Ask a manufacturer what they make and you get an answer in their own words: high-precision ball screws for machine tools, release film for optical applications, brushless motor controllers for industrial fans. The words are exact, and they are the words a buyer would recognise in conversation. They are almost never the words under which that product crosses a border.
A customs declaration does not say "ball screw for machine tools." It carries a classification code, a description written to satisfy the code, a weight, a value and a consignee. The description is often terse, sometimes in a language other than the manufacturer's, and shaped by what the filing agent thought would clear fastest rather than by what the product is for. The same product can appear under several codes depending on how it was declared, and the same code can hold products that have nothing in common beyond a shared material.
Buyer databases have their own vocabulary again. A company is described by its industry, its size, its products in the broadest terms. "Manufacturer of industrial fans" tells you a company might buy motor controllers. It does not tell you the company imports them, in what quantity, or under which description. Each of these three sources, the catalogue, the customs record and the company profile, is accurate on its own terms. None of them speaks the others' language.
Why keyword search fails
The obvious answer is keyword search, and it fails in an instructive way. Search import records for "ball screw" and you find the shipments that happened to be described that way, and miss the ones filed as "linear motion components," "machine tool parts," or a bare code. Search too broadly and you drown in unrelated goods that share a code. The precision that makes a manufacturer's description useful is exactly what makes it a poor query, because the world did not file its imports using the manufacturer's vocabulary.
What is needed is a translation, and translation is harder than it looks because it is not one-to-one. A single catalogue product maps to a set of codes, a set of description patterns in several languages, and a set of buyer characteristics that together say "this company probably imports this thing." Build that set well and the matching afterwards is almost mechanical. Build it badly and every result downstream is wrong in a way that is hard to see, because a list of the wrong buyers still looks like a list of buyers.
A model drafts, a person checks
This is why the first step of NEXPORT is not search. It is building that translation, which we call a tag set, for each manufacturer's catalogue, and it is why we treat it as work that needs judgment rather than volume. A model reads the catalogue and the certifications and drafts the tags: the codes, the description patterns, the buyer traits. That draft is fast, and it is usually mostly right. Mostly right is not good enough for something every later result depends on, so a person checks every tag before it is used to match. They know that "film" in one industry is packaging and in another is optics, that a certification which is decisive in one market is unknown in the next, and that the difference between a product a buyer would switch for and one they would ignore lives in the specification, not the category.
The person is not there because the model is poor. They are there because the cost of a wrong tag is paid by the manufacturer's sales team in wasted calls, and because a wrong tag is invisible from the inside. A wrong buyer on a list looks like a buyer. The only way to catch it is to know the product, and the only way to know the product is to read it the way a buyer would.
What the translation becomes
There is a quieter benefit to doing the translation once and carefully. When the sales team crosses a buyer off the list, and says why, that reason travels back into the tag set. The next list is better, not because a model was retrained on the world, but because it learned this manufacturer's own product a little more precisely. Over time the translation becomes something the manufacturer owns: an accurate account of what they make, in the languages the rest of the world uses to buy it. That is worth having apart from any single list of buyers, and it is the thing a keyword search never produces.