AI Policy: I personally write each draft and final copy on this website. All content reflects my own thinking, ideas, style, and craft.  I do not use AI such as ChatGPT or other LLMs to generate articles. Occasionally, I ask AI (such as Formalizer or Equativ) to summarize or re-state my own ideas and may restructure sections based on the response.

My recent experience with language models for website information architecture redesign

If you’re new here, I’m an information architecture consultant and publish a new article every Tuesday on my website, Little Language Models.

In most content management systems, including mine (WordPress), web administrators can group their posts by categories and tags. Categories broadly group topics; tags specify details.

WordPress comes with a few different post types and offers the option to create custom ones known as custom post types (CPT). (For example, if you want to have a section on Books, it would be better suited to creating a custom post type for them.)

When I first started publishing these articles in March 2025, I kept it simple by creating 3 wide categories: Information Architecture, Small business websites, and Consulting.

Since then, I’ve published 24385 words, started a newsletter, reached readers in 50 countries, and have gotten into a publishing groove. 🪵knock on wood🪵

I wanted to create a more robust categorization system because:

• Website readers can quickly find relevant information and skip articles that are irrelevant to their goals.
• Clearer categories improve search engine optimization (SEO) and large language model optimization (LLMO), making it easier for robots to find and reference relevant articles to people looking for them in search engines or AI interfaces.
• Robust categories help authors more easily find That Thing we wrote, which makes it easier to reference that article in related articles (known as internal linking), which also boosts SEO and creates a nice rabbit hole for readers interested in specific topics.

I used Duck.ai to group articles into categories. Duck.ai allows you to have private conversations with 3rd-party AI chat models, anonymized by DuckDuckGo.

Since this was a free feature, Duck.ai couldn’t find and scan the articles for me. I pasted all my articles (or rather, what fit within the character count) and told it to please categorize the pasted articles below into article categories.

I asked a few 3rd-party AI chat models to categorize my content, which is a key aspect of information architecture work, and this is what they came up with:

1. Information Architecture Consulting
2. Content Audits
3. Stakeholder Management
4. User Research and Testing
5. Best Practices in Information Architecture

This list of categories isn’t bad, but it isn’t great.

The first category includes all the other categories, so in technical terms, there’s a parent-child relation between them. Non-stated hierarchical relations can create problems for engineers in larger content modeling redesigns.

The second and third categories are activities related to information architecture, but the other categories aren’t.

My concerns with using ChatGPT or other language models for website information architecture redesign

AI currently lacks factors such as human intuition (reading the room), creativity, complexity of human behavior, change management, and empathy, which play a crucial role in creating an effective information architecture. There’s so much work done between the lines as a consultant (or employee in a dynamic and challenging environment)—tasks beyond current AI capabilities.

AI tools may occasionally overlook important business context, and giving them all the necessary company information can be time-consuming, which defeats the purpose of reducing time and costs.

AI models tend to hallucinate: they say something is true when it isn’t. A BBC investigation found that over half of AI-generated news have “significant issues” like hallucinations (incorrect factual statements or numbers) and altered quotes.

AI demonstrates bias. Among many demonstrated cases, ProPublica found a criminal justice algorithm mislabeled African-American defendants as “high risk” at nearly twice the rate it mislabeled white defendants. 

Artificial intelligence (AI) has an environmental cost. Beginning with the extraction of raw materials and the manufacturing of AI infrastructure, and culminating in real-time interactions with users, every aspect of the AI lifecycle consumes natural resources – energy, water, and minerals – and releases greenhouse gases. The amount of energy needed to power AI now outpaces what renewable energy sources can provide, and the rapidly increasing usage of AI portends significant environmental consequences.

The Environmental Impacts of AI – Policy Primer from Sasha Luccioni, Bruna Trevelin, Margaret Mitchell (HuggingFace)

Besides the concerns above, current AI privacy risks, such as sensitive data collection, data leakage, and collection of data without the consent of the people or companies from whom it’s being collected, are concerning to me, so I can’t recommend relying on AI for information architecture.

Taxonomy synthesis – A good AI use case in information architecture

In my opinion, it’s more suitable to consider AI as an intern rather than a strategist, which reflects my personal experience in deriving value from it.

AI can assist with menial tasks like taxonomy synthesis.

For large-scale content audits, we can export thousands of categories and tags and custom post types from the CMS and ask AI to group similar-sounding ones.

To take an example from the medical industry, let’s talk about medical simulation. Medical simulation employs real-life-looking mannequins and virtual reality to provide a realistic, secure space for healthcare professionals to practice skills and refine techniques.

Medical simulation services are categorized into several distinct classifications, including disciplines, roles, and phases. Each classification has its own sub-categories, i.e., Phases include Needs Assessment, Operations, Pre-briefing, Facilitation, Debriefing.

For a medical simulation education provider, platform administrators can tag a service to multiple classifications, both within a category (tagging 2 phases: Needs Assessment, Operations) and across categories (different disciplines, roles), so they’re shown to people visiting relevant pages.

Over the years, when linking services and resources to the Needs Assessment phase, different people might have added “Assessing needs” (different phrasing) or “needs assessment” (lower case) or “needs asesment” (typo).

Miro’s clustering works quite well for taxonomy synthesis.

“Use Cluster by keywords to jump-start the organization of thoughts and themes across sticky notes — a perfect tool for brainstorms, large workshops, or user research projects.

Imagine dealing with thousands of variations for every tag, category, post type, and having to manually review them and decide what stays. Now imagine having a helpful intern, AI, who’s eager to help. What a relief.