Product Matching Across Marketplaces: Techniques and Challenges

Cross-marketplace product matching is one of the hardest problems in e-commerce data engineering. Here's how the best teams approach it.

The Challenge

• Different titles ("Apple AirPods Pro" vs "AirPods Pro 2nd Generation - White")
• Different images (studio shots vs lifestyle photos)
• Different identifiers (ASIN on Amazon, Item ID on Walmart)
• Different categorization hierarchies

Matching Strategies

1. Identifier-Based Matching (Most Reliable) The gold standard. If two listings share the same UPC or EAN, they're the same product. Period.

Implementation: Use EcomSource to resolve ASINs to UPCs, then search other marketplaces by UPC.

Accuracy: 99%+ for exact matches Limitation: Not all listings have valid identifiers

2. Title Similarity Matching When identifiers aren't available, compare product titles using NLP techniques.

• TF-IDF vectorization + cosine similarity
• Jaccard similarity on tokenized titles
• Sentence embeddings (BERT, all-MiniLM) for semantic matching

Accuracy: 70-85% depending on category Limitation: False positives from similar but different products

3. Image Matching Compare product images using computer vision.

• Perceptual hashing (pHash)
• CNN feature extraction + similarity
• CLIP embeddings for cross-modal matching

Accuracy: 75-90% Limitation: Different photos of same product may not match; stock photos shared across products

4. Hybrid Approach (Recommended) Combine all three strategies: 1. First, try identifier matching (highest confidence) 2. If no identifier match, use title + image similarity 3. Set confidence thresholds and flag low-confidence matches for human review

EcomSource's Role

EcomSource provides the identifier layer that makes strategy #1 possible at scale. With 1.6B+ verified ASIN-to-UPC mappings, you can automatically match the majority of products without resorting to fuzzy matching techniques.