GDELT machine-reads the world’s news every fifteen minutes in 65 languages. This is the share of all of it that mentioned Indonesia, day by day — on average 1.15% of everything monitored. The flares are the days Indonesia took an unusually large slice of the world’s coverage — which happens both when Indonesia is covered more and when everything else is covered less. Our review found that most of the biggest spikes are partly the second; read the marked days as events worth looking up, not as a ranking of how much the world cared.
Each era below carries its average share of world coverage, its mean tone, and the themes that defined it — all computed, none asserted. Read the share against the two numbers underneath it. Between 2017 and 2026 the articles GDELT saw naming Indonesia fell −71% — from 5,378 a day to 1,572 — while everything it monitored fell −77%. The share rose +29% because the second fell faster than the first. These are five regimes of what Indonesia’s slice looked like, not five regimes of how much the world was watching — a distinction the review article takes apart in detail.
GDELT’s tone runs from roughly −10 (uniformly negative) to +10. Coverage of any country lives below zero — news is what goes wrong — so the line is drawn against its own long-run mean: teal when the world was kinder than usual to Indonesia, coral when it was harsher.
Each dot is a month: how much the world talked about Indonesia against how it sounded. At this resolution the two barely correlate (the r is on the chart) — sustained attention comes from summits and elections as often as from catastrophe. But the answer depends on the timescale, and this is the one scale at which it vanishes. Measured day by day the correlation is negative, and the loudest tenth of days is about a fifth of a tone point darker than the rest: loud days are dark, loud months are not. The review article shows both. Notice also the colour drift: months after 2023 sit warmer — the review decomposes that rise and finds roughly two-thirds is the same outlets writing more warmly and one-third is GDELT reading a different set of outlets.
Each panel is a keyword theme’s share of all Indonesia coverage, month by month, on its own scale. Because both halves of that ratio share the same denominator, this is the one layer on the page the crawl’s size cannot move. Hover to read one month across every theme.
Twenty theme queries were registered; the ones below are those the rate-limited API has returned so far, and each panel states how many months it holds. A panel with a short window is not comparable with a full one — its mean is a mean of those months only.
For each publishing country, the share of its own monitored output that mentioned Indonesia — a measure of interest, not of volume. Neighbours watch closely; the Anglophone giants barely glance, and when they do the tone is different.
Two cautions on this chapter. The intensity below is not size-controlled: a country GDELT barely indexes can float to the top on a handful of articles, so read the large neighbours and ignore the tail. And this measure cannot tell you who supplies the bulk of the coverage. Counting publishers directly in the event layer — which the review article does — shows that 66% of the kept events come from the translated feed and that the great majority of all Indonesia coverage here is published in Indonesia. This is substantially Indonesia’s own press, not the world’s.
Share of Indonesia coverage published outside Indonesia (-sourcecountry:ID). This is the query the API refuses most often, so the curve is still pending; the review article builds the same split from the event layer instead, and finds foreign-published coverage of Indonesia falling across the decade while the share above rises.
Intensity of Indonesia coverage by language — the world’s narrative is not only the English one.
Underneath the articles, GDELT codes discrete events — who did what to whom, where. From the raw 15-minute feed we keep every event that touches Indonesia (as location or as actor), count it against everything the feed carried that day, and type it: cooperation or conflict, words or deeds.
CAMEO root code 14. The 2019 post-election riots and the 2025 protests are the two spikes the pipeline must reproduce unprompted — the anchor-event check.
Event locations on a 0.25° grid; the archipelago draws itself from the data — no basemap. Events geocoded only to “Indonesia” (no province) are excluded and counted below.
Counter-party actors in Indonesia events — CAMEO codes: ISO3 countries plus regions (SEA = Southeast Asia). All years streamed so far.
CAMEO root codes across all kept events.
An independent read of the same archive against the published literature. It finds that 60% of the decade’s apparent rise in attention is predicted by GDELT’s shrinking crawl; that 24 of the 30 loudest days fall on a Saturday, Sunday or Monday, when the world’s newsroom is quiet; that 23 of the 25 largest publishers here are based in Indonesia; and it rebuilds the anchor test on foreign publishers only, which changes 5 of 9 verdicts. The corrections it asked for have been applied to this page.
Read the review article →