I make and sell resin miniatures under Artifex Fabricatum — 746 products on Shopify, built up over several years the way most catalogues are: one product at a time, in between everything else a small business has to do. Nobody had gone back through it looking for what had quietly gone wrong.
| Finding | Count | Share |
|---|---|---|
| No product type set | 400 | 54% |
| Only one photo | 131 | 18% |
| SKU shared with another listing | 31 | 4% |
| Duplicate description | 16 | 2% |
| Title under 20 characters | 8 | 1% |
Duplicate descriptions — the finding this whole tool started with — turned out to already be nearly solved here: the largest cluster left was two products sharing one paragraph, not hundreds. The real story on this shop was somewhere else entirely.
400 of 746 products, 54% of the catalogue, had no product type set at all. A shop’s product type is what a shopper filters by — the category links, the collection pages, the "narrow by" sidebar. A product with none of those does not rank lower in that search. It is not in it. Filter by anything and it is as if the product is not for sale.
That is a direct mechanism for losing a sale, not a tidiness complaint, and it had been sitting there the whole time.
Assay only assigns a type the shop already uses — never a new category invented for the occasion. Artifex Fabricatum has eight real ones in active use: Squad, HQ, Miniature Kit, Pin Up, Vehicle, Battle Pack, Terrain, Army Bundle. Every one of the 400 was checked against that list and nothing else.
389 of 400 were classified confidently. 11 were not, and were left alone. A handful of real examples, exactly as returned:
The 11 declines matter more than the 389 that worked. A wrong category is worse than no category — a shopper who filters by "Vehicle" and finds a bag of loose bits stops trusting the filter, not just that one product. Every one of these was a genuine edge case: an accessory or add-on that does not cleanly belong to any of the shop’s own existing types, so nothing was guessed.
The token counts above are from this exact run. The cost estimate is not a fresh guess at a price per token — it carries over a separately measured run of about the same size on this same shop and model, on the basis that re-measuring the same thing twice is not worth the extra call. See the pricing page for what a catalogue this size actually costs to run start to finish.