How strongly world news coverage claimed each side was gaining, day by day since the invasion, from a sample of up to 1,200 headlines a day. This measures what the press asserted, not what happened on the ground.
Every chart uses the date range selected above, so changing the range changes all of them together. Tick any two and press Compare to overlay them on a shared axis.
Everything here is reproducible. The code, the rubric and the data are public.
Up to 1,200 headlines a day are drawn from GDELT's Global Knowledge Graph, restricted to coverage naming both countries under an armed-conflict theme. Duplicates are removed and no domain may contribute more than three, so a wire story running in forty local papers cannot dominate a day.
About 24 percent make a scoreable claim. The rest are diplomacy, aid, refugee coverage, market reaction, or activity with no stated outcome.
Each headline is read at temperature zero against a fixed rubric, in its original language, and returns six fields: whether it makes a specific claim about either side's position, how strongly it shows Ukraine gaining, how strongly it shows Russia gaining, what kind of claim it is, and whether it comes from the publication itself, a belligerent, or an independent analyst.
Four of the five kinds count toward the index: territory, position, strike and attrition. Each describes something that happened. The fifth, capability, describes what a side holds rather than what changed, and equipment destroyed is already counted as attrition, so in practice it measures deliveries. Ukraine receives most of the aid, which left a standing lean under its line. It has its own chart below.
Excluding it costs a third of the scoreable headlines. In return, agreement with the ground-only measure rises from 0.61 to 0.75, and the months where the index pointed opposite to who was taking ground fall from 11 of 49 to 3.
They do not sum to 100. Both low means nothing is happening; both moderate means both sides are being ground down. A deadlocked front reads differently depending on why, which a single needle could not express.
Attacking is not gaining. An assault with no stated result is not evidence the attacker is succeeding, because captures get reported when they happen. Without this rule the index would measure who holds the initiative rather than who is succeeding, and one side attacks far more often than the other.
If a named attacker gained nothing, that is the attacker's failure and the defender's success. This applies symmetrically: a stalled Ukrainian counteroffensive scores for Russia. Where no attacker is identifiable, the day scores low on both.
Against real events in both directions: the Kharkiv counteroffensive of September 2022 and the Kursk incursion of August 2024 register for Ukraine, the fall of Avdiivka in February 2024 and the Kharkiv-axis offensive of May 2024 for Russia. Scoring was also checked against real multilingual headlines from the dataset itself, not examples written for the test.
Scores carry the model version and a hash of the rubric that produced them. Change the rubric and older scores stop being comparable, so the pipeline refuses to write against a changed rubric rather than mixing the two, and the series is recomputed instead.
This series was scored by gemini-3.5-flash-lite at temperature 0,
sampling up to 1,200 headlines a day, capped at 3 per domain. 1,650 days carry a
score, from 2022-02-24 to 2026-09-17. Backfill last written 2026-09-18.
One row per period. Both scores, the headline count behind each point, and the composition by claim type and source.
Covers the date range currently selected.
Keeps the range and view you have selected.
| Period | Ukraine | Russia | Diff | Headlines |
|---|
Stated here rather than in a footnote, because each one changes how the chart should be read.
Corrections, missing events and methodology objections are all welcome. Or write to directly.
Goes straight to an inbox. Nothing is stored on this site.