How we calculate
your carbon emissions.

Every Carbon Draft report aligns to the GHG Protocol Corporate Accounting and Reporting Standard (2015 Revised Edition) — the framework used by over 92% of Fortune 500 companies that report emissions. Developed by the World Resources Institute (WRI) and the World Business Council for Sustainable Development (WBCSD), it underpins CDP, the Science Based Targets initiative (SBTi), California SB-253, the EU CSRD, and the US SEC climate rule. When a supplier questionnaire or disclosure filing asks for Scope 1, 2, and 3, this is the accounting grammar they mean.

For Scope 3 in particular we additionally apply the Corporate Value Chain (Scope 3) Standard (2013) — and its §7.3.2 economic-value method, which governs how spend-based estimates are constructed, how refunds and credits are netted, and what uncertainty ranges are acceptable.

Spend-based carbon accounting does something simple by design: every dollar spent in a commodity category is multiplied by an industry-average emissions intensity (kg CO₂e per dollar) to estimate the embedded emissions. The EPA USEEIO model supplies these intensities for 405 NAICS commodity codes, derived from supply-chain linkage analysis of the entire US economy.

This trades precision for coverage. An activity-based measurement (therms of natural gas burned, gallons of diesel pumped, kWh of electricity metered) gives ±5–10% accuracy per GHG Protocol guidance; a spend-based estimate from the same data category gives ±25–40%. The tradeoff: activity data is expensive and slow to collect; spend data already sits in your general ledger and can be classified in minutes. For a screening-grade response to a supplier questionnaire or an internal baseline, spend-based is the correct tool. For formal third-party assurance under CSRD or the SEC climate rule, high-impact categories should upgrade to activity data.

Credits, refunds, and returns are netted within their emission category before the factor is applied — per GHG Protocol Scope 3 §7.3.2. This aligns with GAAP ASC 606 and IFRS 15 net-revenue treatment, so category totals reconcile to the general ledger by construction. The Data Quality Summary on page two of every disclosure-aligned report discloses each netted row explicitly.

We combine four published factor sets. No proprietary multipliers, no black-box coefficients. Every rate below is reproducible from public data.

Every vendor in your CSV needs to be mapped to a NAICS commodity code before a factor can be applied. Our classifier runs three layers, in order, with each layer catching what the previous one missed:

1. Regex pre-pass — 36 compiled patterns for accounting-system transaction prefixes (AMZ*, SQ *, PAYPAL *, SHELL-OIL, MSFT AZURE, airline ticket prefixes, ride-share tags). Catches ~22% of messy AP data instantly. ~5 µs per row.
2. Normalized substring — vendor name is stripped of LLC/INC/LTD suffixes, transaction refs (#4421), long digit runs, then matched against a 240-entry known-vendor dictionary. Catches ~58%.
3. Fuzzy match — rapidfuzz Levenshtein with a threshold of 85. Catches typos (gasoln → gasoline), truncations, and OCR scan artifacts. Safety net for the remaining ~10%.

On the paid Carbon Draft tier, unresolved rows additionally get Claude classification for context-aware decisions the keyword layers can't handle — vendor normalization ("SHELL OIL STN #4421" → Shell), ambiguous-category resolution, and non-English supplier names. Each row gets a per-row confidence score (0.0–1.0); the Data Quality Summary discloses what confidence bracket each row landed in.

On our 3,500-row Nightmare Spend Corp test fixture, offline classifier produces 84.2% high-confidence (≥0.80), 9.1% medium, 6.7% low. With Claude enabled on the paid tier, high-confidence rises to ~93%. Low-confidence rows are surfaced explicitly so an auditor or the customer can review them — never silently hidden.

The upload accepts CSVs from any general-ledger export. Two columns are required (vendor/description + amount); everything else is optional. The intake layer auto-detects:

• File encoding — UTF-8 and Windows-1252 (the latter is a common QuickBooks export default that breaks most parsers)
• Decimal convention — US (1,234.56) vs European (1.234,56)
• Negative notation — hyphen (-1234) vs parens-accountant ((1234))
• Localized headers — German (Lieferant → vendor, Betrag → amount), French, Spanish header maps
• Date formats — MM/DD/YYYY, DD/MM/YYYY, YYYY-MM-DD ISO, free-form strings
• Multi-currency — FX conversion to USD at the vintage stamped on the report (current: February 2026 rates)

Every decision the intake layer makes is logged to the Data Quality Summary so auditors can reconcile. If the parser has to guess (e.g., a file with no column headers), the guess is disclosed.

Spend-based accounting is a screening method. The GHG Protocol recognizes this explicitly — per Scope 3 Standard §7.3.2, economic-value data carries ±25–40% uncertainty at the category level. Activity-based data (therms, gallons, kWh) is tighter at ±5–10%. Don't let anyone tell you otherwise.

A screening-grade Carbon Draft is the right tool for:

• Responding to supplier sustainability questionnaires (Walmart Project Gigaton, Microsoft Supplier Sustainability, Apple Clean Energy, CDP Supply Chain)
• Establishing a carbon baseline for internal strategy
• Identifying emission hotspots before making reduction investments
• Preparing for a third-party assurance engagement (the Disclosure-Aligned Report tier eliminates 60–80% of the preparation work an auditor would bill for)

It is not a substitute for:

• Third-party assurance (ISO 14064-3) on regulated filings
• Science-Based Targets initiative (SBTi) validation
• Activity-data-grade precision on regulated reporting (CSRD mandatory disclosures, SEC climate rule formal filings when live)

Every change to the classifier, factor tables, or intake layer is stamped and versioned. Prior reports remain reconcilable on their original engine version — we don't silently rewrite history.