Information Architecture for a World That Doesn't Browse Anymore

02 Oct, 2026

1. The Premise

As someone who relies on the internet for research and general information, I've used a handful of search engines as my go-to tools for finding things. Google, Bing, and Yahoo have been the staples for decades, returning ranked, relevant results within seconds. Search engines are still relevant today, but a new technology is taking over a job they have held for decades: artificial intelligence. Instead of leaving the judgement to the user, AI now provides a summarized or direct response to a query. In this essay, I look at how AI is changing user search behaviour, and how that shift is forcing websites to rethink their information architecture to keep pace with it.

Before Google took over the web, people used a mix of tools to find information. Yahoo! organized websites into human-curated categories, so rather than typing a question or keyword, users browsed a hierarchy. Alongside Yahoo!'s directory, search engines such as AltaVista, Lycos, Excite, Infoseek, WebCrawler, and HotBot let users search an index of web pages — though, in that era, keyword matching was far less precise than it would later become, and many users relied as much on bookmarks and memorized addresses as on search. Google changed the landscape with PageRank: instead of browsing Yahoo!'s human-curated directories, users could type a keyword and get matched results, ranked in part by how many other reputable pages linked to them. That model held until AI answer engines — ChatGPT, Perplexity, Google AI Overviews, Claude — changed how results are crawled, synthesized, generated, and consumed.

In short, AI engines work on a different premise than traditional search. They let a user type a question rather than a keyword, and return a direct response rather than a list of links. The traditional route left the judgement to the user: browse the links, decide which one actually answers the question. AI answer engines instead crawl and chunk information from multiple sources, synthesize it, and cite it in a single response.

Having traced that shift, it's worth being precise about what it does and doesn't mean for information architecture: the discipline doesn't lose its relevance, but it picks up an additional job. Google's own guidance on generative AI search confirms that core SEO fundamentals still apply, since its AI features are built on the same ranking and quality systems as traditional search.¹ What AI answer engines add is a requirement that content also work as something that can be retrieved, verified, and assembled on its own — which is where IA's work begins to change.

2. What changed, concretely

The Google Search era rewarded ranking. PageRank, and the methodology behind it, set off a rush among websites to be found and clicked. Google Analytics gave teams a vocabulary for measuring that performance — page views, active users, acquisition.² Behind it sat a discipline that predates Google: SEO. Google didn't invent SEO, but it became the main force behind its modern logic, shaping how developers structured sites and content so that search bots could crawl, index, and rank pages for keyword queries.

That model shaped information architecture for decades. Content was organized hierarchically — homepages, categories, subcategories, and deeper pages — and navigation (menus, breadcrumbs, sitemaps, contextual links) helped users move through that hierarchy. Labels became descriptive: "pricing plans" instead of "click here," giving both users and Google a clearer signal of intent. The result was a browsing model, built on hierarchy and interconnected pathways, that persists whenever someone types a keyword into Google and gets a ranked list of links to sort through themselves.

AI answer engines work differently. Long documents are divided into passages — chunks — and instead of returning links, the engine synthesizes information from multiple chunks into one output. The question "which database is best for our startup?" might trigger simultaneous searches across pricing, scalability, operational complexity, and benchmarks, assembled into one cited answer. The underlying mental model has shifted: traditional search asks "which pages are relevant?" An AI answer engine asks "what answer can I construct from the relevant information?"

Mechanically, the process runs in three loosely sequential stages. Retrieve: the engine finds potentially relevant content — crawling, indexing, chunking, and running keyword or semantic search across it. Reconcile: the engine interprets the question (often rewriting or expanding it before or during retrieval), compares what it found, and attempts to weigh relevance and apparent authority. Respond: the model generates a natural-language answer from that material, typically with citations or qualifications attached. It's worth being precise about what "reconcile" can and can't do: an engine can compare sources and favour ones it judges more authoritative, but it has no independent way to verify a claim against your actual records or policies — which is exactly the gap a governed source of truth exists to close, a point I'll come back to.

This shift is measurable. In a March 2025 Pew Research Center analysis of roughly 900 U.S. adults' Google searches, an AI-generated summary appeared in 18% of all searches, and in 60% of queries beginning with "who," "what," "when," "where," "why," or "how." When a summary appeared, users clicked a traditional search result in only 8% of those visits, compared with 15% when no summary appeared.³ That doesn't show people have abandoned traditional search — Pew is explicit that it's a snapshot, not a trend line — but it does establish that, at least on Google, a meaningful share of searches now end at a synthesized answer rather than a list of links.

That changes what "findability" means. In the browsing era, findability meant a user could locate the right page. In the AI-search era, it means the correct fragment gets retrieved and attributed correctly — a considerably higher bar, because the reader never sees the alternatives the system discarded, or the context the fragment was lifted from.

3. The bridge from DTH

In keyword search, a human reads a list of links, exercises judgement, and chooses a source. The search engine doesn't always surface the right page — it can return something confusing, or stale, or simply off-target — but that has always been treated as a UX problem with the website, not a failure attributable to the user. In AI-answer retrieval, the model makes that judgement instead, and it does so out of sight. The user doesn't see the messy source material search engines were known to expose; they see only the model's polished answer.

This is where the earlier case study on the corporate website for a media and entertainment company becomes directly relevant. The core finding there was that no single group within the organization owned the decision about which content was authoritative, or what should be published. That was a failure of content governance — the rules, roles, and processes that should have governed the site's restructuring were absent — and it showed up as drift between research insight and implementation, because no one was accountable for the handoff between the marketing and IT teams. The resulting turf war exposed how fragile the architecture actually was, and stalled the redesign.

Even as traditional search engines like Google and DuckDuckGo work AI-generated answers alongside conventional results, the handoff this article is concerned with isn't the one between internal teams. It's the handoff from your content to a machine that represents your brand's answer to a stranger, with no editorial oversight in between.

Consider a concrete version of the failure: a return-window policy is updated on your current FAQ page, but an older blog post from several years ago still states the previous window, and a product page footer references it too. Nobody owns the job of checking whether those other instances were updated. An AI answer engine has no reason to prefer the current version — it will retrieve whichever chunk best matches the query, correct or not, and present it with the same confidence either way.

That's the stakes this piece is making: the cost of ungoverned content has risen sharply, because it's no longer confined to a single confused user clicking the wrong link. It's a machine representing your brand's answer, at scale, with no one catching the error in the moment.

4. Pre-empt the obvious objection

A fair question at this point: isn't this just SEO, or its newer cousin GEO (generative engine optimisation), applied to AI? GEO is about making content more likely to be discovered, understood, cited, or incorporated into AI-generated answers, and there's real overlap with SEO — useful original content, crawlability, authority, and clear writing help with both. Google's own guidance on generative AI search makes the overlap explicit: from its perspective, optimizing for AI search is still SEO.¹

But that overlap covers only part of what this article is arguing. SEO and GEO are both, fundamentally, visibility disciplines — they ask can people, or AI systems, find and select my content? That's a necessary condition for a good outcome, but it isn't a sufficient one. Once content is found and cited, a separate question opens up that visibility work doesn't answer: is what gets surfaced actually correct, current, and traceable to someone accountable for it? That's a governance question, not a visibility one.

SEO → "Can people find my page in search?" GEO → "Can an AI find, understand, and cite my information?" IA governance → "When an AI uses that information, is it correct, current, and owned by someone?"

A page can be excellently optimised for GEO — well-structured, clearly written, frequently cited — and still be wrong, if nobody owns keeping that content accurate once it's being pulled out of context and recombined by a model. GEO improves your odds of being retrieved. It says nothing about whether what gets retrieved deserves to represent you. That's the gap the rest of this piece is about closing.

5. A model for AI-ready IA

Content is how organizations communicate with their audience — the websites, applications, and platforms, and the documents, images, video, and code that make them up. It's also a critical asset: it carries an organization's information, services, policies, and brand to employees, customers, and other stakeholders. Well-managed content improves usability, supports business decisions, and strengthens credibility, but only if it's actively governed — kept accurate, accessible, secure, and compliant through a defined lifecycle: research and assessment, creation, approval and publication, then maintenance, archiving, and eventual deletion.

That governance lifecycle is the foundation this section builds on. Preparing it for AI answer engines adds a layer most organizations haven't built yet, organized around three questions: is the content structured so it can be retrieved (Structural); does it make its relationships to other content explicit (Relational); and is someone accountable for its accuracy over time (Governance)? A full framework for each layer is beyond this piece and belongs in a later one, but the shape is worth stating plainly: Structural → Relational → Governance.

Structural is about whether a piece of writing can be understood on its own, out of context — crawlable, clearly organized, and verifiable, while still following SEO fundamentals. It's worth a caveat here: Google has been explicit that chunking content and adding special markup aren't requirements for its generative AI features to work.¹ The structural layer isn't about satisfying a technical mandate from any one engine; it's about writing content that holds up when a reader — human or machine — encounters a fragment of it with nothing around it.

Relational is about whether an AI system can tell that a piece of information belongs with other pieces — that a product page's pricing, features, and limitations are part of one coherent picture, not four unrelated facts. One practical discipline here is maintaining a single canonical record for a given fact, rather than letting the same information drift across several pages.

Governance is the layer that makes the other two trustworthy over time: content should be researched, vetted, and verified before release, and a layered governance structure should keep anything that doesn't meet that bar from reaching an audience, human or machine.

It's useful to see what each layer replaces. Traditional IA was built around pages: a sitemap mapped the hierarchy, and a nav structure encoded priority and relationship through position — what sat under "Billing," what was buried three clicks deep. Success meant a human could find a page and recognise, once there, that they'd landed somewhere coherent; the surrounding context — breadcrumbs, related links, section framing — did real interpretive work. The whole model carried an assumption that never had to be stated: whoever reads your content will encounter it the way you built it, as a full page, in its own sequence.

AI-ready IA doesn't discard that model so much as add structure the page-based model never needed. Three elements do that work:

A content graph sits alongside the sitemap rather than replacing it, because AI retrieval doesn't climb a hierarchy — it pulls whichever chunk best matches a query, from wherever it lives. The design question shifts from where does this sit in the tree to what is this semantically related to, regardless of where it's published. (It's also why an AI-generated answer can sound confident while linked search results underneath it show a contrasting view — the two systems are drawing on the content differently.)

A source-of-truth registry answers a question a sitemap never had to: when two chunks about the same fact disagree, which one is authoritative? A sitemap left that judgement to the user. An AI answer engine makes that judgement itself, often invisibly, which is precisely why an explicit, maintained answer matters.

A chunk ownership map addresses accountability. A page traditionally had an owner — a team, a CMS entry, a named editor. Once content is chunked and recombined by a retrieval system, the page owner and the owner of a specific chunk aren't necessarily the same person, and almost no organization has had to track that distinction before. It's the direct continuation of the accountability gap from the DTH case: ownership has to follow the fact, not just the page it happens to live on.

Three structural shifts follow from that. First, content has to be locally coherent — a chunk needs to make sense on its own, because a headline or intro that used to disambiguate it may not travel with it anymore. Second, relationships have to be explicit rather than positional: if "Refund Policy" sat under "Billing" in the nav, that placement used to communicate the relationship on its own; an AI answer engine never sees that nav structure, so the relationship has to be stated or tagged directly in the content. Third, ownership has to follow the content, not the page, for the reason above.

Three mechanisms do the work of making relationships explicit:

Stated links — writing the relationship into the prose itself, rather than relying on where a page sits in the nav. "Our refund policy (see Returns and Exchanges) applies to enterprise contracts signed after January 2026" names the relationship directly; in practice, this means writing paragraphs that state their own scope rather than assuming a reader or model will infer it.

Schema properties — structured, machine-readable tags embedded in a page that label what something is and how it relates to other things. An Article type can carry an about property linking it to a specific entity, or a mentions property flagging secondary topics. These properties aren't invented per page; schema.org is a shared, versioned vocabulary originally developed by Google, Microsoft, Yahoo, and Yandex, with ongoing community governance.⁴ The vocabulary is standardised, but applying it — deciding which type a page is, which block is the Question, which is the acceptedAnswer — is a judgement call, and keeping it accurate as content changes is a governance problem, not a technical one. A stale or wrong schema tag is arguably worse than none at all, since it states something false with unwarranted confidence. It's also worth noting that Google has said structured data isn't required for its generative AI features specifically, even while continuing to recommend it as part of a broader SEO strategy.¹ The honest case for schema, then, isn't that it guarantees better AI retrieval — it's that it removes ambiguity for any system reading the content, and keeps your own records internally consistent.

Explicit cross-references — descriptive anchor text ("this depends on our data retention policy") instead of "learn more" or "click here," which leaves the relationship entirely to inference. At scale, this is best backed by a maintained source-of-truth record stating that Chunk A (an updated page) supersedes or relates to Chunk B (a legacy page), so the relationship is tracked centrally rather than left to whoever happens to notice.

6. What this means in practice

An organization's information architecture tends to expand quietly, driven by ordinary business demand rather than any plan. What starts as manageable in the early stages of a site's life rarely stays that way. Content added years ago, long before AI answer engines existed, has no natural owner checking whether it's still accurate — and the same underlying fact, repeated across multiple pages over time, becomes harder to track with every year that passes.

If you lead design or content for an organization, the question of whether your information architecture is ready for AI retrieval is worth raising now, not after an AI engine has already misrepresented you. A few questions worth putting to your own organization:

• If an AI answer engine cited your site tomorrow, do you know which page it would pull from — and is that page owned by anyone?
• What should your information architecture let the business do in the next three to five years that it can't do today?
• Is your information structured and governed well enough for both humans and AI systems to find, understand, and reuse it reliably?
• Where are the biggest gaps or inconsistencies in your content that could limit your ability to adopt AI?
• Who is accountable for the meaning, accuracy, relationships, and lifecycle of your information — not merely for publishing the page?

The question underneath all of them is the one worth asking first: are you designing your information architecture for today's interfaces, or for an information system that can support humans, applications, and AI agents alike?

7. Closing

The world that browses less still needs architecture, only a different kind. When a user asks a question and receives an assembled answer, menus, breadcrumbs, and sitemaps no longer do the interpretive work they once did; the Pew data above shows how often a reader now meets a fragment rather than a page. That shift changes the object of information architecture, not its purpose.

Organizations that treated IA as a one-time deliverable failed in the browsing era, as the DTH case showed: no one owned the rules, roles, and processes behind it. The ones that treat it as an owned, living structure — with an accountable person behind every fact, not just every page — will be the ones AI answer engines can actually trust.

Let's talk about a complex problem.

If you're navigating a transformation, enterprise design challenge, or organisational problem where the answer isn't obvious, let's explore it together.

Get in touch →
Connect on LinkedIn →

Footnotes

#ai #information-architecture