Constrained LLMs deliver scalable, conservative measures of firms' energy-transition disclosures, revealing that although disclosures in Spain rose between 2023 and 2025, only a minority report concrete actions and most disclosures focus on renewables—particularly solar—while integrated decarbonization strategies are rare.

Main Finding

LLM-assisted coding of corporate website disclosures, constrained by a restrictive policy-aligned rubric and transparent decision rules, yields a scalable, reproducible, and conservative measure of firm-level engagement with the energy transition. Applied to a panel of Spanish firms (2023 and 2025), the method shows engagement is rising but remains limited: only a minority of firm-year observations disclose actionable measures, disclosures are asymmetric across dimensions (renewables dominate; solar is concentrated; energy efficiency and decarbonization strategies are less common), and multidimensional, integrated transition strategies are uncommon.

Key Points

• Methodology: Use corporate website text + LLMs constrained by explicit decision rules and a restrictive rubric to code evidence of:
• Energy efficiency improvements
• Decarbonization / emissions‑reduction strategies
• Renewable energy use or procurement (and renewable source when reported)
• Conservatism: The rubric is intentionally restrictive and policy-aligned to avoid counting generic sustainability narratives as implementation (limits false positives / greenwashing).
• Reliability: Authors report that constrained LLMs can extract reliable and reproducible indicators from heterogeneous web content.
• Empirical patterns (Spain, 2023–2025):
  • Disclosure of concrete actions is still limited to a minority of firm-year observations.
  • Renewable energy disclosures are most common; among adopters, reported mixes are heavily skewed toward solar.
  • Energy efficiency disclosures occur less frequently.
  • Explicit decarbonization/emissions-reduction strategy disclosures are rare.
  • Adoption and disclosure vary systematically by firm size and sector.
  • Multidimensional, integrated transition strategies are uncommon—firms tend to disclose in only one dimension.
• Scalability: The approach supports near real-time, large-scale monitoring of disclosed engagement.

Data & Methods

• Data source: Corporate website disclosures from a large panel of Spanish firms observed in two cross-sections (2023 and 2025).
• Coding rubric: A restrictive, policy-aligned set of rules that classifies evidence only when text indicates concrete implementation or procurement actions (not general statements).
• Dimensions coded:
• Energy efficiency improvements (e.g., investments, retrofits, quantified savings)
• Decarbonization / emissions-reduction strategies (e.g., targets, carbon capture, process changes)
• Renewable energy use or procurement (e.g., on-site generation, PPAs) and identification of renewable source(s) (solar, wind, etc.)
• LLM implementation: Large language models used to read heterogeneous web text and apply the rubric; decision rules constrain outputs to limit hallucination and over-interpretation.
• Validation & properties: Emphasis on reproducibility and transparency of decision rules. The approach is framed as conservative—prioritizes specificity over sensitivity to avoid false positives from greenwashing-style text.
• Limitations noted by authors: Measurement captures only disclosed actions (not undisclosed implementation); results reflect disclosure behavior and web presence heterogeneity; sample limited to Spanish firms and two years.

Implications for AI Economics

• Measurement innovation: Demonstrates how constrained LLMs can produce scalable, reproducible firm-level indicators for policy-relevant economic research (e.g., diffusion of clean technologies, firm heterogeneity in transition readiness).
• New datasets for causal and descriptive work: Provides near real-time, high-coverage signals that can be merged with financials, emissions registries, procurement, or energy-consumption data to study:
  • Determinants of adoption (firm size, sector, financial health)
  • Effects of adoption/disclosure on firm performance, investment, and valuations
  • Policy evaluation (impact of subsidies, mandates, disclosure rules)
• Informing policy targeting: Sectoral and size heterogeneity in disclosures can help target interventions to low-adoption segments and monitor policy uptake.
• Behavioral and strategic considerations: Public, automated monitoring may change incentives—firms could increase disclosure without substantive action (disclosure vs. implementation). The conservative rubric reduces but does not eliminate this risk.
• Methodological caveats for economists:
  • Disclosure bias: Indicators reflect what firms report publicly; unreported actions and firms with weak web presences will be missed.
  • Sample and external validity: Results are from Spain and two years; patterns may differ elsewhere or over longer time horizons.
  • Model risks: Constraining LLMs and using explicit decision rules reduces hallucination but requires careful auditing and periodic recalibration as firms’ language evolves.
• Opportunities for extension:
  • Combine LLM-coded disclosures with administrative or sensor data (energy use, emissions) for validation and richer analysis.
  • Use longitudinal monitoring to study dynamic responses to policy changes or shocks.
  • Expand geographic and language coverage to build cross-country comparative panels.
• Ethics & governance: Transparent rules and reproducibility are important for accountability; researchers should document rubric decisions, share code, and assess potential biases in coverage and classification.