Data quality

How good is the funding graph?

How much of the funding graph is connected, where those connections come from, and how often our matching is right. Every figure here is computed from the same published data the API serves, so you can check them yourself. How the matching itself works is on the matching page.

What is actually connected

A grant joins the graph when we can attach its recipient to a specific organization. We show rows and dollars separately because unresolved grants tend to be small ones, so the row share understates how much money is connected.

A grant with no recipient EIN is still published — the amount, the recipient name and the funder are all there. What is missing is the link to a US organization, so it does not appear in the EIN-keyed graph. Foreign NGOs, governments and individuals have no US EIN at all; those are clustered separately into entities with their own pages, rather than left as loose text.

What could be connected, at best

Some of these grants name a recipient that has no US nonprofit EIN to find: a person, an overseas body, or a line where the filer wrote a cross-reference instead of a name. Those cannot be connected by us or by anyone working from these filings. The figures below split the rest into what is already connected and what is reachable but not yet reached.

The cap follows from what the IRS collects, so it applies to every dataset built from these filings. A coverage figure quoted without its denominator is not comparable to this one: it may be counting a different universe, or claiming recipients that cannot be identified from the source. We publish the denominator so the comparison is possible at all, and we set the boundary against ourselves, counting a grant as unreachable only where the filing shows why. Every grant row without a recipient EIN is classified by the FIRST reason it cannot be keyed, against the same candidate index the resolver uses. A row is counted against the structural cap only where the filing itself shows why it is unresolvable; anything merely ambiguous is counted as reachable, so the achievement rate is a floor rather than a flattering estimate.

Every reason a grant is not connected

Kind labels (church / school / government) come from name patterns and are indicative only — they describe a band, they never decide one. Form 990-PF asks filers to state each recipient's foundation status directly; carrying that field into the warehouse would replace this inference with the filer's own answer. Separately, 3 grants (0.0% of all rows) name what reads as a church or a government body that appears in no register. Those are almost certainly unreachable as well, but estimated from names rather than shown by the filing, so it is left out of the figures above.

How much of this did we work out ourselves?

990 filers give the recipient’s EIN on Schedule I. 990-PF filers, private foundations and most of the corpus, give a name and address and no EIN, so those connections are ones we worked out. That is why the accuracy figures below matter.

Recorded per grant at parse time, so this split is exact.

How often is our matching right?

Filers who do supply an EIN give us an answer key. We hide it, run the same matching rules, and compare what comes back.

These figures score two of the eight passes above: the unique-name rule and the ZIP tie-break. They are the oldest and most conservative, and they are the only two the harness re-runs, so the reach below does not move when a newer pass adds coverage. Read them as the accuracy of that part of the matching, not of all of it.

What this does not cover. The registry, model, fuzzy and address passes resolved 1,599,545 grants that this measurement never scores , 18.1% of everything we matched ourselves. They are the newer and less conservative passes, so they are the ones a sceptical reader would most want measured. Extending the harness to re-run every pass is outstanding work, and until it lands the figures above describe 81.9% of our matching.

Precision on its own is easy to inflate, since a matcher that rarely answers can be right almost every time. Reach and the declined share are here so you can see both. Where a name is ambiguous we leave it unresolved rather than picking the most likely match. That lowers coverage on purpose.

Caveat. Ground truth comes from 990 Schedule I, whose recipients skew larger and more formally registered than the 990-PF recipients the resolver actually runs on. Treat this as an estimate on a related population; the true figure on 990-PF names is likely lower.

Every layer that infers something

The figures above cover one step: attaching a grant to a US organization. Several other layers do their own matching, and each one moves the numbers you see elsewhere on the site. Each reports what it attempted, what it resolved, and what it declined, plus which way it fails, because the consequences are not the same. A layer that merges too eagerly invents connections; one that merges too cautiously just splits a thing in two.

Coverage by fiscal year

Recent fiscal years are still filling up. The IRS releases e-file data 12-24 months after the activity, so the latest years are floors that rise with each monthly release, not final totals.

Bars are grant counts; the right-hand figure is the share with a resolved recipient. Grey bars are fiscal years still arriving, so treat them as floors rather than totals.

Check it yourself

The grant data behind these figures is queryable through the API: a free key covers the grant endpoints, and analytical SQL (POST /api/v1/sql) comes with the paid tiers. The resolver measurements are the exception: precision is scored against held-out filer-supplied EINs, which the API does not expose as a separate set, so those figures are ours to justify rather than yours to recompute. The methodology explains how the filings become rows in the first place.

Measured 2026-08-25 against IRS Form 990, 990-EZ and 990-PF e-file filings (public domain). Recomputed on every universe refresh.