Methodology · How we know this

How these figures are built

Every page on this site is assembled from public US IRS filings and government spending records. Nothing is licensed from a directory and nothing is hand-entered. This page is the standing reference for how the raw filings become the grant figures, funder profiles, government-funding layers, board and asset layers, and the co-funder graph, and for the limits we hold ourselves to.

The sources

The spine is the IRS Form 990 / 990-EZ / 990-PF e-file XML: the machine-readable returns the IRS publishes for electronically-filed nonprofit and private-foundation tax returns. To those we add the IRS Business Master File for organization identity (name, EIN, location, ruling), and the Pub 78 and automatic-revocation lists for tax-deductible status. Sanctions screening uses the OFAC Specially Designated Nationals list. Those are all primary government data.

The filings are not the only source, and the rest are not all government. The need overlay draws on Census ACS and SAIPE, CDC PLACES, the USDA Food Access Research Atlas, Eviction Lab at Princeton, and the Opportunity Atlas tables the Census Bureau publishes. Government funding comes from USAspending and nine state transparency portals. Partnerships are extracted by a language model from organization websites and public news, and portfolio themes come from text embeddings. Several of those are modeled estimates rather than measured counts, and one describes a birth cohort from the early 1980s. Every source is listed with its publisher, how it joins, and where it stops being reliable.

Grants and recipients

Grants are read from the grant schedules inside each filing: Form 990 Schedule I and Form 990-PF Part XV. Each recipient is reconciled to its EIN where the filing reports one, which is what lets us link a funder to the recipient’s own filings and to other funders of the same organization.

Where the filer supplies no EIN, we resolve the name ourselves. This is a large share of the corpus: it is the difference between a key join and an inference. The recipient name, city and state are matched against IRS organization records; a match is kept only when it clears a confidence threshold, and below that the recipient stays a named, unlinked record rather than a guess. So the graph is part deterministic and part resolved, and a resolved edge is a judgement we have measured rather than a fact the filer stated. How much of the graph came from each, and the precision of the matching against held-out filer-supplied EINs, are published on data quality. How that matching works, and which grants no resolver could link, is on matching.

Charitable status screens

Separately from the grant graph, the API looks an organization up across seven public IRS, Treasury and California files before a grant is made: exemption status, Publication 78, auto-revocation, OFAC, the Internal Revenue Bulletin, and California registration. Three of those files join on EIN; the rest can only match a name. A source that is not on disk reports not screened, never clear. How the screen works is on the screening page.

Government funding (federal and state)

Alongside private grantmaking, we layer in government funding to nonprofits. Federal awards come from USAspending.gov, the official record of federal spending: prime grants and contracts (and first-tier subawards) to an organization, by agency, program, and year. USAspending identifies recipients by their federal UEI (and legacy DUNS), not by EIN, so the link to a nonprofit’s 990 identity is a confidence-scored name-and-location match rather than a key join: coverage is partial by design and we favor precision over recall, omitting an award rather than attaching it to the wrong organization.

State funding is pulled from individual state-government transparency portals, grants and payments from state agencies to nonprofits, and matched to the recipient by EIN where the portal publishes one. There is no national standard, so each state is a separate source and coverage varies from state to state; an absent state means we have not yet ingested it, not that no funding exists. As with every figure here, government amounts are dated to the award or fiscal year they come from.

Classification and geography

Grants are classified into cause areas using the IRS National Taxonomy of Exempt Entities (NTEE), and a foundation’s service area is resolved against a place gazetteer to county (FIPS) level where the filing and website support it. Both are confidence-tiered; inferences that rest on weak evidence are framed as questions or withheld.

The co-funder graph

Two funders are linked when their filings report grants to the same recipients. A link’s strength is a shrinkage-adjusted co-funding rate, so a single shared grantee does not read as a strong tie. The graph is a map of association, shared grantees, and not a claim of coordination between funders.

Boards and people

Every Form 990 lists an organization’s officers, directors, trustees, and key employees (Form 990 Part VII, and the equivalent schedule on 990-PF and 990-EZ). We read that roster directly, which is what lets us show who governs a funder and, by matching the same person across organizations, where boards interlock. People are matched on normalized name and, where the filing supports it, address; a shared-trustee link is read as association, evidence that two organizations share governance, never an inference about what was decided between them. Common names are matched conservatively and a thin or ambiguous match is withheld rather than asserted.

Assets and investments

For private foundations, the 990-PF balance sheet and investment schedules report what the endowment holds. We read the itemized holdings, named funds, managers, and securities, and group them into asset classes and recurring entities so a foundation’s portfolio can be summarized and compared with its peers. Holding names on a tax form are inconsistent, so the entity grouping is a confidence-scored match, corroborated across filings where possible; an uncertain holding is left unclassified rather than misattributed. These figures describe what the most recent filing reported and are dated to it.

Dating and provenance

IRS e-file data is released on a lag, so a filing for a given fiscal year typically becomes available 12–24 months later. Every figure is dated to the fiscal year it comes from. Where a foundation’s most recent filing names no grant recipient (grants abroad are filed by region only, and grants under $5,000 are not itemized), we show its most recent year of named grant activity and say so. Each displayed figure links back to the specific filing it was read from.

Named individuals

Some 990-PF filers report grants to peoplerather than organizations: a scholarship, a hardship-fund award. We withhold the name rather than publish it beside a dollar amount; the dollars still count in every total on the page. What’s withheld, why, and how to reach us about it is on redactions. What the filings will not support however good the matching gets is on known limitations.

What we don’t claim

We never state that a foundation caused a grantee’s outcome. Grantee achievements come from each organization’s own program-service reporting and are read as association with sustained funding. We do not infer numbers the filings do not contain. And provocative, sector-level findings are shown only as aggregate distributions, never as an unflattering ranking aimed at a named organization. For the data behind these pages, see the free Grants API; for terms of art, the glossary.