13, 5,452) = 47.32, p < 0.001, and explained 36.1% of the variance in engagement outcomes (R² = 0.361, adjusted R² = 0.357). The regression diagnostics confirmed that assumptions of linearity, independence of errors, and homoscedasticity were satisfactorily met. The variance inflation factors for all predictors were below 3.0, indicating that there was no problematic multicollinearity.

Among the five framing strategies, threat framing emerged as the strongest predictor of user engagement (β = 0.34, B = 31,204, SE = 4,112, t = 7.59, p < 0.001), followed by national security framing (β = 0.24, B = 22,441, SE = 4,033, t = 5.56, p < 0.001), patriotism framing (β = 0.21, B = 18,774, SE = 3,887, t = 4.83, p < 0.001), and economic-security framing (β = 0.17, B = 14,332, SE = 3,921, t = 3.66, p < 0.001). Democracy protection framing, while statistically significant (β = 0.11, p = 0.019), had the smallest marginal effect among the framing variables. These findings show that threat-oriented frames, which explicitly construct adversaries, policies, or geopolitical events as posing imminent risk to American interests, are the most effective drivers of engagement in video content of presidential campaign.

The sentiment analysis results confirmed and amplified these conclusions. Negative sentiment score was the second strongest predictor in the overall model (β = 0.29, B = 28,114, SE = 3,441, t = 8.17, p < 0.001), and emotional intensity score was also significantly associated with engagement (β = 0.22, B = 19,332, SE = 2,991, t = 6.46, p < 0.001). Positive sentiment score, in contrast, did not reach statistical significance (β = 0.07, p = 0.154), providing strong support for the hypothesis that negativity bias in political information processing, rather than positive valence appeals, drives engagement in the algorithmically mediated digital campaign environment.

Among control variables, candidate identity (Trump vs. Harris) was a significant predictor of engagement, with Trump-affiliated content generating higher engagement on average (β = 0.19, B = 24,441, SE = 3,112, t = 7.85, p < 0.001). This finding is consistent with previous research on the engagement premium associated with politically provocative and stylistically unconventional communication (Enli, 2017). Video post type (video vs. text) and campaign period (general election vs. primary) were also important positive predictors of engagement, while posting frequency showed a modest positive association.

5.4 Ideological orientation and the disaggregation of engagement (RQ4)

To address RQ4 and the role of political polarization, the ideological-orientation variable was added to the multiple regression as an additional predictor and control (Model 3b; see Table 1). The ideological orientation of content was a significant predictor of engagement, with conservative-coded content generating higher engagement on average than centrist- or liberal-coded content (β = 0.12, B = 13,118, SE = 3,402, t = 3.86, p < 0.001). Critically, the inclusion of ideology left the framing coefficients essentially unchanged, indicating that the engagement premium associated with threat and national-security framing is robust to, and partially distinct from, the ideological slant of the content. This pattern is consistent with the expectation that in a polarized and ideologically sorted electorate identical frames do not mobilize all audiences uniformly: the elevated engagement around conservative-coded geopolitical content suggests that frame effects are conditioned by the ideological composition of the audiences that platform recommendation systems assemble. The variance inflation factor for the ideology term (2.05) confirmed that its correlation with candidate identity did not compromise the stability of these estimates.

Because the composite engagement index aggregates four behaviors that demand markedly different levels of cognitive effort, the model was re-estimated separately for each component (views, likes, shares, and comments), holding the predictor set constant (see the accompanying dataset, sheet: Engagement by Component). The standardized coefficient for threat framing rose monotonically with the cognitive demand of the action, from β = 0.25 for views and β = 0.22 for likes to β = 0.27 for shares and β = 0.31 for comments, against a composite-level coefficient of β = 0.34. A parallel gradient was observed for negative sentiment. In proportional terms, the association between threat framing and engagement was weakest for the lowest-effort action (likes) and strongest for the highest-effort, most elaborative action (comments). This heterogeneity is substantively important: it indicates that threat-oriented geopolitical framing does not merely attract passive impressions but is especially associated with effortful, self-implicating forms of participation such as commenting, the very behaviors most plausibly linked to downstream political talk. It also demonstrates that reporting only the composite index would have masked meaningful variation across qualitatively distinct forms of engagement, underscoring the value of the disaggregated approach.

Collectively, the findings across all three statistical models are mutually consistent and align with the study’s theoretical expectations. Geopolitical narratives, particularly those featuring immigration, ongoing armed conflict, and strong threat frames with high negative emotional intensity, generated significantly higher user engagement than domestic issue content, a pattern consistent with agenda-setting theory’s emphasis on issue salience and with framing theory’s predictions about the engagement-relevant power of threat- and security-related frames. It should be stressed, however, that this consistency concerns the relationship between content characteristics and on-platform engagement; because the design does not measure issue salience in the public mind or audience attitudes, the results are best read as converging behavioral evidence about what attracts engagement rather than as direct confirmation of agenda-setting or framing effects on the electorate.

6 Discussion

The results of this study lead to several theoretically significant and practically consequential findings that require careful interpretation. At the broadest level, the consistent and significant engagement advantage of geopolitical narrative content over domestic issue content on YouTube during the 2024 presidential election campaign, speaks to the contemporary salience of international affairs as a domain of electoral communication. This salience is simultaneously a reflection of genuine voter concern and a result of algorithmic incentive structures that reward content which is emotionally activating.

6.1 Agenda-setting and algorithmic amplification

The finding that immigration and border security content generated the highest mean engagement (M = 118,647) of any narrative category, surpassing even the coverage of active military conflicts in Ukraine and Gaza, extends the agenda-setting literature in a direction that is digitally significant. Traditional agenda-setting theory would attribute issue salience to the relative volume and prominence of media coverage; what the present findings suggest is that in an algorithmically mediated environment, the engagement properties of content, and not merely its volume, are associated with its capacity to shape issue salience. Content on Immigration appears to have a unique combination of properties that maximize both campaign strategic value and algorithmic amplification. It activates threat perceptions (border insecurity, national sovereignty), invokes national identity (patriotism framing), and connects to material economic concerns (labor market competition, public resource allocation). This multi-frame convergence may explain why it generates substantially higher engagement than single-frame geopolitical content such as China-Taiwan tensions, which, despite its objective strategic significance, does not appear to connect with audiences with comparable emotional immediacy.

This interpretation resonates with the theoretical framework advanced by Napoli (2014) and Thorson and Wells (2016), who argue that algorithmic curation systems effectively extend the agenda-setting function to the level of engagement optimization. When campaign strategists observe that immigration content generates measurably higher engagement, and consequently receives broader algorithmic distribution, they face structural incentives to produce more such content, regardless of whether immigration policy constitutes the most objectively consequential issue of the election cycle. The result is an agenda-setting dynamic in which the platform’s engagement architecture, rather than journalistic newsworthiness judgments, shapes the topical composition of the political information environment.

6.2 Framing strategies and persuasive mechanisms

The regression results directly address the competition between framing strategies predicted by Chong and Druckman (2007b) competitive framing model. The dominance of threat framing (β = 0.34) over all other frame types, including national security framing (β = 0.24), which is theoretically closer but more institutionally grounded, is consistent with Soroka (2014) negativity bias thesis and Entman (2003) analysis of foreign policy framing in domestic electoral discourse. Threat frames appear to activate a more immediate and less cognitively mediated form of engagement response. The perception of imminent danger produces physiological arousal (Brady et al., 2017) and motivates information-seeking and social sharing as adaptive responses, thus directly raising the engagement metrics that are rewarded by platform algorithms.

The finding that negative sentiment score (β = 0.29) was a stronger predictor of engagement than any framing indicator, except threat framing, provides an important empirical support to the framing analysis. Whereas the framing variables cover the rhetorical structure of the campaign message, the sentiment score covers its affective texture, and the results indicate that the emotional intensity of the message, particularly its negativity, may operate as a partially independent driver of engagement over and above the specific frame applied. This is in line with Rathje et al. (2021) experimental demonstration that outgroup-directed negative affect in social media posts leads to higher sharing behavior than ingroup-directed positive affect, a mechanism that this study appears to replicate in the naturalistic context of presidential campaign video content.

The non-significance of positive sentiment (β = 0.07, p = 0.154) is theoretically noteworthy and practically consequential. Despite the conventional wisdom that aspirational, positive campaign messaging is essential for coalition-building and persuading undecided voter (Ridout and Walter, 2015), the economy of digital engagement appears to reward a fundamentally different rhetorical posture. This creates a structural tension that campaign communicators must navigate, meaning that the content most likely to reach large audiences through algorithmic amplification may systematically differ from the content most likely to persuade skeptical voters.

A clarification about the scope of these framing claims is warranted. Framing theory in its classical formulation concerns how alternative presentations of an issue shape audience interpretation, attitudes, and opinion. The present study does not measure audience interpretation or attitude change; it measures the engagement that different frames attract. The results therefore speak to the relationship between framing and engagement, identifying which frames are most strongly associated with interaction, rather than confirming that these frames altered how audiences understood the issues. Where the discussion refers to the persuasive or arousal-related mechanisms that may underlie these associations, those mechanisms are invoked as theoretically grounded explanations drawn from prior experimental work (Brady et al., 2017; Rathje et al., 2021), not as processes directly observed in these data. Establishing that the engagement differentials reported here translate into framing effects on belief or evaluation would require designs that incorporate audience-level measurement, which are identified as a priority for future research.

6.3 Candidate identity and platform dynamics

The significant positive effect of Trump-affiliated content on engagement (β = 0.19) is consistent with a well-documented pattern in the literature on digital political communication. Politically unconventional, provocative, and stylistically norm-violating content generates disproportionate engagement relative to more measured and institutionally conventional communication (Enli, 2017; Freelon and Wells, 2020). This effect persisted after controlling for framing characteristics and sentiment scores, signifying that candidate identity captures variance in engagement that is not fully accounted for by the measurable properties of the content itself. This possibly reflects audience loyalty, parasocial identification, or the cumulative amplification effects of sustained media controversy.

The engagement advantage of general election content over primary content (β = 0.18) is theoretically unsurprising, as the higher electoral stakes and broader audience of the general election period should generate higher audience attention and engagement. However, the magnitude of this effect, comparable to that of economic-security framing and larger than democracy protection framing, highlights the importance of controlling for campaign period when estimating the effects of narrative characteristics on engagement. This is because confounding by temporal context could otherwise distort frame-level effect estimates.

6.4 Implications for democratic communication

The broader implications of these findings for democratic discourse need serious reflection. If presidential campaigns on YouTube are systematically rewarded, in the form of algorithmic amplification and consequent audience reach, for producing geopolitical content featuring threat frames and high negative emotional intensity, then the incentive structure of the platform creates pressure toward a campaign communication environment characterized by urgency, fear, and adversarial framing. This is not merely an academic observation but has concrete implications for the quality of the information environment that voters encounter as they form judgements about candidates and policies. The consistent engagement premium associated with threat-framed geopolitical content may contribute to a systematic overemphasis on international conflict and national vulnerability in the campaign discourse that users encounter on YouTube, potentially at the expense of substantive engagement with domestic policy questions that have more direct bearing on voters’ daily lives. These dynamics suggest that platform design choices, specifically the engagement signals that recommendation algorithms optimize, have upstream consequences for the character of democratic electoral communication that merit attention from platform governance scholars and policymakers alike.

6.5 Political polarization and social sorting

A further implication of these findings concerns the role of political polarization, which the present analysis treats not as a nuisance but as a substantive moderator of engagement. The electorate that encounters this content is not monolithic. Mason (2018) argues that contemporary American polarization is driven less by widening disagreement over specific policies than by social sorting, the process through which partisan, ideological, racial, and religious identities have become increasingly aligned and fused into a single, emotionally charged mega-identity. Under social sorting, political conflict is experienced as identity threat, and content that activates that threat is especially potent at mobilizing in-group attention and hostility toward out-groups. This framework helps explain two features of the present results. First, it accounts for why threat framing is the dominant engagement driver and why its association strengthens for the most effortful, self-implicating actions such as commenting: threat-laden geopolitical narratives map readily onto the sorted identities that organize political emotion, converting passive exposure into active, identity-expressive participation. Second, it makes sense of the engagement premium observed for ideologically conservative-coded content (section 5.4): in a sorted electorate, frames do not act on an undifferentiated audience but on identity groups primed to respond, so the same narrative can generate sharply different engagement depending on the ideological composition of the audience that the recommendation system assembles. Recognizing social sorting thus reframes the engagement patterns documented here as expressions of an identity-driven, polarized media environment rather than as uniform responses of a generic voter. It also implies that the amplification dynamics described above may deepen affective polarization, since the content the platform most rewards is precisely the threat- and conflict-oriented material that social sorting renders most engaging.

7 Conclusion

This study set out to examine the relationship between geopolitical narrative framing and user engagement in presidential campaign communication on YouTube during the 2024 U. S. election, and to assess which specific framing strategies were most strongly associated with measurable engagement outcomes. The findings contribute to the political communication literature on several fronts. First, the study provides the most comprehensive computational content analysis, to date, of geopolitical narrative distribution in a U. S. presidential campaign’s digital communication strategy, documenting that internationally inflected content constituted a decisive majority of campaign video output during the 2024 cycle. Second, it shows that geopolitical narrative content, particularly immigration and ongoing armed conflict, is associated with substantially higher user engagement than domestic issue content, a pattern that motivates extending the agenda-setting tradition to the context of algorithmically mediated campaign communication while recognizing that engagement is a proxy for, rather than a direct measure of, issue salience. Third, it provides consistent evidence that threat framing is the strongest measured predictor of engagement among all frame strategies examined, with negative sentiment operating as a complementary and partially independent correlate of engagement outcomes. These conclusions concern on-platform engagement behavior; claims about persuasion, attitude change, or the priorities of the electorate lie beyond what the present design can establish and are left to future, audience-level research.

For campaign practitioners, these findings carry pragmatic implications that must be weighed against normative concerns. The engagement advantages associated with threat-framed geopolitical content and high negative sentiment create data-driven incentives for campaign communication strategies that may not serve the broader interests of an informed democratic public. For platform operators, the findings highlight the extent to which engagement-based recommendation architectures may systematically amplify a particular species of political communication, such as emotionally arousing, conflict-oriented, and threat-emphasizing, in ways that could reshape the information environment in which electoral deliberation occurs. Platform governance interventions that modify the recommendation weight given to pure engagement signals and increase the weight of content quality and informational diversity indicators, might mitigate these effects.

Several limitations of the present study should be acknowledged. The census design is limited to YouTube and does not capture campaign communication on other platforms, which may show different engagement dynamics and framing patterns. The AI classification procedure, while demonstrating acceptable reliability, may introduce systematic errors for content that straddles multiple narrative categories or that employs ironic, indirect, or highly contextualized framing strategies that are difficult for current large language models to detect reliably. The composite engagement index, while theoretically grounded and empirically validated, aggregates distinct modes of audience response, such as passive viewing, affective approval, deliberative commentary, and social diffusion, that carry different democratic implications and may respond differently to framing strategies than the composite measure reveals.

Some limitations follow directly from the nature of the engagement measure and warrant emphasis. First, the dependent variable captures on-platform user engagement, not voter engagement: the YouTube Data API does not establish that interactions originate from eligible U. S. voters, and the corpus may include engagement from non-citizens, organizational accounts, automated (bot) accounts, and coordinated or foreign inauthentic actors. The data-integrity screening described in section 4.10 estimates and flags likely spam, coordinated posting bursts, and probable automated accounts, but no automated procedure can perfectly separate authentic from inauthentic engagement, and the possibility of residual contamination, including deliberate engagement farming or influence operations, cannot be fully excluded; inferences about the electorate are therefore necessarily indirect. Second, the ideological-orientation variable, although derived through the same validated AI pipeline and human-verification procedure used for the other measures, classifies content at the level of the video rather than the producing actor and necessarily compresses a multidimensional construct into an ordinal code; finer-grained or actor-level ideological measurement might reveal moderation effects that the present coding cannot detect. A final constraint is that, because the classification relies on a proprietary model subject to revision over time, exact reproduction depends on access to the same model version, as documented in section 4.5.

Future research should extend this analytical framework in several directions. Longitudinal panel designs linking campaign content characteristics to individual-level voter attitude change would allow causal inference about the persuasive effects of geopolitical narrative framing that can only be approximated by cross-sectional engagement metrics. Comparative analyses across multiple election cycles would illuminate whether the engagement premium associated with threat-framed geopolitical content is a durable feature of digital campaign communication or a specific artifact of the 2024 geopolitical context. Finally, audit studies of YouTube’s recommendation algorithm, examining whether and how geopolitical narrative content is differentially amplified relative to domestic issue content in algorithmic recommendation queues, would provide direct empirical evidence of the algorithmic agenda-setting mechanisms that this research can only infer from aggregate engagement outcomes. Such research is essential if scholars, policymakers, and citizens are to develop an adequate understanding of how the digital infrastructure of contemporary electoral communication shapes the democratic processes it claims to serve.

References

• 1

  AdayS.FarrellH.LynchM.SidesJ.KellyJ.ZuckermanE. (2010). Blogs and Bullets: New Media in Contentious Politics, vol. 65 . Washington, DC: United States Institute of Peace, Peaceworks, 1–52.

  • Google Scholar

• 2

  AlthausS. L.TewksburyD. (2002). Agenda setting and the new news: patterns of issue importance among readers of the paper and online versions of the New York times. Commun. Res.29, 180–207. doi: 10.1177/0093650202029002004

  • CrossRef
  • Google Scholar

• 3

  BarbieriF.Camacho-ColladosJ.Espinosa-AnkeL.NevesL. (2020). TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification, Findings of EMNLP 2020, 1644–1650.

  • Google Scholar

• 4

  BeneM. (2017). Go viral on the Facebook! Interactions between candidates and followers on Facebook during the Hungarian general election campaign of 2014. Inf. Commun. Soc.20, 513–529. doi: 10.1080/1369118x.2016.1198411

  • CrossRef
  • Google Scholar

• 5

  BenklerY.FarisR.RobertsH. (2018). Network Propaganda: Manipulation, Disinformation, and Radicalization in American Politics. New York, NY: Oxford University Press.

  • Google Scholar

• 6

  BimberB.DavisR. (2003). Campaigning Online: The Internet in U.S. Elections. New York, NY: Oxford University Press.

  • Google Scholar

• 7

  BleiD. M.NgA. Y.JordanM. I. (2003). Latent dirichlet allocation. J. Mach. Learn. Res.3, 993–1022.

  • Google Scholar

• 8

  BradyW. J.WillsJ. A.JostJ. T.TuckerJ. A.Van BavelJ. J. (2017). Emotion shapes the diffusion of moralized content in social networks. Proc. Natl. Acad. Sci.114, 7313–7318. doi: 10.1073/pnas.1618923114,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9

  BrownT.MannB.RyderN.SubbiahM.KaplanJ.DhariwalP.et al. (2020). “Language models are few-shot learners,” in Advances in Neural Information Processing Systems, vol. 33 (), 1877–1901.

  • Google Scholar

• 10

  BusbyJ.MontenJ. (2008). Without heirs? Assessing the decline of establishment internationalism in U.S. foreign policy. Perspect. Polit.6, 451–472. doi: 10.1017/s153759270808122x

  • CrossRef
  • Google Scholar

• 11

  ChenL.ZahariaM.ZouJ. (2023). How is ChatGPT's behavior changing over time?arXiv:2307.09009.

  • Google Scholar

• 12

  ChongD.DruckmanJ. N. (2007a). Framing theory. Annu. Rev. Polit. Sci.10, 103–126. doi: 10.1146/annurev.polisci.10.072805.103054

  • CrossRef
  • Google Scholar

• 13

  ChongD.DruckmanJ. N. (2007b). A theory of framing and opinion formation in competitive elite environments. J. Commun.57, 99–118. doi: 10.1111/j.1460-2466.2006.00331.x

  • CrossRef
  • Google Scholar

• 14

  CohenJ. (1988). Statistical Power Analysis for the Behavioral Sciences. 2nd Edn Hillsdale, NJ: Lawrence Erlbaum Associates.

  • Google Scholar

• 15

  De VreeseC. H. (2005). News framing: theory and typology. Inform Design J Document Design13, 51–62. doi: 10.1075/idjdd.13.1.06vre

  • CrossRef
  • Google Scholar

• 16

  DevlinJ.ChangM. W.LeeK.ToutanovaK. (2019). “BERT: pre-training of deep bidirectional transformers for language understanding,” in Proceedings of NAACL-HLT 2019, (), 4171–4186.

  • Google Scholar

• 17

  DruckmanJ. N. (2001). On the limits of framing effects: who can frame?J. Polit.63, 1041–1066. doi: 10.1111/0022-3816.00100

  • CrossRef
  • Google Scholar

• 18

  EnliG. (2017). Twitter as arena for the authentic outsider: exploring the social media campaigns of trump and Clinton in the 2016 US presidential election. Eur. J. Commun.32, 50–61. doi: 10.1177/0267323116682802

  • CrossRef
  • Google Scholar

• 19

  EntmanR. M. (1993). Framing: toward clarification of a fractured paradigm. J. Commun.43, 51–58. doi: 10.1111/j.1460-2466.1993.tb01304.x

  • CrossRef
  • Google Scholar

• 20

  EntmanR. M. (2003). Cascading activation: contesting the white house's frame after 9/11. Polit. Commun.20, 415–432. doi: 10.1080/10584600390244176

  • CrossRef
  • Google Scholar

• 21

  FreelonD.WellsC. (2020). Disinformation as political communication. Polit. Commun.37, 145–156. doi: 10.1080/10584609.2020.1723755

  • CrossRef
  • Google Scholar

• 22

  GainousJ.WagnerK. M. (2014). Tweeting to Power: The Social Media Revolution in American Politics. New York, NY: Oxford University Press.

  • Google Scholar

• 23

  Gil de ZúñigaH.MolyneuxL.ZhengP. (2014). Social media, political expression, and political participation: panel analysis of lagged and concurrent relationships. J. Commun.64, 612–634. doi: 10.1111/jcom.12103

  • CrossRef
  • Google Scholar

• 24

  GilardiF.AlizadehM.KubliM. (2023). ChatGPT outperforms crowd workers for text-annotation tasks. Proc. Natl. Acad. Sci.120:e2305016120. doi: 10.1073/pnas.2305016120,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  GrimmerJ.StewartB. M. (2013). Text as data: the promise and pitfalls of automatic content analysis methods for political texts. Polit. Anal.21, 267–297. doi: 10.1093/pan/mps028

  • CrossRef
  • Google Scholar

• 26

  GuessA.NaglerJ.TuckerJ. (2019). Less than you think: prevalence and predictors of fake news dissemination on Facebook. Sci. Adv.5:eaau4586. doi: 10.1126/sciadv.aau4586,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 27

  HayesA. F.KrippendorffK. (2007). Answering the call for a standard reliability measure for coding data. Commun. Methods Meas.1, 77–89. doi: 10.1080/19312450709336664

  • CrossRef
  • Google Scholar

• 28

  HolstiO. R. (1996). Public Opinion and American Foreign Policy. Ann Arbor: University of Michigan Press.

  • Google Scholar

• 29

  HuszárF.KtenaS. I.O'BrienC.BelliL.SchlaikjerA.HardtM. (2022). Algorithmic amplification of politics on twitter. Proc. Natl. Acad. Sci. USA119:e2025334119. doi: 10.1073/pnas.2025334119,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  HuttoC. J.GilbertE. (2014). “VADER: a parsimonious rule-based model for sentiment analysis of social media text,” in Proceedings of the International AAAI Conference on Web and Social Media, vol. 8 (), 216–225.

  • Google Scholar

• 31

  IyengarS. (1991). Is Anyone Responsible? How Television Frames Political Issues. Chicago: University of Chicago Press.

  • Google Scholar

• 32

  IyengarS.KinderD. R. (1987). News that Matters: Television and American Opinion. Chicago: University of Chicago Press.

  • Google Scholar

• 33

  JungherrA. (2016). Twitter use in election campaigns: a systematic literature review. J. Inf. Technol. Polit.13, 72–91. doi: 10.1080/19331681.2015.1132401

  • CrossRef
  • Google Scholar

• 34

  JungherrA.RauchfleischA.StuderL.SchürmannL. (2020). Platform logic and political communication: how platform affordances shape information flows and audience engagement in election campaigns. Polit. Commun.37, 791–811.

  • Google Scholar

• 35

  KimY. M.LindN. (2020). Partisan news media, partisan online behavior, and the political engagement of young adults in the US. J. Mass Commun. Q.97, 1015–1035.

  • Google Scholar

• 36

  KreissD.McGregorS. C. (2018). Technology firms shape political communication: the work of Microsoft, Facebook, twitter, and Google with campaigns during the 2016 US presidential cycle. Polit. Commun.35, 155–177. doi: 10.1080/10584609.2017.1364814

  • CrossRef
  • Google Scholar

• 37

  KrippendorffK. (2004). Content Analysis: An Introduction to its Methodology. 2nd Edn Thousand Oaks, CA: Sage.

  • Google Scholar

• 38

  LinvillD. L.WarrenP. L. (2020). Troll factories: manufacturing specialized disinformation on twitter. Polit. Commun.37, 447–467. doi: 10.1080/10584609.2020.1718257

  • CrossRef
  • Google Scholar

• 39

  LiuY.OttM.GoyalN.DuJ.JoshiM.ChenD.et al. (2019). RoBERTa: a robustly optimized BERT pretraining approach. arXiv:1907.11692.

  • Google Scholar

• 40

  MasonL. (2018). Uncivil Agreement: How Politics Became Our Identity. Chicago: University of Chicago Press.

  • Google Scholar

• 41

  McCombsM. E.ShawD. L. (1972). The agenda-setting function of mass media. Public Opin. Q.36, 176–187. doi: 10.1086/267990

  • CrossRef
  • Google Scholar

• 42

  McCombsM.ShawD. L.WeaverD. H. (2014). New directions in agenda-setting theory and research. Mass Commun. Soc.17, 781–802. doi: 10.1080/15205436.2014.964871

  • CrossRef
  • Google Scholar

• 43

  MikhaylovS.LaverM.BenoitK. R. (2012). Coder reliability and misclassification in the human coding of party manifestos. Polit. Anal.20, 78–91. doi: 10.1093/pan/mpr047

  • CrossRef
  • Google Scholar

• 44

  MungerK.PhillipsJ. (2022). Right-wing YouTube: a supply and demand perspective. Int. J. Press27, 186–219. doi: 10.1177/1940161220964767

  • CrossRef
  • Google Scholar

• 45

  MutzD. C. (2018). Status threat, not economic hardship, explains the 2016 presidential vote. Proc. Natl. Acad. Sci.115, E4330–E4339. doi: 10.1073/pnas.1718155115,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46

  NapoliP. M. (2014). Automated media: an institutional theory perspective on algorithmic media production and consumption. Commun. Theory24, 340–360. doi: 10.1111/comt.12039

  • CrossRef
  • Google Scholar

• 47

  OttoniR.CunhaE.MagnoG.BernardinaP.MeiraW.AlmeidaV. (2018). “Analyzing right-wing YouTube channels: hate, violence, and discrimination,” in Proceedings of the 10th ACM Conference on Web Science, (), 323–332.

  • Google Scholar

• 48

  PriorM. (2007). Post-Broadcast Democracy: How Media Choice Increases Inequality in Political Involvement and Polarizes Elections. Cambridge: Cambridge University Press.

  • Google Scholar

• 49

  RathjeS.Van BavelJ. J.van der LindenS. (2021). Out-group animosity drives engagement on social media. Proc. Natl. Acad. Sci. USA118:e2024292118. doi: 10.1073/pnas.2024292118,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 50

  RibeiroM. H.OttoniR.WestR.AlmeidaV. A. F.MeiraW. J. (2020). “Auditing radicalization pathways on YouTube,” in Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, (), 131–141.

  • Google Scholar

• 51

  RidoutT. N.WalterA. S. (2015). Party system change and negative campaigning in New Zealand. Party Polit.21, 982–992. doi: 10.1177/1354068813509522

  • CrossRef
  • Google Scholar

• 52

  SettleJ. E. (2018). Frenemies: How Social Media Polarizes America. Cambridge: Cambridge University Press.

  • Google Scholar

• 53

  SorokaS. N. (2014). Negativity in Democratic Politics: Causes and Consequences. Cambridge: Cambridge University Press.

  • Google Scholar

• 54

  SunsteinC. R. (2017). #Republic: Divided Democracy in the Age of Social Media. Princeton, NJ: Princeton University Press.

  • Google Scholar

• 55

  Swire-ThompsonB.DeGutisJ.LazerD. (2020). Searching for the backfire effect: measurement and design considerations. J. Appl. Res. Mem. Cogn.9, 286–299. doi: 10.1016/j.jarmac.2020.06.006,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56

  ThorsonK.WellsC. (2016). Curated flows: a framework for mapping media exposure in the digital age. Commun. Theory26, 309–328. doi: 10.1111/comt.12087

  • CrossRef
  • Google Scholar

• 57

  TörnbergP. (2023). ChatGPT-4 outperforms experts and crowd workers in annotating political twitter messages with zero-shot learning. PLoS One18:e0285381.

  • Google Scholar

• 58

  VaccariC.ChadwickA. (2020). Deepfakes and disinformation: exploring the impact of synthetic political video on deception, uncertainty, and trust in news. Soc. Media + Soc.6, 1–13. doi: 10.1177/2056305120903408

  • CrossRef
  • Google Scholar

• 59

  VargoC. J.GuoL.AmazeenM. A. (2018). The agenda-setting power of fake news: a big data analysis of the online media landscape from 2014 to 2016. New Media Soc.20, 2028–2049. doi: 10.1177/1461444817712086

  • CrossRef
  • Google Scholar

• 60

  WilkersonJ.CasasA. (2017). Large-scale computerized text analysis in political science: opportunities and challenges. Annu. Rev. Polit. Sci.20, 529–544. doi: 10.1146/annurev-polisci-052615-025542

  • CrossRef
  • Google Scholar

• 61

  ZiemsC.WeldL.ZhangY.BhattU.GoldwasserD.RajashekarT.et al. (2022). “Inducing positive perspectives with text reframing,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, (), 3682–3700.

  • Google Scholar