Audience attention and shot-level content analysis in one queryable place — and an honest report of how weakly they actually relate.
Gemini picks one of seven parameterised questions and fills in its
arguments. The SQL is composed server-side and executed by the
mcp-clickhouse MCP server. Every query it runs is listed below the
answer it produced.
Spearman rank correlation with tie-averaged ranks, of each content feature against measured attention, against the position-corrected residual, and against position itself. The floors and the verdict are computed in SQL, not written here. Edge bins excluded.
Both floors are optimistic. The residual is taken against a moving average and is therefore serially correlated, so the effective sample size is below the row count — read them as at least this much noise. The ±8-bin window also caps what is findable: any content effect lasting longer than about two minutes is absorbed into its own baseline, so "nothing in the residual" means nothing on a short timescale.
YouTube's most-replayed heatmap for this upload. Always exactly 100 buckets, min-max normalised by YouTube. It records rewatching, not retention: a peak means people went back, not that people stayed.
Six scores per segment from Gemini, scored from the clip itself with a blind prompt that never sees an attention value and is told not to guess what viewers preferred.
13.3M generated events, because per-viewer playback data is something only a platform owner has. They are derived from the attention curve, so they demonstrate the query workload and cannot corroborate the curve.