Conversation Clusters
Data Visualisation / Product Design
·
2024-2025
ROLE
Sole Product Designer
TEAM
1 PM · 3 data science · 2 backend · 2 frontend · 1 search
TIMELINE
Late 2024 – early 2025
OUTCOME
Removed a one-minute wait on entry, with strong feedback from Customer Success Managers
Network view: clusters ranked by share of conversation, with trend lines and layout options
Conversation Clusters, launched by Talkwalker in 2020, uses machine learning to group social and online conversation into semantic network graphs, helping brand, agency and insights teams see the themes behind a topic. After years of added features, the experience had become slow and hard to read. As the sole designer, I redesigned it.
A workspace table replaced the one-minute auto-load, and setup handles several languages in one flow
AI summaries can be shown or hidden in the list, and editing moved to a dedicated modal
the challenge
Support tickets, usage data and interviews with Customer Success Managers pointed to the same friction:
Slow entry: the product auto-loaded the last session, adding a wait of more than a minute every time users opened it.
Rigid setup: clusters could only be created in one language, timeframes were manual only and queries were capped at 20 clusters.
Hard to read: dense 20-item charts with long labels hid patterns, the side panel was cramped and "Top Keywords" were shown without explaining where they came from.
Design goal: make Clusters fast to start, flexible to set up and easy to read, without removing the depth power users relied on.
Split preview shows both new clusters before the change is applied
Before: 20 clusters with full-length names. After: top 10 views with truncated names and hover
Research. I worked from support tickets and usage data, and ran interviews and feedback sessions with CSMs, who use Clusters with clients every day and could show where analysis broke down.
“Would it be possible to highlight the topic keywords in Conversation Cluster ‘main stories’ snippets? I’m trying to figure out how certain clusters have derived from the topic, and this would speed the process up.” (Customer Success, Professional Services)
What I changed:
Workspace landing table instead of auto-loading the last session, so users can open, duplicate or create a cluster straight away.
Multi-language setup in a single flow, removing the need to build the same query several times.
Readable charts: top 5, 10 and 15 filters, predefined timeframes and a doubled cluster limit.
AI summaries on demand: a show/hide control in the side-panel header.
Keyword traceability: "Top Keywords" highlighted in the network chart and in the results widget, so users can see each keyword in its original post.
Room to work: high-friction editing moved from the cramped side panel into dedicated modals.
A trade-off. I proposed auto-translation in the widget, so teams could build reports in several languages without creating a separate cluster for each. It was deprioritised because of its cost and timeline.
Clicking a top keyword highlights where it appears in the network
conclusion
The redesign changed the experience in two ways:
Faster start: the landing table removed the one-minute wait on entry.
Positive feedback from power users: the CSMs interviewed during the redesign highlighted how much faster clusters were to create, how much easier the visuals were to analyse and how useful split and merge and the keyword highlights in the results widget had become.
Reflection. I'd build an interactive prototype in Cursor early on, so CSMs could test the new flows with real interactions before development, not just Figma prototypes.