Card Sorting for Information Architecture
Why Navigation Decisions Shouldn't Be Made by Committee Guesswork
Website navigation and category structure are frequently decided internally by whoever is most senior in the room, using the organization's own internal mental model of how the business is structured — which often has little to do with how actual visitors think about the same content. Card sorting is a research method that externalizes this decision: participants from your actual target audience group a set of content labels into categories that make sense to them, revealing the mental model you should be designing navigation around instead of the one your org chart happens to produce.
Open vs Closed Card Sorting
In an open card sort, participants create their own category labels and group cards however they see fit — useful early in a redesign when you don't yet know what the right top-level categories even are. In a closed card sort, participants sort cards into categories you've already defined — useful for validating a proposed IA once you have a working hypothesis, checking whether content lands where you expect it to. Running an open sort first to generate candidate structures, then a closed sort with a different participant sample to validate the resulting structure, is the standard two-phase approach for a significant IA redesign.
Card sorting variants
| Variant | When to Use | Output |
|---|---|---|
| Open card sort | Early-stage, no existing IA hypothesis | Candidate category labels and groupings, analyzed via dendrogram/cluster analysis |
| Closed card sort | Validating a proposed IA | Pass/fail rate of cards landing in the expected category per participant |
| Hybrid card sort | Some categories are fixed (e.g., brand/legal requirements), others open | Validated required categories plus discovered additional ones |
Card Wording Itself Can Bias Results
Running the Study: Moderated vs Unmoderated at Scale
Small moderated card sorts (5-8 participants, done over a video call or in person) let you ask participants to think aloud while sorting, capturing *why* they group things a certain way — rich but slow to recruit and run. Unmoderated online card sorting tools (OptimalSort, UserZoom) let you run the same study with 30-50+ participants asynchronously, enabling quantitative cluster analysis across a larger sample, at the cost of losing the verbal reasoning behind each grouping decision. For a significant site-wide IA decision, running both — a small moderated round for qualitative reasoning, followed by a larger unmoderated round for statistical confidence — is the most rigorous approach.
// A simplified similarity-matrix calculation: for each pair of cards,
// what percentage of participants grouped them in the same category.
// High-similarity pairs (>70%) are strong candidates for the same
// navigation section; low-similarity pairs should stay separated.
interface SortResult {
participantId: string;
groups: string[][]; // each inner array is one category's card IDs
}
function buildSimilarityMatrix(results: SortResult[], cardIds: string[]) {
const matrix: Record<string, Record<string, number>> = {};
for (const cardA of cardIds) {
matrix[cardA] = {};
for (const cardB of cardIds) {
if (cardA === cardB) continue;
const together = results.filter(r =>
r.groups.some(g => g.includes(cardA) && g.includes(cardB))
).length;
matrix[cardA][cardB] = together / results.length;
}
}
return matrix;
}Card Sorting Validates Grouping, Not Labeling — Run Tree Testing Separately
What's Next?
Card sorting output feeds directly into the broader information architecture principles that govern how deep a navigation hierarchy should go and how categories should be named — a set of structural decisions worth examining as their own discipline.
I build these systems professionally.
Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.