How the database is built

What happens between a company’s published report and the figures shown on this site.

What the database contains

The database collects public country-by-country reports, in which a multinational group discloses, for each tax jurisdiction where it operates, its revenue, profit before tax, corporate income tax paid and accrued, and number of employees. Some of these reports are mandatory. The EU public country-by-country reporting directive, formally Directive (EU) 2021/2101, requires large multinationals active in the European Union to disclose this information for financial years starting on or after 22 June 2024. The rest are voluntary: most follow the tax transparency standard of the Global Reporting Initiative, known as GRI 207-4, and some appear under national transparency rules, such as the Spanish law on non-financial information, Ley 11/2018.

We do not cover the confidential country-by-country reports that large groups have filed with tax authorities since 2016 under the OECD’s rules against base erosion and profit shifting, known as BEPS Action 13. They become public only when a company chooses to publish its report itself. Payments-to-governments disclosures, which the EU requires separately from oil, gas, mining and logging companies, follow a different regulation and are excluded as well.

1Collect: gathering the reports

We look in two main places. Where publication is mandatory, we go to the official source: the business registers in which companies file their reports. For voluntary reports, we go to each company’s own website, its investor relations and sustainability pages, since that is where groups tend to put them. Because a disclosure may sit deep inside a long annual report, or behind a title in another language, our searches use a vocabulary of report titles, table captions and field labels collected from real reports in many languages: a report called Informe de transparencia fiscal is found as readily as one called Tax Report.

Every company and year we have examined is tracked in a coverage ledger. When nothing turns up, we record that outcome too. An absence from the database means we looked and found no public report, not that we assumed there was none. It cannot rule out a disclosure sitting somewhere we did not look.

2Read: copying each table as printed

Numbers are copied exactly as printed, in the currency and unit of the source document. Reports arrive in every format a company might choose:

• PDF documents, from clean digital files to scanned pages that only exist as images;
• structured register filings in xHTML and inline XBRL;
• ordinary web pages, saved as published;
• spreadsheets, where a company releases its table as a file.

Each format is read in the most reliable way it allows. A register filing tagged in inline XBRL, or a spreadsheet, is parsed by deterministic code, with no interpretation involved. For the rest, above all PDFs, whose tables come in countless layouts and languages, we use artificial intelligence. A blank or unreadable cell stays missing. We never substitute a zero and we never estimate.

3Verify: checking every number against its source

Every value the AI reads is checked by code against the text of the document itself; anything that cannot be confirmed is flagged for human review rather than accepted. Each value is also linked to the file, page and table cell it was printed in, and the original documents are kept unchanged, so every figure can be traced back to its source at any time.

4Standardise: making the reports comparable

Companies name the same variables in many different ways and use different currencies and units. Six operations make the reports comparable:

• variable names, in whatever language, are mapped to a single vocabulary;
• financial years are labelled by one rule: a year takes the calendar year in which it ends, unless it ends before June, in which case it takes the year before, so that a year running from April 2024 to March 2025 is labelled 2024;
• jurisdiction names are mapped to standard three-letter country codes, and rows in which a report lumps several jurisdictions together keep an explicit ‘Other’ label instead of a guessed split;
• the unit scale is read from the report’s own wording and confirmed a second time, because a scale mistake would multiply every amount by a thousand or a million;
• amounts are converted to euros at official reference rates, mainly those of the European Central Bank;
• tax figures follow one sign convention, tax paid positive and refunds negative, checked, where possible, against the group’s audited accounts.

What we add to the reports

A few pieces of information are added alongside what the reports state:

• each group is assigned to one of twenty sectors, based on how the group describes itself and on official industry classifications, and every classification is recorded with its source; the sector describes the whole group, not any single subsidiary;
• each group carries the country of its ultimate parent, the headquarters shown throughout the site;
• the ultimate parent undertaking is identified for reports filed under the EU directive, which name it, but not for the other formats.
• each group-scope report carries a transparency score, defined below.

5Check and publish: testing every release

Before a release the whole database passes around forty automated tests. Among them:

• jurisdiction rows are reconciled with the totals the report states, where it states them;
• we look for duplicates, the same report entering the database twice under different names;
• each company’s figures are compared across its years, so that an implausible jump is reviewed before it is believed.

Where the automatic check against the document text cannot run, the report is read twice, independently, and the two readings must agree. For selected groups we also check the reported totals against the group’s audited consolidated financial statements. Finally, before a new release replaces the previous one, the two are compared cell by cell. Every difference must have a known cause, a newly added report or a documented correction; an unexplained change stops the release.

The transparency score

Alongside the reported values, each group-scope report carries a transparency score, designed to measure the extent to which a multinational discloses financial information across different jurisdictions. The score was developed for the EU Tax Observatory’s report Advancing Corporate Tax Transparency (Aliprandi and Borders, 2024). It is calculated from the disclosure of a set of predefined financial variables across jurisdictions, with higher scores indicating greater transparency: 0 is the lowest score and 100 the highest. The calculation follows the general formula:

where:

• n is the number of financial variables, the ten of the OECD reporting standard;
• w_i is the weight assigned to the i-th financial variable; all variables are equally weighted, with w_i = 1/n;
• J_i is the set of jurisdictions for which the i-th financial variable is disclosed, excluding the aggregated categories;
• x_ij is the value of the i-th financial variable for jurisdiction j, in euros.

A score of one hundred therefore describes a report that publishes, for every jurisdiction separately, the whole set of items that the OECD standard on country-by-country reporting (Action 13) requires groups to file confidentially with tax authorities. That confidential template, made public in full and with no jurisdiction grouped into a residual, is the reference point the scale is built on.

Three rules make the score careful with missing information. A variable the report does not disclose contributes zero, and is never estimated. Absolute values are used, since a disclosed loss locates activity exactly as a disclosed profit does. And the score is computed only for reports whose scope is the whole group: for partial filings, such as a subsidiary publishing without group data, the group share cannot be calculated, and the score is left empty rather than set to zero. An empty score means the calculation is impossible; a zero means a group-wide report that locates nothing at country level.

Whether a report covers the whole group is established from what the report says about itself, as much as from its figures. Filings frequently state their own perimeter: that they cover the entities of a group whose management and tax information is accessible from one member state, that the parent did not provide group data within the filing deadline, or that the group’s own report is published elsewhere. Those statements are read from the document and recorded with the passage they come from. Where a report says nothing about its scope, the figures are used instead: a report naming a single jurisdiction with no aggregated line, or whose rows fall short of the total it prints itself, does not describe the whole group.

How to read a low score. The score describes the document, not the company. It measures how much of a group’s activity is attributed to named countries rather than to an aggregate, and that depends first of all on the format the report follows. A report filed under the EU directive breaks out the member states and the jurisdictions on the European list of non-cooperative territories, and is permitted to group everything else into a single residual line. A worldwide group filing in that format therefore scores low even though it discloses exactly what the law requires, while a voluntary report that itemises every country in which the group operates scores higher. Apple’s 2025 filing is the clearest illustration: it scores 18 on this scale and complies fully with the directive it was filed under. A low score describes the reporting format, not the company. The scale is benchmarked on the OECD template, which no public disclosure law requires, so a group that meets the EU directive in full still does not reach one hundred. A low score is therefore not a finding of non-compliance.

The badge. Each report carries a short label showing where it falls on this scale, with the percentage beside it. Hovering over the label explains what the figure means for that report, and names the regime the report was filed under, because the same percentage carries a different meaning under a rulebook that permits aggregation than under a voluntary report whose author chose its own perimeter.

Corrections

The database is updated regularly and errors are fixed as they surface. If you spot an imprecision, or a report we have missed, write to taxobservatory@gmail.com. When a company reissues or corrects a report, the newer document replaces the older one.

Taxplorer is developed within the EU Tax Observatory. If you use the data, please cite: Aliprandi, G., Borders, K., Gabriel, F., von Zedlitz, G. (2022): “Public Country-by-Country Reports: a new database”, EU Tax Observatory.