AI for Information Architecture: Organize and Test a Structure
Use AI to suggest content groups, compare labels, and challenge navigation. Test the structure with people, then keep a record of what they showed you and what you changed.
Updated 6 min read
Use AI for information architecture to propose ways to group content, generate clearer labels, and identify gaps in navigation. Treat the output as a draft structure, then test whether real people can find what they need. fenbs can keep the research references, naming decisions, and review outcomes available to the next person or assistant working on the structure.
Information architecture is how you organize and name information so people can find it. A neat site map is useful only if its categories match the tasks and language of its audience. Your job is to turn an assistant’s plausible organization into a structure supported by observed behavior.
Where should you start before asking AI to organize content?
- Choose a bounded area, such as the support center or one section of a college website. State whose tasks it must support.
- Inventory the content. Give every item a stable reference, current title, short description, owner, and intended audience.
- Add evidence of user needs: approved research notes, common support questions, and relevant search terms.
- Identify constraints, including required content, naming rules, accessibility needs, and items that need an owner’s review.
Do not ask the assistant to invent an inventory from a few page titles. Titles can hide duplicate material or distinct services that sound similar. Check the source pages first, and mark missing descriptions as gaps. Keep confidential content out of the input unless the tool and use are approved.
How can AI propose groups without hiding missing content?
Ask for two contrasting structures using the same inventory: perhaps one organized around audience tasks and another around subject matter. Require every item reference to appear, with ambiguous placements listed separately. Tell the assistant not to add new pages unless it labels them as proposals.
For example, a hypothetical community college site may contain tuition deadlines, payment plans, refund requests, and scholarship information. An assistant might put all four under “Finance.” That is a candidate label, not evidence that students will look there. Compare it with language from student questions before selecting anything.
Use a simple coverage check: count inventory references before and after grouping, flag duplicates, and inspect everything left unassigned. AI can help prepare the comparison, but verify it against the original inventory. Save the chosen grouping and its reasoning in fenbs, including placements still waiting for research.
What makes a useful navigation label?
A useful label helps the intended audience predict what sits behind it. Ask AI for alternatives in the words users already use, then check for overlap, vague categories, and internal department names that outsiders may not recognize. Shortening a label is helpful only when it keeps the meaning.
| Candidate label | Possible problem | Question to test |
|---|---|---|
| Resources | Its contents are difficult to predict. | What would a student expect to find here? |
| Student Accounts | Students may interpret it as a website login. | Where would someone look to arrange tuition payments? |
| Paying for College | It may combine distinct tasks too broadly. | Can students distinguish scholarships, payments, and refunds within it? |
Keep a glossary of approved terms, including the reason for rejecting tempting alternatives. The W3C menu guidance (opens in a new tab) recommends short, descriptive menu labels and consistent wording, order, and destinations across screen sizes. A menu label should not acquire a different meaning when the phone menu opens.
How do card sorting and tree testing help you check the draft?
Card sorting asks people to group representative content items and, in an open sort, name their groups. Use it to explore how people understand relationships. Digital.gov’s card-sorting framework (opens in a new tab) starts with discovery findings and short descriptions of the information participants need. AI can help prepare consistent descriptions, but check that it has not changed their meaning.
Tree testing asks people to find something within a proposed hierarchy, without the full page design. As Nielsen Norman Group explains (opens in a new tab), it evaluates a structure; card sorting helps generate possible groupings. The two methods answer different questions, so a convincing card sort does not finish your navigation testing.
Write realistic tasks without repeating the category label you are testing. For “Refunds,” ask where someone would learn what happens after withdrawing from a class, rather than asking them to find the refunds page. Ask AI to flag clues in your wording, then pilot the task with a person before the study.
How should you interpret confusing paths and conflicting results?
Keep observation separate from explanation. “The participant first selected Admissions” is an observation. “They think all student services belong in Admissions” is an interpretation that needs more evidence. Ask the assistant to list competing explanations and identify what additional observation could distinguish them.
Review patterns by task and audience instead of collapsing everything into one score. Record failed paths, backtracking, unfamiliar words, and comments that explain choices. Do not claim population-wide success rates from a handful of exploratory sessions. Decide what warrants another round of research with the person responsible for the study.
A short AI brain record can preserve the difference between a finding and a proposal. In fenbs, record the human decision about the next revision, link the supporting research, and leave unresolved explanations open rather than turning the assistant’s preferred story into an agreed fact.
What needs checking before the structure becomes a real menu?
A successful text hierarchy still needs testing in the actual interface. Check whether the menu works on a phone, whether people can identify their current location, and whether important destinations remain reachable. Test keyboard access and assistive technology with the implementation; a tidy diagram cannot demonstrate those behaviors.
Prepare a change list that connects each old content reference to its new location. Review orphaned pages, duplicate destinations, and items with no maintenance owner. Ask AI to compare the mapping with the approved structure, then verify its findings. Keep the original references so a renamed page can still be traced through the review.
Keep the reason behind every naming change
Use fenbs for approved labels, placement decisions, research links, and tasks for unresolved navigation problems. Keep the site map and study material in their working tools, with references in Links and listings. Your connected assistant can read the decision record before proposing another round of changes.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs
- Create a free account and add a project for the navigation work, with the audience, scope, and current structure reference.
- Add approved terms to the Glossary. Record the human choices about grouping and naming in Decisions and rules, with their reasons.
- Keep references to the inventory, research, and site map in Links and listings. Add tasks for confusing paths and the next test, with review criteria in their plans.
- Connect your supported assistant through Integrations. Have it read the project context before drafting changes, then save the checked results and next step before stopping.
fenbs is free for one person, no credit card. It does not run card sorts, conduct tree tests, or publish your navigation. If you only need to sketch a one-time site map, a diagram or document may be enough. fenbs helps when revisions and their reasons must survive several people or AI sessions.
How do you keep later changes from undoing the research?
Give each new page a proposed home and an owner, then check it against the naming decisions. When evidence contradicts an earlier choice, record the replacement decision and why it changed. Save what was tested as well as what was edited, so “menu updated” does not become “navigation validated” during a handoff.
This concerns website content and navigation; architecture diagrams and AI context architecture have separate guides. For a broader review workflow, see using AI at work and the AI at work hub. Start with one confusing journey and retain the evidence behind the improvement.