Information architecture: labelling, navigation and the methods that test them

An information architecture is a claim about how the people who use a site divide their world, and the claim can be tested. This lecture argues that an architecture is a hypothesis about how users categorise the world. It argues that the organisation's own structure is a rival hypothesis which wins by default because it is already built. It also argues that labels and structure must be settled with methods such as card sorting and tree testing rather than defended by intuition. The test itself, however, is an instrument with measurable properties, and recent work shows those properties depend on how the instrument is built. This tenth unit asks how a person finds them.

The organisation already knows its own departments, and the people who write the navigation are usually paid by them. A label that is never tested still appears to work, because a user who fails either abandons the task or leaves without a record.

A hypothesis with a well-funded rival

Rosenfeld, Morville and Arango (2015) treat information architecture as the structural design of shared information environments. The architecture is not the code that renders a menu. It is the set of decisions about what things are called, how they are grouped, and how a person moves between groups. Those decisions can be right or wrong about the people who use them, which is what makes the architecture a hypothesis rather than an aesthetic.

Morville (2005) gives the goal a name, arguing that findability, the quality of being locatable or navigable, is a design aim in its own right. Jul and Furnas (1997) reported a CHI workshop on navigation in electronic worlds that drew together the problems of disorientation and of knowing where one is. Navigation fails in ways the artefact conceals, because a lost user and a confident user can submit the same query.

The rival hypothesis is the organisation chart. Departments, budgets and content ownership already impose a grouping. That grouping is cheaper to ship than a tested one, because nobody has to run a study to defend it.

A university site organised by faculty and department puts its admissions pages under whichever faculty owns the programme. A prospective student looking for how to apply must first guess which part of the institution runs the course. The grouping is faithful to the organisation and wrong about the user, who does not know the faculties, may not know the difference between undergraduate and postgraduate, and may be a mature applicant in later life. A student who cannot find the pages does not file a report.

The label is the mechanism

Pirolli and Card (1999) proposed information foraging theory. It treats a person seeking information as a forager who follows information scent, adapting the optimal-foraging framework from biology. The unit of decision is the link, judged by how much its label appears to promise the thing being sought.

Furnas (1997) supplied the structural counterpart. On his account, effective view navigation requires three conditions to hold at once:

• the views must be small;
• movement between them must not take too many steps;
• the route to any target must be discoverable.

A hierarchy can satisfy the first two and still fail the third, which is the failure a label causes: the route exists but its name does not promise the target. If scent decides the path, a label that carries none for the target is a defect that can be demonstrated rather than a matter of taste.

Card sorting recovers the grouping, and tree testing tests the names

Card sorting surfaces the grouping a set of people already hold. Rugg and McGeorge (2005) give a tutorial on the sorting techniques used in knowledge and requirements acquisition. They cover card sorts, picture sorts and item sorts, and work through one variety, the repeated single-criterion sort. The variants are not interchangeable: a repeated single-criterion sort yields different evidence from an open sort in which a participant builds the groups. Spencer (2009) turns the method into a procedure a team can run. It covers how to plan, run and analyse a sort and how to turn the result into usable categories.

The method has a measurable cost. Dougalis and Katsanos (2025) ran a closed card sort in which 191 participants sorted 45 cards into 7 predefined categories for an e-commerce site. They then used resampling across sample sizes from 1 to 191 to correlate each sample with the whole. They found that 8 to 12 participants were sufficient for reliable closed card sort data, and that further increases were more costly than insightful. This converts a folklore number into a stated range for one site.

Tree testing tests the labels rather than the grouping, by asking people to find items in a text-only hierarchy. Kuric, Demcak and Krajcovic (2025) compared three commonly used tree testing variants against three high-fidelity prototypes across 180 participants and 1800 task completions. They found significant differences between the variants and in how closely each predicted navigation in a real interface. The variant that displayed a user's earlier navigation choices and gave direct controls to reverse them was the most accurate. A tree test is therefore not one instrument but several.

Facets turn a designed vocabulary into navigation

A facet is a dimension along which a collection can be narrowed, and faceted navigation lets a person filter without typing a query. Tunkelang (2009) gives the definitional account of a facet and of how faceted navigation is designed and implemented. This connects the controlled vocabularies of the previous unit to the navigation of this one: each facet is a small vocabulary, and each item is assigned terms from each. Hearst (2009) provides the interface-side reference on what works in search interfaces and what does not.

Yee and colleagues (2003) tested the approach on a collection where keyword search is weak. In a usability study, 32 art history students explored a collection of 35,000 fine arts images through hierarchical faceted metadata with dynamically generated query previews. Ninety per cent preferred the faceted interface overall, 97 per cent said it helped them learn more about the collection, and 72 per cent found it easier than the baseline, despite initial unfamiliarity. The case matters because the user could not name the thing sought, so a query could not have worked.

Contested question: whose structure, and how far can the test be trusted?

The literature disagrees on two connected questions: whether a designed structure beats a computed one, and whether the methods that test a structure are settled instruments.

On the first, Hearst (2006) holds the position that designed categories serve exploration better. She compared automatic clustering with designed faceted categories, and the usability studies she reviewed found participants preferred categories, especially for browsing. Clustering is automatic and works on any collection, but its labels are unpredictable and its groups inconsistent. Yee and colleagues (2003) supply the worked case, and Rosenfeld, Morville and Arango (2015) generalise it: an organisation and labelling system is a design output, not a by-product of the content.

On the second, Kuric, Demcak and Krajcovic (2025) hold the position that the instruments are still being validated. Their three tree testing variants predicted real navigation differently, and Dougalis and Katsanos (2025) ground their range in a single e-commerce case. The scent evidence that underwrites the testing programme is thinner than its influence suggests, because Pirolli and Card (1999) state the theory the methods assume rather than measure it. Furnas (1997) sharpens the point. The requirements for effective view navigation are analytical. A structure can be shown navigable or not without asking anyone, which competes with the empirical methods rather than supporting them.

The position this lecture takes is that the architecture is a hypothesis and the test is a measurement, and a measurement is only as good as the instrument behind it. Designed categories can win on average, and the size of that win can still be unknown because the measuring method varies. What would settle it is an experiment in which the same items are organised two ways, the user model and the organisation model. Findability would be measured on the same tasks and participants. It would also need a replication of the card-sort threshold across sites, with the sort variant held constant. It would need a tree test whose variant is reported alongside its result.

Rate each side by its evidence. Hearst (2006) reports the usability studies she reviewed, and Yee and colleagues (2003) report one study of 32 students, so the designed-categories side rests on measurement inside a single instrument. Kuric, Demcak and Krajcovic (2025) measured three tree testing variants on the same tasks, and Dougalis and Katsanos (2025) resampled a closed card sort, so the instrument side rests on measurement too. The strongest objection is that the difference between the methods can be larger than the effect of the structure they measure, so a study can find a real preference and still not know its size in a live interface. While the dispute is open, a designer names the sort variant and the test variant beside every number.

Where the argument is heading

The direction of the recent work is towards validating the instruments rather than asserting the model. Kuric, Demcak and Krajcovic (2025) propose their cross-methodological comparison as a basis for more accurate early validation. Dougalis and Katsanos (2025) apply resampling to bring a sample-size claim under a stated method. Hearst's (2006) comparison of designed categories with computed clusters is the same move applied to the structure itself. The open question is the one Rosenfeld, Morville and Arango (2015) leave when they treat the architecture as a designed set of systems. Some parts can be settled analytically, as Furnas (1997) argues a navigable view can be. Others can be settled only by asking the people who will use it.

The sources this unit rests on

• Dougalis, T. and Katsanos, C. (2025). How many participants do you need for closed card sorting? A case study of an e-commerce website. Proceedings of the 3rd International Conference of the ACM Greek SIGCHI Chapter. doi:10.1145/3749012.3749050 Ran a closed card sort in which 191 participants sorted 45 cards into 7 categories, then resampled every sample size from 1 to 191. They report that 8 to 12 participants were sufficient for reliable results, so the test has a measurable cost.
• Furnas, G. W. (1997). Effective view navigation. Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/258549.258800 Argues that effective view navigation requires small views, few steps, and discoverable routes to any target, and derives strategies from those requirements. A label can destroy a discoverable route.
• Hearst, M. A. (2006). Clustering versus faceted categories for information exploration. Communications of the ACM. doi:10.1145/1121949.1121983 Compares automatic clustering with designed faceted categories for exploring a result set. The usability studies she reviews found participants preferred categories, especially for browsing.
• Hearst, M. A. (2009). Search User Interfaces. Cambridge University Press. doi:10.1017/CBO9781139644082 A book-length survey of search interface design covering what works and what does not across research and commercial deployment. It connects the retrieval model to the interface where faceted navigation lives.
• Jul, S. and Furnas, G. W. (1997). Navigation in electronic worlds: a CHI 97 workshop. ACM SIGCHI Bulletin. doi:10.1145/270950.270979 Reports a CHI workshop on navigation in electronic worlds that drew together the problems of disorientation and of knowing where one is. It names the failure modes the unit's methods address.
• Kuric, E., Demcak, P. and Krajcovic, M. (2025). Validation of information architecture: cross-methodological comparison of tree testing variants and prototype user testing. Information and Software Technology. doi:10.1016/j.infsof.2025.107740 Compared three tree testing variants against three high-fidelity prototypes across 180 participants and 1800 task completions. The variants predicted real navigation differently, with earlier navigation choices and reversal controls the most accurate.
• Morville, P. (2005). Ambient Findability. O'Reilly Media. https://isbnsearch.org/isbn/9780596007652 Only the ISBN record was verified, so it is cited for the position rather than the text. Argues that findability, the quality of being locatable or navigable, is a design goal in its own right. It supplies the vocabulary that makes labelling a design problem rather than decoration.
• Pirolli, P. and Card, S. (1999). Information foraging. Psychological Review. doi:10.1037/0033-295X.106.4.643 Presents information foraging theory, which treats people seeking information as foragers who follow information scent. It is the basis for the unit's claim that users follow the labels that carry scent.
• Rosenfeld, L., Morville, P. and Arango, J. (2015). Information Architecture: For the Web and Beyond. O'Reilly Media. https://isbnsearch.org/isbn/9781491913529 Only the ISBN record was verified, so it is cited for the position rather than the text. Treats information architecture as the structural design of shared information environments, giving most attention to organisation, labelling and navigation. It carries the unit's framing: an architecture is a set of systems a team designs and defends.
• Rugg, G. and McGeorge, P. (2005). The sorting techniques: a tutorial paper on card sorts, picture sorts and item sorts. Expert Systems. doi:10.1111/j.1468-0394.2005.00300.x Gives a tutorial on the main sorting techniques used in knowledge and requirements acquisition, working through the repeated single-criterion sort. A card sort is one technique among several, and the variety changes what it measures.
• Spencer, D. (2009). Card Sorting: Designing Usable Categories. Rosenfeld Media. https://rosenfeldmedia.com/books/card-sorting/ Only the publisher's description was verified, so it is cited for the method's shape rather than for a trial. A practitioner reference on card sorting, covering how to plan, run and analyse a sort and how to turn the result into usable categories. It shows the method as a procedure a team can run.
• Tunkelang, D. (2009). Faceted Search. Synthesis Lectures on Information Concepts, Retrieval, and Services. doi:10.2200/S00190ED1V01Y200904ICR005 Only the publisher's description was verified, so it is cited for the definition. A short book that explains what a facet is and how faceted navigation is designed and implemented. It links the vocabularies of the previous unit to the navigation of this one.
• Yee, K. P., Swearingen, K., Li, K. et al. (2003). Faceted metadata for image search and browsing. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/642611.642681 In a usability study, 32 art history students explored 35,000 fine arts images through hierarchical faceted metadata. Ninety per cent preferred it overall, 97 per cent said it aided learning, and 72 per cent found it easier than the baseline.