The same generator, eleven shops, and a reversal file for every change

Every figure above was measured on the date in the last column, from our internal log — the checklist written for that shop the day its import was prepared.

Our own run-order document gives two of these totals differently — 13,463 against 13,301 for one shop, 12,287 against 12,286 for another. The difference is exactly the number of products with no title at all, which are handled in a separate file and a separate step. The two documents count different things, and we would rather explain that than quietly pick one.

1. None of these eleven shops was suspended. This case is about what a catalogue looks like before Google notices, not about lifting a suspension.
2. The figures are what the import files change. They are not a claim about what the shops looked like a week later — a catalogue is a moving object, and a measurement is a day.
3. Two of the eleven are missing from the tables. They were run the same way on the same dates, but their figures were never written into a checklist, and a number we cannot point at does not go on a page.

1. In six of the shops, 172 rows would have been left with an empty description once the leaked text was struck. An empty field does not mean 'unchanged' in a WooCommerce import — it deletes the text in the shop. Those rows kept their original text instead, and they are listed separately because they still need reading by a human.
2. In one shop the finished replacement text did not exist at all, so 12,501 rows went through with the code fences removed and nothing swapped in. That is a smaller job than the other ten got, and the checklist says so rather than reporting it as a completed pass.
3. In one shop 1,515 product titles still had words from another language in them after the run. Machine translation is not what that needs, so they were left for a human and counted.
4. Eighteen rows in one shop had a strike reversed, or a claim still standing after the pass. They are in a file called FOR-HAND — by hand — because a rule that gets a sentence wrong eighteen times out of ten thousand is still a rule that gets it wrong.

That first row is also the last column of the table above. A fault is recoverable; a deletion across five thousand products is not, so the rule is that the original text wins whenever the correction would leave the field empty.