Which emission factor belongs on each row of the general ledger?

A group that has to report Scope 3 for its purchases has thousands of accounts and hundreds of thousands of transactions, and each one needs an emission factor. The choice decides both the total and where the group believes its emissions are, and so where it focuses its reduction efforts.

How it can be done

Four methods, four kinds of data

The GHG Protocol's Scope 3 standard, from 2011, and its technical guidance describe four ways to calculate the emissions from purchased goods and services, ordered by how specific they are to the supplier and the product (GHG Protocol, chapter 1). Each method needs its own kind of data, and in practice it is the data available that decides which one can be used.

  • The supplier-specific method uses the supplier's footprint for that particular product, from an environmental product declaration (EPD), a product footprint under ISO 14067 or an exchange format such as PACT.
  • The hybrid method uses the supplier's own Scope 1 and 2 emissions, allocated to what the customer bought, and average data for the rest of the chain.
  • The average-data method, often called an activity-based proxy, multiplies the quantity bought, in kilos, cubic metres or another unit, by an industry average from LCA databases such as ecoinvent or Boverket's climate database.
  • The spend-based method multiplies the amount in the general ledger by a factor per krona or euro for the product group, from input-output models.
Supplier-specific
The product's footprint, from the supplier
Hybrid
The supplier's scope 1 and 2, plus averages for the rest
Average-data
Quantities from the ERP system, times an average per unit
Spend-based
Amounts from the ledger, times an average per krona

The four methods, the most specific at the top. The bar is schematic: the more specific the method, the fewer rows it usually covers. The spend-based method covers them all.

Higher up the ladder means more specific, but not automatically more accurate. A figure from the supplier can be less accurate than an industry average, for example when the supplier has had to allocate its emissions to the products the customer buys, as in the hybrid method. That allocation can bring large uncertainty. The guidance therefore separates how specific a figure is from how accurate it is (GHG Protocol, chapter 1, box 1.1).

The standard is being revised. Under the plan from July 2026 it will be merged with the Corporate Standard, and the new standard is expected to be finished in the fourth quarter of 2028. Until then the 2011 rules apply (GHG Protocol, Standard Development Plan).

The whole chain, but calculated from the top down

The spend-based factors come from environmentally extended input-output models (EEIO). They calculate from the top down: a country's emissions are allocated to product groups according to how money flows between industries in the national accounts. The most widely used are EXIOBASE (Stadler et al., 2018), the US EPA's EEIO model for the United States and Watershed's CEDA (Watershed). EXIOBASE has its own tables for 44 countries and five regions for the rest of the world, and is licensed by XIO Sustainability Analytics.

Such a factor gives the average emissions in the whole chain behind one euro of a product group from a country, from raw material to factory gate. No upstream source is left out, and that is the model's strength. Process-based methods, which the more specific approaches rely on, work from the bottom up instead. They follow the chain step by step and have to stop somewhere. Lenzen showed that what is then left out can be in the order of 50 percent (Lenzen, 2000).

The factor is an average for the product group and says nothing about a particular product or supplier. Expensive and cheap steel get the same factor per euro, so a higher price looks like higher emissions, and a switch to a supplier with lower emissions does not show up if the price is the same. The guidance itself says that the method makes it hard to show a reduction (GHG Protocol, technical guidance). The factor is also calculated in a given year's prices, for example euros at 2022 values. Purchases from a later year therefore need to be adjusted for inflation, and converted from kronor to euros, before they are multiplied. It also applies to the price the producer receives, not the price the buyer pays, which is discussed under Where it gets difficult.

The factors are uncertain in absolute terms but more stable in how they rank. Different databases give different figures, and they differ most for individual product groups. When Li and Suh derived factors for goods imported by the United States from three databases, the results were on average between 8 percent below and 44 percent above the domestic factors, depending on the database (Li and Suh, 2025). Kilian, Owen and Hoekstra compared four databases and found that they agreed well for most countries, and that the same industries came out largest in all four (Kilian et al., 2026). So the choice of database affects the total more than what is large and what is small. The ranking is what the next step needs: deciding where better data makes the most difference.

Separating large from small

Starting coarse and putting the detail where it matters is both the standard's advice and an older idea. The standard says the company should first estimate everything with simpler data, noting that input-output data is particularly useful here, rank from largest to smallest and collect better data where it matters most (Scope 3 Standard, section 7.1). In life cycle assessment the same thing is called contribution analysis, and ISO 14044 counts it as the first step of interpretation (ISO 14044). Huang, Weber and Matthews showed that a company can capture a large share of its upstream footprint with complete data from a few direct suppliers (Huang et al., 2009). That matches what we have seen.

The analysis is done in four steps:

  1. The whole general ledger is calculated with spend-based factors, so that nothing is missing.
  2. The accounts are ranked by estimated emissions and plotted as a curve, which shows how few accounts carry how much. The ranking is by emissions, not by spend, since the two do not always track each other, something the standard itself warns about.
  3. A threshold is set. The standard's example takes the suppliers that together make up 80 percent of spend, any other supplier above one percent and those that matter for other reasons (Scope 3 Standard, box 7.3).
  4. Better data is collected for what is above the threshold. The analysis is repeated every year, since the ranking shifts.

Where the threshold should go also depends on who is asking. A buyer can change supplier, a design engineer can change material and a production manager can change the company's own processes, and each needs to see the part of the chain they can influence. In life cycle assessment this is called the actor's perspective: the results are organised around the decisions that are actually made (Löfgren et al., 2011; Baumann et al., 2011). The standard points the same way, since it lets reduction opportunities and business goals decide where better data is needed, not size alone.

The tail, what lies below the threshold, keeps its spend-based factor. It cannot simply be dropped, since the standard requires all Scope 3 emissions to be reported and every exclusion to be disclosed and justified (Scope 3 Standard, section 6.2). With spend-based factors in the tail the total is complete at a low cost, and the uncertainty ends up where it matters least.

The matching

In practice the rows are linked to product groups manually by consultants, or with rules that match account numbers and keywords. Research has begun to test language models for the task. Jain et al. fine-tuned models on examples they created themselves. The models place descriptions from the general ledger in the EEIO model's product groups, and according to the authors they beat simpler methods and reached the level of a domain expert (Jain et al., 2023). ATLAS, a benchmark with 10,000 synthetic purchase rows, was published in 2024 on the grounds that there was no shared way to compare such solutions (ATLAS, 2024).

Where it gets difficult

There are three kinds of difficulty: knowing what was bought, making sure the factor fits the purchase and being able to compare the figures from year to year.

Knowing what was bought

An account rarely holds one kind of purchase. An account for consumables can contain gloves, cutting fluid and workwear from forty suppliers. The account name and the purchase texts are short, and the supplier's name rarely says what it sells. Two knowledgeable people split the same account differently, and both can defend their choice. So there is no right answer to check against, only choices that can be judged.

An error here can hide. If a large purchase lands in a product group with a low factor, it looks small and ends up in the tail, where nobody looks closer. The matching therefore has to be good even for rows that will never get better data.

Making sure the factor fits the purchase

Three things can make a spend-based factor fit worse than it seems.

The country. The factor can apply to the country where the product is made or the country where it is bought. With the producing country's factor, imported steel is counted as if it were Swedish, made with Swedish electricity. Where a product was made is rarely clear from a general ledger, and then XIO recommends an average for what is bought in the country, including imports (XIO, Choosing the right emission factor).

The price. The factors apply per euro of what the producer is paid. The general ledger shows what the buyer paid, with trade margins, transport and tax. If the whole amount is counted as if it were the product, the margin is counted as manufacturing, and some products come out too high while others come out too low (XIO, Prices, inflation and time).

The detail. The models are built to be right for whole countries and the whole economy, not for every product in every country. A single factor can therefore be unreasonable, and for some products in some countries it is missing altogether (XIO, Data quality and reliability).

Comparing from year to year

A factor higher up the ladder can be missing parts of the chain. The spend-based factor includes everything the industry buys, at every step. The supplier's EPD and the proxies in LCA databases are instead based on life cycle assessments of the processes, and they leave parts of the chain out for two reasons.

The first is the rules. A product footprint only needs to include what becomes the product, makes it or transports it, and not machinery and facilities, administration or staff travel. The GHG Protocol's product standard lists what may be left out (Product Standard, section 7.3), and industry rules such as the automotive industry's Catena-X do the same (Catena-X, section 4.2). The second is that the analysis has to stop somewhere, since in practice it is not possible to map every step back through the chain. That is the gap Lenzen described.

When a large account switches to such a factor, emissions can therefore look lower without anything having changed at the supplier.

The input-output models behind the spend-based factors also change. Each version of EXIOBASE, for example, contains the whole time series from 1995 and is recalculated in full. A new version replaces projected years with measured ones, corrects input data and sometimes changes the method, for all years at once (EXIOBASE, version 3.11.2). The spend-based factors then change even for purchases that are unchanged, and only the part that comes from new measured data reflects a real change. In one case we have calculated, the median factor fell by about a fifth, because the database had revised the size of the economy upwards. Sometimes what is counted changes: in a later version a row was added for carbon dioxide from burning biofuel, and it is carried into the factor unless someone takes a position on it.

That makes it hard to track a target. Every switch moves the figures, whether it is a new version of the database or a move from amounts and a spend-based factor to quantities and the supplier's own factor, so both the result and the baseline it is compared with shift. With spend-based factors, emissions also only fall if you spend less.

The base year is the year the target is measured from, and it stays fixed. The figure for the base year is another matter, and there the rules differ. The GHG Protocol requires it to be recalculated when a changed method or better factors affect it significantly, but not when the change reflects a real reduction. If only the latest year is recalculated with better data, emissions appear to have changed more than they have (EY, Greenhouse Gas Protocol: Interpretative guidance, section 2.7.2.1; Corporate Standard, chapter 5). SBTi sets the threshold for significant at five percent (SBTi, criterion C27).

ESRS is less clear. The target's baseline value should only change if the target or the reporting boundary changes significantly, chiefly which companies are included. A new method or new factors are not mentioned (ESRS E1, AR 25). At the same time, the undertaking is to consider the GHG Protocol when calculating emissions (ESRS E1, AR 39), and it requires recalculation. We have not found any guidance from EFRAG or the audit firms on how the two should be read together. In their sustainability statements for 2025, H&M, Schneider Electric and SKF recalculate the base year when the method changes, without addressing what ESRS says.

Our approach

The method began in Peab's general ledger, which we have worked with since 2024 together with Peab's own experts and the LCA consultants at Miljögiraff (the customer case). Since then it has been developed with customers in other industries and now works for more kinds of data than accounts: transactions, product lists and bills of materials. It separates large from small, and uses process-based factors for what dominates, from an LCA database such as ecoinvent or from the suppliers' own data.

Each row is first broken down into what it contains: which product or service, which materials and in what proportions. The factors are linked to the materials, not directly to the account. Two purchases on the same account can then have different contents, the same method works for every kind of data, and the breakdown can be reviewed on its merits. A row is first matched against what has already been judged, by name and then by meaning, so that "wellpapplåda 320x225 mm" and "corrugated cardboard box" end up on the same item. If there is no reliable match, a language model proposes the contents. The proposal is marked as an estimate until a person has approved it, and every row is saved with method, confidence and reasoning. What has once been approved is used again the next time the same purchase turns up.

ROWPRODUCT OR SERVICEMATERIALFACTORAccount 4010 Building materialsAccount 4010 Building materials31 t CO2eAccount 5460 ConsumablesAccount 5460 Consumables9 t CO2eBill of materials, switchboardBill of materials, switchboard5 t CO2eTransaction, machine hireTransaction, machine hire5 t CO2eAccount 6990 OtherAccount 6990 Other2.5 t CO2eConcrete C30/37Concrete C30/37Reinforcing barReinforcing barScrews and fastenersScrews and fastenersWorkwearWorkwearSwitchboardSwitchboardMachine hireMachine hireOffice supplies, proposedOffice supplies, proposedAI, not approvedCementCementAggregateAggregateSteelSteelCopperCopperTextileTextileService, not broken downService, not broken downPaperPaperCementCementecoinventGravel and crushed stoneGravel and crushed stoneecoinventReinforcing steelReinforcing steelEPDFabricated metal productsFabricated metal productsEXIOBASECopperCopperecoinventTextilesTextilesEXIOBASERenting of machineryRenting of machineryEXIOBASEPaper productsPaper productsEXIOBASE

Five rows, three accounts, a bill of materials and a transaction, broken down into product, material and factor. The steel gathers three rows and gets two factors: the supplier's EPD where it matters, EXIOBASE in the tail. The rows and values shown are made up.

Where the volume of data is large, the model is trained on the experts' judgements. A sample of rows is drawn so that rows with a larger estimated impact are more likely to be picked, the experts judge them, and the model is retrained until the difference between two versions is below a threshold set in advance. So the model learns round by round, not once as in Jain et al.

The model reads more rows than anyone has time for, but the choice of method, system boundaries and validation stay with people. The model does not decide which question the figure should answer, or for whom.

For each factor table the choices are written down, so that it is possible to see how country and price have been handled: database and version, price base and whether the amounts are purchaser prices, whether the factor applies to the producing country or the market where the product is bought, which gases are included and how biogenic carbon has been handled.

So that a target can be tracked, the factor version is locked per reporting year. A new version is tested once a year by applying both the old and the new one to the same year's purchases. The difference is split into the effect of the new year and corrections to earlier data, and only the first reflects a real change. The corrections are added up year by year (Corporate Standard, chapter 5). As long as the sum stays below the threshold, the switch can wait, and several changes can be made in one step. When it reaches the threshold, the version is switched and the comparison year and the base year are recalculated in the same year. Targets are set on the suppliers' own figures, not on the spend-based ones.

Separating large from small means that the factors change precisely where they matter most, and that is where targets are tracked. Every time a large account gets more specific data the figure moves, and the new factor can be missing parts of the chain. The switch is therefore shown account by account, and if the difference is significant the base year is recalculated with the new factor.

Often that is not possible, because the switch involves more than the factor. Instead of amounts from the general ledger, quantities are used, for example kilos from digital delivery notes, and instead of a factor per krona the supplier's EPD per kilo is used. For the base year there are neither quantities nor an EPD, only amounts. There are then two ways forward: estimate the base year's quantities from the amounts and apply the new factor to them, assuming that prices and suppliers were the same then, or leave the base year as it is and disclose when the new data began to be used and how much of the reduction it explains (EY, section 2.7.2.1). The first erases a real improvement at the supplier. In the second, better data shows up as part of the reduction, even if it is disclosed. The share of purchases calculated with the suppliers' own figures can also be tracked separately, alongside the emissions. None of this solves the problem, and it is one of the open questions below.

What we saw

Emissions are concentrated, in every dataset we have calculated. In one group's general ledger, three percent of the accounts made up a quarter of the emissions. In the open supplier ledgers of seven Swedish municipalities, 3 to 23 suppliers make up half of the footprint and 23 to 118 make up four fifths, calculated with the National Agency for Public Procurement's environmental spend indicators (Upphandlingsmyndigheten, miljöspendanalys). That is where better data makes a difference, since it is needed to find the causes behind the emissions and see where action is possible.

020406080100120020406080100number of suppliers, largest firstcumulative share of CO2e (%)3 to 23 suppliers account for halfof 1,800 to 4,000 per municipality

Cumulative share of the climate footprint against number of suppliers, largest first, in seven municipalities in 2025. One line per municipality, ending where the municipality passes 80 percent. The dot is where it passes half. Calculated with the National Agency for Public Procurement's environmental spend indicators on the municipalities' open supplier ledgers.

Spend and emissions do not track each other. In the municipalities, food, electricity and district heating are a few percent of the spending but a quarter to a third of the emissions, while purchased services such as schools and care are large in money but not in climate impact.

An account's name says little about what is bought. What matters is in the suppliers and the rows behind the account, and there are too many of them to read by hand. A model can read them all and finds things the account text hides, as when an account for freight turns out to be public transport. It also makes mistakes that sound reasonable, as when workwear from Björnkläder (literally "bear clothes") became fur goods, and only someone who knows what the account is for will spot them. That is why people review, and why the reasoning is saved on every row.

Which factor applies?

It depends on how much the row contributes, and for whom. For most rows the answer is a spend-based factor for the right product group, country and price base, with the version locked and the choice justified, so that the total is complete and can be reviewed. For the few rows that carry most of the emissions, that is not enough. There a factor is needed that says something about the particular product or supplier, since that is where the causes are and where action is possible, and it needs to be read knowing what it leaves out. None of these choices has a single right answer. A choice holds up when someone who knows the subject can follow the reasons.

How far that answer goes in practice depends on four questions the field has not yet answered.

How accurate is a choice of factor? The differences between databases are known for countries and industries but not for a single row in a general ledger. Today the choices are judged by the company's own experts, on a sample of rows. We know of no comparison of how different experts, in different companies, choose for the same row.

Where is the line between tail and detail, and for whom? The ranking shows what is large for the group. For a subsidiary, a buyer or a design engineer something quite different can be large, depending on their targets and what they can influence. We do not yet know how to rank the same ledger for each of them, or how much better data would be worth to each.

How does a real reduction show up? The reduction is what the group should steer towards. With spend-based factors it is barely possible, for the reasons above. With the other methods it works better but is still hard, since they all rest on assumptions that are hard for the buyer to see through: industry averages that do not show what a particular supplier has done, and system boundaries and cut-offs in every EPD. The rules for version changes exist, but separating what has happened in the chain from what follows from the assumptions is still unsolved.

How do the suppliers' own figures come in? For what matters most, no proxy is enough. The figures need to come from the suppliers, and that takes more than a way to share them. It also takes a shared way to calculate, where every step calculates its own emissions with the same factors and allocates them to its customers, so that the figures can be added up along the chain without anything being counted twice or left out. That is what MASSIV+ is working on.

Sources

  • GHG Protocol. Corporate Value Chain (Scope 3) Accounting and Reporting Standard, 2011. Sections 6.2, 7.1 and 7.3. ghgprotocol.org
  • GHG Protocol. A Corporate Accounting and Reporting Standard, revised edition, 2004. Chapter 5. ghgprotocol.org
  • SBTi. Corporate Near-Term Criteria, version 5.3.1, 2026. Criterion C27. sciencebasedtargets.org
  • European Commission. Delegated Regulation (EU) 2023/2772, ESRS. ESRS E1, AR 25 and AR 39. eur-lex.europa.eu
  • EY. Sustainability reporting developments: Greenhouse Gas Protocol, interpretative guidance, August 2026. Section 2.7.2.1. ey.com
  • Sustainability statements for 2025: H&M, Schneider Electric, SKF
  • GHG Protocol. Corporate Standard - Standard Development Plan, version 2.0, 29 July 2026. ghgprotocol.org
  • Boverket. Klimatdatabas. boverket.se
  • GHG Protocol. Technical Guidance for Calculating Scope 3 Emissions, 2013. The introduction on EEIO data and chapter 1: Purchased Goods and Services. Full guidance, chapter 1
  • Stadler, K. et al. EXIOBASE 3: Developing a time series of detailed environmentally extended multi-regional input-output tables. Journal of Industrial Ecology, 2018. doi.org
  • Watershed. CEDA, Comprehensive Environmental Data Archive, and the free version Open CEDA. watershed.com
  • Lenzen, M. Errors in conventional and input-output-based life-cycle inventories. Journal of Industrial Ecology, 2000. doi.org
  • Kilian, L., Owen, A. and Hoekstra, R. Quantifying the impacts of multi-regional input-output data on emissions from imports. Economic Systems Research, 2026. doi.org
  • Li, M. and Suh, S. Comparing CEDA, GLORIA and EXIOBASE for greenhouse gas emissions embodied in imports by USEEIO model. Conference paper, International Input-Output Association, 2025. Pdf
  • ISO 14044. Environmental management, Life cycle assessment, Requirements and guidelines. iso.org
  • Huang, Y. A., Weber, C. L. and Matthews, H. S. Categorization of Scope 3 emissions for streamlined enterprise carbon footprinting. Environmental Science & Technology, 2009. doi.org
  • Löfgren, B., Tillman, A.-M. and Rinde, B. Manufacturing actor's LCA. Journal of Cleaner Production, 2011. doi.org
  • Baumann, H., Berlin, J., Brunklaus, B., Lindkvist, M., Löfgren, B. and Tillman, A.-M. The usefulness of an actor's perspective in LCA. In Towards Life Cycle Sustainability Management, Springer, 2011. doi.org
  • Jain, A. et al. Supply chain emission estimation using large language models. arXiv 2308.01741, 2023. arxiv.org
  • ATLAS: A spend classification benchmark for estimating scope 3 carbon emissions. Climate Change AI, NeurIPS 2024. Pdf
  • GHG Protocol. Product Life Cycle Accounting and Reporting Standard, 2011. Section 7.3. ghgprotocol.org
  • Catena-X. Product Carbon Footprint Rulebook, version 4, 2025. Section 4.2. catena-x.net
  • Upphandlingsmyndigheten. Miljöspendanalys: beräkna och analysera inköpens klimatpåverkan. Miljöspendindikatorerna, version 4, 2025. upphandlingsmyndigheten.se
  • XIO Sustainability Analytics. Knowledge hub: Choosing the right emission factor; Prices, inflation and time; Data quality and reliability. Accessed October 2026. xio-sa.com
  • Icons Of. Customer case: Expert judgement from Peab, scaled to thousands of accounts. iconsof.se