Trade intelligenceyou can defend.

Named buyers come from US customs bill-of-lading records, so the companies we can put a name to are US importers. That is the honest boundary of the discovery engine, and it is better stated here than found after you sign in.

Naming buyers outside the US needs a commercial customs licence we do not hold yet. Market-level intelligence reaches much further: demand by destination and clearing unit values are built from Indian customs aggregates, and which markets are growing fastest comes from UN COMTRADE, which reports each participating economy’s own imports of an HS code. Party names in both of those sources are masked at source — so they tell you which market to go after, not which company to email.

Discovery starts from customs records. A product search returns the overseas suppliers who ship it, and the buyers are the importers those suppliers provably shipped to.

A supplier only survives the filter when the commodity’s customs chapter dominates its shipments, so a tool manufacturer with one incidental spice consignment never becomes a spice lead. Shipment counts, dates, HS codes and supplier lists are then stored exactly as the source recorded them — no language model rewrites a fact to make a profile look complete — and every record keeps its provenance: customs evidence, directory or enrichment data, model-inferred, or sample data. Those labels stay distinct.

No. It is a fixed 0–100 formula over five published components: product match against the HS headings on the buyer’s own shipments (40 points), shipment volume (25), recency (15), which side of the trade they sit on (10), and market fit (10).

Every score is shown with the reasons that produced it, so you can disagree with one factor instead of the whole number. Two rules are deliberate: a company with no customs evidence scores low rather than average, because an empty field is never quietly counted as a positive; and a high score is a queue order, not a forecast. It says contact this one first — not that they will buy, and not that they are good for the money.

Sometimes — and the product tells you which, every time.

Contact discovery escalates cheapest-first: contacts already on the record, then a crawl of the company’s own website, then an enrichment provider once one is connected. The limits are real and worth knowing before you rely on them. Many customs profiles carry no website at all, and a crawl returns nothing on JavaScript storefronts or sites with no crawlable contact page. When a tier finds nothing you get the reason — “no website on record” — instead of a silent zero, and every match carries a reachability marker so you can work the buyers you can actually reach first. Pattern-guessed addresses stay switched off unless explicitly enabled, and even then survive only when a verifier confirms the address is deliverable.

No. Companies, messages, contacts, deals, quotes, documents and search results are scoped in the database by row-level security to your organisation, and there is no cross-workspace read path for any of it.

Two layers are shared on purpose, and neither one contains another workspace’s work. The first is a global cache of public customs facts about a company — shipment counts, HS codes, supplier lists — so the second person to look a buyer up gets it instantly rather than paying to fetch the same public record again. The second is tenant-anonymous behaviour: plain counters for how a buyer tends to respond, holding no workspace identity, no message text, no contact and no price. Nothing there is published from fewer than two contributing workspaces, and no percentage is published below a minimum sample size.

No. It drafts; a person reviews and sends.

Outreach copy, follow-ups and quote language are generated as drafts and stay drafts until someone in your workspace sends them, and the Copilot’s chat tools are read-only apart from a prospect scan you explicitly ask for. Sending has its own guards: email carries an unsubscribe footer and honours your suppression list, and a buyer flagged by denied-party screening raises a compliance warning at the moment of send.

Discovered companies are screened against the US Consolidated Screening List — OFAC’s SDN list, the BIS denied-persons and entity lists, and the rest of that consolidated file.

The matcher is deliberately conservative, because falsely flagging a legitimate buyer would poison trust in every other number on the page: it only flags when the normalised names genuinely align. If a check fails or has not run yet, the record reads “not screened” — it is never quietly recorded as clear. Treat a clear result as a prompt to look closer, not as legal clearance.

Report it or merge it, without leaving the product.

Any field visible on a company dossier can be flagged as wrong. The report is stored on the record as your claim and travels with it through a merge, and it deliberately does not edit the underlying figure — the source record has to stay the source record. Duplicates thrown up by customs OCR variants can be merged by an owner or admin: the merge repoints everything attached to the duplicate before deleting it, and fills only the fields the surviving record left empty. Matching errs in the other direction too — two similarly named rivals stay two records rather than being fused into one.