How Video Personalisation Actually Works
Personalisation gets described as magic and built as a ranking problem. What these systems actually do, and how they reliably fail.
Video personalisation is sold as understanding the viewer. Mechanically it is a ranking problem: given a slot, a visitor and a library, order the candidates by predicted engagement and show the top one.
What the system knows
- Context — the page, the section, the referrer, the device, the time
- Session behaviour — what this visitor has read or watched in this visit
- Aggregate performance — how each asset has done in comparable slots
- Content similarity — what each asset is actually about, if you have analysed it
Notably absent from most deployments: durable individual profiles. Publishers largely personalise on context and session rather than identity, and the results are closer to the sophisticated version than the marketing implies.
The cold-start problem
New assets have no performance history, so a pure engagement ranker never surfaces them, so they never acquire history. Left alone, the system converges on a small pool of proven videos and quietly stops using the rest of the library.
The fix is deliberate exploration — reserving a share of impressions for under-tested assets. It costs a little engagement now to avoid a catalogue that shrinks to twenty videos.
Where personalisation fails
- Feedback loops. Surfacing what performed reinforces what was already surfaced.
- Optimising the wrong signal. Ranking on clicks produces clickbait; ranking on completion favours short clips.
- Thin libraries. With forty videos, personalisation is rounding error. Fix the catalogue first.
- No content understanding. Without knowing what is in each asset, the ranker is matching tags someone typed at upload.
A reasonable target
Do not aim for a system that reads minds. Aim for one that reliably avoids showing an irrelevant video, explores enough to keep the catalogue alive, and can explain why it chose what it chose.