Paid Voices vs. Public Feeds: Interpretable Cross-Platform Theme-Based Analysis of Climate Discourse

Technical Summary: Paid Voices vs. Public Feeds

Problem Statement
Climate discourse online shapes public understanding and policy but unfolds across structurally distinct environments. Paid advertising platforms (e.g., Meta) host strategic, institutionally produced messaging designed for targeted persuasion, while public social media (e.g., Bluesky) reflects organic, user-driven discussion. Prior computational work has largely examined these environments in isolation, focusing on tasks like topic modeling or stance detection within single platforms. This leaves a gap in understanding whether thematic patterns are shared across communication environments or specific to the medium. Furthermore, existing methods often rely on predefined topic inventories or seed sets, which may not capture the emergent, platform-specific nuances of climate discourse.

Methodology
The authors present a comparative analysis of climate discourse across paid advertisements on Meta and public posts on Bluesky from July 2024 to September 2025. The study utilizes a dataset of 17,026 deduplicated Meta ads and 19,182 deduplicated Bluesky posts, collected via curated climate-related keywords.

The core contribution is an interpretable thematic discovery pipeline that operates without predefined topic inventories or seed sets. The methodology proceeds as follows:

1. Embedding and Clustering: Texts are embedded using SentenceBERT (all-MiniLM-L6-v2) with L2 normalization, followed by dimensionality reduction via PCA and UMAP. Semantically similar texts are grouped using HDBSCAN.
2. Coherence Filtering: Initial clusters are filtered for semantic coherence using an LLM (Mistral-Large-Instruct-2407), which judges whether representative texts within a cluster reflect a single underlying theme.
3. Summarization and Labeling: For retained clusters, the top-k representative texts are used to generate concise summaries via the LLM. These summaries are then merged based on embedding similarity and assigned human-interpretable theme labels (1–3 words).
4. Assignment: Texts are assigned to themes via two strategies: (i) matching text to the most similar cluster summary, or (ii) directly assigning text to the generated theme label.

The framework was evaluated against standard baselines (LDA, BERTopic, TopicGPT) using both human annotators and an LLM judge. Additionally, the utility of the discovered themes was validated through downstream tasks: stance prediction and theme-guided retrieval.

Key Results
The analysis reveals systematic divergences between paid and organic climate discourse across thematic structure, rhetorical register, and temporal responsiveness:

• Thematic Structure: Paid advertising (Meta) concentrates on the strategic promotion of specific solutions (e.g., "Solar incentive urgency," "Offshore wind momentum") and diverse promotional topics. In contrast, organic discourse (Bluesky) is heavily concentrated on systemic critique of corporate and political actors (e.g., "Corporate climate sabotage," "Greenwashing exposed"). A single critique-oriented theme on Bluesky accounted for more text volume than the top four Meta-dominant themes combined.
• Rhetorical Register: Meta ads exhibit a formal, authoritative, and solution-oriented register, scoring higher on "Analytic" and "Clout" (LIWC) and emphasizing individual actors and future framing. Bluesky posts are more personal, crisis-oriented, and scientifically grounded, scoring higher on "Authentic" and "Crisis" framing, with a focus on institutional actors and systemic critique.
• Stance Alignment: The induced themes demonstrate coherent stance alignment. Pro-Climate themes align with Pro-Climate stances, and Pro-Energy themes with Pro-Energy stances. The authors' methods produced more concentrated correlations than baseline topic models (LDA, BERTopic), which yielded diffuse correlations.
• Temporal Responsiveness: The platforms respond to major events in opposite directions. Following the November 2024 U.S. election, paid climate advertising contracted significantly across all major themes, while organic discourse expanded, with critique-oriented themes surging. Similarly, during the January 2025 California wildfires, organic discourse expanded with increased causal attribution and negative tone, whereas advertising patterns shifted differently.

Significance and Contributions
The paper makes three primary contributions:

1. Dataset: It constructs and releases the first multi-platform dataset enabling direct comparison between paid climate advertising and public climate discourse.
2. Framework: It develops an interpretable thematic discovery pipeline that generalizes to cross-platform analysis without requiring predefined topic inventories or seed sets. The induced themes are shown to be more coherent and stance-consistent than standard topic-modeling baselines.
3. Empirical Findings: It provides a descriptive characterization of how climate discourse differs across heterogeneous communication environments. The findings illustrate that paid and public discourse are not merely variations of the same conversation but reflect fundamentally different communicative orientations: strategic persuasion versus systemic critique.

The authors note that while the framework generalizes to other domains, the analysis is descriptive rather than causal. They acknowledge limitations regarding the generalizability of findings to other platforms due to differences in user demographics and moderation policies, and the potential for biases inherent in the pre-trained LLMs used for theme labeling. The work does not claim to infer speaker intent or causal effects of platform incentives but rather offers a robust method for comparative thematic analysis.