The brief and the reality
Viewers cannot find anything, engagement is flat, and a recommendation system is going to fix it. A model is chosen, an engineer assigned, and a proof of concept looks convincing on the recent catalogue.
Then it meets the archive, and the programme slows down and does not recover.
1. The model is downstream of the description
A recommender works with what content is described as. Thin, inconsistent or wrong labels are not recoverable by modelling — the system will confidently recommend on the basis of bad data.
In most media businesses the labels are bad for structural reasons. Content was ingested over fifteen years by different teams under different schemes, some acquired with a third party’s metadata attached, some tagged by whoever had five minutes.
2. What “bad metadata” looks like in practice
- Genre applied inconsistently. The same programme is drama in one batch and crime in another, because two people made reasonable different calls.
- Free text where structure was needed. Contributor names spelled four ways, so “more from this presenter” cannot work at all.
- Missing on the archive — which is exactly the content discovery is meant to surface. Recent titles are well described; they were also easy to find.
- Rights data separate from descriptive data, so something can be surfaced that is not licensed to show.
- No structural relationships. Series, episodes, seasons, brands and clips related by naming convention rather than by data.
3. Fix the foundation
One taxonomy with an owner. Someone decides what a genre means, and it is enforced at ingest rather than corrected afterwards. A taxonomy without an owner diverges within a year.
Structured entities for people and titles, with stable identifiers, so a contributor is a thing rather than a string.
Rights and availability as queryable data — windows, territories, platform rights — so what can be shown is a fact the system knows rather than a rule applied at the end.
Tagging at ingest, validated. Every item entering untagged is a future gap somebody will pay to fill.
Provenance on every tag. Who or what applied it, when, and with what confidence. This is what lets you re-derive later without destroying human work.
4. Where AI genuinely helps
The archive is too large to tag by hand, and this is a good use for models: automatic tagging, transcription, and semantic indexing of content that has no useful description at all.
Two conditions:
- Generated tags marked as generated, so confidence is known and they can be reviewed or regenerated when a better model exists.
- Populating the same taxonomy as human tagging, not a parallel one. Otherwise you have created a second metadata problem alongside the first.
Transcripts are the highest-value artefact here. They make archive content searchable by what was actually said, which is frequently how people remember it.
5. Search before recommendation
Search gets the budget left over, despite being what people reach for when they have a specific intent — the moment they are least tolerant of failure.
Search traffic in a media product is dominated by:
- Exact titles, misspelled or half-remembered.
- People — an actor, a presenter, a pundit. Frequently the strongest intent signal there is, and frequently unsupported.
- Descriptive queries. “That documentary about the trawlers.”
- Franchises and series, where someone wants a brand and gets a single episode from 2019.
A title-string match handles the first, badly, and none of the rest.
Search failures are silent
Recommendation failing is visible — nothing appealing on the home screen. Search failing produces an empty page and the person leaves.
Zero-result queries are the most useful log a media business has, and almost nobody reads them. They are a direct statement of what an audience expected you to have, in their own words. Some is content you do not own, which is buying intelligence. Much is content you do own, described in a way that did not match how anyone thinks about it.
6. What good search needs
Mostly the same foundation as recommendation, which is why it is the right first project: typo tolerance and partial matching; structured entities for contributors and series; synonyms and colloquial names — what the audience calls something rather than what the rights holder calls it; availability applied to results; and semantic matching over transcripts and descriptions.
7. The interface is not exempt
None of this helps if the surface it lands on is hard to use. Media products are used habitually, for hours, by people who are not concentrating, and increasingly through four arrows and a select button from three metres away.
Friction compounds. An extra tap to resume is nothing once and is the reason somebody drifts away over a year. The specific things that matter: resume that is correct across devices; controls that survive imprecision; content that does not move between sessions, because reordering destroys the muscle memory habitual users rely on; undo rather than confirmation; and quiet recovery from buffering and dropped connections.
Measure time from opening to playing something. In a habitual product that overhead is paid every single day, by everyone.
8. The order to do it in
- Taxonomy and ownership. Decide what the labels mean and who arbitrates.
- Structured entities and relationships. People, titles, series.
- Rights as data, joined to the descriptive record.
- Ingest validation, so the problem stops growing.
- Backfill the archive, using models, with provenance recorded.
- Search, measured by zero-result rate and time-to-play.
- Recommendation, on a foundation that can support it.
A business that fixes search usually finds the discovery problem was smaller than it looked — and arrives at recommendation with the data a recommender actually needs.
Planning discovery work? Get in touch, or read about our entertainment and media work.