Returning users make up a huge proportion of traffic to sites and products that have customer account functionality.
Are customers of these businesses served with speed and efficiency when they try to log in?
Or is a single word or visual treatment sending them the wrong message?
I wanted to find out. I tested 3 design variations of log in action triggers against 3 tasks that would require logging in, with 60 participants each.
This gives us a total participant number of 540, which is a good range for a web UI usability study like this.
A customer not feeling good about an online interaction is a big risk for anyone focusing on engagement.
Some of the insights in this report are not intuitive at all. And it was a fun report to do, because the data proved me wrong – more than once!
I've broken down the first three tests here, to give you an idea about what the Pro version of this report contains. The Pro version includes all raw data, a 43 minute walkthrough of the insights, and Miro board access for comments and discussion.
Let's get stuck into some data!
For this report, our test site will be the homepage of Bank Australia, which is a socially-aware bank based in Victoria, Australia.
I know plenty of Art Directors and Product Managers who wouldn't cope with the level of UI complexity in this trigger for a simple log in action.
There are "too many words", it "looks clumsy", or "takes focus away from our marketing".
This is understandable feedback. It's not intuitive that the average user would need so much context for a simple log in action.
But let's hold our judgement for a minute, and see what the data says.
To see how well the initial design performs at helping people log in, I set up a task test on UsabilityHub.
I uploaded the 'Control' variation screenshot (which is just a snap of the existing site above), and set the 60 participants a simple task:
Imagine you are a Bank Australia customer. Where would you click to access your customer profile?
I used this wording so that participants wouldn't be able to scan the page for the exact wording in the task.
To get a deeper understanding of how each variation performs, I also tested these two tasks (which you can find in the Pro version of this study):
Let's start with a basic success heatmap for the control test.
I've filtered out any responses from current or past customers of Bank Australia, as they have a head start with this task.
There was only a single BankAust customer in this first cohort of 60 participants, so filtered that gives us n=59.
58 of 59 participants completed the task successfully. This is the ideal result for a test like this. Even for simple tasks, there is room for customer error, or bot noise. So this is a gratifying result for Bank Australia, if nothing else.
I made a histogram of the successful clicks, which reveals more good news.
If you're not familiar with them, histograms show the spread of time-based data in a set. Each column (or "bucket") shows the number of users who completed the task at the time intervals along the x axis.
So in this test, there were:
...which counts as a really quick set of results. In total, 45 participants succeeded in less than 15 seconds.
There are definitely outliers, with one participant taking upwards of 2 minutes to finish the task.
With this kind of test, there's always variation in the speed that participants complete tasks.
At this scale, time-to-click alone is not a great quantitative measure because it's difficult to objectively determine the complexity of the task.
There might also be other factors – distraction, technology issues, an unclear task, and so on.
Despite this, I think it's useful to compare time-to-click data between different test variations in a qualitative fashion, i.e. getting a feel for the shape of the responses, and using that in conjunction with success rate, as well as a 'soft' metric – in this case, a single ease question.
This spread of metrics helps paint a full picture of the performance of each design variation.
Here's the single ease question responses for the Control test:
For the uninitiated, a single ease question (or SEQ) is a measurement of how hard or easy the task was for a participant.
As you can see, it's subjective, as it asks about how the participant perceived the difficulty or ease of the task. This means that the data is interesting from an attitudinal standpoint, as the customer is reporting this data themselves.
Directly comparing these kind of SEQ results in practice is hard because you need big numbers before the results stabilise.
The mean (average) of these results is difficult to directly compare with good statistical accuracy without testing with n=200-300 or more. The maths are just too fuzzy.
So again, similar to the time-to-click histogram, we're looking at these SEQ results to see the shape of the response spread.
Here, with the high number of 6 and 7 responses towards the 'Extremely easy' end of the scale, it looks like a positive result.
The 98% success rate, healthy time-to-click histogram shape, and attractive grouping of SEQ scores show that the control variation works well in this test setting.
Now it's time to throw some design variations in, and see if they work just as well.
The next variant is a much simpler treatment of the log in trigger: a single text element that says simply "Log in", replacing the avatar and longer text.
I ran this text-only variation with exactly the same test setup as the control, changing only the image used.
2 of the 60 participants in this test were BankAust customers, leaving us with n=58.
The successful click rate held fairly steady, with only a small drop to 50/58 participants or an 86% success rate. There's some noise from bots in the results, so this is a good result.
The time-to-click histogram tells a slightly different story:
There's not nearly as many results to the left of the graph compared to the Control test.
The Control had 45 sub-15-second results, whereas this variation with just basic text has only 20.
This is not a reliable number from a statistical standpoint, and I would never make critical decisions based only on this metric.
But it's curious that this treatment would slow people down, even if only slightly.
What if this is statistical anomaly? Given that we're not testing with 600 or 6,000 participants per test, we can't say for sure that this is a worse performance on time-to-click alone.
That's where the SEQ comes in – it's a great sanity check to any more fuzzy metrics you might be collecting.
Here's SEQ for the text-only variant:
This graph isn't as nice as the Control SEQ results, which were very heavy on 6 and 7 results, indicating extreme levels of ease. Here, there's a slight spread of results to the 'neutral' middle of the graph.
It's results like this that make me grateful to have something soft like SEQ in each test, alongside the more cold time-to-click data.
You could forgive the text-only version a slightly worse success rate, and a chunkier histogram in the 15-30 sec results range.
But adding in this customer sentiment result, the results are becoming a bit more clear. This variation seems to perform slightly worse than the Control.
In my opinion, leaving the text off entirely would be too much.
But that's the whole point of this kind of experimentation – to see if your gut intuition is correct, and to hopefully expand your knowledge, regardless of your personal opinion for a particular design variation.
Nobody in this third cohort was a Bank Australia customer. So with low expectations (personally), here's the heatmap of all 60 participant clicks:
BAM. You're looking at a 90% successful click rate – 54/60 participants got it right.
To me, this seems crazy, but let's not get carried away just yet. What about time-to-click for these 54 successful people?
This is a great result, in the same league as the Control test. I'm shocked.
I was certain that taking the text off the log in trigger element would make it harder for participants to find the correct answer. But here we are!
Let's check the SEQ to make sure we're not infuriating people with this new variant:
No! We're NOT infuriating people. It's only marginally worse than the Control test.
There's still a couple of numbers down the 'difficult' end of the scale, so it's clearly not a win for everyone.
But that was also true in the Control test – a similar number of responses were below the 'neutral' middle line.
I'm still shocked by the great performance by the avatar-only version. But also, it makes sense.
It's such a strong pattern to have this kind of trigger at the top right of the screen (for native-English web properties, anyway).
So maybe how we treat this log in action trigger, at least visually, doesn't matter for this user task?
That sounds nuts. Really?!
"But what about something more real?!", I hear you cry. "Nobody is trying to find their customer profile - that's jargon! They're doing real world stuff like changing details, looking at usage, or contacting support!"
Agreed! Expanding our analysis to more specific tasks will add more depth to our testing of these log in trigger variations.
Once I'd seen these first task results, I did just that – testing a 'change password' and an 'update address' task with n=60 in each further test on UsabilityHub.
And the results are very different!
Here's an example of a successful time-to-click histogram from one of the more specific tasks, still with n=60:
Wow. And if you think that's bad, check out this SEQ result – the worst in the dataset – which came from the third set of task tests:
To find out more about which of these design variants performed well (or badly!), purchase the Pro version of the report now!
The Pro report includes:
If you want to know when new reports like this are released, join the free mailing list below! ✨