AI Product MatchingAI Product Matching › Algorithms
How do product matching algorithms work?
A product matching algorithm first narrows each product to a short list of candidates (blocking), then scores each pair with several signals: string similarity of titles, embeddings that capture meaning, image similarity and attribute checks. A classifier or weighted rule turns those signals into one confidence score, and a threshold decides what counts as a match.
Updated 10 November 2026 by the ScrapeWise team.
Quick picks
For a quick start
Blocking plus string similarity
Simple and explainable.
For mixed wording and languages
Embeddings plus attribute rules
The modern default.
To skip building it
ScrapeWise (our product)
EAN, text and image matching with a score per match.
Compared
| Technique | What it does | Strength | Weakness |
|---|---|---|---|
| Blocking | Groups candidates by brand, category or keyword | Makes matching fast at scale | Misses matches in the wrong block |
| String similarity | Compares characters or words (e.g. Jaccard, edit distance) | Fast, explainable | Fails on synonyms and reordering |
| Text embeddings | Turns titles into vectors that capture meaning | Handles different wording and languages | Can call similar products identical |
| Image similarity | Compares product photos | Works when titles are poor | Can't see size or model year |
| Attribute rules | Checks size, colour, pack, model | Stops costly false matches | Needs clean attributes |
| Classifier | Learns how to weigh the signals from labelled pairs | Best overall accuracy | Needs labelled training data |
Putting it together
- Identifiers first
- If EAN or MPN match, accept the pair without the AI steps.
- Signals second
- Compute text, image and attribute scores for the rest.
- Decide
- Combine into one score; auto-accept above the high threshold, reject below the low one, review in between.
- Learn
- Feed reviewed pairs back as labelled data to improve the classifier.
Frequently asked questions
What is a product matching algorithm?
The set of steps that decides whether two listings are the same product: choosing candidates, scoring them on text, image and attributes, and comparing the combined score with a threshold.
Which algorithm is best for product matching?
No single one. The strongest setups combine identifiers, embeddings, image similarity and attribute rules, weighted by a classifier trained on labelled pairs.
What is blocking in product matching?
Grouping products into small candidate sets, for example by brand, so the algorithm compares each product only with likely matches.
Do I need machine learning for product matching?
Not always. Barcodes and simple rules handle many products. Machine learning helps when titles differ and barcodes are missing.