2.1 ESG reporting and variation in environmental communication

ESG reports have become a primary channel through which firms communicate their environmental commitments, risks, and strategies. However, reporting practices differ markedly across firms and sectors. Companies vary in the level of detail they provide, the structure of their reports, and the environmental issues they prioritise [2,3]. Recent systematic reviews confirm that even under established standards, reporting practices vary widely in structure and emphasis, suggesting that firms retain considerable discretion over their environmental communication [4]. Empirical studies also find that topic emphasis differs substantially across countries and industries, reinforcing the idea that ESG reporting remains far from standardised [5].

Concerns about credibility have intensified as ESG reporting has expanded. Much ESG information remains unaudited, and recent analyses show that the relationship between disclosed policies and underlying environmental performance is often weak [6]. Persistent divergence in ESG ratings further reflects methodological inconsistencies in scope, measurement choices, and aggregation rules [7]. Recent work concludes that these divergences have structural origins and are unlikely to fully converge without substantive changes to data quality and verification processes [8]. Recent financial research also shows that environmental signals embedded in ESG disclosures can affect market risk perceptions by demonstrating, for instance, that firms perceived as environmentally friendly influence sectoral crash risk asymmetries, suggesting that environmental characteristics embedded in ESG communication can shape investors’ risk assessments [9].

Although this research documents substantial variation and credibility concerns, it offers little insight into whether the specific environmental issues firms choose to emphasise correspond to measurable improvements in environmental KPIs. Understanding this link matters because stakeholders often rely on the content of ESG reports to evaluate whether firms are addressing environmental priorities in a meaningful way.

2.2 Greenwashing and the persistence of the word–action gap

A central theme in the literature is the gap between firms’ sustainability communication and their environmental behaviour. Research documents numerous mechanisms of greenwashing, including selective disclosure, focus on positive achievements, and omission of negative outcomes [10,11]. Recent reviews indicate that research on greenwashing has expanded rapidly, with growing attention to symbolic disclosure strategies and impression management [12]. Modelling work further demonstrates that symbolic communication can be a rational choice for firms facing pressure to appear responsible in contexts with limited verification [13].

Textual analyses reveal that some reporting practices, such as vague statements, repeated boilerplate content (i.e., keywords and stable common phrases that can be applied in several contexts without needing significant modifications), or broadly framed environmental claims, may signal symbolic rather than substantive communication [14–18]. These findings suggest that ESG reports often contain a mix of substantive and symbolic elements. However, such analyses typically evaluate credibility or symbolic tendencies at the firm level, treating the word–action gap as a general characteristic of a company’s reporting.

These insights help explain why the word–action gap persists but do not reveal whether discrepancies differ by environmental issue. Some topics may entail more measurable commitments and thus be more closely tied to performance, whereas others may be easier to discuss symbolically. Identifying whether the alignment between word and action varies by environmental topic would allow for a more precise understanding of the credibility of ESG reporting.

2.3 Regulation and growing expectations for meaningful disclosure

Regulatory initiatives such as the EU’s Corporate Sustainability Reporting Directive (CSRD) and Sustainable Finance Disclosure Regulation (SFDR) seek to improve the reliability and comparability of ESG reporting [19–21]. Similar developments in the United States and other regions reflect growing expectations that enhanced reporting will provide stakeholders with more decision-useful information on environmental risks and performance [22–24].

Although these frameworks require firms to disclose more structured and detailed information, companies still retain substantial flexibility in selecting which environmental issues to highlight. Disclosure requirements often mandate reporting on strategies, risks, and governance structures but do not directly verify whether the issues emphasised in reports correspond to real changes in environmental KPIs [25–27]. This leaves considerable room for variation in how firms frame their environmental efforts.

This evolving regulatory landscape highlights the need for evidence on whether the content of ESG disclosures relates to environmental performance. Such evidence is essential for assessing how firms respond to new reporting requirements and for evaluating whether regulatory reforms improve the environmental transparency of ESG reporting.

2.4 Analysing ESG reporting content through topic modelling

Among most frequent machine learning (ML) methods for ESG reports in the literature is the so-called Topic Modelling (TM) methodology. Different TM algorithms have been employed to analyse ESG reports, mostly Latent Dirichlet Allocation (LDA) and its variations [28,29]. Another approach is taken by Schimanski et al. [30] who used supervised ML models, RoBERTa and DistilRoBERTa, to determine whether a sentence relates to any of the three pillars of ESG. Similarly, the study by Funk et al. [31] classifies ESG reports from companies in the STOXX Europe 50 to more specific categories within the three ESG pillars. Recent methodological reviews highlight the increasing use of distribution-sensitive econometric techniques to capture heterogeneous economic relationships, including quantile regression approaches that analyse effects across different parts of a distribution rather than only mean outcomes [32].

Yet, as a pretrained model, its capacity to provide new topics or trends that might emerge in the ESG domain in the future is limited. Furthermore, it possesses solely the ability to categorise the specific pillar of ESG reporting to which it pertains, yet it lacks the capacity to delve further into that classification.

The key methodological difference between topic modelling and the two studies above-mentioned is that TM performs an unsupervised classification, i.e., one does not have a given list of terms or sentences to be searched for by the algorithm. Dyer et al. [8] study the trends in 10-K reports during the period of 1996–2013 in the United States employing LDA finding that the presence of boilerplates and redundant texts in the documents has increased substantially during the studied period.

Yang and Yang [6] study the correlation between the information present in the ESG reports from 97 companies from the Fortune 500 during the 2010s and their financial key performance indicators like the ROA or Tobin’s Q. The authors find that companies addressing customer-oriented capabilities in their ESG disclosures are associated with better financial outcomes, which collaborates to the literature perception that ESG efforts being directly related with market value and profitability.

Baier et al. [33] focus on 10-K reports published by US companies. The sample is constituted by the documents published by the 25 largest companies in the S&P100 index during 2015–2019. These documents contain not only non-financial information but rather the global activities executed by a company throughout the year. The authors use LDA to reveal trends in these disclosures. Across all the reports included, the share of ESG-related words represented 3.7% of the total. The company with the lowest presence was Berkshire Hathaway (specialised in financial services) and the one which presented the highest proportion of ESG words was Johnson&Johnson.

Bowen and Min [34] study over 4500 articles from national newspapers and 456 academic papers related to ESG, published between 2012 and 2022 using LDA. For academic papers, they found a prevalence of topics related to what they labelled “Environmental Protection”, followed by “Sustainability” and “Financial Performance”. The key difference with the non-academic articles is the dominance of the topic “Carbon Emissions and Climate Neutrality” in newspapers.

Yet, across these contributions, one key shortcoming persists: a lack of evidence connecting ESG communication themes with companies’ real-world environmental performance. This disconnect limits our ability to distinguish between substantive ESG engagement and symbolic disclosure. The present study addresses this by directly linking ESG report content to changes in firms’ environmental KPIs.

3.1 Data

The present study focuses on the ESG reports released between 2010 and 2023 by companies listed in the STOXX Europe 600 index. We concentrate on Europe for two reasons. First, the EU is one of the global leaders in decarbonizing its economy with many policies stimulating companies to reduce their emissions, such as the European carbon market (the EU-ETS). Second, it is one of the largest stock markets in the world covering companies from many countries and industries ensuring that the dataset will contain a diverse sample of companies.

We collected the ESG reports from the Sustainability Reporting Navigator (https://www.sustainabilityreportingnavigator.com/). At the time of data collection (8th of August 2024), the database encompassed a total of over 11,409 documents from 787 companies across 18 European countries not all of which listed in the STOXX Europe 600 index. It included five types of publications:

• Annual reports that provide financial information about the firm and general managerial decisions (in total, n = 6306 documents)
• ESG reports that are the ones which highlight the companies’ efforts in terms of ESG performance (n = 4056)
• Integrated reports that combine the above-mentioned two types (n = 248)
• Carbon disclosure project reports which are documents that are more specific to carbon emissions related efforts (n = 19)
• Other reports published by the organisations (n = 780).

We preselect only ESG reports to avoid more general reports the companies publish primarily on their financial situation. Full texts of 1489 reports that match companies from the STOXX Europe 600 were retrieved in a PDF format, which we subsequently digitalised. We successfully transformed 1477 documents while 12 files were corrupted (i.e., one could not extract text from the PDF) and had to be dismissed. In doing that we extracted only the textual data from the reports and omitted tables and graphs which is in line with the previous literature [25,35]. We did not limit the sample to a specific type of ESG reporting (e.g., GRI-based), as there is currently no consensus in the literature on which reporting format is superior in quality or reliability. Additionally, narrowing the scope to a particular standard would have significantly reduced the size and diversity of our dataset. Instead, we aimed to retain a heterogeneous sample that includes companies exhibiting different dynamics on environmental KPIs, allowing us to examine how the content of ESG reports is associated with actual performance. See the number of reports per country and per year in Table A1 in the Appendix (S1 File).

The ESG reports released are mostly written in English, but a few are in German, French or Italian. To ensure comparability of the textual data across the reports, we translated the non-English reports using the DeepL translator service. Translating documents into a common language prior to topic modelling is standard practice in cross-national computational text analysis, as it ensures corpus comparability while avoiding sample bias from excluding non-English documents [36]. Topic modelling relies on word co-occurrence structures rather than stylistic features, and prior research has shown that automatic translation tools provide reliable and statistically comparable results for text-as-data applications [37–39].

The final dataset encompasses 1477 reports from 268 distinct companies from STOXX Europe 600. The average length of reports is 47,528 words, with a minimum of 411 and a maximum of 1,616,575 (left chart in Fig 1). The dataset covers 16 countries with the UK, France and Germany leading by a big margin (see right chart in Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Distribution of textual documents in terms of length (left chart) and across countries (right chart).

https://doi.org/10.1371/journal.pone.0350762.g001

Furthermore, it is worth stressing that the final dataset is an unbalanced panel: while some companies are observed for all 14 years (for instance, Endesa, Infroma and Nokia), for some other firms we only have 1–2 reports (e.g., Ryanair and Bouyges). Additionally, for all companies in the dataset we obtained information on several environmental KPIs from the Refinitiv database such as CO₂ emissions, energy use, ESG score, emissions score, environmental innovation score. This KPI data, however, is only available from 2019 to 2023. Several of these measures (like CO₂ emissions and energy use) we subsequently normalise in relation to revenue to account for company’s size thus obtaining emission and energy intensity of firm output (see Table A2 in the Appendix (S1 File)).

The firms in our dataset belong to different industrial sectors (see Fig 2). To reduce the number of sectors and avoid any sector having too few observations (like Forestry & Agriculture covering less than 1% of observations, see Table A3 in the Appendix (S1 File)) we group these sectors into seven sector groups as explained in Table 1 below. The primary aim of this grouping is to put together firms similar in their core business activity, like producers of transport vehicles and companies exploiting those vehicles to provide services (e.g., logistic services).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Sector clusters and their prevalence in the dataset.

https://doi.org/10.1371/journal.pone.0350762.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Number of reports per sector.

https://doi.org/10.1371/journal.pone.0350762.g002

Plotting the number of reports over time in the final dataset we see a significant upward trend in the number of reports published peaking in 2020 (Fig 3). In 2021 the number of reports drops which may be due to COVID-19 pandemic. Another drop in 2023 may be because not all reports for this year being collected by the Sustainability Reporting Navigator by the time we accessed the data.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Number of documents per year included in the dataset.

https://doi.org/10.1371/journal.pone.0350762.g003

This study relied exclusively on publicly available corporate ESG reports and secondary firm-level environmental KPI data. No human participants, personal data, animal subjects, surveys, or experiments were involved. Therefore, ethics committee approval and informed consent were not required.

3.2 Methods

To uncover latent topics from the ESG reports, we use the Topic Modelling (TM) approach. The main advantage of this approach compared to keyword search is that it allows words to change their meaning dependent on the context they are presented in [40]. TM clusters words in topics based on how often those words co-occur across different documents in our sample of ESG reports [41]. For example, if the words “emissions”, “carbon”, “factory”, “production”, “gas”, and “GHG” appear in a topic labelled “GHG emissions” (see Section 4), we can conclude that the topic is about GHG emissions.

The main advantage of TM is the ability to quickly process very large data sets that would be very costly to manage using manual coding [42]. Another advantage is that the classification of topics is not based on the programmer’s own judgment, but is objective in the sense it is based on the textual data of the ESG reports, and this approach is transparent and reproducible [33,43,44]. More specifically, we use the Structural Topic Modelling (STM) method. The advantage of STM compared to LDA, is that it can use metadata about the underlying texts (in our case, the sector group of the company the ESG report corresponds to and the year of publication). Thanks to this STM is proved to generate semantically more coherent (i.e., easier to interpret) and interpretable topics [45]. This further allows us to formally study how these themes differ across sectors and evolve over time, within one coherent modelling framework. The STM algorithm we used in this study is based on the corresponding package in R by [46].

Before analyzing the ESG reports with STM we had to pre-process the textual data. We transformed all capital letters to the lower case, deleted non-Latin characters, stop words (such as “we”, “and” or “like”) and words being shorter than three letters. Afterwards, all terms included in the dataset were lemmatised transforming words to their dictionary form. Words that appeared in less than three reports were eliminated, and stable two-word sequences called bigrams were additionally formed (e.g., “corporate_governance”, “ghg_emissions”, “hazardous_waste”). As a result, the final dataset contains 17,270 unique words and 6,727,732 total word occurrences for building a topic model. All these preprocessing steps follow standard practice for bag-of-words topic models and are appropriate for STM, which relies solely on textual information (e.g., 8,28,29,33). In the STM literature, such steps are treated as necessary normalization procedures rather than robustness dimensions to be systematically varied. Therefore, removing non-textual elements and applying lemmatisation or stop-word filtering does not meaningfully reduce the information available to the model [46,47].

Subsequently, despite individual documents often being very long, the total number documents (only 1477) is rather limited, particularly given the size of the data in terms of number of unique terms. Since STM, like any other TM method, attributes words to topics based on how often these words co-occur in the same textual documents, running STM on the 1477 documents considering each of them as unique observations leads to rather vague topics that are hard to interpret. To address this challenge, we follow Savin and et al [48] by splitting 1477 ESG reports into shorter texts of no more than 10000 words each, resulting in a total of 4709 separate texts Following this approach, and consistent with common practices in computational linguistics (e.g., [49]), the 10,000-word threshold increases the number of observational units while still capturing co-occurrences that may occur farther apart in substantially longer ESG reports and prevents a few exceptionally long reports from dominating the analysis. Topic prevalences were then aggregated back to the original report level using segment-length weights, such that each segment contributes proportionally to its word count in the original document, thereby reconstructing the report-level topic distribution, mirroring Savin et al. [50] and preserving the integrity of subsequent regressions. As a robustness check, we also tried a segmentation threshold of 3000 words, and the results remain robust (see Appendix B in S2 File).

To choose the appropriate number of topics, we follow the earlier literature [46,51,52] and compare STM models with three to 50 topics on such criteria as out-of-sample prediction accuracy (held-out log-likelihood), the extent to which popular words from topics overlap (exclusivity), and the degree to which words from the same topic appear in the same texts (semantic coherence). Fig A1 summarises the results indicating that 34 topics exhibit the maximum prediction accuracy and close to maximum exclusivity, while maintaining a relatively high semantic coherence (S1 File).

Preprocessing steps such as stop-word removal, lemmatization, and bigram construction follow standard practice in bag-of-words topic modelling and are intended to enhance interpretability rather than to define alternative model specifications [40,45,53]. Topic-number selection was based on a comparison of models ranging from 3 to 50 topics using held-out likelihood, semantic coherence, and exclusivity, consistent with established STM guidance (Roberts et al., 2019).

Once the topics are formed and their prevalence are estimated, we proceed by analysing factors explaining the variation of topic prevalence among the ESG reports. To this end, we regress topic proportions over the covariate variables we used in building our topic model [50]. In particular, a linear regression model was specified for each of the 34 topics (indexed by k) containing year of report publication and dummy variables for each of the sector groups as follows:

(1)

Results of these regressions on the whole data sample covering 1477 reports are then reported to demonstrate the distribution of topics over time and across sector groups.

Next, we cluster companies in terms of the environmental topics addressed, i.e., we group them based on prevalence of these topics in their ESG reports using hierarchical clustering. We focus on the topics related to environment (see Table 3) and limit the analysis to the period between 2019 and 2023. The reasons for selecting this timeframe are as follows. First, the dataset in unbalanced with only few companies covered in early years, and some of them not observed in the later years. Since topic prevalences changed considerably over time (see Fig 4 and Fig A2 in S1 File), comparing those companies with firms observed only in the later years would be misleading. Second, the number of reports published during these years is significantly higher compared to earlier periods (see Fig 3). Hence, for the clustering we consider only firms which published at least three reports 2 during this period to ensure the comparison is more balanced as 1–2 reports filed in the earlier or later years may still bias our conclusions. The resulting sample includes 125 companies covered in the period 2019–2023.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Aggregation of 34 ESG reports topics into six thematic groups.

https://doi.org/10.1371/journal.pone.0350762.t003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Trends in topic popularity over time.

Note: Values generated by a regression where the outcome variable is topic prevalence and year of publication is the predictor. Confidence intervals plotted as dashed lines indicate the 95% uncertainty range and include both regression and measurement uncertainties associated with the STM model. ***, ** and * denote 0.1%. 1%, and 5%, significance, respectively.

https://doi.org/10.1371/journal.pone.0350762.g004

After identifying topics addressed in ESG reports, their distribution over time and across sectors, and forming company clusters based on how (dis)similar they are in environmental topics addressed in their ESG reports, we study how prevalence of certain topics is associated with firms’ actual efforts for climate change mitigation captured by environmental KPIs. In doing that we distinguish between KPIs provided directly by the companies (such as amount of CO₂ emissions and energy used in proportion to revenues generated) and the KPIs assigned by Thomspon Reuters/Refinitiv (like ESG score, emissions score, environmental innovation score). Given that KPI data are available only for 2019–2023, our analysis should be interpreted as evidence of contemporaneous alignment between reporting emphasis and environmental performance rather than long-term dynamics.

To this end, we perform a regression analysis putting change in each of the KPIs on the left-hand side of the equation and change in the prevalences of the six topics related to environment on the right-hand side. This is meant to capture whether companies increasing discussion of certain environmental topics tend to improve their KPIs. Moreover, next to the change we also add the level of the prevalence of the topic since companies stressing environmental issues in their ESG reports in absolute terms may tend to improve their environmental KPIs more. All six environmental topics are included simultaneously, so coefficients represent conditional associations—i.e., relationships net of competing reporting themes—rather than isolated correlations.

(2)

In addition, we add year () and company () fixed effects (see Equation 2). Inclusion of the company and year fixed effects is meant to capture idiosyncratic differences between companies (e.g., smaller firm in the digital sector producing less emissions than a large company in the energy sector) and changes in the KPIs specific for a particular year (e.g., COVID-19 pandemic). Inclusion of these fixed effects also captures such stable company characteristics like size, age and sector. We also apply Driscoll and Kraay standard errors [54] that account for heteroskedasticity serial correlation and cross-sectional dependence.

Including both the level and the change of topic prevalence follows the logic used in empirical studies such as Dosi et al. [55], who show that levels and changes of an explanatory variable capture different mechanisms: levels reflect persistent, structural heterogeneity across firms, while changes capture dynamic adjustments that are more directly linked to performance outcomes. Put differently, the level captures a firm’s stable, long-run emphasis on a given environmental theme, whereas the change captures short-term adjustments in reporting focus, enabling us to separately identify structural heterogeneity and dynamic effects. In our context, the level of a topic indicates a company’s stable emphasis on an environmental theme, whereas the change reflects year-to-year shifts in disclosure. Omitting either would force us to ignore one of these channels or absorb it into fixed effects. Including both is therefore appropriate, even if non-standard, because it allows us to distinguish persistent disclosure strategies from dynamic adjustments in ESG reporting—exactly the distinction that motivated the dual specification in Dosi et al. [55].

To ensure that including both level and change terms does not introduce multicollinearity concerns, we computed variance inflation factors (VIFs). All VIF values are within the interval between 1 and 3 which well below the conventional threshold of 5, indicating that collinearity does not materially affect the estimates.

Two main hypotheses associated with Equation 2 that we are testing are:

H1: Companies with higher share of environmental topics in their ESG reports are expected to improve their environmental KPIs.

H2: Companies with increasing share of environmental topics in their ESG reports are expected to improve their environmental KPIs.

The corresponding null hypotheses are implying that there are now such associations. The regression analysis described in Equation (2) is used to empirically test H1 and H2 against these null hypotheses.

The motivation for the abovementioned hypotheses is the expectation that companies stressing certain environmental topics in their ESG reports reflect their actual efforts in improving their performance in this area [6]. Alternatively, if empirical data will not be able to provide any evidence in support of the hypotheses, we may conclude that companies’ discussion of environmental issues in their ESG reports may be interpreted as a sign of greenwashing.

4.1 Main topics

34 main topics (denoted by the prefix “T”) we identified in the ESG reports are presented in Table 2 with their label, prevalence and most frequent and exclusive words. The largest topic is Inclusive leadership (T1) with the share of 8.47% of all the ESG reports. This topic is mostly related to various actions and activities enhancing employee performance and their well-being. This is in line with Sarkar and Searcy [56] and Goloshchapova et al. [28] who find employee well-being as one of the key subjects in ESG disclosures. The second most popular topic is Supervisory board (T2, 6.73%). This topic covers the communication around the different internal politics of the organisation, their main governance modifications and other matters that might affect the upper-level administration of the firm. Closely behind, the next most prevalent topic is GHG emissions (T3, 6.62%). This is the most prevalent topic related to environment, but not the only one.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. 34 main topics in ESG reports.

https://doi.org/10.1371/journal.pone.0350762.t002

To facilitate a more streamlined analysis of emerging trends, we group the topics in six thematic groups (designated with the prefix “G”, see Table 3), reporting also the number of ESG reports dominated by each topic. The latter allows to see that some topics can be dominant in many reports despite their not so big overall prevalence (e.g., T8 on sustainable VCs), while others are rather complementary despite occupying rather large prevalence overall (e.g., T6 on corporate disclosure). The first thematic group (G1) is Organisation governance which includes several topics that address management of the organisation and the stakeholders that do have direct decision power. It is by far the most prevalent thematic group (33.22%) and includes the two most frequent topics.

The second group by order of prevalence is the one that comprehends the discussion related to Environment (G2). It accounts for 18.48% in the ESG reports which demonstrates that it holds a major position in the list of priorities for companies in terms of non-financial aspects. This thematic group covers topics related to emissions (T3), value chains (T8), electric vehicles (T14), renewable energy (T20), energy sources (T27) and subsidies related to environmental KPIs (T28).

As the third thematic group (G3) we have a category labelled as Sector-specific issues corresponding to more specific issues that might be isolated from the rest of ESG subjects that appear because of the corresponding company belongs to a certain industry such as healthcare, real estate or telecommunications.

G4 encompasses topics related to the social pillar of ESG, with a strong focus on diversity, inclusion, and the equal treatment of minorities. It also addresses the relationship between companies and their clients, considering customer characteristics and specific needs. It covers 13.07% of the ESG reports. Thus, this area (S) results to be the last among the ESG pillars ranked by prevalence.

G5 includes predominantly financial topics, which continue to be reflected in ESG disclosures. Lastly, G6 covers topics related to companies’ interactions with public and regulatory institutions, focusing on regulatory compliance and accreditation.

With these results, the governance pillar (associated with G1, G5 and G6) gathers the most attention from organisations, since it encompasses the majority of topics and most of the associated textual data (52.08% total prevalence). Surpassing the prevalence of the two other pillars combined, this result is in accordance with the findings from Baier et al. [33] and Sarkar and Searcy [56].

4.2 Topic popularity over time

After obtaining the topic clusters and their prevalence, we proceed to study their evolution over the thirteen years period included in the dataset. The output is visualised in two figures, Fig 4 showing the statistical linear trends of the topics and their significance, and Fig A2 displaying the actual variation in topic prevalence over time (S1 File).

The most prevalent topic studied, Inclusive leadership, has experienced a significant upward trend from a 5% prevalence to 15%, being the most popular topic in ESG reports. In contrast, topics T2 and T5 on supervisory board and young customers lost popularity from around 10% to less than 5%.

Zooming on the topics related to the environmental pillar of ESG, T3 on GHG emissions is confirmed as a subject that has gained a lot of attention over the last 15 years, reaching its peak in 2021 (15%). In a similar manner, prevalence of Electric vehicles (T14) also increased across the period observed. The theme covers different characteristics, needs and possibilities of these vehicles. Correspondingly, T20’s (Renewable energies) prevalence distribution over time exhibits a significant upward direction. After hitting a low in 2016, its prevalence steadily increased. Prevalence of T28 on subsidies and environment also increased, particularly over the last five years. On the other hand, T27 on various energy sources (from coal, oil and gas to nuclear) has lost prevalence among companies’ sustainability disclosures, which reflects the shift in attention from fossil fuels to renewable energy (T20).

These tendencies align closely with the growing stringency of climate policies in the EU, but also with the introduction of the “double materiality” concept, This concept encourages organisations not only evaluate how climate and other risks and opportunities impact organisations, but also assess how companies’ actions impact the planet. The concept started to gain traction and was officially introduced in the corporate reporting landscape with the release of the EU Non-Financial Reporting Directive (NFRD) in 2014. However, it became widespread only in 2019 when the European Commission formally described the in the context of ESG reporting [57].

4.3 Topic concentration across sectors

After identifying topics and their prevalence over time, we distinguish which topics are most associated with different sector groups that we formed earlier (see Table 1). Fig 5 visualises the topics with either a positive or negative significant association with the different sector groups. We obtain a significant positive association of T5 (Young customers), T11 (Risk management), T18 (Cultural identity), T23 (Real estate), T32 (Legal proceedings) and T33 (Financial risks) with the Finance sector. While T11, T23, T32 and T33 are logical topics that are easy to link with the industries related to finance, T5 and T18 emerge as unexpected subjects to have attired attention from companies. Contrarily, there is a negative significant association with T17 (Natural resources), T27 (Energy sources) and T30 (Aquaculture), which, again, appears to be logical based on the nature of the sector.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Topic concentration across sector clusters.

Note: Only results significant at 5% level are reported. The plots show mean difference in topic proportions between the sector cluster under consideration and the rest of the sample (a positive value on the X-axis indicates a larger prevalence in the corresponding sector).

https://doi.org/10.1371/journal.pone.0350762.g005

In the case of the Consumer goods sector, the analysis shows a strong positive association with T12 (Consumables) and T24 (Financial disclosure), as well as slight associations with T4 (Corporate governance) and T5 (Young customers). Conversely, this sector demonstrates negative associations with the same topics as the Finance sector (T17 and T27), with the addition of T8 (Sustainable value chains). While the majority of the associations align logically with the sector priorities, it is possible that T8 appears less than expected beforehand, what could suggest a need for greater emphasis on the role of global supply chains within these industries.

The Digital sector has a significant positive association with T5 as well (which may be explained by the role of young customers as main consumers) and other related topics such as T29 (Telecommunications) and T14 (Electric vehicles). T25 (Shareholder rights) is also strongly present in this sector. Contrary, T15 (Firm growth) as well as T17, T27 and T30, on average are stressed much less in ESG reports of companies in this sector.

T1 (Inclusive leadership) and T17 (Natural resources) tend to be stressed more in the Resources cluster, while for T13 (Asset value) the situation is the opposite. This presents an interesting tendency to social value over the financial one.

T14 is the most positively associated topic in the Industrial sector which can be explained by a large share of automotive companies in this sector. T8 also tends to be stressed by industrial firms more. The slight positive association with T9 (Women rights) deserves to be highlighted as it is one of the issues that have been brought back to attention of the STEM companies at several occasions because of the lack of female employees in the industry [58,59].

Housing sector is significantly positively associated with T4 (Corporate governance), T10 (Affordable housing) and T26 (Public utilities), which is logical as the issue is directly linked to these companies’ activities. T20 (Renewable energy) is also associated with the ESG reports published by companies in this sector, as the matters included in this topic are connected with the sector activities (think of solar photovoltaic panels installed on house roofs). The last sector cluster, Pharma, tends to stress stronger T4 (Corporate governance) as well as T22 (Healthcare).

4.4 Clustering firms based on topics related to environment

Estimating average prevalence for each of the six topics (T3, T8, T14, T20, T27 and T28) for each of the 125 companies in the sample for the period 2019−2023 and applying hierarchical clustering, we obtain the dendrogram reported in Fig A3 in Appendix (S1 File) and decide for 6 clusters highlighted with different colours. Left plot in Fig 6 presents this dendrogram in a radial form to enhance its readability. The six clusters (designated with the prefix “C”). have the following sizes: C1 is an outlier cluster with a single company, C2 - 16, C3 – 11, C4 – 9, C5 - 9 and C6 (the largest cluster) includes 79 companies. Right plot in Fig 6 compares the clusters in terms of the six topics related to environment. The clustering exercise is exploratory in nature and complements the regression analysis by illustrating how firms group according to their thematic emphasis rather than establishing predictive or causal classifications.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Six company clusters (left plot) and their focus on the topics on environment.

Note: Left plot is a radial dendrogram clustering companies based on topic prevalence in their ESG reports. Right plot shows average prevalence of each of the six topics related to environment among the identified clusters of companies. Note that the y-axes of the six graphs have different scales. Dashed lines indicate the mean for each topic across the five clusters. Colours are consistent on the charts of the figure.

https://doi.org/10.1371/journal.pone.0350762.g006

The outlier cluster (C1) comprising the only company, Klepierre (French real estate firm), displays below-average prevalence of all topics except T28 on subsidies & environment (40%) and T3 on GHG emissions (over 20%). Exceptionally high focus on T28 makes this company stand out from the rest of the sample.

Cluster 2 (C2, including companies like L’Oreal) has a strong focus on GHG emissions (almost 35%), considerably above the average (10%). Contrarily, cluster 5 (including among others energy companies Enel and Repsol) does not emphasise this topic (only 5% prevalence).

Sustainable VCs (T8) are primarily addressed by 11 companies in cluster 3 (on average, prevalence is above 30%). Among these companies are industrial and energy companies like Siemens and RWE. On average, the topic covers 35% of the content of the 11 companies clustered in this group. The topic share among other four clusters is significantly lower (under 5%).

T14 on Electric vehicles tends to be stressed much stronger (on average, over 25%) among companies in cluster 4 which includes many companies specialised in automotives (like Ferrari) and transport services (e.g., Georg Fischer). Cluster 3 (which includes several manufacturers of transports components, such as Bechtle, MTU Aero Engines or Continental) also shows an above-average emphasis on the subject, although at a much lower level (around 7%).

T20 on renewable energy is predominantly present (average share of 25%) in ESG reports of companies in cluster 5, largely composed by energy companies (Enel, Repsol, EDP). It is worth noting that the share of this topic in other clusters is very low (below 5%). Finally, T27 on Energy sources logically is also stressed by cluster 5 the most.

The largest cluster (C6) shows a below-average prevalence for most of the environmental topics. The only exception is T27, stressing the role of different fossil and non-fossil fuels (e.g., nuclear), but not providing any concrete solutions on how to reduce emissions.

Fig 7 demonstrates sector composition of the six clusters, while Fig A4 in Appendix (S1 File) their country composition. Clusters 2 (C2) and 6 are rather diverse both in terms of country and sector presence. C3, in contrast, is dominated by German companies from Digital and Resources. C4 is mainly concentrated in Industrial sector with companies predominantly from France and Switzerland. C5 is composed of companies in Resources, Housing and Finance based in Southern Europe (Spain, Italy and Portugal).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Company clusters based on prevalence of environmental topics.

Note: Wordclouds showing the different company clusters based on similarity in prevalence of the environmental topics they address in ESG reports. Font size in a cloud reflects the revenue of the company in 2023, while colour indicates the sector: Finance, Resources, Housing, Pharma, Digital, Industrial and Consumer Goods. Companies can be compared by revenues from one cloud to the other.

https://doi.org/10.1371/journal.pone.0350762.g007

The identified clusters can be interpreted as reflecting distinct strategic orientations: for example, firms emphasizing renewable energy and energy sources (Cluster 5) are predominantly concentrated in energy-intensive sectors and appear more directly exposed to regulatory transition risks, whereas firms in Cluster 3 emphasizing sustainable VCa may reflect supply-chain restructuring strategies. In contrast, the largest cluster shows limited emphasis across environmental themes, suggesting a more compliance-oriented or less strategically differentiated disclosure approach.

4.5 Nexus between environmental topics and climate mitigation efforts

Table 4 provides results of the linear regression of environmental KPIs on topic prevalences including year and company fixed effects as explained in Equation 2 (Section 3). The first two columns report results for the indicators reported directly by the companies (emission intensity and energy intensity), while the latter three models analyse scores provided to the companies by ESG rating agency (in our case, Refinitiv). The relatively low R² values for some KPIs reflect the fact that year-to-year changes in environmental indicators are driven by many factors beyond ESG report content. Nevertheless, the significant coefficients indicate that specific topics still contain meaningful explanatory information even within this noisy and highly multifactorial setting.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Regression of environmental KPIs on topic prevalences in ESG reports.

https://doi.org/10.1371/journal.pone.0350762.t004

Our results demonstrate that companies that increase discussion of sustainable value chains (ΔT8) tend to improve their performance both in self-reported measure of emission intensity and Emissions score from Refinitiv. This coefficient on ΔT8 thus captures a short-run effect: in years when a company increases the share of its ESG report devoted to sustainable value chains, it tends to experience a small but significant reduction in emission intensity (6 kg of CO2 per m USD) and an improvement in its emissions score (by 0.002 points), compared to years in which its reporting focus remains unchanged. Furthermore, more focus in ESG reports on renewable energy (T20) and less on a more general topic of energy sources (T27) is positively and significantly associated with reduction in CO₂ emissions indicating that companies focusing on renewable and not fossil fuels are actually reducing their emissions. Here, the coefficients on the levels of T20 and T27 reflect long-run reporting orientations: firms that systematically devote a larger fraction of their ESG reports to renewable energy tend, on average, to achieve significant reductions in emission intensity (1 tone of CO2 per m USD for 1% higher share of the topic), whereas firms that place relatively more emphasis on general energy sources (including fossil fuels) tend to increase their emission levels.

Furthermore, we find that companies increasing their discussion of renewable energies in their ESG reports (ΔT20) tend to increase their performance in emissions and environmental innovation scores. This indicates a dynamic component: years in which a firm increases its emphasis on renewable energy (a higher ΔT20) are also years in which its emissions and environmental innovation scores improve, suggesting that shifts in reporting towards renewables coincide with incremental progress on these KPIs. The latter is also intuitive since renewable energy technologies are still relatively new compared to technologies based on fossil fuels and require considerable R&D investments for their implementation [60,61].

The topic 28 on subsidies and environment, in contrast, seems to demonstrate a mixed relationship with ESG and emissions scores of the companies. The negative coefficients on the level of T28 imply that firms whose ESG reports are systematically more focused on environmental subsidies and taxonomy-related issues tend to exhibit lower ESG and emissions scores, whereas the positive coefficient on ΔT28 shows that short-run increases in attention to this topic are associated with slight improvements in the emissions score, consistent with firms seeking external support while undertaking emission-reduction efforts. Topics T3 on GHG emissions and T14 on electric vehicles, do not have any significant statistical association with environmental KPIs. Taken together, the level coefficients capture how firms with a persistently stronger emphasis on particular environmental topics differ in their KPI trajectories, while the change coefficients (ΔT) capture how year-to-year shifts in reporting emphasis co-move with short-run KPI changes. The results allow us to conclude that the empirical evidence partially supports H1, as higher levels of certain environmental topics—most notably Renewable energy (T20)—are significantly associated with lower emission intensity, whereas others show no detectable relationship. Regarding H2, we find support for the idea that increases in discussion of specific topics (e.g., T8 on Sustainable VCs; T20 on Renewable energy) correspond to improvements in selected KPIs, particularly emissions-related scores. For the remaining topics, the estimated effects are statistically insignificant, meaning we fail to reject the null hypotheses. These results indicate that the hypotheses hold for some environmental themes but not universally across all topics.

Some coefficients in Table 4 exhibit signs that may appear counterintuitive, for example, the positive level coefficient of T8 (Sustainable VCs) for emission intensity. However, these estimates are statistically insignificant and therefore do not provide evidence of a systematic association. We refrain from interpreting such coefficients.