Information Architecture

IA = structuring, organizing, and labeling content so people can find and understand it. The classic frame (Rosenfeld & Morville, the "polar bear book"): organization systems, labeling systems, navigation systems, search systems — designed for the trio of users, content, context.

Core concepts

• Information scent (Pirolli & Card's information-foraging theory, PARC): users follow cues (links, labels) that smell like their goal; weak scent = pogo-sticking and abandonment. Every label is a scent decision — specific beats clever, user vocabulary beats org vocabulary.
• Organization schemes: exact (alphabetical, chronological, geographical — easy, only good when users know the item's name) vs. ambiguous (by topic, task, audience — harder, usually more useful). Hybrid schemes on one level create confusion; keep levels internally consistent.
• Breadth vs. depth: moderately broad-and-shallow generally outperforms deep hierarchies (Larson & Czerwinski 1998); every added level is a decision + a chance to lose scent.
• Polyhierarchy & tagging: items can live in multiple categories; faceted classification (filter by several independent attributes) is the workhorse for large catalogs (Ranganathan's facet theory → ecommerce filters).
• Labeling: front-load keywords; match user language (mine search logs, support tickets, interviews); test comprehension — a label that needs explanation has failed.

Process

1. Content inventory/audit → what exists, what's ROT (redundant, outdated, trivial).
2. Open card sort (users group and name content) → candidate categories reflecting user mental models.
3. Draft hierarchy → tree test (findability tasks on the bare structure, no visual design) → iterate until success rates satisfy.
4. Define navigation + search together (users switch strategies; both must work — berry-picking behavior, Bates 1989).
5. Validate in-context with first-click tests on real pages.

Search-system notes

Support the query users type (synonyms, typos, plurals); results show scent (titles + snippets that differentiate); faceted refinement for big sets; "no results" is a design surface (see Onboarding, First-Run & Empty States).

Search-first products

When search is the primary access path (docs, marketplaces, large catalogs), the query log IS your IA research: failed and reformulated queries reveal missing content, missing synonyms, and vocabulary gaps. Browse structure still earns its keep — it teaches what exists (you can't search for what you can't name), disambiguates results (breadcrumbs and category badges in result lists are scent), and powers facets. Maintain a synonym ring / controlled vocabulary so "sign in," "log in," and "login" resolve identically.

Taxonomy governance

A taxonomy without an owner rots. Practice conventions: name a steward; define change control (who may add/rename/merge terms, with what review); keep a controlled vocabulary with preferred terms, synonyms, and scope notes (ANSI/NISO Z39.19 is the formal standard for this); schedule periodic audits against search logs and new content; log every change so downstream systems (search, navigation, analytics) can follow. Uncontrolled folksonomy tagging is fine for discovery input, not as the system of record.

Structured content & headless CMS

Model content as typed, reusable chunks with explicit fields and metadata — not as pages ("COPE: Create Once, Publish Everywhere," NPR's Jacobson 2009). Consequences for IA: the taxonomy lives in the content model (fields, references, tags), navigation is assembled per channel, and labeling errors propagate everywhere at once — governance matters more, not less. Author experience is part of the IA: if the model is too granular to author in, people will stuff everything into rich-text blobs and the structure dies.

LLM-era findability (emerging — practices unsettled)

A growing share of "visits" are AI agents and answer engines consuming content to answer users elsewhere. Emerging observations, honestly labeled: structure that helps humans scan (descriptive headings, self-contained sections, one idea per chunk, tables for tabular facts) also helps machines extract and cite accurately; schema.org metadata remains the established machine-readable layer. The llms.txt proposal (Howard/Answer.AI, 2024) — a curated Markdown index for LLM crawlers — has real but limited adoption as of mid-2026 and is a community convention, not a W3C/IETF standard; "generative engine optimization" is an emerging practice area with thin evidence. Don't restructure your IA around any of this yet; do keep content chunked, labeled, and factual — that serves both audiences.

Sources

• Rosenfeld, L., Morville, P. & Arango, J. (2015). Information Architecture: For the Web and Beyond (4th ed.). O'Reilly.
• Pirolli, P. & Card, S. (1999). "Information foraging." Psychological Review, 106(4).
• Larson, K. & Czerwinski, M. (1998). CHI '98 — breadth vs. depth.
• Bates, M. (1989). "The design of browsing and berrypicking techniques." Online Review, 13(5).
• Spencer, D. (2009). Card Sorting; Covert, A. (2014). How to Make Sense of Any Mess.
• ANSI/NISO Z39.19 — controlled vocabularies standard.
• Jacobson, D. (2009). "COPE: Create Once, Publish Everywhere." NPR (programmableweb write-up).
• llms.txt proposal — llmstxt.org (Howard, 2024; community proposal).