Limitations · What this cannot tell you
What this data cannot tell you
Six limits, each a property of what a Form 990 records rather than a gap we are closing. Methodology covers how the figures are built and data quality measures how well the matching works. This page is the other half: the conclusions the filings will not support however good the matching gets.
1. The most recent fiscal years are still arriving
The complete analysis window ends at FY2023. FY2024–2026 are loaded and shown, and nothing on this site reads a trend across them.
The IRS releases e-file data 12 to 24 months after the activity it describes, one bundle a month. A fiscal year therefore fills over several years, and the newest year in the corpus is always a handful of early filers: FY2026 held about eleven thousand grant lines when this was written, against nearly three million for FY2023.
Read as the end of a series, a year like that produces false and quite specific conclusions — a collapsing funder base, a sudden jump in concentration, a fall in giving. So every derived figure on this site stops at the last year whose filings have finished arriving, determined by the same rule the API publishes as last_complete_fiscal_year: walk back from the newest year while each holds under 90% of the year before it. Later years are drawn unshaded and marked, because a reader looking for a recent grant should be able to see it.
Electronic filing became mandatory for tax years beginning after July 2019, and this corpus is built from e-filed returns only. FY2017–2019 therefore hold fewer filers than the sector contained — funder count rose 21.7% in FY2020 alone against 2–5% in the years after — so a year-over-year comparison that spans FY2020 measures the filing mandate as much as it measures giving. Compare FY2020 onward.
2. About a quarter of grant lines carry no recipient identity
15.1 million of 19.8 million grant lines resolve to a recipient EIN. The rest are real grants whose recipient we would not stake an identity on.
Schedule I asks for a recipient’s name and address, and most filers give both; a good number give a name we cannot uniquely place, and some give one an organization would not recognize as its own. Where the filer states the EIN we use it. Where they do not, a layered resolver works from name, address, ZIP and a registry, and each rule carries its own measured precision — published on data quality and explained on matching.
A filer’s own schedule is also where a bad match is most visible. In 71,090cases a funder writes the same recipient name on two lines and the two are matched to different organizations — 561,309 grant lines across 10,473 funders, where at least one of the two is wrong by construction. Those are flagged on the page rather than quietly picked between: a funder whose identity we would not stake the page on is marked, with the reason, and still shown.
The consequence for a page: a recipient-keyed figure is computed over the resolved 15.1million lines, so an organization’s “funders on record” is a floor. A funder that named it in a way we declined to match is missing from its page, and a page will not tell you which. Aliases and former names are the usual reason: an organization that renamed itself, or one whose funders write a programme name instead of the legal entity.
3. A donor-advised fund is a custodian, not a donor
24% of the dollars in the resolved graph arrive through 2,402 donor-advised or pass-through sponsors.
When a sponsor sends money to a charity, the sponsor files the grant and the sponsor’s name and address are what the filing carries. The person who chose the recipient does not appear anywhere. Fidelity Charitable is in Massachusetts and Foundation for the Carolinas in North Carolina; neither fact tells you where the donor behind a given gift lives, or whether they would give again.
These payments count in every total on this site, because the money is real. What they are not allowed to do is stand in for a donor: a sponsor’s address is excluded from the funder-geography map on an organization’s page, and a “who funds you” page whose income is mostly sponsored says so rather than describing a reach it cannot see. The regranting layer as a whole is mapped on intermediaries and donor-advised funds.
4. Some payments are internal transfers between arms of one institution
20%of the dollars in the resolved graph move between two organizations that share a governing board, carry the same EIN on both sides, or are named as related parties by the filer’s own words.
A university endowment vehicle paying its university, a hospital system paying its own foundation, a supporting organization paying the organization it supports: each is one balance sheet moving money to another part of itself, and each sits on Schedule I because that is where the form puts it. Left in, they dominate exactly the figures a reader trusts most, because they are large and they repeat every year.
These are excluded from every external-funding figure — largest funder, concentration, retention, geography, co-funding, prospects — and are named on the page rather than removed quietly, so a reader who knows the relationship can see we did too. Three signals find them: 61,514 relationships where the two organizations share a governing board, 2,132 carrying the same EIN on both sides, and 510 where the filer says so in its own words.
Detection is deliberately partial, and we can put a number on what it misses: a separate screen built from Schedule R and Schedule A finds related-party ties worth roughly $160billion more than these three signals reach. It under-detects rather than over-detects, because a real funder deleted from a fundraiser’s page is the worse error.
5. Not every Schedule I line is a grant programme
616 filers report payments on Schedule I that no applicant could ever apply for, worth 3% of the dollars in the resolved graph.
An athletic conference distributing media-rights revenue to its member universities files those distributions as grants. So does a pharmaceutical company’s patient-assistance foundation booking individual patients on one aggregate line, and a fundraising arm remitting to the hospital it exists to fund. The shape is consistent enough to detect: a closed set of recipients, a median payment in the millions, and one identical stated purpose on nearly every line.
Those payments are excluded from funding analysis and named where they are excluded. The broader point survives the filter: 28%of the dollars in the graph come from organizations whose filings do not say what kind of payment they are making — a hospital paying a local charity, a university paying another university’s research group. Those stay in, labelled as unclassified rather than described as philanthropy. Under half the relationships in the graph, and 25% of the dollars, are payments we can positively call grantmaking.
6. Some relationships are reported and some are inferred
A grant is reported. A cluster, a look-alike, a prospect and a co-funder tie are inferred, and they carry the uncertainty of the inference.
Anything read directly from a filing links to that filing. Everything computed on top — which funders resemble each other, which organizations sit near yours in a portfolio, which funder might plausibly fund you next — is a model output. Those are useful precisely because they say something the filings do not, which is the same reason they can be wrong in ways a filing cannot.
Two boundaries we hold to. We never state that a funder caused an outcome at a grantee; sustained funding and a later result are reported as association, because a 990 cannot support more. And a low-confidence match is suppressed rather than guessed at, which is why a page sometimes shows less than you expect: the alternative is a page that shows a stranger.
If a page looks wrong
Every page names the filings behind it, so a figure that disagrees with what an organization knows about itself can be traced to the return that produced it. If one still looks wrong — or right for a reason we have not accounted for — write to data@useplinth.com. Corrections are applied to the published data, and the ones that turn out to be a rule rather than a row change what every page does. That has happened more than once, and the six limits above are partly the record of it.
Coverage FY2017–FY2026, complete through FY2023. Shares are measured over EIN-resolved relationships. Related: methodology · data quality · what can be matched · what we withhold