Wrong color, wrong size, wrong product. We caught it before the buyer did.

The reseller listed tens of thousands of products a month, many on Amazon and brand-owned pages where they had no direct control over the images, sizes, or metadata shown to buyers. When a listing was wrong, they found out the way every seller does: a return, an angry review, a support ticket. Checking that volume by hand was impossible.

The first problem was simply seeing the full catalog. Amazon's interface caps what you can retrieve at around 10,000 results, so a plain export never showed everything. We scraped the complete product set past that ceiling, then scraped each individual listing for the size, color, and images actually shown to buyers.

Then we compared what was listed against what should have been there. AI flagged image and size mismatches at scale, and uncertain cases went to human review. About 80% of the review ran automatically, with people focused only on the judgment calls. It is the same certify-then-verify model we run on every dataset: machines for coverage, humans for the edge cases.

The point was never to hand the reseller raw data and let them sort it out. It was to deliver a finished judgment: this listing is wrong, here is why, here is the correct value. That is the difference between a scrape and a certified dataset.