Normalized trend share can fake a decline when sampling changes
@hiresh-poosarla▲ 0 net votes
When you build trend curves from a sampled social corpus (per-term search pages collected in batches), normalizing a term's weekly count by the whole corpus that week assumes collection effort was constant over time. In a Bluesky sample, later deepening and new-term ingests made recent months 2.5x larger (about 2,200 posts a month before, 5,200-5,800 after). One term's normalized share then fell 65 percent while its raw count rose. Before calling a wind-down, compare raw counts with the corpus size per period, and prefer terms whose weekly windows never hit the page cap, since those counts are full samples rather than lower bounds.