Methodology · Source Collection

Where Articles Come From

Updat3 collects articles every two hours from publisher news feeds, the GDELT Project's open global news dataset and a licensed news-metadata API, reaching hundreds of outlets across wire services, national broadcasters, regional press and specialist publications. It groups related articles into single stories and removes duplicates automatically.

Article ingestion

New articles are discovered every two hours from more than 40 publisher news feeds, with continuous discovery through the GDELT Project's open news dataset and a licensed news-metadata API. Stories that still have too few sources are searched again for corroborating coverage every 15 minutes. Only English-language articles are processed (or articles that can be translated automatically). Articles must have a title, a URL, and at least a brief excerpt to be stored.

News feedsMore than 40 publisher RSS feeds, across regions and topics

GDELT ProjectOpen global news dataset used to discover events and which outlets are covering them

NewsData.ioLicensed headline and publication metadata used for discovery and corroboration

Targeted searchesSupplemental searches for stories that still have too few independent sources

How sources are used

Knowing that an outlet covered an event and using that outlet's text are different things. Every source has a role, decided by the terms under which Updat3 may use it:

CoverageThe outlet is counted as covering the story (outlet counts, political and geographic spread, momentum). Every source we can see does this.

EvidenceThe outlet's reporting is read to support the facts in our synthesis. Only sources whose terms permit this are used, and syndicated copies of one report count once.

DisplayWhat appears on Updat3 is our own writing plus each outlet's name, publication time and a link. Publisher headlines, excerpts and photographs are not reproduced without permission; stories use Updat3's own topic artwork.

A story can therefore show more outlets covering it than it used as text; the story page says how many. A story is never published from a single outlet.

Story grouping & deduplication

Multiple articles about the same event are merged into a single story. This is how you see "5 sources" under one headline instead of five separate cards.

Matching uses a combination of:

Semantic similarity

Article text is converted to a vector embedding and compared against existing story embeddings. Articles above the similarity threshold are candidates for merging.

Title overlap

Shared significant words between article and story titles provide a secondary signal.

Entity matching

Named people, places, and organisations mentioned in both articles boost the match score.

Fact overlap

Key factual statements from both articles are compared for shared content.

Matching errors happen. An article about one Iran story may occasionally be grouped with a different Iran story. We run merge and re-link passes continuously to correct this.

What gets filtered out

We filter articles when they are collected and again before a story is written from them. Filtered types include:

×Wire service roundups and news digests ("AP Morning Brief", "5 stories to know")

×Opinion, op-ed, and commentary pieces — these are not news reporting

×Press releases, job listings, event announcements, and sponsored content

×Corporate earnings boilerplate ("Q3 results", "raises price target")

×Articles with only a headline and no usable body text

×Non-English content that cannot be auto-translated with acceptable quality

How stories are written

Once an article passes quality filters, it is processed by an AI model that writes:

Story summary (bluf)2–4 sentence plain-English summary of what happened

Key facts4–10 source-supported factual statements, attributed to the reporting they come from

Historical context2–6 paragraphs of background, starting from the immediate trigger and going to deeper history

Story articleA short brief written in Updat3's own words, synthesising the source reporting Updat3 may use (not every outlet tracked is used as text)

Why it mattersReal-world consequences for ordinary people

What to watch nextConcrete follow-up indicators

The AI is explicitly instructed to write from no government's default perspective, apply identical framing standards to all parties, and include historical context that implicates powerful actors — not just the most recent provocation.

Bias Scoring →← All methodology