Understanding the problem (2-3 weeks)

I started by interviewing 12 researchers and reviewing existing session recordings in Pendo. The most important thing I learned was that researchers weren't struggling with the visualisation because it was ugly they were struggling because they couldn't tell which relationship paths were worth investigating and which were noise. When a query returned dozens of connections, everything looked equally important, so users defaulted to exporting data to spreadsheets and doing their own manual filtering. This told me the redesign wasn't just about making the visualisation look better it was about changing the underlying information model so that evidence strength and confidence were visually encoded into the diagram itself, letting researchers prioritise paths at a glance instead of reading every one.

Exploring approaches (2-3 weeks)

I explored 4 different approaches to the visualisation. Our first approach would have introduced a completely new chart type that gave us more control over layout and filtering, but it meant abandoning the Sankey format that researchers were already familiar with and would have required significantly more engineering effort. The second approach kept the Sankey but added a separate filtering panel alongside it. The third added colour-coding and confidence labels directly onto the existing paths. The fourth was a hybrid design. It was a simplified Sankey for overview with a linked detail table below. We went with colour-coding the paths by confidence level combined with interactive filtering, because it improved readability without forcing users to learn a new visualisation type and was feasible within our engineering capacity. The key constraint was that the Sankey library had limited support for custom styling per path, so engineering needed to extend it which meant we had to keep the visual changes focused rather than attempting a full rebuild.

Testing and iterating (2-3 weeks)

We tested with 8 users. The biggest surprise was that the numerical confidence scores we'd initially kept as tooltips were still confusing even as a secondary detail, users kept asking what a score of 0.73 actually meant in practical terms. For example, one researcher said they'd ignore any path unless someone told them the threshold for "reliable," which defeated the purpose of showing confidence at all. Based on this, we replaced numerical scores entirely with High, Medium, and Low labels tied to the colour-coding, and added a brief explanation of what each level meant when users hovered over the label.

Shipping and measuring

We launched to all users in November 2023. We tracked engagement with Cause and Effect Paths, session depth, and return usage using Pendo. The results are above.