What is the difference between a centroid and a median?

A centroid averages each measure; a median picks the middle value. They part company when outliers pull the mean.

Which distance metric should I use?

Straight-line distance for scaled numbers; cosine similarity for text and embeddings.

Is a centroid only used in clustering?

No. It appears wherever a group needs one representative point, including search and recommendation.

The real data point closest to the group's center, used when an actual example is preferred to a synthetic average.

AI Glossary / AI and machine learning / Intermediate

Centroid

Definition

A centroid is the center point of a group of data, found by averaging the group's values on each measure, and used to represent the whole group with one typical point.

What it means

A group of customers, sensor readings or documents can be a cloud of thousands of points. The centroid is its balancing point: average spend, average visits, average of every measure, combined into one representative location.

Centroids power centroid-based clustering. Methods such as k-means assign each point to its nearest centroid, move the centroid to the center of its members, and repeat until the groups settle.

In text and image systems, the centroid of a group's embeddings, the numerical vectors behind modern AI, acts as an average meaning, handy for summarizing what a cluster is about.

The centroid is an average, so it inherits the weaknesses of averages. A few extreme members drag it, and for a group shaped like a ring or a banana, the center can sit where no real member lives.

For a business, centroids are how a segment gets its pen portrait: the typical mid-value, monthly, urban customer is usually a centroid wearing clothes, and it deserves the same suspicion as any average.

In one dimension the centroid is just the mean; the idea becomes interesting across several measures at once, where the center is a point nobody can picture without the maths.

How it works

Compute the centroid by averaging each measure across the group's members: mean spend, mean visits, mean age, one number per measure.

In k-means, alternate two steps: assign every point to its nearest centroid, then recompute each centroid from its members, until assignments stop changing.

Scale the measures first. If spend runs to thousands while visits run to tens, spend dominates every distance unless each measure is standardized.

Choose a distance that suits the data. Straight-line distance is the default; text and embeddings often use cosine similarity instead.

Check the centroid against typical members. If no real customer looks like it, the group may be the wrong shape for an average to describe.

Some methods use medoids instead: the actual member nearest the center. A medoid is slower to compute but always real, which helps when a synthetic average would mislead.

Three practical examples

Customer segmentation: the average profile of each segment, used for personas and campaign design.
Logistics: the center of a store's delivery orders, a first guess for a depot location.
Text analytics: the average embedding of a cluster of reviews, used to label what the cluster discusses.

Think of it this way

Think of the balancing point of a tray of glasses. It is a real place, but it is not one of the glasses, and one heavy glass pulls it sideways.

The limit: if the glasses stand in a ring, the balancing point floats in the empty middle. An average can describe a place nothing occupies.

Worked example

profiling a customer segment

Suppose a retailer clusters 2,000 loyalty members on four measures: annual spend, visits, average basket and share of online orders.

One segment's centroid reads: spend 340, 22 visits, basket 15, 80 percent online. The team writes the persona from that centroid and designs a campaign around it. Figures are illustrative.

Before launch, an analyst checks the ten members nearest the centroid. They match the persona well, so the average is honest for this group.

A second segment's centroid matches nobody, because the group mixes two kinds of shopper; it is split before use.

Illustrative case study

A fictional gym chain clusters members by check-in times and spend, and plans classes around each cluster's centroid.

One club's centroid is dragged toward midnight by a handful of shift workers, making the typical member look nocturnal. The chain trims outliers and reruns, and the schedule finally matches the real crowd.

The invented lesson: a centroid is only as trustworthy as the group it summarizes. Look at the members before you look at the average.

For board reporting the chain keeps the medoid club alongside each centroid, because a named real example makes the segment easier to discuss.

Common misconceptions and mistakes

The centroid is a real member of the group. It is usually a synthetic point that no actual record matches exactly.
A centroid is always meaningful. For curved or multi-lobed groups the center can describe nobody.
Mean and median are interchangeable here. The median resists outliers; the mean, and so the centroid, does not.

Business use

Use centroids to write segment summaries and personas, but sanity-check them against real members near the center before acting.

Standardize measures before averaging across different units, and investigate outliers before trusting any centroid.

When a report shows an average customer, ask which measures were averaged, whether they were scaled and how outliers were handled; three small questions that catch most centroid mistakes.

Check your understanding

Why must measures be scaled before computing a centroid across them?

Answer: Unscaled, the measure with the biggest units dominates every distance, so the centroid describes that one measure and ignores the rest.

Related terms

Learn next

AI for Business Owners shows how to turn customer data into segments you can act on: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S2] Google, Machine Learning Glossary; [S8] Jurafsky and Martin, Speech and Language Processing (3rd ed. draft). Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Centralized computing Next: Centroid-based clustering