How we test the converter
A product import with a wrong price is worse than no import: it goes live and nobody notices until a customer pays it. So the rule is simple. A missed row is acceptable; a wrong price is not. Anything the converter isn't sure of goes into needs_review.csv with the page it came from, and the price is left blank instead of guessed.
Here is how that rule is checked, and what the numbers are, including the parts that still go wrong.
1. Real supplier price lists, checked by hand
11 real, public supplier price lists (grills, patio heaters, lab equipment, fabrics, spirits) in PDF and Excel. Every row's SKU, name and price was written down from the source file independently of the converter, then compared field by field.
| Rows checked | Rows found | SKU correct | Price correct |
|---|---|---|---|
| 8,426 | 99.7% | 8,398 of 8,398 | 8,220 of 8,220 |
"Correct" counts every value the converter wrote out. The 0.3% of rows not found were reported as missing, not filled in.
2. 1,000,000+ generated rows across 9 layouts
Real catalogs are the best test but there aren't a million of them. So we also generate supplier-style price lists with known answers, in the layouts suppliers actually use: plain tables, ruled grids, section headings, side-by-side lists, product cards, boxed part lists with no header row, multi-page tables, wrapped descriptions, and separate dollar-sign columns. Then every one is converted and scored.
| Rows | Files | Rows found | Wrong SKUs | Wrong prices |
|---|---|---|---|---|
| 1,004,403 | 4,445 | 99.83% | 0 | 0 |
These rows are generated test catalogs, not real supplier files; only section 1 is real. Product names, codes and prices are sampled from real catalogs so they look like the real thing (codes with spaces, prices like $13,666.77, names with inch marks).
3. Scanned PDFs
The same real price lists, printed to images (slight tilt, noise, blur, JPEG) and read back.
| Rows checked | Rows found | SKU correct | Price correct |
|---|---|---|---|
| 6,768 | 94.0% | 6,363 of 6,365 | 6,198 of 6,198 |
Scans are harder: more rows go to the review list, and 2 product codes were read wrong (look-alike characters). Prices stayed correct. If your file is a scan, expect a longer review list.
What still goes wrong
- Missed rows. Usually rows whose text is printed over other text. They're reported, never guessed.
- Product names on card layouts that read exactly like a spec line ("Vent Kit: Includes…"). The product gets its SKU as the title, and you'll see it in the review list.
- Scanned pages with photos and spec grids find fewer rows than clean PDFs.
Every conversion still comes with a 5-product test file to import first, and everything imports as Draft.