Per-title encoding thresholds that actually pay back

The promise

Per-title encoding measures each asset and builds a ladder from its actual complexity instead of applying one ladder to the whole catalogue. On mixed content it reliably removes bytes.

The bill nobody quotes

The analysis costs compute, and its results are only valid for the encode settings they were measured at. Every time a preset, a ladder or a packaging target changes, you either re-analyse the asset or keep running a ladder that no longer matches it.

Members that re-ran analysis on every rendition change found the engineering overhead — not the compute — was the larger cost.

Where the crossover sits

• Short-tail catalogues with a few thousand titles that are rarely re-encoded: analysis amortizes well.
• Long-tail catalogues with continuous ingest and frequent preset changes: automate re-analysis or accept drift.
• Live and event traffic: per-title analysis does not apply to content you have not seen yet. Only a well-shaped default ladder does.

A rule that held up

Analyse once, then hold the ladder stable for the life of that rendition. If a preset changes, treat the ladder as unvalidated and re-measure on a sample before rollout.

What to report

Bytes per viewing minute against delivered quality, measured on the same population before and after. Anything else compares two different things.