GDELT News Search & Monitor – Global News in 65 Languages, Only New Articles

Search worldwide online news through the GDELT DOC 2.0 API and get one clean JSON record per article: headline, link, site, language, publisher country, the moment GDELT first saw it, the social image and the search that found it. No API key, no proxy, no browser.

Scheduled, it delivers only the articles it has never delivered before. It remembers every article it hands over (in a key-value store whose name you choose), so the same article is never returned twice, across runs, across days and across the searches of one list. A run with nothing new delivers nothing and costs nothing.

Source	GDELT Project, DOC 2.0 API (api.gdeltproject.org), the last 3 months of online news in 65 languages, updated every 15 minutes
Search	keywords, "exact phrases", (A OR B), -excluded, plus GDELT operators: domain: sourcecountry: sourcelang: theme: tone< tone>
Window	"how far back" (90min to 3m) or a start and end date; 3 months is GDELT's limit
Depth	up to 10,000 articles per search: GDELT gives 250 per request, so the window is read in pieces (stated below)
Only new	persistent memory per monitor name, first run can be a silent baseline, nothing delivered is charged twice
Order	newest first (default), oldest first, relevance, most positive / most negative tone
Polite	one request at a time, 5.5 seconds after each answer; when GDELT answers 429 the request is retried every 20-30 seconds for up to 8 minutes
Price	$0.0008 per article delivered ($0.80 per 1,000). Searching, empty answers and errors are free

Typical uses: brand and competitor monitoring · PR and crisis alerts · news research and media analysis across countries and languages · feeding an LLM, a RAG pipeline or an AI agent with fresh headlines · daily or hourly digests to Slack, e-mail, Google Sheets.

This Actor returns headlines and links, not article text: the text belongs to the publishers and GDELT does not carry it.

Quick start

1. Write one or more searches in Searches, one per line.
2. Choose How far back (default 24h) and Max articles per search.
3. Run. Schedule it: every later run delivers only what appeared since.

{
  "queries": ["\"artificial intelligence\" regulation", "domain:bbc.co.uk economy"],
  "timespan": "24h",
  "maxArticlesPerQuery": 20
}

Recipes

Brand alert, scheduled every hour (first run only builds the baseline, so the first alert is a real one):

{
  "queries": ["\"Acme Corp\" -careers", "\"Acme Corp\" sourcelang:german"],
  "timespan": "3h",
  "maxArticlesPerQuery": 50,
  "firstRunDelivers": false,
  "monitorName": "acme"
}

What does the French press say about a topic over the last day (a real run on 20 October 2026 with "timespan": "6h" returned 15 articles, from radiofrance.fr, humanite.fr, laprovence.com…):

{
  "queries": ["macron"],
  "sourceCountries": ["france"],
  "sourceLanguages": ["french"],
  "timespan": "24h",
  "maxArticlesPerQuery": 500,
  "onlyNew": false
}

Italian, Spanish or German press: write the keywords in English (GDELT matches on its English translation of the article; a real run on 20 October 2026 with the search below returned 25 articles, 24 of them in Italian, from dagospia.com and others):

{
  "queries": ["\"artificial intelligence\""],
  "sourceCountries": ["italy"],
  "timespan": "24h",
  "maxArticlesPerQuery": 25,
  "onlyNew": false
}

Very negative coverage on named news sites, between two dates, most negative first (a real run on 20 October 2026 delivered 11 articles from bbc.co.uk and theguardian.com: 3 for economy, 6 for war, 2 for election; very negative articles on two sites are few, so keep the search broad):

{
  "queries": ["economy tone<-5", "war tone<-5", "election tone<-5"],
  "domains": ["bbc.co.uk", "theguardian.com"],
  "startDate": "2026-10-01",
  "endDate": "2026-10-07",
  "sortBy": "ToneAsc",
  "maxArticlesPerQuery": 10,
  "onlyNew": false
}

Searches: what you can write

GDELT searches the headline and the text of the article. Write the keywords in English, also to find articles in other languages (see the FAQ). All words must appear (AND). Every word needs at least 3 letters: "artificial intelligence", not AI.

You want	Write
a phrase	"climate change"
either word	(tesla OR spacex) (OR needs parentheses; a bare a OR b is wrapped for you)
leave a word out	tesla -recall
one site	domain:bbc.co.uk
one publisher country	sourcecountry:france (names without spaces: unitedkingdom; 2-letter codes are GDELT's FIPS codes, not ISO: Germany is GM)
one language	sourcelang:italian
a GDELT theme	theme:ENV_CLIMATECHANGE
negative or positive coverage	climate tone<-5 · economy tone>5

The fields Only sources from these countries, languages, Only these sites and Never these sites add the same operators to every search without writing them (several values mean "any of them").

Input

Field	Default	Meaning
queries	—	One search per line (see above). Up to 200; empty, duplicated (also Airbus and airbus) or malformed lines are skipped and listed under ignoredQueries in SUMMARY and in the log
timespan	24h	90min (at least 15), 6h, 24h, 7d, 2w, 3m (m = months, at most 3)
startDate, endDate	—	Instead of timespan: when startDate is set, timespan is not used (the log and SUMMARY say so if you changed it). 2026-10-01 or 2026-10-01T08:00:00Z, UTC unless a zone is given; a bare end date includes that whole day. Older than 3 months is cut to 3 months
maxArticlesPerQuery	100	Articles delivered (and charged) per search, 1-10,000
maxArticlesTotal	0	0 = no extra limit; otherwise stop the run after this many
sortBy	DateDesc	DateDesc, DateAsc, HybridRel (relevance), ToneDesc, ToneAsc
onlyNew	true	Remember delivered articles and never deliver or charge them again
firstRunDelivers	true	false: the first run of a search only builds a baseline (free) and you get what appears afterwards
monitorName	default	Runs with the same name share one memory (key-value store gdelt-news-monitor-<name>)
sourceCountries, sourceLanguages, domains, excludeDomains	—	Filters added to every search
timeoutSecs	90	Per request; GDELT often needs 15-25 seconds to answer

Output

One record per article. Whatever GDELT does not give is null; nothing is guessed.

{
  "title": "Volkswagen bet on Xpeng to save its China business . It not going to plan",
  "url": "https://www.asiaone.com/money/volkswagen-bet-xpeng-save-its-china-business-its-not-going-plan",
  "domain": "asiaone.com",
  "language": "English",
  "sourceCountry": "Singapore",
  "seenAt": "2026-10-20T09:00:00.000Z",
  "socialImage": "https://media.asiaone.com/sites/default/files/styles/article_top_image/public/original_images/Oct2026/volkswagen%20car.JPG?itok=1thYIN67",
  "query": "\"electric vehicles\" sourcelang:english",
  "citation": "Source: The GDELT Project (https://www.gdeltproject.org/)"
}

(A real record, from a search for "electric vehicles" sourcelang:english over 24 hours on 20 October 2026.)

Field	Meaning
title	Headline as GDELT stores it. Its tokenizer sometimes puts spaces around punctuation (7 %, 497 . 82) and the Actor leaves it as it is
url	Link to the article on the publisher's site
domain	Publisher domain
language	Language of the article (English, Italian…)
sourceCountry	Country of the publisher, not necessarily the country the article is about
seenAt	When GDELT first saw the article, ISO 8601 UTC, in steps of 15 minutes. It is not the publication time: GDELT does not give that
socialImage	The article's sharing image, when it has one
query	The search of your list that found it (the first one, when several found it)
citation	Source: The GDELT Project (https://www.gdeltproject.org/), as GDELT's terms of use require, in every record

Besides the dataset, the SUMMARY record of the key-value store lists for every search: the exact text sent to GDELT, the window, requests made, articles read, delivered, skipped because already delivered, held back by a limit, warnings and the error text if the search failed.

Only new: how it works

• Every article you receive is remembered by a short fingerprint of its link (www., http/https, #fragment, trailing slash, tracking parameters such as utm_* and fbclid ignored). The same story reached through two searches, or through a link with ?utm_source=, is one article and is charged once.
• The memory lives in a named key-value store (gdelt-news-monitor-<monitorName>), so it survives between runs. Use another monitorName for an independent schedule. Delete the store to start from zero.
• Only what was delivered (or baselined) is remembered. If a limit stops the run (your spending limit, maxArticlesPerQuery, maxArticlesTotal), the rest stays new and arrives in the next run.
• The first run is per search, not per monitor: a search the monitor has never run delivers what GDELT lists now (firstRunDelivers: true) or, with firstRunDelivers: false, silently remembers it as a baseline and delivers only what comes later. With false, adding a search to an existing list never floods you with its old articles.
• An article nobody listed for 100 days is forgotten (GDELT itself keeps 3 months). The memory is capped at 50,000 articles.
• Each run reads the window from the newest article backwards until the window is exhausted or your limit is filled, skipping the ones already delivered. Skipping is free but each 250 articles read is one request: for a frequent schedule use a short window (every hour: 3h; every day: 2d).
• Widening the window later (24h → 7d) on a search that already ran will deliver the older articles of the wider window once; they were never delivered.

Speed and GDELT's limits: read this before you schedule

• GDELT allows one request every 5 seconds per client. The Actor sends one request at a time and waits 5.5 seconds after each answer.
• GDELT is slow: opening the connection alone takes about 15 seconds (measured on 19 October 2026), so a request takes 15-25 seconds and there is no way around it. A search of up to 250 articles is one request.
• GDELT returns at most 250 articles per request. For more, with newest-first or oldest-first order the window is read in pieces: each piece ends where the previous one stopped, nothing is lost or repeated, one request per 250 articles. With relevance or tone order a request cannot be continued: the window is cut into equal time slices and GDELT's best 250 of each slice are read (the order holds inside a slice, not across them; the SUMMARY says so).
• GDELT often answers 429 ("too many requests"), and it does so at random. The Actor does not give up: it asks again after 20-30 seconds (a random wait, never less than 20 s, longer if GDELT sends a Retry-After) and keeps trying for up to 8 minutes per request, about 19 attempts. Waiting longer between attempts does not help (direct tests on 20 October 2026 with requests 60 s apart and 20 s apart gave about the same share of answers, one in three to one in five), so many short waits beat a few long ones. If a request is still refused after 8 minutes, that search fails with a clear message (nothing delivered, nothing charged, the next run tries it fresh); if three searches in a row fail like that, the run stops trying the rest (they are listed as skipped and the next run takes them) instead of waiting for hours.
• Measured on 20 October 2026 (one day, one ordinary home connection, 9 real runs, 18 searches, 85 requests to GDELT): 67 of the 85 requests (79 %) were answered 429 and every one of the 18 searches still delivered its articles (or an honest zero). It is the data of one day: on a quieter day there are fewer 429s, and GDELT's behaviour can change.
• How long does a run take? One search with GDELT answering at once: about 30 seconds (28 s measured). With 429s in between, one search took 78 s, 236 s and 335 s in our runs (2, 6 and 7 refusals before the answer). Lists: 3 searches took 5.6 and 6.4 minutes, 6 searches took 22 minutes (32 refusals). Rule of thumb: allow 4 minutes per search, and more for searches that read several pages of 250 articles. A bad day can take longer: up to 8 minutes per search before it is given up.
• Set the run timeout to match (Run options → Timeout): at least 4 minutes × the number of searches, and 1 hour or more for lists of 10 searches or more. If the timeout comes first, the Actor stops 2.5 minutes before it, writes the SUMMARY, keeps what it delivered and lists the searches it did not reach as skipped; with onlyNew the next run takes them.
• Moments with more than 250 matching articles sharing the same 15-minute stamp cannot be paged any further; the SUMMARY warns when it happens.

Pricing

One event, article ($0.0008): one article delivered to the dataset. No fee per run, none per search; an empty answer, a refused query, an error, an article skipped because it was already delivered and the silent baseline run all cost $0. The platform usage of the run is included, as with every pay-per-event Actor. If a run reaches your spending limit it stops delivering, keeps the rest as new for the next run, and says so in SUMMARY.

Use	Articles	Cost
5 brand names, hourly, 10 new articles a day in total	~300 / month	~$0.24 / month
Daily digest of 20 topics, 40 new articles a day	~1,200 / month	~$0.96 / month
One-off research: 3 topics × 1,000 articles	3,000	$2.40
Backfill: 10,000 articles of the last 3 months	10,000	$8.00

Cost cap per run: maxArticlesPerQuery × number of searches (or maxArticlesTotal), and your own spending limit.

FAQ

Is this the full text of the article? No. GDELT indexes headlines, links and metadata; the text belongs to the publishers. Use the url with a crawler or a content extractor if you need it.

Why seenAt and not a publication date? GDELT's article list gives the time it first saw the page (to the nearest 15 minutes). For fresh news it is within minutes of publication; for an old article first seen late, it can be later than the real date. The Actor does not guess.

Why fewer articles than I asked for? Because GDELT has fewer: the window is short, the filters are narrow, or the query needs fewer words ("AI" is refused, GDELT wants 3 letters). The SUMMARY says how many articles were read and why the search ended.

What does a refused query look like? GDELT answers with a plain sentence ("Your search contained a keyword that was too short."). The Actor reports it for that search, with a hint, at no cost, and carries on with the other searches.

Why did my Italian, Spanish or German search return nothing? In our measurements GDELT finds articles in other languages through English keywords, and an original-language phrase finds nothing (we think because it searches an English translation of each article; GDELT does not say so on the search page). Measured on 20 October 2026, sent straight to GDELT without this Actor: "intelligenza artificiale" and "Künstliche Intelligenz" returned an empty list, while "artificial intelligence" sourcecountry:italy returned 250 articles, "climate change" sourcelang:spanish 163 and "climate change" sourcecountry:germany 80. Write the keywords in English and choose the country or language with the filters; the articles come back in their own language. Some scripts (for example Japanese) are refused by GDELT as "too short": the Actor reports it, at no cost. The Actor reports exactly what GDELT answers ("Nothing new matched") and does not guess; the SUMMARY shows the exact text sent.

How far back can I go? 3 months. Older dates are cut to 3 months and the log says so.

Why only 6 articles for "Formula 1" in 12 hours? Because that is what GDELT returned for that phrase in that window (measured on 20 October 2026, the Actor delivered every one of them). GDELT indexes a selection of the world's online news, not every page, so counts from it are a sample of coverage, not a census.

Why does seenAt sometimes fall 15 minutes after the end of my window? GDELT works in 15-minute batches and its window edges are not exact to the minute: a window ending at 11:00 returned articles stamped 11:15 (seen on 19 October 2026). The Actor keeps what GDELT returns and does not trim it.

Does GDELT cover every outlet? No. It monitors tens of thousands of online news sources worldwide, unevenly: large and national outlets are well covered, small local sites less. Treat counts as a sample of coverage, not a census.

Can I use it commercially? GDELT's terms (checked on 20 October 2026 at gdeltproject.org/about.html) make the data available "for unlimited and unrestricted use for any academic, commercial, or governmental use of any kind without fee", and require that "any use or redistribution of the data must include a citation to the GDELT Project and a link to this website (https://www.gdeltproject.org/)". Every record carries that citation; keep it when you pass the data on.

Is this affiliated with GDELT? No. It is an independent client of GDELT's public API.

Use the output

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: process.env.APIFY_TOKEN });
const run = await client.actor('far003/gdelt-news-monitor').call({
    queries: ['"electric vehicles" sourcelang:english'],
    timespan: '6h',
    monitorName: 'ev-news',
});
const { items } = await client.dataset(run.defaultDatasetId).listItems();
for (const article of items) console.log(article.seenAt, article.domain, article.title);

from apify_client import ApifyClient

client = ApifyClient("<APIFY_TOKEN>")
run = client.actor("far003/gdelt-news-monitor").call(run_input={
    "queries": ['"electric vehicles" sourcelang:english'],
    "timespan": "6h",
    "monitorName": "ev-news",
})
for article in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(article["seenAt"], article["domain"], article["title"])

Scheduling and integrations

Create a Schedule for the Actor (for example every hour with timespan: "3h"), keep the same monitorName, and add a Webhook or an integration (Slack, e-mail, Google Sheets, Make, Zapier, n8n) on "run succeeded": each run's dataset holds only the new articles, so an empty dataset means nothing new. The Actor also works as a tool for AI agents through the Apify MCP server.

Attribution

News data from The GDELT Project. GDELT is a free, open platform for research and analysis of global society. This Actor is not endorsed by or affiliated with GDELT.