ESG Scores Should Read the Evidence, Not the Brochure
Most ESG assessments start with a questionnaire the company fills out about itself. That is the wrong starting point. A score worth trusting starts with what independent sources say instead.
In short
5 things that decide this
- 01A sustainability report is the company describing itself, so an ESG score built mainly from it is closer to marketing review than research.
- 02A defensible score reads laws, certifications, academic studies, and news coverage from sources the company does not control.
- 03Controversies and certifications both need checking against a real registry or a real news record, not a claim on a company page.
- 04Coverage has to be wide enough to matter. A scan touching a handful of topics misses the one violation that changes the picture.
- 05We built this architecture for Greenlight, an ESG platform that scores companies from 10 to 15 independent sources per topic, across 50-plus topics.
A questionnaire is not evidence
Ask a company how sustainable it is, and it will answer well. That is not dishonesty. It is the nature of self-reporting. The company picks which numbers to publish, which incidents to omit, and which certifications go on the front page.
Many ESG scores still start there. A form goes out, the company fills it in, and the answers become the bulk of the score. A buyer reading that score is mostly reading the company's own summary of itself, dressed up as an outside review.
Greenwashing lives in that gap. A company can hold a genuine certification in one area and stay silent about a regulatory fine in another. A self-reported score has no way to notice the omission. The fix is reading sources the company cannot edit.
What a defensible score actually reads
A score built on outside proof pulls from legal filings, court records, studies, certification lists, and news. None of those sources answer to the company being scored. That is what makes them worth reading.
Coverage has to be wide, and outside sourcing alone is not enough. A scan of five topics can clear a firm on all five and still miss the one flaw in a sixth. Environment, social, and governance each split into dozens of narrow checks: labor rules, emissions, board makeup, and supply chains.
Certifications need the same scrutiny as controversies. A badge on a website is a claim. Checking it against the issuing body's own list is proof. Treat a claimed badge as a proven one, and you are trusting the brochure again.
Weighting matters as much as sourcing. Count self-reported material equally with independent findings, and a polished report can outscore a thin compliance record. The score has to lean toward evidence the company did not write.
- 01Regulatory filings and enforcement records, not the company's own compliance summary.
- 02Certifications checked against the issuing body's registry, not a badge on a webpage.
- 03News and controversy coverage, so a fine or a lawsuit cannot go unrecorded.
- 04Wide topic coverage, so a clean scan on five topics cannot stand in for the whole picture.
How we built this for Greenlight
We built Greenlight, an AI ESG and sustainability research platform, around this exact architecture. Enter a company name, and it runs a scan across more than 50 ESG topics. Each topic carries 10 to 15 independent sources with live, clickable citations. You can check any finding yourself, instead of taking a summary on faith.
The research pipeline runs on GPT-4 and Perplexity Sonar-Pro with retrieval grounding. Every finding stays tied to a source you can click and check. Scoring weights independent findings above self-reported material. Controversies get flagged on their own. Certifications get confirmed against real registries, not a company's own claim.
Grounding retrieval in real sources is what keeps an AI research system honest. Ask a model to sum up a company with no citation trail, and it will write a smooth answer either way, true or not. The hard engineering sits in retrieval, not in the write-up.
That discipline is the argument for any serious research automation, ESG or otherwise. Define the topics exhaustively. Require independent sources per topic. Weight the score toward evidence, and keep a citation trail a skeptic can follow. A single well-sourced answer beats a fast one nobody can check.
Questions this raises
01How does AI-driven ESG research differ from a self-reported questionnaire?
A questionnaire scores what a company chooses to share about itself. AI-driven research reads outside sources instead: legal filings, certification lists, studies, and news coverage. A claim has to survive a source the company does not control.
02What makes an ESG scoring methodology defensible?
Four things: wide topic coverage, multiple independent sources per topic, weighting that favors evidence over self-reporting, and a citation trail readers can follow. A score with none of those is a summary of the company's own marketing.
03How do you verify sustainability data at scale without losing accuracy?
Ground every finding in retrieval against real, checkable sources rather than letting a model summarize from memory. Greenlight backs each of its 50-plus ESG topics with 10 to 15 independent sources and live citations, so accuracy holds as the number of companies scanned grows.
