Insider stories: analyzing internal sustainability efforts of major US companies from online reviews

The study addresses the challenge of determining whether companies support internal sustainability efforts (ISEs) such as gender equality, diversity, and staff welfare, an assessment hampered by the lack of operational methodologies and honest data sources. The authors aim to conceptualize and measure ISEs by deriving a six-dimension framework from the UN Sustainable Development Goals (SDGs), and to operationalize it using a deep-learning approach applied to over 350,000 employee reviews from 104 major US companies (2008–2020). The research investigates whether companies’ commitments to ISEs are detectable in employee-generated text and whether such commitments are associated with indicators of company success, including employee ratings and stock growth. Three research questions guide the work: RQ1 validates the ISE scoring method; RQ2 examines associations between ISEs and company success; RQ3 explores how ISEs vary across industry sectors.

The paper situates its contribution within corporate sustainability research, noting that prior work often emphasizes external sustainability outcomes and treats sustainability as a monolithic construct. It builds on the UN SDGs as a comprehensive foundation and on literature highlighting links between work-life balance, diversity, and well-being (e.g., Chung and van der Lippe, 2020; Lyonette, 2015), as well as the business value of social-impact and DEI initiatives. Prior studies also examine relationships between social performance and stock performance (Ziegler et al., 2007) and identify sectoral differences in sustainability signaling between B2C and B2B contexts (Hoejmose et al., 2012). This work advances the field by focusing on internal, employee-centered sustainability dimensions and by proposing a two-factor structure (staff welfare vs. financial benefits) emerging from employee narratives.

Data: The authors collected over 350,000 employee reviews for 104 major US companies spanning 2008–2020 from a company review platform, along with employee-provided ratings (balance, career, culture, management, overall) and company sector labels. Stock data for 84 of the 104 companies (2009–2019) were obtained from Yahoo Finance. ISE framework definition (mixed-method, three steps): - Step 1 (Pre-selection): Starting from the 17 UN SDGs, three independent annotators unanimously discarded goals not applicable to the corporate context, reducing to 13 SDGs. - Step 2 (Unsupervised discovery): An unsupervised deep-learning framework based on Sentence-BERT (SBERT) scored reviews for relevance to the 13 goals. For each goal, the top five most relevant reviews were manually assessed by three annotators. Goals were retained only if at least four of five reviews were marked relevant by a majority; inter-annotator agreement was high (Fleiss’ κ = 0.83). Five goals—primarily environmental (e.g., clean water, climate action)—were dropped, leaving eight goals. - Step 3 (Consolidation): The three annotators assessed semantic overlap among the eight goals in the context of company reviews and merged two pairs, yielding six ISEs: Monetary, Health, Education, Diversity, Infrastructure, and Atmosphere. Each ISE maps to an original UN SDG (e.g., Monetary→Decent work and economic growth; Health→Good health and wellbeing; Education→Quality education; Diversity→Gender equality; Infrastructure→Industry, innovation and infrastructure; Atmosphere→Peace, justice and strong institutions). Scoring ISEs at company level: For each company u and ISE i, the score s(u,i) is the fraction of reviews with SBERT similarity to the ISE phrase exceeding thresholds: a global threshold of 0.31 (mean similarity across validated goals from step 2) and an ISE-specific threshold equal to the 95th percentile of that ISE’s similarity distribution (as in prior work). Only the “pros” proportions of reviews were considered to focus on positive initiatives. Companies were ranked by s(u,i) per ISE. Validation (RQ1): Face validity was assessed by extracting top n-grams (1–4 grams) from reviews flagged for each ISE and computing TF-IDF scores per ISE document, verifying that discriminative keywords aligned with expected semantics (e.g., “salary” for Monetary; “health benefits” for Health; “opportunities learn” for Education; “flexibility” for Diversity; “technology” for Infrastructure; “positive work environment” for Atmosphere). Semantic relationships among ISEs were examined using PCA on company-level s(u,i) scores. Modeling success outcomes (RQ2): - Employee ratings: OLS regressions (selected via stepAIC) predicted company-level ratings (balance, career, culture, management, overall) using the two PCA components (PC1: staff welfare; PC2: financial benefits) as predictors, controlling for total number of reviews. - Financial performance: For each company, the geometric mean of stock growth over 2009–2019 was computed to handle heavy-tailed distributions. Associations with sustainability were inspected by plotting stock growth against rankings on PC1 and PC2; number of reviews was also plotted to assess the role of popularity. Sector analysis (RQ3): MANOVA tested differences in the two sustainability facets across industry sectors; distributions and company-level scatterplots illustrated sectoral and within-sector variability.

- Validity of ISE detection (RQ1): TF-IDF keyword analysis showed strong alignment between extracted keywords and intended ISEs (e.g., “salary,” “pay good” for Monetary; “health,” “health benefits” for Health; “opportunities learn,” “program” for Education; “flexibility” for Diversity; “industry,” “technology” for Infrastructure; “positive/friendly work environment” for Atmosphere). Some cross-ISE overlap (e.g., work-life balance appearing in Health and Diversity) reflected meaningful semantic relatedness. PCA on company-level ISE scores revealed two principal components explaining 88% of variance (PC1=73%, PC2=15%). All ISEs except Monetary loaded strongly on PC1 and weakly negatively on PC2, while Monetary correlated moderately with both, indicating two main facets: staff welfare (PC1) and financial benefits (PC2). - Association with employee ratings (RQ2): The two sustainability facets explained substantial variance in company ratings (Adjusted R2 up to 0.645 for overall rating). Staff welfare (PC1) was strongly positively associated with all rating facets (balance, career, culture, management, overall). Financial benefits (PC2) showed mixed associations: positive for overall rating, weaker or negative for some subratings (e.g., career, management), per Table 3 coefficients (e.g., staff welfare coeff. ~0.766*** for overall; financial benefits coeff. ~0.313*** for overall). - Association with financial performance (RQ2): Companies ranking highly on both staff welfare and financial benefits exhibited higher geometric mean stock growth, with staff welfare showing a stronger association with stock growth than financial benefits. High stock growth was not observed for companies focusing solely on financial benefits. - Sectoral patterns (RQ3): MANOVA indicated significant differences across sectors. Industrials and Information Technology led on staff welfare; Financials followed; Health Care displayed high variability; Consumer Staples and Consumer Discretionary lagged on both facets. Within-sector heterogeneity was evident: e.g., in IT, Microsoft/Google/Apple scored high on both facets, whereas Infosys/IBM/Cognizant were high on staff welfare only. In Consumer sectors, some firms (e.g., Dollar General, Kmart) scored low on both; Costco scored low on staff welfare but high on financial benefits. Reviewer role composition influenced company-level scores (e.g., Amazon’s low ISE scores likely reflect warehouse worker reviews despite high stock growth).

The findings confirm that internal sustainability efforts can be operationalized from employee-generated text and meaningfully summarized into two core facets: staff welfare and financial benefits. These facets address the research questions by demonstrating: (1) methodological validity of detecting ISEs in reviews; (2) substantive links between ISE engagement and company success, with staff welfare consistently associated with higher employee ratings and stronger stock growth; and (3) sectoral variations that highlight where sustainability efforts are leading or lagging. Conceptually, the study reframes internal sustainability as multifaceted rather than monolithic, emphasizing the need to balance financial benefits with broader welfare initiatives (health, diversity, infrastructure, atmosphere). Practically, the results suggest that managers should pursue holistic ISE strategies to enhance both employee satisfaction and financial performance, while policymakers can support policies incentivizing welfare-oriented practices. The framework also provides scholars with a validated, scalable method to analyze ISEs across large textual corpora.

This work introduces and validates a six-dimension framework of internal sustainability efforts derived from the UN SDGs and operationalized via a deep-learning approach applied to over 350,000 employee reviews from 104 major US companies. It demonstrates that employee-centered sustainability can be distilled into two principal facets—staff welfare and financial benefits—that together relate strongly to both employee ratings and stock growth, with staff welfare showing the stronger association. The study advances theoretical understanding by decomposing internal sustainability, offers a reproducible analytical framework and dataset for scholars, informs policymakers about the economic relevance of welfare-oriented practices, and provides managers with actionable insights to balance financial and welfare initiatives. Future research should expand the ISE taxonomy where needed, cover more companies (including those with fewer reviews via qualitative methods), establish causal directions between ISEs and outcomes, and further examine sectoral and role-specific dynamics.

- Potential incompleteness of ISE dimensions: Internal sustainability has been less systematically studied than external sustainability, and the resulting six ISEs may omit relevant aspects despite the mixed-method derivation and validations. - Temporal data constraints: The review platform begins in 2008, limiting the ability to account for earlier economic events (e.g., dot-com bubble) that might affect findings. - Sample size and coverage: Only 104 major companies were analyzed due to review count requirements; results may not generalize to companies with fewer reviews. - Lack of causal inference: The study establishes associations, not causality, between ISEs and socio-economic outcomes (e.g., stock growth). - Representativeness and potential bias: Sectoral review imbalances and possible self-selection in reviewers could bias results, though representativeness checks across sectors, state populations, and headquarters distributions suggest reasonable coverage; analyses were restricted to companies with at least 1,000 reviews to improve robustness.