EXPERIMENTS · BUILT 2 Oct 2026, 00:34 UTC

Is the clustering cut in the right place?

Articles are grouped into stories at one similarity cut. Moving it shows what splits and what merges.

Question: Is the clustering cut in the right place?
Method: Re-cluster the latest day's articles at cuts from 0.30 to 0.60 with Google News groups hidden, and score how many groups survive whole.
Result: At the current cut 0.42, 10% of groups survive whole on headlines alone; the largest cluster holds 2% of articles.
Decision: Keep the cut: production keeps Google's groups whole, and a higher cut starts swallowing topics.

How it was measured

Articles become stories by clustering: articles that say similar things are joined, and the joining stops at a cut, a cosine distance of 0.42. Too low a cut splits one story into several; too high a cut swallows a whole topic into one blob.

A Google News item lists the outlets covering one story, and the site never splits such a group. Here those groups are hidden from the clustering and used only to score it, at every cut from 0.30 to 0.60, on the latest day's articles:

Completeness: the share of Google's groups whose articles all land in one cluster.
Largest share: the biggest cluster's share of all articles. A jump means a topic is being swallowed.
Merged: the share of clusters holding three or more of Google's groups.

The numbers

2026-10-02 · 2500 articles, 1158 of them in a Google News group.

Cut	Clusters	Largest	Completeness	Merged
0.3	1704	1%	3%	0%
0.35	1514	1%	5%	0%
0.4	1360	2%	8%	0%
0.42 (now)	1290	2%	10%	0%
0.45	1185	2%	12%	0%
0.5	1028	2%	23%	1%
0.55	866	4%	31%	2%
0.6	714	4%	37%	3%

Limits

Headlines word one story many ways, so completeness stays well below 100% at any sensible cut; the site leans on Google's groups for that reason.
Google's grouping is a yardstick, not the truth: it makes its own mistakes.
One day's articles. The best cut can differ on a day dominated by one huge story.