Separate the sure from the unsure
Where a barcode or part number appears in both files, match on it and count those first. Only the rest needs attribute matching. A model is for the pairs software rules cannot settle.
Pack sizes are the trap
A 12-pack and a 6-pack of the same drink look nearly identical in text. Test for it with planted look-alike pairs, and send every doubtful pair to an unsure list. A wrong link puts price or stock on the wrong product.
- Brand, size, pack and colour should all agree.
- No many-to-one link unless it is listed.
- Every identifier in the output must be in your files.
Which job fits
The old-to-new matching job builds the mapping and tests it on 30 pairs you label, half true pairs and half look-alikes. If a feed also lacks size, colour or material, the attribute job fills what descriptions state and leaves the rest 'unknown'. For a Shopify store there is a separate version with a metafield-ready import file. Neither applies anything to your store.
Sources and limits
- Google Merchant Center: GTIN Checked 2026-10-12.
- Says to submit only valid GTINs, not to guess or make up a value, and that a product with an incorrect GTIN is disapproved; it marks google_product_category as optional.
- Zenodo: Cost-Aware Evaluation of LLMs for Ambiguous Product Entity Matching Checked 2026-10-12.
- A single-author preprint (8 May 2026) on 2,518 ambiguous product pairs reports a fuzzy rule baseline at F1 0.445 and a small language model at F1 0.958 for USD 1.30 per 1,000 pairs.