Information Architecture Research: The Complete Guide (2026)

Information architecture (IA) research is the practice of studying how users categorize, label, and look for information so you can structure a product, website, or app the way they actually think. Done well, it is the difference between a navigation users glide through and one they abandon. The core toolkit is small and proven: card sorting to discover users' mental models, tree testing to validate a structure, first-click testing to check entry points, and qualitative interviews to understand the why behind the patterns.

Why Information Architecture Research Matters

When structure does not match users' mental models, they leave — fast. The business cost is well documented:

• 37% of users abandon a website specifically because of poor navigation or layout (PR Newswire / UX research).
• Users who can navigate efficiently complete tasks at dramatically higher success rates; when findability breaks down, task success can fall to the 44–46% range (UXmatters).
• Users frequently give a page only 10–15 seconds to prove it has what they need before leaving — and structure is what makes the answer obvious or invisible.

As Jakob Nielsen of Nielsen Norman Group has long argued, navigation and findability are not cosmetic — they are core to whether a digital product works at all. Good IA is invisible; bad IA is the reason a support queue fills with "I couldn't find…" tickets.

The Goal: Mental Models, Not Org Charts

The central mistake in IA is structuring content around how your company is organized (departments, internal jargon, product lines) instead of how users think about the domain. IA research exists to close that gap. Every method below is a way to extract the user's mental model and then test whether your structure honors it.

Core Information Architecture Research Methods

1. Card Sorting (Generative)

In a card sort, participants group labeled cards — representing content, features, or topics — into categories that make sense to them. The result reveals how users naturally cluster information and what they would call each group.

• Open card sort: participants create and name their own categories. Best for designing a new IA from scratch.
• Closed card sort: participants sort cards into your predefined categories. Best for validating labels.
• Hybrid: a mix — participants use your categories but can add their own.

Sample size: Nielsen Norman Group recommends about 15 participants per user group for card sorting, while the influential Tullis & Wood (2004) study — which tested 168 participants — found that 20–30 participants yield results nearly identical to those from hundreds (NN/G).

2. Tree Testing (Evaluative)

Tree testing flips card sorting around: you give participants your proposed structure (as a text-only "tree," with no visual design to bias them) and ask them to find where they would go to complete a task. It measures whether your IA actually works.

As NN/G frames it, "card sort studies help shape information architectures; tree-testing studies evaluate them." Because tree testing is quantitative, it needs more participants — NN/G recommends around 50 per tree test, with 50–150 being ideal for stable findability scores.

3. First-Click Testing

First-click testing measures where users click first when given a task. It matters because research consistently shows users who get the first click right are far more likely to complete the task successfully. It is a fast, high-signal check on whether your top-level navigation points people in the right direction.

4. Top-Tasks Analysis

Top-tasks analysis identifies the handful of things users actually come to do, so you can prioritize IA around them instead of burying them under rarely-used content. IA optimized for the top 5% of tasks beats IA that treats all content equally.

5. Mental-Model and Findability Interviews (The "Why")

Card sorts and tree tests tell you what users do and where they struggle. They rarely tell you why. To understand the reasoning — the vocabulary users expect, the categories that feel "wrong," the moments of hesitation — you need conversation. This qualitative layer is what turns a similarity matrix into a redesign you can defend.

The Modern Workflow: Combine Quantitative IA Tests With AI Interviews

Traditionally, the quantitative layer (card sort, tree test) and the qualitative layer (interviews) were separate, slow exercises. You would run an unmoderated card sort in one tool, then spend weeks scheduling moderated interviews to understand the patterns. AI-native research collapses that gap.

With Koji, you can run the qualitative half of IA research at the scale and speed of the quantitative half. After (or alongside) a card sort or tree test, launch an AI-moderated interview behind a single link. The AI interviewer asks every participant about their reasoning and probes follow-ups automatically — "You put 'Invoices' and 'Receipts' in different groups; what makes them different to you?" — capturing the mental-model rationale that a static tool never could. Interviews can run as voice or text, are automatically transcribed, and are synthesized into themes in real time, so the why behind your navigation data lands the same day.

Using Structured Questions to Quantify Mental Models

Koji's six structured question types — open_ended, scale, single_choice, multiple_choice, ranking, and yes_no — let a single study blend IA measurement with reasoning:

• A ranking question — "Order these labels by where you'd expect to find pricing information" — produces a quantified preference order across participants, much like a lightweight tree test.
• A single_choice question — "Which of these category names is clearest?" — settles labeling debates with data.
• A scale question — "How confident were you that you'd find it there?" — captures the hesitation that predicts navigation failure, with an anchored follow-up asking why.
• open_ended questions capture the exact words users use for categories — the raw material for navigation labels.

Where legacy survey tools like SurveyMonkey can only collect static answers, an AI-native platform turns each response into a probed conversation and an aggregated theme — giving you both the numbers and the narrative behind your information architecture.

A Practical IA Research Sequence

1. Discover — Run an open card sort and mental-model interviews to learn how users group and name things.
2. Design — Draft an IA based on the dominant clusters and the vocabulary participants actually used.
3. Validate — Run a tree test and first-click test on the draft structure.
4. Understand failures — Use AI-moderated interviews to probe the tasks where findability scored low.
5. Iterate — Adjust labels and hierarchy, then re-test. IA research is cheap to repeat and expensive to skip.

Common Information Architecture Research Mistakes

• Designing around the org chart instead of user mental models.
• Skipping validation — designing an IA from a card sort but never tree-testing it.
• Under-powering tree tests — they need ~50 participants, far more than a card sort.
• Collecting structure without reasoning — numbers without the "why" lead to changes you cannot defend to stakeholders.
• Testing labels in isolation — always test within realistic task scenarios, not as abstract words.

Conclusion

Information architecture research is a tight, repeatable loop: discover mental models with card sorting and interviews, design a structure from them, validate it with tree testing and first-click testing, and probe the failures with conversation. The bottleneck has always been the qualitative half — slow to schedule and slow to analyze. With Koji's AI-moderated interviews and structured questions, you can capture the reasoning behind your navigation data at the same speed you capture the data itself, and ship an IA that matches how users actually think.

How to Analyze Information Architecture Research

Each method produces a distinct, actionable artifact:

• Card sort → similarity matrix and dendrogram. The similarity matrix shows how often any two items were grouped together; the dendrogram visualizes the resulting clusters. Pair these with the category names participants invented — that vocabulary is the raw material for your navigation labels.
• Tree test → success rate, directness, and time. Success rate is the share of participants who found the right location; directness is the share who got there without backtracking. Low directness on a task with otherwise high success still signals a confusing structure.
• First-click test → first-click success rate per task. A first click into the wrong branch reliably predicts task failure, making it an early-warning metric for a broken top level.

The trap is treating these numbers as the whole story. A dendrogram tells you that users split "Billing" from "Account settings," but only a conversation tells you the mental model behind the split — and that reasoning is what lets you defend the redesign to stakeholders. This is exactly why pairing quantitative IA tests with AI-moderated interviews matters: Koji aggregates the why across every participant automatically, so your similarity matrix arrives with its explanation attached instead of weeks later.

Related Resources

• Structured Questions Guide — the six question types that quantify mental models
• Card Sorting Guide — the generative method for designing IA
• Tree Testing Guide — validate a proposed structure
• First-Click Testing Guide — check whether navigation points users the right way
• Top-Tasks Analysis Guide — prioritize IA around what users actually do
• Mental Models Research Guide — understand the reasoning behind user expectations