How NeuralNews groups matching news coverage
By NeuralNews · 4 October 2026. Prepared with AI assistance from our grouping implementation and regression tests. Examples below are illustrative, not measured publisher data.
What Related coverage means
Related coverage connects stored stories whose headlines match after normalization. It helps readers compare the linked sources without losing each story’s permalink. It does not establish that the sources independently verified a claim, or that the articles contain different reporting.
The matching rules
Our groupStories function normalizes Unicode text, converts letters to lowercase, and replaces punctuation with spaces. A normalized headline needs at least five words to qualify. Stories must share a category and have usable publication or collection timestamps within 72 hours of the group’s representative. Different headlines remain separate, even when they describe the same event.
A concrete example
Imagine two AI stories titled “OpenAI releases a new developer tool” and “OPENAI RELEASES A NEW DEVELOPER TOOL!” on different URLs one day apart. They qualify for a shared group. A story titled “OpenAI launches another developer tool” does not: the words differ. An otherwise matching story a week later stays separate. This conservative rule avoids some false matches but misses paraphrased coverage.
Grouping does not delete the archive
Groups are presentation data. The underlying stored stories and their URLs remain intact. On an article page, that article is used as the representative so readers can find its matching siblings. A separate ingestion step handles duplicate URLs; headline grouping is not a replacement for that step.
Source balance is a different operation
The balanced feed rotates available sources within each UTC publication day, keeping newer days ahead of older ones. This prevents one busy source from filling the start of that day’s feed. It does not assign credibility scores or guarantee equal representation. The briefing is narrower: recent eligible reporting and official announcements, with at most one pick per source and up to five picks.
What our tests check
The regression checks exercise matching headlines across sources, older stories that must remain separate, distinct headline versions, and preserved permalinks. We also check balanced pagination for repeated items. These checks validate the rules; we have not measured grouping precision across the full archive or demonstrated an increase in reader engagement.
How to use the links
Compare the original sources and dates before treating repeated headlines as corroboration. If two unrelated stories appear together, send us both NeuralNews URLs and explain the mismatch. Clear examples can help us improve the rule without silently merging more of the archive.