Designing Information Architecture for an AI-Mediated Web

What an AI Assistant Needs Before It Can Cite a Site

When an AI assistant is going to serve as the front door for part of a Web site’s audience, the practical question becomes: What does the AI need from the site’s content to select and quote it accurately? Figure 2 breaks down the four things that I have found matter most.

To select and accurately quote a site’s content, an AI requires the following essential contextual information:

• clear entity labels—The labels state plainly what a page is about, using language that matches how people actually ask questions, not a company’s internal jargon. For example, if a returns page were titled “Post-Purchase Logistics,” the AI model would have to work much harder to connect the page to the user’s question: Can I return this? Even worse, the AI might instead take the user to a competitor’s more clearly labeled page.
• self-contained answers—These answers really matter because a citation pulls a chunk of a page’s content entirely out of its surrounding navigation and context. Thus, a paragraph that uses an “as mentioned above” reference or a table that makes sense only after a user has read three prior sections won’t translate well into an AI assistant’s answer. Content that gets cited well is content that would still make sense if you handed someone just that one paragraph and nothing more.
• structured data markup—Using schema.org markup, structured data such as frequently asked questions (FAQs), and clear subheading hierarchies gives an AI model an explicit, machine-readable signal about what a piece of content is and how it relates to the rest of a page rather than making the AI infer that context from the page layout alone.
• stable, linkable URLs—The uniform resource locators (URLs) matter because a citation is useful only if it resolves properly when someone clicks a link. Sites that restructure their taxonomy every year without providing proper redirects not only break links for search engines, they create AI citations that quietly stop working, with no obvious signal to the team that this has happened.

What This New Model Does and Doesn’t Change About IA Fundamentals

It would be a huge mistake to read any of what I’ve said here as: Information architecture no longer matters. Optimize for AI instead. The fundamentals of information architecture—grouping content by user goals, maintaining a coherent taxonomy, and providing consistent navigation across platforms—are exactly what makes content legible to both humans and AI models. An inconsistent information architecture that would confuse users would result in confusing, inconsistent citations. However, while the same IA fundamentals still hold true, what changes is what parts of the information architecture must do double duty, as shown in Table 1.

The information architecture’s taxonomy and labels still exist to group content in the ways users think about it. That work doesn’t go away, and now those labels are also the entity signals that an AI model uses to match a user’s query to a page. Internal linking still signals what is related and important, but now it also shapes the citation graph a model builds when deciding what content to trust. Page hierarchy matters somewhat less at the user’s exact entry point, since a citation can land a user three levels deep with no visible trail, but that hierarchy still matters the moment that user wants to explore further. The page’s structure, which has always existed to guide scanning and reading, now carries the additional burden of having to make sense to users with no surrounding context.

None of this requires a new discipline that is being bolted onto information architecture. Information architecture is the same discipline, but it is being evaluated against a new kind of reader that happens to arrive first, on behalf of a human who arrives second—or not at all.

A Practical Starting Point

You don’t need to redesign an entire Web site to accommodate this change. Start by redesigning the pages that can most likely answer discrete, well-defined questions—for example, questions about return policies, pricing tiers, specification sheets, and troubleshooting steps, then audit those pages against the four requirements I’ve laid out in Figure 2. Does the page state clearly what it’s about in the first two sentences? Would the key paragraph make sense if it were lifted out of context? Is there structured markup describing what kind of content is on the page? Does the URL look like something that will still exist in two years?

Redesigning a few pages to meet these requirements is a small, bounded piece of IA work in comparison to a full IA overhaul. Plus, it is the piece that is most likely to pay off first because you’re focusing on exactly the kind of content that users are already asking AI assistants to provide.

Conclusion: Designing for a Reader That Isn’t Human

Information architects have traditionally made every IA design decision on behalf of a human user who is trying to find something on a Web site. This is no longer true. What has changed is that, for a growing share of a site’s audience, the first reader of the information architecture is a model that is deciding whether the content deserves to be the answer to a user’s question. Only after the AI has made that decision does a person ever see the page. Designing an information architecture that serves both readers well is in no way a departure from the discipline. It is the same discipline, but we must now apply it to an AI audience that reads everything but clicks nothing.