Signal clusters

A cluster is what many signals look like when you stand back from them. One signal is a single customer sentence. A cluster is the shape a hundred of those sentences form: “slow dashboard loads”, “requests for SSO”, “confusion in onboarding”. Clusters are where a complaint stops being anecdote and starts being evidence. For how clusters fit between signals and insights, see The evidence pipeline.

​

Where clusters live

Signal Clusters is the third tab on the Insights page, next to Insights and Data Signals. A cluster gets a CL- reference (CL-12) and a theme label once Squad has named it.

​

How a cluster gets its name

You do not create clusters. Squad forms them as signals arrive, comparing each new signal against the existing clusters of the same type and either adding it to the closest match or starting a new one. Naming waits for evidence. A cluster stays as Pending label until it reaches three members. At that point Squad reads a diverse sample of the members and writes a short theme label plus a confidence score for how well the grouping holds together. Around 0.9 means one clear theme. Around 0.6 means mostly coherent. If the members are too jumbled to name honestly, Squad writes nothing rather than inventing a label. Labels get refreshed as clusters grow, and a nightly pass revisits clusters whose membership has changed. A grouping that has drifted too far apart is dissolved and its signals go back into the pool to be regrouped.

​

The clusters list

One row per cluster:

• Theme. The CL- reference, the label (or Pending label), and a type badge where one applies.
• Data Signals. How many signals belong to the cluster.
• Confidence. A bar showing how sure Squad is about the label.

Filter by Data Signals count (presets 3, 5, 10, 25), by Confidence (0.2, 0.4, 0.6, 0.8), or by Type. Sorting and filters live in the URL, so a filtered view is shareable. Click a row to open the cluster. Types are the same seven a signal can carry: pain point, feature request, bug report, praise, churn risk, competitive intel, and agent insight. A cluster only ever contains one type, so a bug report never lands in the same group as a feature request even when the wording is close.

​

Inside a cluster

Members. The signals that make up the theme. The five most recent show as quote cards so you can read the real feedback without clicking through. Larger clusters get a paginated table below with the rest. Every quote links back to its signal. Info. Type, signal count, how many insights the cluster has produced, label confidence, and when it was last labelled. Insights. The insights Squad synthesised from this cluster. This is what tells you whether the theme has become something you can act on. More than three, and View all opens the full list. A young cluster will show none yet. The page stays live. New members and refreshed labels appear without a reload.

​

How to read a cluster

Use the Members quotes to sanity-check that the grouping holds. If the quotes are talking about different things, the confidence score has usually already said so. Treat the Insights panel as the exit. A large, confident cluster with no linked insight means the evidence is there but the conclusion has not been drawn yet.

​

Cookbook

Find out what is driving a spike in complaints. Open the Signal Clusters tab, filter Type to pain point, and sort by Data Signals descending. Open the largest cluster and skim the member quotes. If there is a linked IN- insight, follow it to see the score and what Squad recommends. Sanity-check a theme before you quote it. Open the cluster, read its confidence in Info, then read the member quotes. If the label says “billing confusion” and half the quotes are about export speed, do not build a case on it.