Leaders Insights
Leaders Insights

Rester au meilleur niveau, un peu chaque jour.

DomainesMarketingDataFinanceIA
RessourcesApprendreTestOutilsBlogGlossaire
© 2026 Leaders Insights — Tous droits réservés.
Formations/Data in media/Data landscape, quality and metrics/Metadata as infrastructure: how tagging and content taxonomies determine what gets measured
4/5+150 XP

Data landscape, quality and metrics

5Mapping the media data landscape: sources, silos and standard datasets+1506Why your viewership numbers disagree: reconciling panels, census and self-reported data+1507Data quality audits: catching bots, duplicate IDs and broken pipelines before they skew decisions+1508Metadata as infrastructure: how tagging and content taxonomies determine what gets measured+1509Benchmarking what 'good' looks like: industry-standard metrics for reach, retention and content performance+150

Metadata as infrastructure: how tagging and content taxonomies determine what gets measured

# MetadataMetadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). as infrastructure: how tagging and content taxonomies determine what gets measured

A subscriber searches for "The Office" on a streaming app and gets the UK version, three fan compilations, and no results for the US original because it's tagged under its production banner name instead. Nobody touched an algorithm to cause this. Somebody just left a metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). field blank, or filled it inconsistently, months or years earlier. Multiply that by a catalog of 20,000+ titles and you get broken search, irrelevant recommendations, and a content report that tells executives a licensing deal is underperforming when really it's invisible.

This lesson looks at (the descriptive data attached to content, like genre, cast, runtime, language, and rights windows) as the foundation everything else in media data sits on.

metadata
metadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition).

What metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). actually is, and why it's not optional

MetadataMetadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). is "data about data": structured information describing a piece of content rather than the content itself. For a single film, that typically includes:

  • Descriptive metadata: title, synopsis, genre, cast, crew, keywords, mood tags
  • Technical metadata: resolution, aspect ratio, audio codec, subtitle languages
  • Rights metadata: territories licensed, window start/end dates, platform exclusivity
  • Administrative metadata: internal content IDs, production company, ingest date

Every layer downstream, search, recommendation engines, royalty payments, regulatory reporting, depends on these fields being complete, consistent, and correctly structured. This is why practitioners increasingly call metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). "infrastructure": it's not a nice-to-have content descriptor, it's the plumbing that determines whether a $50 million licensing deal actually reaches viewers.

The industry standard schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.Voir la définition complète → for describing media assets is EIDR (Entertainment Identifier Registry, a nonprofit that assigns unique IDs to films and episodes so different companies can refer to the "same" title unambiguously). Without a shared identifier system like EIDR, a studio's internal title ID for a film won't match a distributor's ID, which won't match a streamer's catalog ID, and reconciliation becomes manual, slow, and error-prone.

The core datasets in play

Media companies typically manage three interlocking metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). datasets:

1. Content/catalog metadata, the descriptive and technical layer. Sources: production studios, post-production houses, metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). vendors like Gracenote (owned by Nielsen) or Rovi/TiVo data feeds.

2. Rights and licensing metadata, territory, window, exclusivity, and format rights, usually stored in a rights management system. This is the layer that, if wrong, causes a title to play in a country where it isn't licensed, a real compliance exposure.

3. Usage and engagement metadata, timestamps of what was watched, for how long, on what device, tagged back to the content ID. This is what analytics and recommendation systems actually run on, but it's only useful if it joins cleanly to the content metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). above.

The joining step is where most quality problems surface. If "content ID 4471" in the viewing log doesn't match the same ID in the rights table, you cannot legally or accurately report what was watched under which license, which matters for both royalty payments and regulatory obligations like European content quotas.

Why Europe adds a layer: the AVMS Directive

In the EU, the Audiovisual Media Services Directive (AVMSD), enforced by national regulators and coordinated via the European Commission, requires on-demand platforms to ensure at least 30% of catalog content is European works, and to give that content "due prominence" in how it's presented. (European Commission summary)

That 30% figure cannot be measured without accurate country-of-origin metadata at the title level. A platform that mistags Spanish co-productions as "US" content will misreport compliance, even if its actual catalog composition is fine. MetadataMetadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). quality here is a direct regulatory dependency, not just an operational nuisance.

Data qualityData qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → metrics that matter

Media metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). quality is typically tracked on four dimensions, each with a practical, computable metric:

| Dimension | What it measures | Example metric |

|---|---|---|

| Completeness | Are required fields populated? | % of titles missing genre tag |

| Consistency | Same entity described the same way everywhere | % of cast names matching a canonical spelling across systems |

| Accuracy | Does the field reflect reality? | % of runtime values verified against source file |

| Timeliness | Is metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). updated before it's needed downstream | Days between rights expiry and metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). update |

Worked example. Say a streamer has 18,000 titles. An audit finds 1,260 titles missing at least one required descriptive field (genre, cast, or synopsis).

Completeness rate = (18,000 − 1,260) / 18,000 = 0.93, or 93%.

A 93% completeness rate sounds fine until you learn that most industry metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). leads treat anything below roughly 97 to 98% as a search and recommendation risk (this threshold is a common practitioner benchmark, not a universal regulatory standard, and should be treated as an estimate). The missing 7% isn't randomly distributed either: it tends to cluster in library/catalog titles and international acquisitions, exactly the content a recommendation engine needs the most metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). to surface, since it lacks the promotional buzz of new releases.

How bad metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). breaks measurement, concretely

  • Search: a title tagged "Comedy" instead of "Dark Comedy, Crime" won't surface for users filtering by subgenre, undercounting real demand.
  • Recommendations: collaborative filtering and content-based recommenders both rely on tags as features. Inconsistent tagging (one system uses "Sci-Fi," another "Science Fiction") fragments what should be one signal into two weaker ones.
  • Reporting: a title split across two internal IDs due to a re-ingest will show two rows of small viewership numbers instead of one meaningful one, making it look like a flop in exec dashboards.
  • Royalties: talent and rights-holder payments are frequently calculated from usage reports keyed to content ID. A mismatched ID can mean a musician or actor's residual (a royalty payment for reuse or streaming of previously released content) is undercounted or missed.

A simplified version of a completeness check, the kind a metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). QA analyst might actually run:

python
import pandas as pd

catalog = pd.read_csv("catalog_metadata.csv")
required_fields = ["genre", "cast", "synopsis", "runtime_minutes", "country_of_origin"]

missing_report = catalog[required_fields].isnull().sum()
completeness_rate = 1 - (missing_report.sum() / (len(catalog) * len(required_fields)))

print(f"Overall completeness: {completeness_rate:.1%}")
print(missing_report.sort_values(ascending=False))

This kind of script is often the first thing a data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → team runs against a newly acquired catalog, before any licensing deal metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). is trusted for reporting.

Vérification des acquis

1. In the streaming app example where a search for a US show returns only foreign versions and fan compilations, what is the underlying cause of the failure?

2. Why does the lesson describe metadata as 'infrastructure' rather than as a simple content descriptor?

3. A content report shows a licensing deal 'underperforming' with low viewership. Based on the lesson's framing, what should an analyst investigate first?

CHOIX MULTIPLES

4. Select ALL correct answers: Which of the following are examples of metadata categories described in the lesson?

Sélectionnez toutes les réponses correctes.

CHOIX MULTIPLES

5. Select ALL correct answers: According to the lesson, what are consequences of poor or inconsistent metadata at scale (e.g., across a catalog of 20,000+ titles)?

Sélectionnez toutes les réponses correctes.

Governance: who owns the taxonomy?

A taxonomy here means the controlled vocabulary and hierarchy used to classify content (e.g., a fixed list of approved genres, rather than free text). Governance questions that determine whether metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). stays usable at scale:

  • Who has authority to add a new genre or mood tag, and under what approval process?
  • Is there a single system of record (the authoritative source when the same field exists in multiple systems) for rights windows?
  • How often is the taxonomy audited against actual usage (are 40% of titles landing in a generic "Other" bucket because the taxonomy is too rigid)?

Netflix's well-documented internal practice of using thousands of ultra-granular "altgenres" (reported publicly in outlets like The Atlantic's 2014 investigation) is the extreme version of solving this: rather than a small fixed genre list, they built a deep tagging system precisely because recommendation quality is bottlenecked by tagging granularity.

Key Takeaways

  • MetadataMetadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). (genre, cast, rights windows, technical specs) is the infrastructure layer beneath search, recommendations, reporting, and royalty payments in media; quality problems here cause downstream failures that look like algorithm or business problems but aren't.
  • Track quality on four dimensions: completeness, consistency, accuracy, timeliness. A completeness rate below roughly 97 to 98% is a commonly cited practitioner warning threshold (estimate, not a formal standard).
  • Shared identifier systems like EIDR and vendor feeds (Gracenote, Rovi/TiVo) exist specifically to prevent the same title being described differently across studios, distributors, and platforms.
  • In Europe, the AVMS Directive's 30% European-content quota is unenforceable without accurate country-of-origin metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). at the title level, making metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). a regulatory compliance dependency, not just an operational one.
  • Governance (who owns the taxonomy, what's the system of record) determines whether metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). quality holds up as a catalog scales into tens of thousands of titles.

Précédent

Data quality audits: catching bots, duplicate IDs and broken pipelines before they skew decisions

Suivant

Benchmarking what 'good' looks like: industry-standard metrics for reach, retention and content performance