Methodology

This page explains how the ECGT Ready risk score is built. The aim is to be specific enough that you can challenge an output and confident enough to use the score as input to a real decision.

1. Inputs

• The URL or pasted page text you submit.
• The category you select (apparel, food, electronics, beauty, other).
• Optional: market (LU, FR, DE, ...) which affects which national rules layer on.

2. The rule set

The engine checks each claim on the page against rules sourced from:

• Directive (EU) 2024/825 on Empowering Consumers for the Green Transition (ECGT).
• Directive 2005/29/EC on Unfair Commercial Practices, including Annex I item 4a (generic green claims).
• Article 7a UCPD on environmental labels and certifications.
• Article 1 amendments on future-environmental-pledges.
• National guidance (CNPD, DGCCRF, BEIS) where it adds enforceable detail.

3. How the score is computed

3.1 Detection

The page text is split into claim-bearing sentences. A claim is a sentence that asserts an environmental property (e.g. eco, green, carbon-neutral, recyclable, sustainable, plastic-free, biodegradable, 100% natural).

3.2 Classification

Each detected claim is classified as one of: explicit (a measurable property), implicit (a vibe-ier formulation), or label/cert (named certification).

3.3 Risk weighting

• Generic green words without substantiation: high.
• Future commitments without an interim plan: high.
• Self-issued certification labels: high.
• Explicit claims with measurable substantiation linked: low.
• Recognised third-party labels (EU Ecolabel, Blue Angel): low.

3.4 Score

Page-level score is a weighted average across detected claims, capped and clipped to a 0-100 scale, with a 0-100 interpretation guide: 0-30 likely fine, 31-60 needs work, 61-100 high risk.

4. What the score does and does not mean

• It is an indicative risk assessment, not a legal certification.
• A low score does not guarantee no regulator will challenge the page.
• A high score does not mean the page is illegal; it means the language is risky and should be substantiated or rewritten.
• We do not publish a passing label. There is no ECGT Ready badge for stores.

5. Challenging a flag

Every flag in the report has a Disagree button. Clicking it logs your rationale and updates the engine's training data. We review contested flags weekly. Patterns that recur lead to rule revisions, which are noted in the changelog at the bottom of this page.

6. Limits and known weaknesses

• The engine reads HTML; image-only claims (text baked into a banner) are missed.
• Multi-language pages are scanned as the language they are written in. Mixed-language pages can throw off classification.
• Industry-specific certifications outside the recognised list default to medium risk while we expand the catalogue.

7. Drafted fixes

For every finding the engine returns a drafted paragraph the merchant can paste into the relevant page. The drafted text is generated by the same large language model and is intended as a starting point for the merchant to adapt, not a finished legal text. The drafted fix sits next to the finding in the report and is also exported with the dossier.

8. Verification re-scan

After pasting a fix the merchant can trigger one verification re-scan per finding at no extra charge. The re-scan re-fetches only the page the finding was on and re-runs the diagnosis. The original finding is then marked passed (the issue no longer appears) or failed (it does). Additional re-scans require a fresh full scan against the monthly quota.

9. AI involvement

Both the classification step and the drafted fix step use a large language model. See our AI Act notice for the transparency disclosures and the model provider.