How to Build a Content Taxonomy: A Step-by-Step Guide

A content taxonomy is an agreed set of terms, organised into a hierarchy or a set of facets, that an organisation uses to label its content consistently. You build one by deciding what it is for, studying the content and the words people actually use, choosing a structure, fixing one preferred label per concept with its synonyms, and putting a small group in charge of changes.

Done well, a taxonomy makes search filters, topic pages, related content, analytics and AI retrieval work across the archive. Done badly, it becomes another uncontrolled tag list.

This guide walks through the process, drawing on the US standard for controlled vocabularies (ANSI/NISO Z39.19) and the W3C's SKOS model, and ends with a worked example of facets for a publisher.

Key takeaways

• Start from the jobs the taxonomy must do and the people who will apply and use it, not from a blank list of topics.
• Base terms on evidence: the words in your content, the words your users search with, and what your organisation needs to report on.
• Facets (separate lists for subject, audience, type, format and place) usually beat one deep tree, because each item can be described along several independent dimensions.
• Record one preferred label per concept with its synonyms and a scope note, using a model such as W3C SKOS so the taxonomy can be exchanged and mapped to standards.
• Treat the taxonomy as a maintained product: owners, a change process, retired rather than deleted terms, and regular testing against real content.

What a taxonomy is, and what it is not

ANSI/NISO Z39.19, Guidelines for the Construction, Format, and Management of Monolingual Controlled Vocabularies, describes four kinds of controlled vocabulary. It defines a taxonomy as a collection of controlled vocabulary terms organised into a hierarchical structure, where each term has one or more broader or narrower relationships to others.

The standard gives the purpose of vocabulary control as consistent description and better retrieval, achieved by defining the scope of each term, linking synonyms, and distinguishing homographs (such as "Mercury" the planet and the element). Most content taxonomies end up between a taxonomy and a light thesaurus. The current edition of Z39.19 dates from 2005 and was reaffirmed in 2010; NISO lists a revision among its active projects.

Plan: purpose, users and a content audit

Step 1: Define the purpose and the users

Write down, in a few sentences, what the taxonomy is for: search filters and browse pages, related content, analytics by topic or audience, syndication feeds, curriculum alignment, or filtering for AI retrieval. Then list who applies tags (editors, librarians, an AI system, or a mix) and who uses the results. A taxonomy built for analysts can tolerate technical labels; one shown to readers cannot. Set the scope too: if it must describe fifty years of material, it needs terms for things that no longer exist.

Step 2: Audit the content and existing tags

Sample every content type and decade you need, and export the labels that already exist: tags, categories, sections, folder names and search logs.

Z39.19 calls the evidence for a term its warrant, and names three kinds:

• Literary warrant: the term appears frequently in the content itself.
• User warrant: people use the term when they search or ask for information.
• Organisational warrant: the organisation needs the term because of its own context and priorities, such as a reporting category or a product line.

A good audit uses all three: content shows what you hold, search logs how people ask for it, and stakeholder interviews what the business must count.

Design: terms, structure and vocabulary control

Step 3: Gather and normalise candidate terms

Pull candidate terms from existing tags, search queries, key phrases and entities extracted from the content (AI helps here), and relevant standards such as IPTC Media Topics for news (see our guide to IPTC Media Topics), EDItEUR's Thema subject scheme for books, or the curriculum you publish for.

Then normalise:

• Group synonyms and pick one preferred label for each concept.
• Separate different kinds of thing. People, organisations and places are usually better handled as named-entity lists than as topics.
• Set label rules. Z39.19 recommends that count nouns normally take the plural ("elections"), while IPTC Media Topics labels include singular forms such as "festival". Either works if applied consistently.
• Write a one-line scope note for any term that two people might interpret differently.

Step 4: Choose a structure

Z39.19 traces facet analysis to the work of Ranganathan and the Classification Research Group, and describes it as particularly useful for interdisciplinary subjects, cases where several hierarchies are needed, and electronic content where physical location does not matter, which describes most digital archives.

Polyhierarchy is sometimes the honest answer: Z39.19's example places pianos under both stringed and percussion instruments. IPTC Media Topics takes the other route: in its September 2026 release, no current concept has more than one broader term, so each sits in exactly one place in the tree.

Keep the hierarchy shallow enough for people to apply. When IPTC set requirements for Media Topics in 2007, one goal was at most 800 terms, excluding sports competitions, so journalists could scan it quickly.

Step 5: Controlled vocabulary or folksonomy?

NISO's Understanding Metadata primer describes a controlled vocabulary as a predetermined list of terms that identifies one preferred word or phrase for each concept, sometimes maps other terms to it, and often defines relationships among terms. A folksonomy is the opposite: free tags added by staff or users, with no agreed list.

In a hybrid, editors add free keywords, someone reviews them on a schedule, and the ones with real warrant are promoted into the controlled vocabulary.

Step 6: Record labels, synonyms and definitions with SKOS

SKOS, the Simple Knowledge Organization System, is a W3C model for representing thesauri, taxonomies, classification schemes and similar vocabularies, published in August 2009. Using SKOS, even only as an export format, makes a taxonomy portable between tools and mappable to standards. The main parts:

• skos:prefLabel: the preferred label. A concept can have only one per language tag, and SKOS recommends that no two concepts in the same scheme share a preferred label in the same language.
• skos:altLabel: synonyms, near-synonyms, abbreviations and acronyms.
• skos:hiddenLabel: forms used for search but never shown, such as common misspellings.
• skos:broader, skos:narrower and skos:related: hierarchy and association.
• skos:scopeNote, skos:definition and skos:changeNote: meaning and change history.
• skos:exactMatch, skos:closeMatch and skos:broadMatch: links to concepts in other schemes.

@prefix skos: <http://www.w3.org/2004/02/skos/core#> .
@prefix ex: <https://example.com/taxonomy/> .

ex:climate-change a skos:Concept ;
    skos:inScheme ex:topics ;
    skos:prefLabel "Climate change"@en-GB , "Athrú aeráide"@ga ;
    skos:altLabel "Global warming"@en-GB , "Climate crisis"@en-GB ;
    skos:hiddenLabel "climate chnage"@en-GB ;
    skos:broader ex:environment ;
    skos:scopeNote "Causes, effects of and responses to long-term climate change. For daily conditions use Weather."@en-GB ;
    skos:closeMatch <http://cv.iptc.org/newscodes/mediatopic/20000418> .

The last line links the house concept to IPTC's "climate change" Media Topic. closeMatch is used because the two definitions are similar but not identical; IPTC's own guidance is to use exactMatch only when two concepts are truly equivalent.

Govern, test and roll out

Step 7: Set up governance and change management

Z39.19 is blunt about maintenance: language changes, new concepts emerge and terms become obsolete, so a controlled vocabulary is never frozen. Plan for change from the start:

1. Name an owner and a small review group with editorial and technical members.
2. Publish a request process. Proposals state the warrant and example content, and criteria for adding, merging, splitting and retiring terms are agreed in advance.
3. Retire, never delete. IPTC marks outdated Media Topics as retired, usually with a note on what to use instead, so older tagged content stays interpretable. Do the same.
4. Keep a change log and history notes. Z39.19 describes history notes for tracking how terms change over time.
5. Release on a schedule. IPTC releases Media Topics updates every two to three months; a quarterly cycle suits most organisations.
6. Plan re-tagging. Splitting a term means existing content must be reclassified; merging only needs a mapping.

Step 8: Test with real content

Before rollout, test on a few hundred items not used to build the taxonomy:

• Tagging test. Ask two or three people to tag the same items independently. Where they disagree, a scope note is unclear or two terms overlap. Our guide to evaluating AI tagging explains how to measure agreement with Cohen's kappa.
• Coverage and balance. Count items that fit no term (a gap), terms that attract no items (probably unnecessary) and terms that take a large share of the sample (probably too broad).
• Findability. Give users realistic search and browse tasks and watch where they go wrong.

Step 9: Roll out with AI-assisted tagging

A taxonomy is only useful once the archive is tagged with it, and back-tagging years of content by hand is rarely realistic:

1. Turn the taxonomy into instructions. Labels, scope notes and a few example items per term give an AI model far better guidance than labels alone.
2. Tag a reviewed sample first and measure precision and recall per term.
3. Set confidence thresholds: apply high-confidence labels automatically and send uncertain ones to editors.
4. Put suggestions in the editing workflow so editors accept or reject tags on new content at publication.
5. Monitor. Watch for terms that suddenly grow or shrink, and re-test after each taxonomy release.

A worked example: facets for an educational and trade publisher

Imagine a publisher of school textbooks, workbooks, audio and posters, plus some trade non-fiction. A single tree such as "Maths > Fourth class > Workbooks > Fractions" would need a branch for every combination of subject, level and resource type. Facets keep each list short:

A workbook page on comparing fractions is then Subject: Fractions; Level: Fourth class; Resource type: Workbook; Format: PDF; Language: English. Each facet filters on its own, and one subject list serves every level and format.

For trade titles, EDItEUR's Thema, the subject category scheme for a global book trade, reached version 1.7 on 28 October 2026. It is published in many languages, some only partly translated, and as a SKOS concept scheme. For school content, Schema.org's LearningResource type offers learningResourceType, educationalLevel and teaches for publishing the same facets on the web. Our guides to metadata for publishers and AI in educational publishing cover both in more depth.

How Druid Learning helps

Druid Learning is an AI archival processing platform, based in Dublin and New York, that turns content archives into AI-ready datasets. Once a taxonomy is agreed, its Standardization Tool provides real-time content tagging and labelling, and Archive Optimization turns archives in any format into an indexed, labelled dataset. For the Irish educational publisher CJ Fallon, Druid processed 50,000 resources and mapped them to the curriculum. For NPR, it applied 20+ labels across 8 dimensions, including user needs, article type, location, topic and community, to more than 2 million assets.

Frequently asked questions

What is the difference between a taxonomy and a controlled vocabulary?

A controlled vocabulary is any agreed list of terms with one preferred label per concept. A taxonomy is a controlled vocabulary organised into a hierarchy of broader and narrower terms. ANSI/NISO Z39.19 describes four kinds of controlled vocabulary: lists, synonym rings, taxonomies and thesauri, which add equivalence and associative relationships to the hierarchy.

How many levels should a content taxonomy have?

As few as the job allows. Each extra level makes tagging slower and less consistent, because people must choose between finer distinctions. Several shallow facets, such as subject, audience, type and place, often work better than one deep tree. As a reference point, IPTC Media Topics has 17 top-level topics in a tree IPTC extended from three levels to five, and its designers originally aimed for at most 800 terms so journalists could scan it quickly.

Should we use a standard taxonomy or build our own?

Usually both. Adopting a standard such as IPTC Media Topics for news, Thema for books or a national curriculum for education gives you tested terms and easier exchange with partners. Then extend it with house terms for places, products, series or audiences the standard does not cover, kept in your own scheme and linked to the nearest standard term with a mapping relation.

What is SKOS and do we need it?

SKOS, the Simple Knowledge Organization System, is a W3C model for representing taxonomies, thesauri and classification schemes as linked data. It defines preferred, alternative and hidden labels, broader, narrower and related links, notes, and mappings to other schemes. You do not need to manage a taxonomy in SKOS, but exporting it as SKOS makes it portable between tools and easy to map to standards such as IPTC Media Topics or Thema.

How do we keep a taxonomy up to date?

Give it an owner, a published request process and a release schedule, such as quarterly. Require evidence for new terms from content or search logs, and log every change. Never delete terms: mark them retired with a note pointing to the replacement, as IPTC does for Media Topics, so older content stays interpretable. When a term is split, re-tag the affected content.

Sources

• ANSI/NISO Z39.19-2005 (R2010): Guidelines for the Construction, Format, and Management of Monolingual Controlled Vocabularies — NISO
• Information Creation and Curation Topic Committee — NISO
• SKOS Simple Knowledge Organization System Primer — W3C
• Understanding Metadata: What is Metadata, and What is it For? — NISO
• NewsCodes Guidelines — IPTC
• Media Topic NewsCodes vocabulary — IPTC Controlled Vocabulary Server
• Thema — EDItEUR
• LearningResource — Schema.org