The process of UI design has always blown my mind.
Experiments that have generated tens of millions in uplift have often surfaced from designs by one person, based on brand guidelines, experience and vibes, and were the first shot at what this could look like.
No data feeds into the process, we just go from design to live to hoping for an uplift.
That's in stark contrast to what comes before it - filling up a roadmap off UXR sprints, deep analytics research, interviews, lab sessions. Really well defined problems, sent to one person and success or failure ultimately hinges on how good a job that one person does. With no support.
This is where I feel we at Webtrends Optimize can help.
There are several layers of prediction that you can run on top of any given design asset. Visual Attention, Fixation, Scanpaths, Mouse movement and Clickthroughs. These are all things we can predict, and turn into heatmaps or more generally data overlaid on top of your design work.
Predictive Analytics for Design
Here, we're not talking about heatmaps from Hotjar or Clarity, which is the product of what users are actually doing on the website. And that's not so useful or feasible, when we're trying to design something new.
Instead, we have developed a predictive heatmap tool that allows for some rapid iteration.
The way this works is a designer can pull together something that feels reasonable as a first draft, and pull a predictive heatmap from Webtrends Optimize. The heatmap might show that attention is being distracted on something else - not because you mean to, but maybe it's the same colour as your CTA, or it's sitting on a part of the page where it feels like other elements are pointing at it.
So they go and make a few tweaks. Little nudges, softening a few colours or moving things around a little. And generate another heatmap.
It's not enough, and so they go a bit further, and generate one final heatmap.
Better now - people notice the few bits you really want them to, and clickthrough prediction includes your main CTA - great.
So that becomes the version they submit to be built as part of an experiment.
Webtrends Optimize Sovereign AI also ensures your data is always secure by using different models for different purposes.
Where have experiments failed?
AB Tests are pretty limited in their ability to tell you why they failed. And so many do - by most estimates, 33-85% of experiments aren't successful.
We can assume some of these are due to bad ideation, some due to bad execution by dev or your testing platform, and some are the result of bad design.
So even marginal increases in success rate in this area boost your overall experiment success rates, which is good for you, and good for us as a platform.
A reason to run AB/Ns
Velocity is an important metric for testing programmes, and this feature also unlocks a stream in this direction too.
Control vs. The designers pure view vs. a Data bolstered view - test all three.
There are probably times where each of those could win, sometimes for reasons that again won't be obvious. But your testing becomes a wider pipe than usual.
You test more variants, and maybe stumble onto a winner where you wouldn't have before, or a stronger winner which is still a win too.
Back to the Problem
The problem people face is that we're researching with data, we're coding with support, we're analysing with data, but we're not designing with data. And so where so many experiments can fail, that happening even 5% less is a great win all round. Plus having tools to actually support your designers, and so the weight of designing things that could have massive flaws they didn't see reduce.
All feeding into a better testing velocity and better win-rate.
To us, that's AI with purpose.
