Card sorting is a research method where users group a product’s features and settings into categories they name themselves, and for a navigation decision you run it on 30 to 50 cards with at least 15 participants [1].
What a sort gives back is a set of groups and the reasoning behind them. What it doesn't give back is depth, because users can create only one level of categorization, not subcategories [1].
So the result says which features belong together and what people call the pile. Flat or nested stays a judgement call, made with the sort in front of you rather than by it.
Five steps run the method: build the card list, pick open or closed, recruit and run the sessions, count the agreement per card, then read the groups into a structure.
Everything below is written for a product instead of a website. Your cards are features, states and settings, and most of them never get a URL.
What is card sorting, and what does it decide?
Card sorting decides grouping and labels. The method is plain: study participants place individually labeled cards into groups according to criteria that make the most sense to them [1], and each participant gets the unsorted set of cards [1] before the sorting starts. Nothing else is prescribed.
That definition was written with websites in mind, where the set of cards can include pages, links, images, or descriptions of pages that users would navigate to on your website [1].
Inside a product the inventory looks nothing like it. Your cards are features, screens, states, permissions and settings, and most have no URL, no sitemap entry and no place in the primary navigation.
Many exist only while something else is true: a trial is expiring, a seat limit is hit, an integration has gone quiet.
A card sort is the one moment where that whole inventory sits in front of a user at once. That's the value, and it's also the catch, because a pile of cards on a table holds no state, so the sort tests your vocabulary and your groupings, never the product itself.
What is the card sorting technique?
Running a sort is one task, repeated with enough people to count. A participant gets an unsorted set of labeled cards, sorts the items into groups that make sense to them, then names each group in their own words [1].
Two switches change what comes out. A moderated sort has a facilitator present for study sessions and is run for qualitative insights, while an unmoderated sort has no facilitator, runs in a remote tool and is run for quantitative data collection [1].
Cards can be physical, where users organize physical index cards into groups [1], or rows in a remote tool. Material matters less than the question you ask once the sorting stops, which is why these cards went together.
How to do card sorting?
Run it in five steps, in this order. Each one has its own section below.
- Build the card list: one card for each thing a user wants to get done, 30 to 50 cards in total [1].
- Pick the sort type: open while you're deciding a structure, closed while you're checking one [1].
- Recruit and run: at least 15 participants for a qualitative sort, 30 to 50 when the result has to generalize [1].
- Count the agreement: for each card, how many participants put it in the same group.
- Read the groups into a structure, then test that structure before anyone designs a screen [5].
What card sorting will not tell you
One limit shapes everything you do with the result. Users can create only one level of categorization, not subcategories [1], so a sort hands over groups and never a hierarchy. Depth comes later.
Three more limits sit in the same list [1]. Context is missing, because card sorting focuses on isolated content items and doesn't take into account the broader context of the site's images, structure, and links that carry significant information scent for users. Paths are missing too, since a sort doesn't reveal the full paths and hierarchies that users may go through to get the content they need.
The card count has a ceiling as well, because there is a limited number of cards that you can fit in a session before users get tired [1]. Above 50 cards, participants may create a large Miscellaneous category and stuff cards in it [1], and then exhaustion looks exactly like confusion.
Step 1: which features and settings become cards?
A card is one thing a user wants to get done, written in their words, and 30 to 50 of them make a sort [1].
That range exists to prevent user fatigue and encourage thoughtful groupings from participants, and the source hedges it in the same breath: the number of cards depends on the complexity of your content, study goals, and user familiarity with your site [1].
It also answers the objection that 30 cards feels like a lot. This range may seem high, but once users conceptualize a few groupings, the rest become easier to place [1].
For a closed sort, the published instruction is to create a card for each of your navigational items and key pages, then arrange them to mimic your current website's hierarchy [3].
Pages are the wrong unit inside a product, so what follows is our translation rather than the source's: write a card for each job, not for each screen. Export an invoice is a card. The Billing tab is not, because that tab is what the sort decides.
Labels are where a sort quietly fails. A card carrying your internal feature name tests whether users recognise your vocabulary, which is a different study, so borrow the words the support queue already uses. Their words win. Shared naming is design-system work, and our guide to design systems covers where it belongs.
| Where to look | What becomes a card | What to leave out | Why |
|---|---|---|---|
| The primary navigation | Each working area users name as a destination | Containers called More, Other or Advanced | A container is not a job, and it is part of what the sort should decide |
| Every settings tab | Each setting a user changes on purpose | Controls only your own staff touch | Internal switches pull participants into a structure they never see |
| The help centre contents | Each task an article explains how to do | Articles about pricing or policy | Those are company facts, and no navigation has to hold them |
| The support queue, last 90 days | Each where do I find it question, in the asker’s words | One-off bugs and incidents | A repeated question is a card, a single one is noise |
Step 2: open, closed or hybrid card sort?
Open while the structure is still being decided, closed while an existing structure is checked, hybrid when one or two groups are settled and the rest are not [1]. Those are the three variations.
In an open sort, no categories are predefined by the researcher, and the purpose is to understand how users naturally categorize information [1]. In a closed sort, all categories are predefined by the researcher, and the purpose is to validate an existing IA [1].
A hybrid sort has some categories predefined by the researcher, for the case where you are confident about one or two categories in your IA [1], which in a SaaS product usually means Billing and Account.
| Sort type | What you predefine | What it answers | When a SaaS team picks it |
|---|---|---|---|
| Open | Nothing, participants name every group themselves | How users would group the product on their own | A new feature area, or a restructure with the shape still open |
| Closed | Every group name, participants only place cards | Whether existing groups hold the cards users put in them | A proposed navigation whose labels are already agreed |
| Hybrid | One or two group names, participants add the rest | Where the undecided cards land once the settled areas are fixed | Billing and Account are decided, the working areas are not |
The closed sort carries a warning the tool pages tend to drop. Its own purpose line adds that in most cases, tree testing is better for that purpose [1].
So an open sort is the default while a structure is open. If all you want is a verdict on the structure you already have, skip to step 5 and run the test described there instead.
Step 3: how many participants, moderated or remote?
Recruit 15 for a qualitative sort, 30 to 50 when the result has to generalize, and keep any single moderated session between 2 and 8 people [1] [3].
| The question being asked | What the source says | The number | What it means for a SaaS team |
|---|---|---|---|
| A qualitative sort, run to hear the reasoning | At least 15 participants for this approach [1] | 15 | The normal case for a navigation decision |
| A quantitative sort meant to generalize | At least 30 to 50, so results generalize to a broader population [1] | 30 to 50 | Only when the grouping has to hold for a whole user base |
| One moderated session | Keep the number of people small, ideally 2 to 8 [3] | 2 to 8 | A group size per session, so several sessions add up to the 15 |
| How close a small sort gets to the full result | 0.75 correlation at 5 users, 0.90 at 15 [2] | 0.75 and 0.90 | Five is a usability-test number and it is not enough here |
| Where extra participants stop paying | 0.95 at 30 users, 0.98 at 60 [2] | 0.95 and 0.98 | Diminishing returns, and 60 is called wasteful by the same source |
| The published recommendation | 15 users on most projects, 30 on big ones, with Tullis and Wood’s 20 to 30 reported alongside [2] | 15, or 30 | Budget 15 and treat 30 as the funded version |
Fifteen is the working answer, and both ends of the range have evidence under them.
Five is the number most teams carry over from usability testing, and for a card sort there's only a 0.75 correlation between the results from 5 users and the ultimate results, which the source calls not good enough [2]. At the other end, you have to test 60 people to reach 0.98, and doing so is definitely wasteful [2].
The three participant numbers in that table answer three different questions, so merging them would produce a figure nobody published.
Fifteen is a study total. The 2 to 8 is a group size for one session, kept small to maintain focus and enable meaningful discussion [3]. Four sessions of four participants is a 16-person study, which clears the 15 without ever putting more than four people in a room.
Moderated or remote follows from what you need. A facilitator gives you qualitative insights and an unmoderated remote tool gives quantitative data collection [1], and here the reasoning is the asset: much of the value from card sorting comes from listening to the users' comments as they sort the cards [2]. A spreadsheet of counts with no notes beside it is half a study.
One caution about the correlation figures. They come from a single 2004 study at one company, which tested 168 users and then analysed random subsets of that dataset [2].
The article names its own two weaknesses: it's only one study, and the analysis was purely quantitative, focusing on a statistical analysis of similarity scores and ignoring user comments and other qualitative data [2]. Its author adds that he doesn't recommend designing an information architecture based purely on a card sort's numeric similarity scores [2].
Step 4: how do you read the clusters?
Count how many participants put each card in the same group, then read what they said while doing it.
Tools do the counting and draw the views. A similarity matrix visually represents the relationship between two cards, and a card-sorting tool such as OptimalSort also provides a standardization grid and dendrograms [1].
Without a tool you would have to manually identify frequencies of card groupings [1], which is what the example below does by hand.
Counts alone won't carry the decision. Knowing why people place certain cards together gives deeper insight into their mental models than the pure fact that they sorted cards into the same pile [2], so the session notes sit beside the numbers. Bare counts mislead.
A worked example: 12 cards, 15 participants
The numbers in this example are made up to show the arithmetic. Cluma ran no card sort for this guide, and no real product appears in it. This is method, not evidence.
Real sorts use 30 to 50 cards [1]. This one uses 12 so the whole table fits on a page, with 15 imagined participants.
Agreement is the number of participants who put a card in the dominant group, divided by 15 and rounded to whole percentages. Every count is printed below, so you can redo all of it.
| Card | Group most participants chose | Participants who agreed, out of 15 | Reading |
|---|---|---|---|
| Invite a teammate | Members | 14 | Settled, 93 percent |
| Change who can delete a project | Members | 13 | Settled, 87 percent |
| Download an invoice | Billing | 15 | Settled, 100 percent |
| Update the card on file | Billing | 14 | Settled, 93 percent |
| Change my password | My account | 13 | Settled, 87 percent |
| Switch to a dark theme | My account | 12 | Settled, 80 percent |
| Connect Slack | Integrations | 11 | Settled, 73 percent |
| Set a weekly digest | Notifications | 9 | Contested, 60 percent |
| Mute a project’s email alerts | Notifications | 8 | Contested, 53 percent |
| Rotate an API key | Integrations | 6 | Contested, 40 percent |
| Delete the workspace | No dominant group | 5 | Orphan, 33 percent |
| Export all data | No dominant group | 4 | Orphan, 27 percent |
Three readings come out of that table, and the thresholds behind them are our own rule of thumb rather than a published benchmark. No source in the reference list sets an agreement level, so we set one and say so.
Ours is two thirds or more, which is 10 of 15, for a settled card. Six to nine participants in the top group means contested, which says two homes for that card looked plausible to the room. Under six is an orphan.
Settled cards, seven of the twelve here, go where the participants put them. Download an invoice at 15 of 15 and Invite a teammate at 14 of 15 need no further discussion in a design review.
Contested cards split, and the cause is usually the label. Mute a project's email alerts drew 8 of 15 for Notifications, so the other 7 read it as something set on the project itself, and both readings are defensible. One label is doing two jobs, which means either the control appears twice or the name picks a side.
Orphan cards have no dominant group at all. Export all data peaked at 4 of 15, which is 27 percent, and a card that low is usually a product problem instead of a naming one, because nobody's sure what the feature is for.
Read as a priority list, the table is upside down. Those two orphans are the cards to argue about first, and the seven settled ones don't need the meeting.
Step 5: does the result mean flat or nested navigation?
Not on its own, no. The sort gives groups and never depth [1], so flat or nested is a judgement made with the result instead of a number read off it. That call stays a design decision.
Three things come out of a sort: how many groups your users expect, how large each group is, and which cards nobody could place.
Depth is the separate question. Content is more discoverable when it's not buried under multiple intervening layers, and all other things being equal, deep hierarchies are more difficult to use [4].
That tradeoff cuts both ways. In deep hierarchies, when there are only a few categories on each level, they tend to be more generic and, thus, more confusing, while in broad hierarchies with a very large number of items, there is often some conceptual overlap between at least a few of the categories [4].
Like most design questions, there's no single right answer, and going too far to either extreme will backfire [4].
Read your own sort against that sentence. Few groups, each one large and each one named without hesitation, points flat, since flat hierarchies tend to work well if you have distinct, recognizable categories [4].
Many groups, several of them small and several of them overlapping, points nested. Fix the overlap before you nest anything, because contested cards are where overlap shows up first.
If you do go deep, there is a mitigation most teams skip. For deeper hierarchies it's critical to provide alternative navigation options, shortcuts that lead directly to lower levels [4], which covers search, recents, favourites and a command palette.
The three structures a product can pick from, flat, nested and hub and spoke, sit side by side in our guide to information architecture for SaaS products, with what each one costs the person using it.
At one point depth stops being a tradeoff, and that is a setting people want and cannot find. Cancellation and privacy controls are where it goes wrong most visibly, which our piece on dark patterns in SaaS products covers, so an orphan card next to one of those deserves a second look.
Whatever structure comes out, test it before anyone designs a screen. Card sorting often has ambiguous results, meaning it gives you ideas but not a single clear correct IA structure, so you should always test the hierarchy that emerged from a card-sorting study [5].
A tree test does that by having users find the locations in the tree where specific resources or features can be found [5], and it allows you to test multiple IA options without any design, coding, or content development [5].
Where this method breaks
Six limits, and we'd rather hand them over than have you find them later.
The worked example is an illustration. Its counts are invented to show the arithmetic, and the only real numbers in this guide are the ones carrying a reference marker.
Settled, contested and orphan are our own thresholds. No source here publishes an agreement level, so treat the three as a reading habit and not as a finding.
The correlation figures come from one study, at one company, in 2004, which analysed similarity scores and set the comments aside [2]. The article lists both of those weaknesses itself.
Four of the five sources below are Nielsen Norman Group. That's what published research on this method looks like, and it means this guide leans on one organisation's reading of it.
Both method sources were written for websites. The closed-sort instruction is to mimic your current website's hierarchy [3], and the translation to features, states and settings is ours.
A sort can't tell you depth. Step 5 is a judgement this guide supports with evidence, never a result the method returns.
Cluma designs SaaS product interfaces for a living, so we have a stake in you treating structure as work worth paying for. Everything above is checkable against the five sources listed, and the rest of our playbooks sit on the Cluma guides page.
Some teams run the sort themselves and want the structural pass handed over whole. Cluma does that work as a SaaS UI UX design agency, covering UX research, user flows and the information architecture a sort points to, with progress moving forward roughly every 48 hours.
References
- Nielsen Norman Group, Card Sorting: Uncover Users’ Mental Models, by Samhita Tankala and Katie Sherwin, published 02.02.2024, read 03.11.2026. Source of the definition, the open, closed and hybrid sort types, the moderated and unmoderated split, the 30 to 50 cards, the participant minimums, the analysis views and the four limitations. It is written for websites and content inventories rather than for a product’s settings, so the translation to features, states and settings in this guide is ours.
- Nielsen Norman Group, Card Sorting: How Many Users to Test, by Jakob Nielsen, published 18.07.2004, read 03.11.2026. Source of the correlation figures at 5, 15, 30 and 60 users, of the recommendation of 15 users on most projects and 30 on big ones, of Tullis and Wood’s 20 to 30 as reported there, and of the advice to read participants’ comments alongside the counts. The study behind the figures is a single 2004 study at one company and analysed similarity scores only, which the article says itself.
- Yale University, Usability and Digital Accessibility, Card Sorting, read 03.11.2026. Source of the 2 to 8 participants in a single moderated session and of the instruction for building a closed sort. Written for websites, and the page carries no publication date.
- Nielsen Norman Group, Flat vs. Deep Website Hierarchies, by Kathryn Whitenton, published 10.11.2013, read 03.11.2026. Source of the depth tradeoff you weigh after the sort, of the condition on distinct categories and of the shortcut advice for deep structures. Published in 2013 and argued from three hospital websites, so it gives a principle rather than a measured effect for software.
- Nielsen Norman Group, Tree Testing: Evaluate Menu Labels and Categories, by Page Laubheimer, published 06.08.2023, reviewed 19.08.2026, read 03.11.2026. Source of the step that comes after a sort and of the point that a card sort gives ideas rather than one correct structure. It gives no participant minimum and no success benchmark for tree testing. This guide therefore states neither.











