The content registry that television measurement runs on covers 1.84 per cent of television

Nobody watches an abstraction

Every serious content registry distinguishes levels of abstraction. There is the work as an idea, the title you would name in conversation. There is the edit, a particular cut with its own running time: theatrical, broadcast-standards, a thirty-second advertising version against a fifteen. There is the manifestation, a specific encoding with a resolution and an audio configuration, which is the thing a stream actually delivers.

This matters for measurement in a way it does not matter for a catalogue. A viewing event attaches to a manifestation on a service. It never attaches to an abstraction, because an abstraction has no duration, and duration is what completion, frequency and reach are computed from. Resolve an event only to title level and the loss is invisible the moment the rows are aggregated.

EIDR encodes this correctly. Every EIDR record declares a StructuralType: Abstraction, Performance, Digital, Composite. The difficulty is that the identifier does not carry its level in its syntax. 10.5240/ followed by twenty hexadecimal characters and a check character looks identical whether it names a title or a cut, and Wikidata's own property definition says so plainly: EIDR content ID is an "identifier for a film or television work, edit or manifestation". One property, three levels, and no way to tell which you have without asking the registry.

So we asked. We took every work in the open graph carrying more than one EIDR identifier, 522 of them, and resolved all 1,179 identifiers involved against EIDR's public resolution service, which is genuinely open and answered every request. 285 of those works, 54.6 per cent, turn out to hold identifiers at more than one level at once: 281 pairing an Abstraction with a Performance, three pairing an Abstraction with a Digital record, one holding two Abstractions and a Performance together.

That is the shape of the problem stated in data. Two identifiers hang off one node as though they were alternative names for the same thing. They are not. One names the film and one names a particular cut of it, and a measurement pipeline that joins on the first gets every version merged while one joining on the second gets a single cut. Both answers are defensible and they are different numbers. Nothing in the record says which you have.

The registry that does not reach television

The larger finding is simpler and harder to design around. We counted, for each kind of work in the open graph, how many carry an EIDR identifier at all.

Film, 53.44 per cent. Television series, 1.84 per cent. Television film, which is a film, 27.45 per cent. The pattern is consistent and it is about format rather than recency or popularity.

Be precise about what this measures. It measures the open graph, not EIDR's internal registry, which certainly holds more. That is deliberate, because the open graph is what an organisation can obtain without a commercial membership, and the practical question is not how complete EIDR is but how complete the reachable identifier fabric is. On that question the answer is that anyone building cross-screen measurement on open content identity has a spine for cinema and almost nothing for television.

The season row deserves separate attention, because it fails on the other side too. IMDb covers 78.48 per cent of films and 84.20 per cent of episodes, and 0.78 per cent of seasons, for the straightforward reason that IMDb mints no season-level identifier. A join routed through IMDb identifiers cannot express season structure at all. Seasons are how rights are licensed, how catalogues are merchandised, how release patterns are planned and how a great deal of measurement is reported, and the most widely used identifier scheme in the business has no name for one.

One identifier in thirty-six that your validator will condemn

EIDR identifiers carry a check character, so a corrupted value can be caught before it reaches a join. The algorithm is ISO 7064's hybrid MOD 37,36 computation, which is the same family used for other alphanumeric identifiers, and implementing it from the standard is straightforward.

We implemented it and tested against 2,000 real identifiers. It matched 97.05 per cent. That is a suspicious number: a correct algorithm should match everything except genuinely broken data, and three per cent broken would itself be a large finding. The failures were not random. Every one of them was the same case, and in every one the recorded check character was 0 where our implementation computed the supplementary symbol.

The computation yields a value from 1 to 36. ISO 7064 maps the thirty-sixth to a supplementary character, conventionally an asterisk, because the check alphabet only has 36 symbols and the modulus is 37. EIDR does not use an asterisk. It maps that value to the digit 0, which gives a clean 36-symbol check alphabet with no special case in the identifier syntax. With that one mapping corrected the algorithm validates 224,577 of 224,577 distinct identifiers, 100 per cent, and the check characters turn out to be near-uniformly distributed across all 36 symbols, which is independent evidence the algorithm is right rather than accidentally permissive.

The practical consequence is a quiet one. A validator built on a stock ISO 7064 library rejects 6,252 identifiers in this corpus, 2.78 per cent, as corrupt. Every one is perfectly valid. That is a one-in-thirty-six false positive rate inside a data quality gate, which is exactly the sort of defect that gets absorbed as background noise, blamed on upstream partners, and never traced.

Worth stating the null result alongside it, because it cuts against the pattern we have found in other registers. Across all 224,577 identifiers, zero fail their check character. In company and fund registers the equivalent measurement reliably turns up hand-keyed values that cannot exist, including transpositions and letter-for-digit slips sitting in official records. This data has none of that, which is consistent with its having been machine-imported rather than typed. We report it as a null result rather than quietly dropping a finding that did not appear.

The advertising industry's vocabularies have no semantics

The IAB Tech Lab publishes the Content Taxonomy, the Audience Taxonomy and the Ad Product Taxonomy. They are the shared controlled vocabularies of programmatic advertising: what a page is about, what an audience segment means, what a creative is selling. They are used for contextual targeting, brand safety and data transparency across the industry.

The repository holds 21 tab-separated files and three markdown files. It contains no RDF, no SKOS, no OWL, and no URIs. This is not an oversight of format only. Without URIs there is no way to state that one taxonomy's concept is broader than another's, no way to attach a mapping to a concept rather than to a row number, and no way for two organisations to agree they are talking about the same segment except by agreeing on a spreadsheet.

The files already contain a concept scheme. Each row carries a unique identifier, a parent identifier, a name and the tier path it sits on, which is a SKOS hierarchy written in TSV. So we converted them: 2,845 concepts across Content 3.1, Audience 1.1 and Ad Product 2.0, with the upstream Creative Commons Attribution 3.0 licence and source carried into every generated concept scheme. As far as we can establish this is the first SKOS serialisation of these taxonomies.

The conversion reports defects rather than repairing them, which is the part worth arguing for. Ad Product Taxonomy 2.0 contains two concepts that declare themselves as their own parent, a cycle in what is meant to be a tree, and eight rows whose name column disagrees with its own tier column, mostly casing but including Video Games and Consoles against a tier reading Video Game Consoles. Content 3.1 and Audience 1.1 are structurally clean. A converter that silently fixes its input is how a defect survives for years in every downstream copy, each consumer patching it privately and none of them telling the publisher.

One correction we made against ourselves. The first run reported 1,555 defects in the Audience Taxonomy. That was our bug, not theirs. The Audience file's name column is headed Condensed Name (1st, 2nd, Last Tier) and is a deliberate concatenation, so comparing it against the deepest tier flags every row in the file. Corrected, Audience 1.1 has zero defects. We mention it because a checker that reports a thousand defects in someone else's data is much more likely to be broken than the data is.

Nine findings, and what each one costs

All figures computed on 16 August 2026 from a live Wikidata harvest and same-day EIDR resolution. Both are living systems, so a later run gives different totals while the method reproduces exactly. The findings are computed twice, once set-based and once as SPARQL over the built graph, and agree exactly.

What the ontology does about it

An observation is not the thing observed. An AttentionEvent is something that happened in the world. An Observation is a record produced by a measurement method that purports to be evidence of it. Automatic content recognition, a panel meter, return path data and a server-side ad log watching the same household in the same hour produce four observations that will disagree about duration, about attribution and sometimes about what was on screen. The ontology holds all four, with their method, their operator and how the content was identified, rather than forcing a single number that erases the disagreement at ingestion. Each method has a specific blind spot worth recording: fingerprint matching recognises only what its reference library contains, so a non-match is evidence of absence from the library rather than absence of viewing, and a programme guide lookup resolves what was scheduled rather than what was transmitted, which makes it systematically wrong precisely during live sport and breaking news.

Identifiers are reified assertions, and scope is declared on the scheme. An EIDR identifier observed in Wikidata is Wikidata's claim about EIDR, and keeping that distinct from EIDR's own record is what made the coverage measurement possible at all. The registry states that IMDb title identifiers are work-scoped, that Ad-ID is creative-scoped, and that EIDR is multi-level, meaning a bare EIDR identifier does not tell you what it denotes until you resolve it. "Can I join these two datasets on this column?" is answered once in the registry rather than assumed separately by every consumer.

Identity links are graded, never owl:sameAs. A probabilistic or embedding match is evidence about identity, not identity. Embeddings place sequels, remakes, dubbed versions and same-franchise titles very close together, which is exactly the neighbourhood where media entity resolution has to be most careful. The 133 registry collisions are loaded as links with status proposed, a recorded method and a confidence, because a shared identifier is a reason to look rather than a licence to merge. A shape rejects any probabilistic link promoted to accepted below a confidence threshold, since accepting one at low confidence is how two households become one.

Derived attributes are auditable by construction. A genre, brand or topic affinity carries its window, its method, its version and the number of observations it rests on. A shape fails any affinity scoring above 0.8 that rests on fewer than three observations, which is arithmetically defensible and substantively meaningless, and the single most common way an audience graph acquires confident nonsense. A separate rule fails any derived attribute computed after the subject withdrew consent, because in a domain governed by video privacy statutes "which attributes depend on a withdrawn consent" must be one traversal rather than an investigation.

The shapes are not decorative. A probe graph contains exactly one violation of each of the eight rules and a test fails if any rule stops firing, because a SHACL rule that silently stops matching is worse than no rule: the run still passes and the gate still looks green. One rule genuinely did not fire on first run, because rdf:type inside a SPARQL constraint matches only direct types rather than walking the subclass hierarchy the way sh:targetClass does. It is fixed, and a regression test guards it.

The identifier discipline here was first proved on the United States fund universe in our Investment Fund Ontology, then on European insurance in the Insurance Register Ontology, and then on American academic standards in the Learning Standards Ontology. Media is its fourth instantiation, and the transfer without redesign is itself the finding: identifier fabrics fail the same way whatever they name.

Why this matters if you work in media measurement

Cross-screen deduplication. Deduplicating reach across linear, streaming and digital requires deciding that an exposure on one screen and an exposure on another were to the same content. If the content identifier is at title level on one side and cut level on the other, the deduplication is wrong in a direction nobody can see from the output. The 285 cross-level works here are in open data; a commercial catalogue assembled from multiple suppliers will not be cleaner.

Building a graph on content identity. Any team standing up a media knowledge graph has to choose a spine. The measurement to run before choosing is the one on this page: for your own catalogue, what proportion of each work type carries the identifier you were planning to join on, and at which level was it minted. If the answer for television is anything like 1.84 per cent, the spine is title strings with an identifier decoration, and it is much better to know that at design time.

Affinity and audience segments. Genre, brand and topic affinities are the product in audience intelligence. An affinity with no recorded window, method version or evidence count cannot be explained to a client, audited by a regulator, or debugged when it drifts. Making that structure mandatory in the schema costs nothing at write time and is impossible to retrofit once a billion of them exist.

Consent that has to be traceable. Television viewership is among the most regulated categories of behavioural data there is. The question a regulator asks first is not whether consent was captured but which downstream artefacts depend on a consent that has since been withdrawn. That is a graph traversal if consent is modelled at party level and a forensic project if it is not.

Grounding language models on a media graph. Every measurement business in 2026 wants models and agents reading its content and audience graph. A model grounded in a graph where one identifier names both a film and a particular cut of it, or where two works share a registry identifier, answers confidently and wrongly. The confidence is the dangerous part. Validation is not a step before the interesting work; it is what makes the interesting work safe to ship.

Common questions

Open, reproducible, and free to use

The ontology, the SKOS registries, the IAB taxonomy conversion, the eight-rule SHACL suite, the harvester, the EIDR resolver, the graph builder and the query library are public. Ontology and documentation under CC BY 4.0, code under MIT. The build report lists what could not be obtained as carefully as what could, including a research strand that was cut short and whose unverified output we therefore refused to use anywhere, because a study that only reports its successes is marketing.

One open question is stated rather than answered. Findings on collisions and level conflation count what appears in Wikidata, which is a third-party mirror maintained by editors and bots. How much is a mirroring artefact and how much reflects the registries themselves cannot be settled from open data, and we have deliberately not guessed. If you can resolve it, the issue tracker is the right place.

Related work: our machine-validated ontologies study measures when an ontology can reject a wrong statement, and the industrial ontology crosswalks measure how often standards crosswalks can detect a wrong mapping at all.