Research
These are the findings the pipeline is built on. Each one is a measurement with a stated n, not an estimate. Anything we have not measured is not here.
30%
n = 380, uniform sample
of sampled videos contain speech or on-screen text worth processing
Catalogue music, reused viral audio and silent photo carousels account for the rest. Classifying this before transcription is what makes full-population coverage affordable.
6%
n = 603, uniform US sample
of videos fall inside the monetisable topical scope
Politics, economics, work, technology, health and social issues. Entertainment, lip-sync and performance dominate the platform by volume and are excluded by design.
28%
n = 500, with live controls
of a political-by-construction corpus carried a political platform label
What the platform labels as political is genuinely political, so the field is useful for inclusion. It is useless for exclusion: commentary shaped like a rant or a late-night segment is filed under entertainment.
26%
n = 500, controls interleaved
of a 2023–24 political corpus no longer existed by 2026
Live positive controls separate deletion from blocking. A labeled series of public discourse has to be collected as it happens; it cannot be reconstructed later at any price.
Tell us the market and the question. We return a real slice — schema, sample sizes and confidence intervals included — not a brochure.
