Skip to main content
GDELT Cloud sells two different things through one API, and the difference decides how far you should trust a given number.

​
Core — Endpoints we produce & fuse

We code the events, cluster the stories, resolve the entities and score the tone. The upstream GDELT article stream is an input to that pipeline, not a dataset we resell — every event is re-coded under our own taxonomy and every article re-clustered. Nothing in Core is passed through.

​
Open Feeds — Raw open sources we ingest — many fused into Core

Published public registries, ingested and cleaned. We did not create these and we do not claim to; the value we add is that they are resolved onto the same entity spine as Core, so one entity id reaches an event, a filing, a sanctions listing and a physical asset. Each row below names its originating authority and its licence.

​
Intel — Atlas indices and intelligence products built on the data

Products built on both. Atlas indices are computed from our own coded events; briefs are assembled from whatever the question needs.
Coverage figures were last measured against production on 2026-07-20. Every limitation below is measured, not estimated — if one stops matching production the number gets fixed, not the wording.

​
What we produce

The core datasets are GDELT Cloud’s own output — we code the events, cluster the stories, resolve the entities and score the tone. The raw article stream they are coded from is an input, not a dataset we resell.
Coded from: GDELT article stream — a raw feed we code from, not a dataset we resell. None of the upstream provider’s own derived products is an input.Every event is coded by GDELT Cloud under our own CAMEO+ taxonomy, not taken from an upstream event feed. History before March 2026 is being backfilled and is incomplete. Event geography is city-level at best — most events resolve to a country or regional centroid, so they should not be treated as point locations.
Coded from: GDELT article stream — a raw feed we code from, not a dataset we resell. None of the upstream provider’s own derived products is an input.Independently coded using ACLED’s published methodology and taxonomy. Not sourced from, endorsed by, or affiliated with ACLED, and not comparable row-for-row with ACLED’s own dataset. Beta.
Coded from: GDELT article stream — a raw feed we code from, not a dataset we resell. None of the upstream provider’s own derived products is an input.Article clusters we build ourselves. They are re-evaluated as coverage arrives, so a story’s article set can change within its first day.
Coded from: GDELT article stream + public registries — a raw feed we code from, not a dataset we resell. None of the upstream provider’s own derived products is an input.Our resolved entity spine, cross-walked to public registry identifiers where they exist. An entity with no registry match resolves on name alone and is weaker, and an unresolvable name returns no match rather than a guess.
Coded from: GDELT article stream + Bluesky posts — a raw feed we code from, not a dataset we resell. None of the upstream provider’s own derived products is an input.Scored by us. Tone is measured only where an entity is actually covered — an absent bucket means no measured coverage, never a neutral reading. Evidence is news coverage plus, for some entity-days, recent Bluesky posts: the model reads both in one pass and returns a single blended score, so a score is not separable into a news and a social component after the fact. Social is attached only to entities already present in the news stream and never scored alone, but where it is present it is often the majority of the evidence. The evidence mix is recorded per score in the warehouse, but /api/v2/entity-tone serves DAILY ROLLUPS which do not yet carry it — so there is currently no way to request a news-only reading through the API. Early access.
Coded from: Our own coded events and stories — our own output, not an ingested feed.Experimental. The API and MCP tools are gated on the Intelligence plan (can_use_intelligence) — they are no longer Admin-only — but the baseline is provisional because the underlying event history only begins in March 2026. Country-level coverage is the binding limit: measured over 107 days, only 2 countries clear the per-country coverage floor on 90+ days and 12 clear it on most days, so every other country is returned as null rather than a fabricated score. Treat world and continent as usable and country as preview. Not derived from, or comparable to, the Federal Reserve’s published GPR index.

​
What we fuse

Multiple public registries joined into one keyed surface and resolved against the entity spine.
About 99% carry coordinates; the remainder are pipelines and oil & gas fields, which are lines and areas rather than points. Roughly half carry an owner resolved to the entity spine — strongest for heavy industry and data centers, weakest for power. Ports carry no owner at all: the World Port Index is a navigational dataset with no ownership layer, so this is a boundary of the source, not a backlog.

​
Open feeds we ingest

Public datasets we ingest, clean, and — where the source supports an identifier — resolve onto the same entity spine as everything else.
Prime awards only. Sub-award chains are not ingested, so exposure through subcontractors is not represented.
Reflects what registrants filed. Non-registration is not observable, so absence is not evidence of no relationship.
US registrants only — a private or foreign-only company has no EDGAR footprint. Supplier and subsidiary relations extracted from filing text are our interpretation of the document, not statements by the filer.
Epoch’s data-center registry is a curated set of the largest known clusters, not a census. Coordinates are geocoded from published addresses, not supplied by the source.
Capacity units differ by tracker — power is MW while heavy industry is tonnes per annum — so capacity is not summable across asset classes.
Annual and lagged — V-Dem v16 carries a 2025-03-10 vintage and the World Bank indicators trail by one to two years, so these describe a country’s standing condition and can never move on the news. V-Dem indices are expert-coded survey aggregates with published uncertainty intervals; we store the point estimate only. V-Dem publishes no formal licence — its FAQ grants free use for anyone without permission and asks to be cited, which is how these rows are marked.
Series whose upstream terms restrict redistribution are filtered out of every query rather than served with a caveat. GDELT Cloud is not endorsed or certified by the Federal Reserve Bank of St. Louis. Observations cannot predate 1970.
Analytical coverage, not an audit-grade compliance control — use a compliance-grade provider for screening decisions. Country filtering is derived from address text on the US lists and is not complete; two State Department lists carry no country or address at all and cannot be filtered by geography. A list we do not carry returns an explicit error, never an empty result.
Terrestrial AIS only — no satellite feed — so coverage is strongest near shore and thins in open ocean. Accrues forward from June 2026; there is no history before that date.
A periodic research release, not a live feed. It ends in 2021 and will not reflect anything more recent until AidData publishes again.

​
Reaching a feed

Each Open Feed is gated by its own entitlement flag. An endpoint your workspace does not carry returns an explicit error naming the flag it needs — never an empty result, which would be indistinguishable from “there is no data for you”. The flag is stated on each family in the API reference, and the error itself is in the error reference.