Information is structured by grouping, ordering, and relating pieces of content so that people (or machines) can find, understand, and use it. Whether you’re organizing a website, a database, a research paper, or a simple filing system, the same core principles apply: you choose a scheme that matches how your audience thinks, then arrange content within that scheme to minimize confusion. The specific methods range from simple alphabetical lists to complex networks of meaning, but they all serve the same goal of reducing the mental effort required to process what’s in front of you.
Five Universal Ways to Organize Anything
In his 1989 book Information Anxiety, information architect Richard Saul Wurman proposed that there are exactly five ways to organize any body of information. He called them LATCH: Location, Alphabet, Time, Category, and Hierarchy. A museum arranges exhibits by location (room or wing). A dictionary uses alphabetical order. A news feed sorts by time. A grocery store groups items by category (dairy, produce, bakery). An org chart uses hierarchy, showing who reports to whom.
These five schemes aren’t just theoretical. Every time you face a pile of unsorted information, you’re choosing one of them, even if unconsciously. A photo library might be organized by date (time), by event type (category), or by where the photos were taken (location). Picking the wrong scheme for your audience creates friction. Organizing a recipe collection alphabetically by dish name, for instance, is less useful than grouping by meal type or cuisine, because that’s how most people think when deciding what to cook.
How Your Brain Handles Structure
The reason structure matters so much comes down to how human memory works. In 1956, psychologist George Miller published a landmark paper arguing that short-term memory holds roughly seven items, plus or minus two. That number became famous, but more recent research has revised it downward. When you need to hold several pieces of information in mind at once, the practical limit is closer to three or four items for most adults.
Miller’s most lasting insight wasn’t the number itself but the concept of “chunking.” Your brain can treat a group of related items as a single unit. The letter string FBICIAUSA is nine characters, which would overwhelm short-term memory, but if you recognize FBI, CIA, and USA as familiar acronyms, it collapses into just three chunks. Good information structure works the same way. It groups related content into recognizable clusters so your brain can process more with less effort.
This connects directly to a concept called cognitive load. Your working memory has a fixed capacity, and everything competing for that capacity falls into one of three buckets. Intrinsic load is the unavoidable complexity of the material itself. Extraneous load is unnecessary mental effort caused by poor design, like confusing navigation or cluttered layouts. Germane load is the productive effort of actually learning and forming mental models. Well-structured information reduces extraneous load, freeing up mental resources for the thinking that actually matters. Research in medical education has shown that restructuring lecture content along these principles measurably improved students’ comprehension and motivation while reducing their perceived mental burden.
Hierarchies, Facets, and Relational Models
When people talk about information structure in a technical sense, three patterns come up repeatedly.
A hierarchy (or taxonomy) is the most intuitive. It looks like a tree: broad categories at the top, narrower subcategories branching below. Think of how a library organizes books, from broad subjects like “Science” down through “Biology” and then “Marine Biology.” In a strict hierarchy, every item has exactly one parent. A polyhierarchy relaxes that rule, allowing a single item to belong to multiple branches. An online retailer might place the same wireless speaker under both “Electronics” and “Home Office,” which makes browsing easier but adds complexity behind the scenes.
A faceted structure takes a different approach. Instead of placing items in a single tree, it tags each item with multiple descriptive attributes. A clothing site might let you filter by size, color, material, and price simultaneously. This is close to how people naturally think about things, because real objects have many characteristics, not just one classification. Facets give users the power to slice through information from whatever angle matters most to them at the moment.
A relational (database) model goes further still. It stores information in tightly defined tables with rows and columns, where relationships between tables are explicitly mapped. This is the backbone of most business software, banking systems, and anything that needs precise, searchable records. The trade-off is rigidity: you must define your structure before you start filling it with data, and changing that structure later can be difficult.
Structured vs. Flexible Data Storage
The tension between rigid and flexible structure shows up clearly in how databases are designed. Traditional relational databases (often called SQL databases) organize data into predefined tables. Every record follows the same format. This makes them excellent for structured data where relationships between entities are well understood, like customer orders linked to product inventories linked to shipping addresses. Scaling these systems typically means upgrading to more powerful hardware.
Non-relational databases (often called NoSQL) allow different types of data to coexist without a fixed schema. You can store a text document alongside a graph of social connections alongside a simple list of values. There’s less upfront planning required, and it’s easier to add new types of information later. These systems scale by adding more servers rather than upgrading existing ones, which makes them popular for applications that handle massive, varied datasets like social media platforms or content recommendation engines.
Neither approach is inherently better. The choice depends on what kind of information you’re working with and how predictable its shape is. Highly structured, relationship-heavy data fits relational models. Rapidly evolving, varied data fits non-relational ones.
Taxonomies, Ontologies, and Folksonomies
Classification systems exist on a spectrum of formality. At the simple end, a taxonomy is a hierarchical list of categories, like the subject headings a library uses to catalog books. It captures “this thing is a type of that thing” but not much else.
An ontology goes deeper. It maps not just categories but the relationships between them: functional connections, physical properties, cause-and-effect links. Medical ontologies, for example, can represent that a certain virus causes a specific disease, which affects particular organs, which are treated by certain procedures. Ontologies power decision-support systems, data integration across organizations, and knowledge-enhanced search tools.
A folksonomy sits at the opposite end of the formality scale. It emerges when thousands of individual users tag content with their own keywords, with no central authority dictating the vocabulary. Social media hashtags are a folksonomy. The advantage is that the structure reflects how real people actually describe things. The disadvantage is inconsistency: one person tags a photo “NYC,” another tags it “New York,” and a third uses “Manhattan.” There’s no built-in mechanism to recognize these as related.
How Machines Structure Meaning
Modern AI systems have introduced a fundamentally different way of structuring information. Instead of sorting content into categories or tables, they convert text into numerical representations called embeddings, essentially long lists of numbers that capture the meaning of words, sentences, or entire documents.
These numerical representations place similar concepts near each other in a mathematical space. A search for “canine companions” can return results about “dogs” or “pets” even if those exact words never appear, because their numerical representations point in similar directions. Context matters too: the word “bank” gets a different representation depending on whether the surrounding text is about rivers or finance. This approach moves beyond rigid keyword matching and accounts for synonyms, related concepts, and shifts in meaning that traditional structures can’t handle.
Structuring Information for Screens
On the web, information structure directly shapes how easy a site is to use. A common guideline suggests users should reach any piece of content within three clicks, but usability research has found no correlation between the number of clicks and how easy navigation actually feels. What matters more is whether each step feels clear and predictable. A five-click path where every choice is obvious beats a two-click path where the user has to guess.
Mobile screens add extra constraints. With less visible space, deep hierarchies that require many taps to reach important content become frustrating quickly. Current best practices favor progressive disclosure: showing only the most relevant options first and revealing deeper content based on what the user does. This means collapsible menus, contextual sub-navigation that appears when needed, and priority-based layouts that surface frequently used items at the top. The goal is to flatten the structure wherever possible so that the most important content isn’t buried.
Testing Whether Your Structure Works
If you’re designing an information structure for other people, the only reliable way to know it works is to test it with real users. Card sorting is the standard method. You write each piece of content on a card (physical or digital), hand the deck to participants, and ask them to group the cards in whatever way makes sense to them. Listening to why they place certain cards together reveals how they mentally model the information, which is often different from how the designer imagined it.
Nielsen Norman Group, a leading usability research firm, recommends testing with 15 participants for most projects. That number produces correlation scores around 0.90, meaning the patterns in the data are stable enough to act on. For large, high-stakes projects like an intranet serving 100,000 employees, testing 30 users pushes the reliability higher. Useful results have been achieved with as few as 12 participants when budgets are tight, though the patterns become less certain.
Card sorting is a generative method, meaning it’s designed to create a new structure from scratch rather than validate an existing one. Because the goal is to discover how people think rather than confirm assumptions, sampling more participants gives richer, more representative results.