Building Systems That Match Things to Other Things

Where matching beats search

Keyword search misses the candidate whose CV says “client account management” when the brief says “customer success”, and the supplier whose datasheet describes the same specification in different words.

Semantic matching compares meaning rather than strings, which is exactly the gap that costs businesses opportunities.

Rank, do not decide

Present a ranked shortlist with reasons, never a decision. The people using it know things the data does not: who is genuinely available, who fell out with that manager, which supplier was late last time.

Systems that filter automatically reliably discard good options that looked unpromising on paper, and nobody ever sees what was lost.

The matching stack

1. Exact identifiers first where they exist — product codes, registration numbers. Reliable and often missing.
2. Normalised text matching on the structured attributes you do have
3. Semantic similarity for descriptions that differ in wording but not meaning
4. A human review band in the middle, with the decision remembered so it is never asked twice

Realistic expectations

On messy real-world catalogues or CV data, expect 60–80% matching automatically with high confidence, and the remainder needing review. Anyone promising fully automatic matching across untidy data has not tried it on yours.

The remembered decisions are what shrink the review band over the first few months.

What it costs

A matching application over one data set with a review interface typically £20,000–£50,000. Catalogue size and data quality drive the number far more than the algorithm does.

Frequently asked questions