Similarity as the only guide
Clustering is the core technique of unsupervised learning: taking a set of data points and grouping the similar ones together, without ever being told in advance what the groups should be or how many there should be. Feed a clustering algorithm a dataset of customer purchase histories and it might surface a group of frequent, high-spend buyers, a group of occasional bargain shoppers, and a group that buys once and never returns, all without anyone labeling a single customer beforehand.
The algorithm doesn't know what a 'segment' means in a business sense. All it can measure is similarity, typically distance between data points in feature space, where features like purchase frequency, average order size, and recency get turned into numbers. Points that sit close together in that space get grouped together. The groupings that come out are only as meaningful as the features you fed in, which is why choosing good features matters just as much here as it does in supervised learning.