Do fossil fuel firms reframe online climate and sustainability communication? A data-driven analysis

The study addresses how fossil fuel firms may use subtle online misinformation strategies—such as reframing and redirection akin to greenwashing—to influence climate and sustainability discourse among key watchdogs: NGOs and IGOs. Given the vast reach of social media and the prevalence of climate misinformation, the paper investigates whether and how industry messaging affects the broader online information ecosystem. It formulates three hypotheses: H1 posits stakeholders have greater ability to direct conversation in areas of their domain expertise (e.g., NGOs on climate action, IGOs on policy, and industry on corporate sustainability). H2 posits fossil fuel firms will be more successful than NGOs/IGOs at redirecting online conversation. H3 posits fossil fuel firms’ online communication behavior is influenced by exogenous factors like extreme weather events and stock market performance. The research empirically examines interactions among eight major fossil fuel firms, 14 NGOs, and eight IGOs on Twitter from 2014–2021 using topic-sentiment modeling and time-series methods.

Prior work documents fossil fuel firms’ social media use for greenwashing, framing, and creating echo chambers or filter bubbles that reinforce firm-driven narratives. During COP-26, 16 major fossil fuel companies were linked to over 1700 climate misinformation ads on Facebook with 150 million interactions, highlighting virality. Studies show firms often emphasize consumer responsibility for climate change and publicly pledge decarbonization without matching investment actions, evidencing greenwashing and reframing. Social media also shapes climate activism but can amplify extremism and misinformation, fueling polarization and organized campaigns. Research has focused largely on individual-level echo chambers rather than collective behavior across stakeholder groups. There is also evidence that Twitter sentiment can be associated with financial markets, and climate anomalies can influence tweeting rates, motivating controls for exogenous stock returns and extreme weather in analyses of stakeholder interaction. The paper builds on this literature by modeling system-level interactions among industry, NGOs, and IGOs, and by testing whether external events rather than strategic reframing drive observed communication patterns.

Data and scope: The authors used Twitter (X) v2 API to collect daily time-series tweets from January 2014 to September 2021 for 8 investor-owned, high-emitting fossil fuel firms (BP, Chevron, BHP, ConocoPhillips, ExxonMobil, Peabody Energy, Shell, TotalEnergies), 8 IGOs (e.g., IPCC, UNFCCC, WMO, UNEP), and 14 NGOs (e.g., Greenpeace, WWF, NRDC, 350.org). Accounts were required to have ≥10,000 followers and a 7-year tweet history. The corpus reported includes up to 728,967 tweets; analyses commonly reference n=668,826; after preprocessing there were 330,412 tweets with 1009 unique tokens. The cumulative follower base was ~9.6 million. Preprocessing: Standard text-as-data procedures included tokenization, stemming, lemmatization, stopword removal, and n-gram extraction. Tweets with fewer than three unique words were dropped; tokens appearing in <0.005% or >50% of tweets were removed. Sentiment was initially estimated using sentimentr (embedding valence shifters). Joint Sentiment-Topic Modeling (JST): JST was applied to uncover 30 topics each conditioned on 3 sentiment categories (positive, neutral, negative), yielding 180 sentiment-topic combinations plus overall sentiment probabilities per tweet. Hyperparameters (α, β, γ) were set in a weakly supervised manner with a small prior, assuming tweets focus on few topics. Topic number (30) was chosen via coherence optimization. Outputs include per-tweet probabilities over sentiment-topics, used to characterize each group’s topical propensities and sentiment distributions. Vector Autoregression (VAR) and Impulse Response Functions (IRFs): For each identified topic, the authors estimated a VAR over daily average probabilities that each stakeholder group (industry, IGO, NGO) discussed that topic. Fossil fuel firms’ average daily stock returns (from CRSP) were treated as an endogenous variable; extreme weather events (EM-DAT: droughts, extreme temperatures, storms, wildfires causing >$1M damage) were included as exogenous controls. Given bounded [0,1] topic probabilities, a logit specification and stationarity checks (ADF tests) were used. IRFs were computed to estimate hypothetical changes in one group’s propensity to discuss a topic in response to a one-standard-deviation increase in another group’s propensity, with bootstrapped 95% CIs. This framework tests who leads/follows, domain expertise effects, and the influence of exogenous shocks on messaging dynamics.

• Stakeholder interactions and domain expertise (H1 supported): IRFs show these groups respond to one another online, particularly within their domain areas. For industry impulses (Fig. 2a), about 30% of IGOs’ IRFs and about 13% of NGOs’ IRFs are statistically significant. A one-standard-deviation increase in industry discussion of supporting STEM and corporate sustainability is associated with ≈5 percentage-point increases in NGOs’ discussion of those topics (95% CI). Industry mentions of stopping XL pipeline, opposing Trump drilling policies, divesting from fossil fuels, and criticizing the Trump EPA are linked to ≈1 pp NGO increases (95% CI). IGOs respond to industry by ≈3 pp increases for corporate sustainability and climatological changes. For NGO impulses (Fig. 2b), ≈57% of IGOs’ IRFs and ≈20% of industry IRFs are significant. IGOs increase discussion of biodiversity protection by ≈11 pp and address governance/city policy. IGOs also drive NGO communication on climate action by ≈14 pp; renewables (positive), jobs (neutral), corporate sustainability (positive), and media engagement (neutral) influence NGOs by ≈12–16 pp. For IGO impulses (Fig. 2c), ≈53% of NGOs’ IRFs and ≈33% of industry IRFs are significant. Industry responds on PR/advertising themes: industry supports STEM (≈8 pp, positive sentiment) and gas company advertising (≈5 pp, positive). IGOs influence NGOs on climate action (≈10 pp, neutral) and renewables (≈12 pp, positive).
• Relative influence (H2 not supported): Contrary to expectations, NGOs exhibit greater sway over online conversations than fossil fuel firms and IGOs within the VAR framework. IGOs are also notably responsive to NGO activity, especially on climate action topics. Industry elicited responses mainly on corporate sustainability, STEM-support, and media/advertising-related themes.
• Limited role of exogenous factors (mixed for H3): Overall, stock market performance and extreme weather events are not strongly correlated with messaging patterns. IRFs with stock returns show no significant first-period effects on industry, IGOs, or NGOs (Fig. 3). Weather-related IRFs are generally null or weak (Fig. 4), with exceptions: droughts increased IGOs’ discussion of gas company advertising by ≈10 pp; storms increased IGOs’ gas station customer service by ≈7 pp and industry company advertising by ≈4.5 pp; wildfires increased NGO biodiversity protection by ≈5 pp and industry gas company advertising by ≈5.5 pp. These suggest that when extreme events are salient, industry messaging gravitates toward advertising/PR themes.
• Topical and sentiment structure: Fossil fuel firms frequently tweeted neutral-toned media/public engagement and positively about air pollution mitigation; IGOs frequently engaged neutrally on Twitter engagement and climate action, with negative tones on Trump-era drilling and endangered species; NGOs showed heterogeneous coverage across topics. Many industry mentions of climatological changes, divestment, extreme weather, and gas station service had negative sentiment; IGOs’ negative topics included endangered species and extreme weather; NGOs’ negative topics included toxic/plastic waste, extreme weather, divestment, and Trump EPA initiatives.

Findings show measurable, directional interactions among fossil fuel firms, IGOs, and NGOs in online climate discourse. Consistent with H1, groups influence each other primarily within domains of expertise: NGOs and IGOs on climate action and policy, industry on corporate sustainability and PR. Contrary to H2, NGOs—despite fewer resources—exert greater agenda-setting effects than industry and IGOs, potentially reflecting follower bases and engagement dynamics. The results align with prior work on corporate reframing and greenwashing: industry catalyzes discussions when emphasizing sustainability, advertising, and STEM support, indicating strategic narrative management rather than direct engagement with contentious issues. Regarding H3, the largely null relationships between stock performance and messaging suggest corporate public communications follow long-term strategic narratives, not day-to-day financial outcomes; extreme weather generally does not drive cross-group dynamics, with limited exceptions steering industry toward advertising/PR messaging and NGOs toward biodiversity protection. Overall, the analysis advances understanding of how powerful stakeholders shape the online climate information ecosystem and how watchdog organizations (NGOs, IGOs) both influence and are influenced within this space.

This study provides a system-level, data-driven account of online climate and sustainability communication among major fossil fuel firms, NGOs, and IGOs (2014–2021). Using JST and VAR with IRFs, it demonstrates reciprocal influence aligned with topical expertise, reveals NGO prominence in steering online debates (contrary to expectations of industry dominance), and shows limited overall impact of stock returns and extreme weather on messaging dynamics. The work contributes to climate accountability research by quantifying reframing behaviors and their propagation across stakeholders. Policy and platform implications include the potential for regulatory and design interventions to enhance transparency, reduce subtle reframing, and support user evaluation of information. Future research directions include linking macro-level discourse shifts to public opinion or behavior change; developing typologies of climate misinformation based on stakeholder interactions; mapping the value chain of climate and sustainability information; and assessing advanced NLP methods (e.g., transformer-based and generative models) for topic/sentiment dynamics analysis.

Key limitations include: (1) reliance on lexicon-based sentiment tools, which may miss nuanced meanings despite valence shifter adjustments; (2) demographic and behavioral biases inherent to Twitter data, including bot/malicious activity and asymmetric participation; (3) selection biases from focusing on English-language, high-follower accounts; (4) assumptions that follower counts proxy influence; (5) potential latent temporal relationships not fully captured by the VAR, despite stationarity checks and flexible lag structures; and (6) model-dependence in topic discovery (choice of JST, topic number), though coherence optimization and weak supervision were employed. These factors may affect interpretation and generalizability of the results.