Heavy engineering importers · updated 14 September 2026

US importer search for heavy engineering equipment: coverage across 5 product lines

Product lines

5

Average coverage

0.956

Quality

1.000

Time, all runs

35m 35s

What a research run is

Where a run looks

Five kinds of source, in order of how much each one proves. The first records a purchase that happened. The last cannot show that anybody buys anything.

1. 01

   US customs records of bills of lading

   a bill of lading is the shipping document that travels with each cargo, and US customs keeps a record of it: who took delivery, from which country, on what date. It is the only source here that names a buyer against a purchase that really happened. Read through a paid subscription, because the free sites hide the buyer's name.

2. 02

   Utility and public-power procurement

   the requests for bids, contract awards and spending plans that power utilities (private, city-owned and rural co-ops) and state agencies publish, and federal notices on SAM.gov, the US government's contracting site.

3. 03

   EPC contractors and project awards

   EPC stands for engineering, procurement and construction: the firms hired to design and build a project and buy its equipment. Their award notices and project pages are read here. The contractor buys the equipment long before the project's owner is named anywhere, and is usually far easier to reach.

4. 04

   Certification and duty sources

   what a seller must meet to sell here: the product standard buyers ask for, any required lab approval, the federal rule on the product itself, the rules on US-made content for federally funded projects, and the import tax (duty) charged today on goods from this country.

5. 05

   Trade directories

   member lists of trade associations, distributor lists and trade-show exhibitor lists. They are used only to find a contact at an organisation already found some other way. Being listed is not evidence that anyone buys.

How a run is marked

Each finished document is marked against 18 checks, written before the run, in four groups. The groups fail for different reasons. The first catches a run that wrote the wrong document. The next two catch a run that did not go somewhere it should have. The last catches a run that went, and then claimed more than it found.

• Shape — 4 checks

  the finished document has the four parts asked for: a Summary, an Importers table, a section naming the leads that were found and ruled out, and a checklist of reference documents.

• The source set — 5 checks

  all five kinds of source appear, even where nothing was found; none is left empty without a word; every ticked item names a document rather than a description; every “could not get it” says what refused; and every saved file the document points to is really saved.

• Each class — 5 checks

  one per kind of source. Customs records and certification sources must hold a document that was actually opened. Procurement, project awards and directories must be settled: either something was saved, or the document says what was searched and what refused.

• Claims that need a careful reader — 4 checks

  the customs section accounts for every shipment its own Total line says the source reported; that Total names the search it came from; every row in the table carries dated evidence, an origin and a gate (the certification or import duty that decides whether a sale can happen); and no cell copies a source's “hidden” notice where a real value should be.

How each run scored

Coverage is how much of that checklist a run met, from 0 (none) to 1 (all). Buyers and the exporter are not named. Each row is a product line, because the product decides how much there is to find: customs records exist for a transformer and barely exist for a steel structure built to order. The dashed line marks 0.95, the score we aim for.

• ·Average coverage 0.956 across all 5. The fairer reading is 3 of 5 at or above 0.95. An average of three perfect runs and two short ones is not five runs at 0.95.
• ·Quality 1.000 on every scored run. Quality is a second score: the share of sources the document cites that are really where it says they are. Coverage is how much was gathered; quality is whether the claims about it hold.
• ·Three of the five started from an empty workbook, which is harder: there was no earlier version to build on, and no correction from a past run to carry in. One of those scored 1.000.

Time and cost

Time is how long each run took from start to finish. Cost is counted in tokens, the small pieces of text that an AI model reads and writes. Read is everything the model took in; written is everything it produced.

What can still go wrong

• An edit can lose what a line was carrying. That is what the two short runs had in common. Asked to correct a count, one run moved its header from 7 to 12 when the answer was 13. In rewriting that line it also dropped the HS code (the customs code for the product type) and the period that said where the count came from, so a check that had been passing failed. The research was done; the record of it was not kept straight.
• Five product lines is a small test. Each figure is the last run for that product, and the page is updated as the test is run again. It is a record of where this stands, not a claim that the work is finished.
• Runs stopped by hand are left out. Every number is the last run that ended on its own terms: delivered, or out of time. A run someone stopped partway records a score like any other, and those are excluded.