K-Nearest Neighbours (KNN)

K-nearest neighbours (KNN) is a simple algorithm that makes a decision about a new item by looking at the k most similar items it already knows. For classification, as MLforSEO’s classification lesson puts it, it classifies a point based on the majority of its nearest neighbours, letting the closest points vote. The same idea, finding the nearest items to a query, also powers vector search in modern search and AI retrieval.

How it works

KNN has three ingredients:

• A representation: each item becomes a set of numbers, for text usually an embedding.
• A distance or similarity measure, such as Euclidean distance or cosine similarity.
• A value for k: how many neighbours to consult. A larger k smooths out noise but blurs the boundaries between classes, so the right value depends on your data.

For classification, the new item takes the most common label among its k neighbours, optionally weighting closer neighbours more heavily. KNN builds no general model during training; it simply stores the labelled examples and does its work at prediction time, which makes it the classic example of the “lazy” learners in the eager versus lazy distinction MLforSEO’s classification post flags. It is a supervised method, which makes it a different thing from k-means clustering, an unsupervised method that also uses a k.

In search, nearest neighbour search returns the k closest vectors to a query vector. Checking every vector is exact but slow on large indexes, so vector databases and engines such as Elasticsearch use approximate nearest neighbour (ANN) methods, such as HNSW graphs, that trade a little recall for much lower latency.

How it’s used in SEO and marketing

• Assigning new keywords to existing categories or tagging social posts, the uses MLforSEO’s course highlights.
• Finding the nearest match for every keyword, title or URL in a list, as MLforSEO’s fuzzy matching template does when it returns each item’s nearest neighbour and score.
• Internal linking and redirect mapping: suggesting the most similar page for each source page.
• Retrieval in AI search and RAG: the first stage of the retrieval funnel keeps the content chunks whose vectors sit closest to the query.

Example

You have 2,000 keywords already labelled with a topic and 300 new ones. Embed all of them with Sentence Transformers, then for each new keyword find its five nearest labelled neighbours by cosine similarity. If four of the five are labelled “trail running”, the new keyword gets that label, and a 3-2 split flags it for human review.

Limitations and things to watch

• Prediction gets slower as the dataset grows, because every query is compared with stored examples unless you use an index.
• In very high-dimensional data, distances become less meaningful, so the quality of the embeddings matters more than the algorithm.
• Features on different scales distort distances; normalise numeric data first.
• Imbalanced classes bias the vote towards the most common label.
• Closest is not the same as correct: a semantically near chunk can still be factually wrong.

Learn it on MLforSEO

• What Machine Learning Classification Is, and Where It Fits in SEO
• How modern search actually works: the five-stage retrieval funnel
• What Fuzzy Matching Is, and Where It Fits in SEO
• Introduction to Machine Learning for SEO