Auto-Segmentation
Discover natural customer groups using K-Means clustering and convert them into audiences.
Last reviewed July 13, 2026
Auto-Segmentation uses BQML K-Means clustering to discover natural customer groups in your data. Instead of defining segment rules by hand, you let the algorithm find patterns, review the results, and convert any interesting cluster into an audience you can activate.
You can find Auto-Segmentation on the Customers page, under the Discover tab, via the Discover Segments button.
Prerequisites
Auto-segmentation runs against your Feature Store. The Feature Store must be enabled and have active feature definitions before you can discover segments — if it is not set up yet, the Discover tab prompts you to enable it first.
The features available for clustering are your Feature Store feature definitions, grouped by category: commerce, engagement, and lifecycle.
Overview
The workflow has four steps:
- Choose a goal or edit the feature selection manually.
- Configure the number of clusters.
- Review the discovered clusters.
- Convert any cluster into an audience for activation.
Behind the scenes, Vendo creates a BQML K-Means model, trains it in your BigQuery warehouse, and returns the cluster results. No data leaves your project.
Getting Started
- Navigate to the Customers page and open the Discover tab.
- Click Discover Segments.
- Select a goal preset, or expand the feature selector to pick features yourself.
Goal Presets
Presets weight the clustering toward a feature category:
| Preset | Emphasis | Best For |
|---|---|---|
| Find High-Value Customers | Commerce features (purchase behavior, revenue, order frequency) | Identifying top spenders and loyal buyers |
| Identify Engagement Tiers | Engagement features (session activity, page views, interaction) | Separating highly engaged users from passive ones |
| Lifecycle Stages | Lifecycle and commerce features (tenure, recency, purchase history) | Mapping customers to lifecycle phases |
| Use All Features | No weighting — all available features equally | Letting the algorithm find patterns across all your data |
There is no separate “custom” preset — to customize, open the feature selector and edit the selected features directly. You can search and toggle any active feature definition.
Choosing Features
Tips for feature selection:
- More features is not always better. Start with 4—8 features that are relevant to the question you are trying to answer.
- Avoid redundant features. If two features measure the same thing (e.g.,
total_revenueandtotal_spent), pick one. - Mix categories. Combining commerce features with engagement features often produces more useful clusters than using one category alone.
Number of Clusters
Choose how many groups the algorithm should create, from 2 to 15 (the default is 5). If you are unsure, start with 3—5. You can always re-run with a different number.
Understanding Results
After training completes, the results view shows a card for each discovered cluster with:
| Detail | Description |
|---|---|
| Cluster name | Auto-generated label — use Name Clusters with AI to generate descriptive names |
| Cluster size | Number of customers in the cluster and share of the total |
| Top defining features | The features where this cluster differs most from the overall population, with ↑ (higher than average) or ↓ (lower than average) badges |
Silhouette Score
The results also report a silhouette score, a measure of how well-separated the clusters are:
| Score | Interpretation |
|---|---|
| ≥ 0.7 | Strong cluster separation — segments are well-defined |
| 0.5 — 0.7 | Moderate separation — segments are reasonable |
| 0.25 — 0.5 | Weak separation — consider fewer clusters |
| < 0.25 | Very weak — try different features or fewer clusters |
Interpreting Clusters
Look for clusters with clear marketing characteristics:
- A cluster with ↑ order count and ↑ recency is your loyal, active segment.
- A cluster with ↑ historical revenue but ↓ recency may be at-risk churners.
- A cluster with ↓ revenue across the board may be window shoppers or one-time buyers.
Converting to Audiences
Once you identify a cluster worth targeting:
- Click Create Segment from Cluster on the cluster card. A dialog lets you name the audience before clicking Create Segment.
- Vendo creates an audience whose rules approximate the cluster boundaries.
- Click View Segment to open the new audience’s detail page, where you can edit it like any other audience.
To convert everything at once, use Create All N Segments to create an audience from every discovered cluster in one click.
Once created, audiences can be synced to your connected destinations — see Audiences for activation options.
This conversion bridges the gap between unsupervised discovery and marketing activation: the algorithm finds the groups, and you decide which ones to activate.
Related
- Audiences — The audiences that auto-segmentation creates, and how to sync them
- BQML Models — The underlying K-Means model powering auto-segmentation