Fourteen sources can describe one product, and not one of them carries every field. Enrichment reads all of them, plus the pack and the listing, so the attribute you need is present rather than missing.
Three things separate filling a gap from guessing at one.
A missing attribute is found by reading the places it is actually written down: the pack copy, the ingredient statement, the nutrition panel, the product page. It is not predicted from what similar products tend to be, which is the difference between a filled field and a plausible one.
Every value carries a score, so you set the threshold you are willing to act on rather than accepting the whole file at one quality. A team using the data for a planogram and a team using it for a claim audit do not need the same bar.
Each filled field keeps a line back to what it was read from. When somebody challenges a number in a meeting, the answer is the page it came from rather than an assurance that the model is usually right.
Fourteen sources for one UPC, and the field each one carried. The last row is the record enrichment returns.
A filled attribute is only useful if the product it sits on is the product you think it is. Resolution comes first, then depth.
Twelve sources, eleven names, one row. See how records get resolved before they get filled.
Thirty minutes on your categories. How the record gets built, what it holds, and the questions your team could put to it.