How to get started with tree testing 🌱

Cards have been created, sorted and sorted again. The participants are all finished and you’re left with a big pile of awesome data that will help you improve the user experience of your information architecture. Now what?Whether you’ve run an open, hybrid or closed card sort online using an information architecture tool or you’ve run an in person (moderated) card sort, it can be a bit daunting trying to figure out where to start the card sort analysis process.

‍

‍

About this guide

‍

This two-part guide will help you on your way! For Part 1, we’re going to look at how to interpret and analyze the results from open and hybrid card sorts.

• In open card sorts, participants sort cards into categories that make sense to them and they give each category a name of their own making.

• In hybrid card sorts, some of the categories have already been defined for participants to sort the cards into but they also have the ability to create their own.

‍

Open and hybrid card sorts are great for generating ideas for category names and labels and understanding not only how your users expect your content to be grouped but also what they expect those groups to be called.In both parts of this series, I’m going to be talking a lot about interpreting your results using Optimal Workshop’s online card sorting tool, OptimalSort, but most of what I’m going to share is also applicable if you’re analyzing your data using a spreadsheet or using another tool.

‍

‍

Understanding the two types of analysis: exploratory and statistical

‍

Similar to qualitative and quantitative methods, exploratory and statistical analysis in card sorting are two complementary approaches that work together to provide a detailed picture of your results.

‍

• Exploratory analysis is intuitive and creative. It’s all about going through the data and shaking it to see what ideas, patterns and insights fall out. This approach works best when you don’t have the numbers (smaller sample sizes) and when you need to dig into the details and understand the ‘why’ behind the statistics.

‍

• Statistical analysis is all about the numbers. Hard data that tells you exactly how many people expected X to be grouped with Y and more and is very useful when you’re dealing with large sample sizes and when identifying similarities and differences across different groups of people.

‍

Depending on your objectives - whether you are starting from scratch or redesigning an existing IA - you’ll generally need to use some combination of both of these approaches when analyzing card sort results. Learn more about exploratory and statistical analysis in Donna Spencer’s book.

‍

‍

Start with the big picture

‍

When analyzing card sort results, start by taking an overall look at the results as a whole. Quickly cast your eye over each individual card sort and just take it all in. Look for common patterns in how the cards have been sorted and the category names given by participants. Does anything jump out as surprising? Are there similarities or differences between participant sorts? If you’re redesigning an existing IA, how do your results compare to the current state?If you ran your card sort using OptimalSort, your first port of call will be the Overview and Participants Table presented in the results section of the tool.If you ran a moderated card sort using OptimalSort’s printed cards, now is a good time to double check you got them all. And if you didn’t know about this handy feature of OptimalSort, it’s something to keep in mind for next time!The Participants Table shows a breakdown of your card sorting data by individual participant. Start by reviewing each individual card sort one by one by clicking on the arrow in the far left column next to the Participants numbers.

‍

‍

From here you can easily flick back and forth between participants without needing to close that modal window. Don’t spend too much time on this — you’re just trying to get a general impression of what happened.Keep an eye out for any card sorts that you might like to exclude from the results. For example participants who have lumped everything into one group and haven’t actually sorted the cards. Don’t worry - excluding or including participants isn’t permanent and can be toggled on or off at anytime.If you have a good number of responses, then the Participant Centric Analysis (PCA) tab (below) can be a good place to head next. It’s great for doing a quick comparison of the different high-level approaches participants took when grouping the cards.The PCA tab provides the most insight when you have lots of results data (30+ completed card sorts) and at least one of the suggested IAs has a high level of agreement among your participants (50% or more agree with at least one IA).

‍

‍

The PCA tab compares data from individual participants and surfaces the top three ways the cards were sorted. It also gives you some suggestions based on participant responses around what these categories could be called but try not to get too bogged down in those - you’re still just trying to gain an overall feel for the results at this stage.Now is also a good time to take a super quick peek at the Categories tab as it will also help you spot patterns and identify data that you’d like to dive deeper into a bit later on!Another really useful visualization tool offered by OptimalSort that will help you build that early, high-level picture of your results is the Similarity Matrix. This diagram helps you spot data clusters, or groups of cards that have been more frequently paired together by your participants, by surfacing them along the edge and shading them in dark blue. It also shows the proportion of times specific card pairings occurred during your study and displays the exact number on hover (below).

‍

‍

In the above screenshot example we can see three very clear clusters along the edge: ‘Ankle Boots’ to ‘Slippers’ is one cluster, ‘Socks’ to ‘Stockings & Hold Ups’ is the next and then we have ‘Scarves’ to ‘Sunglasses’. These clusters make it easy to spot the that cards that participants felt belonged together and also provides hard data around how many times that happened.Next up are the dendrograms. Dendrograms are also great for gaining an overall sense of how similar (or different) your participants’ card sorts were to each other. Found under the Dendrogram tab in the results section of the tool, the two dendrograms are generated by different algorithms and which one you use depends largely on how many participants you have.

‍

If your study resulted in 30 or more completed card sorts, use the Actual Agreement Method (AAM) dendrogram and if your study had fewer than 30 completed card sorts, use the Best Merge Method (BMM) dendrogram.The AAM dendrogram (see below) shows only factual relationships between the cards and displays scores that precisely tell you that ‘X% of participants in this study agree with this exact grouping’.In the below example, the study shown had 34 completed card sorts and the AAM dendrogram shows that 77% of participants agreed that the cards highlighted in green belong together and a suggested name for that group is ‘Bling’. The tooltip surfaces one of the possible category names for this group and as demonstrated here it isn’t always the best or ‘recommended’ one. Take it with a grain of salt and be sure to thoroughly check the rest of your results before committing!

‍

‍

The BMM dendrogram (see below) is different to the AAM because it shows the percentage of participants that agree with parts of the grouping - it squeezes the data from smaller sample sizes and makes assumptions about larger clusters based on patterns in relationships between individual pairs.The AAM works best with larger sample sizes because it has more data to work with and doesn’t make assumptions while the BMM is more forgiving and seeks to fill in the gaps.The below screenshot was taken from an example study that had 7 completed card sorts and its BMM dendrogram shows that 50% of participants agreed that the cards highlighted in green down the left hand side belong to ‘Accessories, Bottoms, Tops’.

‍

‍

Drill down and cross-reference

‍

Once you’ve gained a high level impression of the results, it’s time to dig deeper and unearth some solid insights that you can share with your stakeholders and back up your design decisions.Explore your open and hybrid card sort data in more detail by taking a closer look at the Categories tab. Open up each category and cross-reference to see if people were thinking along the same lines.Multiple participants may have created the same category label, but what lies beneath could be a very different story. It’s important to be thorough here because the next step is to start standardizing or chunking individual participant categories together to help you make sense of your results.In open and hybrid sorts, participants will be able to label their categories themselves. This means that you may identify a few categories with very similar labels or perhaps spelling errors or different formats. You can standardize your categories by merging similar categories together to turn them into one.OptimalSort makes this really easy to do - you pretty much just tick the boxes alongside each category name and then hit the ‘Standardize’ button up the top (see below). Don’t worry if you make a mistake or want to include or exclude groupings; you can unstandardize any of your categories anytime.

‍

‍

Once you’ve standardized a few categories, you’ll notice that the Agreement number may change. It tells you how many participants agreed with that grouping. An agreement number of 1.0 is equal to 100% meaning everyone agrees with everything in your newly standardized category while 0.6 means that 60% of your participants agree.Another number to watch for here is the number of participants who sorted a particular card into a category which will appear in the frequency column in dark blue in the right-hand column of the middle section of the below image.

‍

‍

‍

From the above screenshot we can see that in this study, 18 of the 26 participant categories selected agree that ‘Cat Eye Sunglasses’ belongs under ‘Accessories’.Once you’ve standardized a few more categories you can head over to the Standardization Grid tab to review your data in more detail. In the below image we can see that 18 participants in this study felt that ‘Backpacks’ belong in a category named ‘Bags’ while 5 grouped them under ‘Accessories’. Probably safe to say the backpacks should join the other bags in this case.

‍

‍

So that’s a quick overview of how to interpret the results from your open or hybrid card sorts.Here's a link to Part 2 of this series where we talk about interpreting results from closed card sorts as well as next steps for applying these juicy insights to your IA design process.

‍

‍

Further reading

• Card Sorting 101 – Learn about the differences between open, closed and hybrid card sorts, and how to run your own using OptimalSort.
• Which comes first: card sorting or tree testing? – Tree testing and card sorting can give you insight into the way your users interact with your site, but which comes first?
• How to pick cards for card sorting – What should you put on your cards to get the best results from a card sort? Here are some guidelines.

We love getting stuck into scary, hairy problems to make things better here at Trade Me. One challenge for us in particular is how best to navigate customer reaction to any change we make to the site, the app, the terms and conditions, and so on. Our customers are passionate both about the service we provide — an online auction and marketplace — and its place in their lives, and are rightly forthcoming when they're displeased or frustrated. We therefore rely on our Customer Service (CS) team to give customers a voice, and to respond with patience and skill to customer problems ranging from incorrectly listed items to reports of abusive behavior.

‍

The CS team uses a Customer Relationship Management (CRM) system, Trade Me Admin, to monitor support requests and manage customer accounts. As the spectrum of Trade Me's services and the complexity of the public website have grown rapidly, the CRM system has, to be blunt, been updated in ways which have not always been the prettiest. Links for new tools and reports have simply been added to existing pages, and old tools for services we no longer operate have not always been removed. Thus, our latest focus has been to improve the user experience of the CRM system for our CS team.

‍

And though on the surface it looks like we're working on a product with only 90 internal users, our changes will have flow on effects to tens of thousands of our members at any given time (from a total number of around 3.6 million members).

‍

The challenges of designing customer service systems

‍

We face unique challenges designing customer service systems. Robert Schumacher from GfK summarizes these problems well. I’ve paraphrased him here and added an issue of my own:

‍

1. Customer service centres are high volume environments — Our CS team has thousands of customer interactions every day, and and each team member travels similar paths in the CRM system.

‍

2. Wrong turns are amplified — With so many similar interactions, a system change that adds a minute more to processing customer queries could slow down the whole team and result in delays for customers.

‍

3. Two people relying on the same system — When the CS team takes a phone call from a customer, the CRM system is serving both people: the CS person who is interacting with it, and the caller who directs the interaction. Trouble is, the caller can't see the paths the system is forcing the CS person to take. For example, in a previous job a client’s CS team would always ask callers two or three extra security questions — not to confirm identites, but to cover up the delay between answering the call and the right page loading in the system.

‍

4. Desktop clutter — As a result of the plethora of tools and reports and systems, the desktop of the average CS team member is crowded with open windows and tabs. They have to remember where things are and also how to interact with the different tools and reports, all of which may have been created independently (ie. work differently). This presents quite the cognitive load.

‍

5. CS team members are expert users — They use the system every day, and will all have their own techniques for interacting with it quickly and accurately. They've also probably come up with their own solutions to system problems, which they might be very comfortable with. As Schumacher says, 'A critical mistake is to discount the expert and design for the novice. In contact centers, novices become experts very quickly.'

‍

6. Co-design is risky — Co-design workshops, where the users become the designers,  are all the rage, and are usually pretty effective at getting great ideas quickly into systems. But expert users almost always end up regurgitating the system they're familiar with, as they've been trained by repeated use of systems to think in fixed ways.

‍

7. Training is expensive — Complex systems require more training so if your call centre has high churn (ours doesn’t – most staff stick around for years) then you’ll be spending a lot of money. …and the one I’ve added:

‍

8. Powerful does not mean easy to learn — The ‘it must be easy to use and intuitive’ design rationale is often the cause of badly designed CRM systems. Designers mistakenly design something simple when they should be designing something powerful. Powerful is complicated, dense, and often less easy to learn, but once mastered lets staff really motor.

‍

Our project focus

‍

Our improvement of Trade Me Admin is focused on fixing the shattered IA and restructuring the key pages to make them perform even better, bringing them into a new code framework. We're not redesigning the reports, tools, code or even the interaction for most of the reports, as this will be many years of effort. Watching our own staff use Trade Me Admin is like watching someone juggling six or seven things.

‍

The system requires them to visit multiple pages, hold multiple facts in their head, pattern and problem-match across those pages, and follow their professional intuition to get to the heart of a problem. Where the system works well is on some key, densely detailed hub pages. Where it works badly, staff have to navigate click farms with arbitrary link names, have to type across the URL to get to hidden reports, and generally expend more effort on finding the answer than on comprehending the answer.

‍

Groundwork

‍

The first thing that we did was to sit with CS and watch them work and get to know the common actions they perform. The random nature of the IA and the plethora of dead links and superseded reports became apparent. We surveyed teams, providing them with screen printouts and three highlighter pens to colour things as green (use heaps), orange (use sometimes) and red (never use). From this, we were able to immediately remove a lot of noise from the new IA. We also saw that specific teams used certain links but that everyone used a core set. Initially focussing on the core set, we set about understanding the tasks under those links.

‍

The complexity of the job soon became apparent – with a complex system like Trade Me Admin, it is possible to do the same thing in many different ways. Most CRM systems are complex and detailed enough for there to be more than one way to achieve the same end and often, it’s not possible to get a definitive answer, only possible to ‘build a picture’. There’s no one-to-one mapping of task to link. Links were also often arbitrarily named: ‘SQL Lookup’ being an example. The highly-trained user base are dependent on muscle memory in finding these links. This meant that when asked something like: “What and where is the policing enquiry function?”, many couldn’t tell us what or where it was, but when they needed the report it contained they found it straight away.

‍

Sort of difficult

‍

Therefore, it came as little surprise that staff found the subsequent card sort task quite hard. We renamed the links to better describe their associated actions, and of course, they weren't in the same location as in Trade Me Admin. So instead of taking the predicted 20 minutes, the sort was taking upwards of 40 minutes. Not great when staff are supposed to be answering customer enquiries!

‍

We noticed some strong trends in the results, with links clustering around some of the key pages and tasks (like 'member', 'listing', 'review member financials', and so on). The results also confirmed something that we had observed — that there is a strong split between two types of information: emails/tickets/notes and member info/listing info/reports.

‍

We built and tested two IAs

‍

‍

‍

After card sorting, we created two new IAs, and then customized one of the IAs for each of the three CS teams, giving us IAs to test. Each team was then asked to complete two tree tests, with 50% doing one first and 50% doing the other first. At first glance, the results of the tree test were okay — around 61% — but 'Could try harder'. We saw very little overall difference between the success of the two structures, but definitely some differences in task success. And we also came across an interesting quirk in the results.

‍

Closer analysis of the pie charts with an expert in Trade Me Admin showed that some ‘wrong’ answers would give part of the picture required. In some cases so much so that I reclassified answers as ‘correct’ as they were more right than wrong. Typically, in a real world situation, staff might check several reports in order to build a picture. This ambiguous nature is hard to replicate in a tree test which wants definitive yes or no answers. Keeping the tasks both simple to follow and comprehensive proved harder than we expected.

‍

For example, we set a task that asked participants to investigate whether two customers had been bidding on each other's auctions. When we looked at the pietree (see screenshot below), we noticed some participants had clicked on 'Search Members', thinking they needed to locate the customer accounts, when the task had presumed that the customers had already been found. This is a useful insight into writing more comprehensive tasks that we can take with us into our next tests.  

‍

What’s clear from analysis is that although it’s possible to provide definitive answers for a typical site’s IAs, for a CRM like Trade Me Admin this is a lot harder. Devising and testing the structure of a CRM has proved a challenge for our highly trained audience, who are used to the current system and naturally find it difficult to see and do things differently. Once we had reclassified some of the answers as ‘correct’ one of the two trees was a clear winner — it had gone from 61% to 69%. The other tree had only improved slightly, from 61% to 63%.

‍

There were still elements with it that were performing sub-optimally in our winning structure, though. Generally, the problems were to do with labelling, where, in some cases, we had attempted to disambiguate those ‘SQL lookup’-type labels but in the process, confused the team. We were left with the dilemma of whether to go with the new labels and make the system initially harder to use for staff but easier to learn for new staff, or stick with the old labels, which are harder to learn. My view is that any new system is going to see an initial performance dip, so we might as well change the labels now and make it better.

‍

The importance of carefully structuring questions in a tree test has been highlighted, particularly in light of the ‘start anywhere/go anywhere’ nature of a CRM. The diffuse but powerful nature of a CRM means that careful consideration of tree test answer options needs to be made, in order to decide ‘how close to 100% correct answer’ you want to get.

‍

Development work has begun so watch this space

‍

It's great to see that our research is influencing the next stage of the CRM system, and we're looking forward to seeing it go live. Of course, our work isn't over— and nor would we want it to be! Alongside the redevelopment of the IA, I've been redesigning the key pages from Trade Me Admin, and continuing to conduct user research, including first click testing using Chalkmark.

‍

This project has been governed by a steadily developing set of design principles, focused on complex CRM systems and the specific needs of their audience. Two of these principles are to reduce navigation and to design for experts, not novices, which means creating dense, detailed pages. It's intense, complex, and rewarding design work, and we'll be exploring this exciting space in more depth in upcoming posts.

Usability guru Jared Spool has written extensively about the 'scent of information'. This term describes how users are always 'on the hunt' through a site, click by click, to find the content they’re looking for. Tree testing helps you deliver a strong scent by improving organisation (how you group your headings and subheadings) and labelling (what you call each of them).

‍

Anyone who’s seen a spy film knows there are always false scents and red herrings to lead the hero astray. And anyone who’s run a few tree tests has probably seen the same thing — headings and labels that lure participants to the wrong answer. We call these 'evil attractors'.In Part 1 of this article, we’ll look at what evil attractors are, how to spot them at the answer end of your tree, and how to fix them. In Part 2, we’ll look at how to spot them in the higher levels of your tree.

‍

The false scent — what it looks like in practice

‍

One of my favourite examples of an evil attractor comes from a tree test we ran for consumer.org.nz, a New Zealand consumer-review website (similar to Consumer Reports in the USA). Their site listed a wide range of consumer products in a tree several levels deep, and they wanted to try out a few ideas to make things easier to find as the site grew bigger.We ran the tests and got some useful answers, but we also noticed there was one particular subheading (Home > Appliances > Personal) that got clicks from participants looking for very different things — mobile phones, vacuum cleaners, home-theatre systems, and so on:

‍

‍

‍

The website intended the Personal appliance category to be for products like electric shavers and curling irons. But apparently, Personal meant many things to our participants: they also went there for 'personal' items like mobile phones and cordless drills that actually lived somewhere else.This is the false scent — the heading that attracts clicks when it shouldn’t, leading participants astray. Hence this definition: an evil attractor is a heading that draws unwanted traffic across several unrelated tasks.

‍

Evil attractors lead your users astray

‍

Attracting clicks isn’t a bad thing in itself. After all, that’s what a good heading does — it attracts clicks for the content it contains (and discourages clicks for everything else). Evil attractors, on the other hand, attract clicks for things they shouldn’t. These attractors lure users down the wrong path, and when users find themselves in the wrong place they'll either back up and try elsewhere (if they’re patient) or give up (if they’re not). Because these attractor topics are magnets for the user’s attention, they make it less likely that your user will get to the place you intended. The other evil part of these attractors is the way they hide in the shadows. Most of the time, they don’t get the lion’s share of traffic for a given task. Instead, they’ll poach 5–10% of the responses, luring away a fraction of users who might otherwise have found the right answer.

‍

Find evil attractors easily in your data

‍

The easiest attractors to spot are those at the answer end of your tree (where participants ended up for each task). If we can look across tasks for similar wrong answers, then we can see which of these might be evil attractors.In your Treejack results, the Destinations tab lets you do just that. Here’s more of the consumer.org.nz example:

‍

‍

Normally, when you look at this view, you’re looking down a column for big hits and misses for a specific task. To look for evil attractors, however, you’re looking for patterns across rows. In other words, you’re looking horizontally, not vertically. If we do that here, we immediately notice the row for Personal (highlighted yellow). See all those hits along the row? Those hits indicate an attractor — steady traffic across many tasks that seem to have little in common. But remember, traffic alone is not enough. We’re looking for unwanted traffic across unrelated tasks. Do we see that here? Well, it looks like the tasks (about cameras, drills, laptops, vacuums, and so on) are not that closely related. We wouldn’t expect users to go to the same topic for each of these. And the answer they chose, Personal, certainly doesn’t seem to be the destination we intended. While we could rationalise why they chose this answer, it is definitely unwanted from an IA perspective. So yes, in this case, we seem to have caught an evil attractor red-handed. Here’s a heading that’s getting steady traffic where it shouldn’t.

‍

Evil attractors are usually the result of ambiguity

‍

It’s usually quite simple to figure out why an item in your tree is an evil attractor. In almost all cases, it’s because the item is vague or ambiguous — a word or phrase that could mean different things to different people. Look at our example above. In the context of a consumer-review site, Personal is too general to be a good heading. It could mean products you wear, or carry, or use in the bathroom, or a number of things. So, when those participants come along clutching a task, and they see Personal, a few of them think 'That looks like it might be what I’m looking for', and they go that way.Individually, those choices may be defensible, but as an information architect, are you really going to group mobile phones with vacuum cleaners? The 'personal' link between them is tenuous at best.

‍

Destroy evil attractors by being specific

‍

Just as it’s easy to see why most attractors attract, it’s usually easy to fix them. Evil attractors trade in vagueness and ambiguity, so the obvious remedy is to make those headings more concrete and specific. In the consumer-site example, we looked at the actual content under the Personal heading. It turned out to be items like shavers, curling irons, and hair dryers. A quick discussion yielded Personal care as a promising replacement — one that should deter people looking for mobile phones and jewellery and the like.In the second round of tree testing, among the other changes we made to the tree, we replaced Personal with Personal Care. A few days later, the results confirmed our thinking. Our former evil attractor was no longer luring participants away from the correct answers:

‍

Testing once is good, testing twice is magic

‍

This brings up a final point about tree testing (and about any kind of user testing, really): you need to iterate your testing —  once is not enough.The first round of testing shows you where your tree is doing well (yay!) and where it needs more work so you can make some thoughtful revisions. Be careful though. Even if the problems you found seem to have obvious solutions, you still need to make sure your revisions actually work for users, and don’t cause further problems. The good news is, it’s dead easy to run a second test, because it’s just a small revision of the first. You already have the tasks and all the other bits worked out, so it’s just a matter of making a copy in Treejack, pasting in your revised tree, and hooking up the correct answers. In an hour or two, you’re ready to pilot it again (to err is human, remember) and send it off to a fresh batch of participants.

‍

Two possible outcomes await.

• Your fixes are spot-on, the participants find the correct answers more frequently and easily, and your overall score climbs. You could have skipped this second test, but confirming that your changes worked is both good practice and a good feeling. It’s also something concrete to show your boss.
• Some of your fixes didn’t work, or (given the tangled nature of IA work) they worked for the problems you saw in Round 1, but now they’ve caused more problems of their own. Bad news, for sure. But better that you uncover them now in the design phase (when it takes a few days to revise and re-test) instead of further down the track when the IA has been signed off and changes become painful.

‍

Stay tuned for more on evil attractors

‍

In Part 1, we’ve covered what evil attractors are and how to spot them at the answer end of your tree: that is, evil attractors that participants chose as their destination when performing tasks. Hopefully, a future version of Treejack will be able to highlight these attractors to make your analysis that much easier.

‍

In Part 2, we’ll look at how to spot evil attractors in the intermediate levels of your tree, where they lure participants into a section of the site that you didn’t intend. These are harder to spot, but we’ll see if we can ferret them out.Let us know if you've caught any evil attractors red-handed in your projects.