Method

How we used Sayari MCP

What we learned from Sayari, how well it worked, and every query we ran. Open a case below to replay its queries. Every link on the site points back to one of them.

What the tool taught us

Confidence ratings are explained in How we rate confidence.

Name filters can silently fail. The shipper_name and receiver_name filters were ignored twice and returned 10,000 unrelated rows (s010, s012). Putting the name in the free-text query worked (s017).
Rollups can be blind where rows aren't. Supplier rollups for HS 3801 from Indonesia to the US returned 0 (s015, s016), yet 19 shipment rows existed.
Company filings fill registry gaps. Where Sayari had no owner, stock-exchange filings did: BTR's 60% of its Indonesian plant, and Gotion's wholly owned US subsidiaries. Sayari plus filings beats either alone.
What worked well: profiles (ships_to lists), upstream mapping by HS code, and free-text shipment search gave a dated, HS-coded trail from China to five US cell plants.
9 more lessons
One plant, five records. Each customs source spells the shipper differently, so Sayari keeps separate records. The street address was what tied them together (s005–s007, s022, s023).
Ownership coverage is uneven. No shareholders for the Indonesian plant (s009), even though Sayari shows BTR owning a different Indonesian company (s020).
Screening hits two hops out are mostly forwarders. All ten watchlist paths ran through freight or trading firms (s021). This matches Sayari's own guidance (s001).
"OR" queries return nothing. Boolean queries like BTR OR Putailai silently returned 0 in Cases 3 and 5. Search one name at a time.
Free text ranks; it doesn't filter. A quoted name in query puts matching rows first, but totals still count everything. Read the rows and don't trust the count (Case 2).
Corridor buckets can be mislabelled. 3,088 Morocco → US "artificial graphite" rows were mostly fertilizer and chemicals, and 917 Oman → US rows were mostly Indian cargo relayed through Omani ports (Cases 2 and 3).
Registry coverage depends on the country. Ownership was visible for plants held through Singapore (Morocco, Malaysia) and in China's registry, but not for the Indonesian PT (Cases 1, 2, 3 and 5).
Query syntax matters. Adding a country word to a quoted name, or putting two quoted names in one query, returned 0. A page token can't be combined with limit, and long shortest-path searches can time out (Cases 7 and 8).
Registries lag behind real changes of owner. Japan's TDK bought Linergy Power in June 2026, but in October Sayari still showed the old Chinese (CATL/Ampace) chain. Company filings caught the change (Case 6).

Latest re-check: 5 Nov 2026

We re-ran every case, buyer scan and policy date. Ownership held in all twelve cases, but several published claims needed correcting, and the trail has moved on since October.

  • Case 5 was wrong. We said Putailai's exports couldn't be found. A search on the full name of its anode company, Jiangxi Zichen, finds 2,313 rows, including 10 to US buyers in 2024–25.
  • Case 12 used a page count as a total. The 87 StarPlus rows were one page of results. The real figure is at least 448, to Oct 2026.
  • The "GM" row of Sep 2026 went to LG. Its delivery address is LG Energy Solution in Holland, Michigan. Sayari's General Motors record mixes several companies, so we now read the address on each row.
  • BTR graphite keeps reaching Ford. Marshall is now 11 rows and about 1,160 t, to 7 Oct 2026. Rows naming BTR Indonesia into the US are now about 100, not 55, and Hyundai's Georgia venture is a new buyer.
  • Sayari retired several records. Syrah Vidalia, Gotion Illinois and Ultium Warren moved to new IDs. Syrah's new record now shows an owner.
  • Policy. The USITC anode vote is under appeal (single source: one company filing). China's pause on tighter graphite checks for the US ends 27 Nov 2026, and no notice has moved the 10 Nov end of its anode-control suspension. Treasury has published nothing new on PFEs.

Not checked: Ford's Q3 10-Q, the court docket, and two hand-read samples in Case 2. Rows into Malaysia and Oman stop months ago, so "no new rows" there can mean no newer data. Full notes on GitHub

Earlier: re-checked on 18 Oct 2026

Before publishing to a wider audience, every headline claim was re-run against Sayari, and every policy date against primary sources (52 calls, v001–v052 in the log below). Most claims were confirmed. These were corrected:

  • Case 1, shipment count. About 55 rows, not 19. A second record holds about 31 rows recorded as loaded in Singapore, so a search filtered to "departing Indonesia" missed them. Shipments to Ultium Cells continue to Nov 2025, and one reaches GM in Sep 2026.
  • Case 1, main record. Sayari merged our anchor record after 14 Oct; its tax ID and LEI now sit on another record, which still shows no owners.
  • Case 2, first graphite. BTR's Tangier plant received 1,240 t of artificial graphite from BTR Indonesia on 19 Aug 2026. It was previously reported as having no graphite trade.
  • Case 7, "via Jamaica". Kingston is the shipping line's transshipment hub. These are Chinese shipments, not a Jamaican route.
  • Policy. EVE Energy's 1260H listing (8 Jun 2026) is confirmed. China's truce now runs to 10 Jan 2027, but the legal suspension of its graphite controls still ends 10 Nov 2026.

Lesson: Sayari's records change over time, and filters shape answers. Re-run key queries before relying on them. Full verification report

Sayari capability scorecard

Worked = the call answered its question. Partial = it answered with noise or only part of the answer. Failed = 0 results, or the filter wasn't applied.

Loading…

Red-flag rubric

Adapted from Sayari's investigation guidance (s001, s004) and mapped to the US prohibited-foreign-entity rules. "Detected" means a Sayari record shows it; "suggested" means signals point to it without a record; "not visible" means we couldn't test it; "checked: absent" means we tested it and the data shows it isn't there.

Loading…

Every query, by case

Open a case to replay its queries in order: the question, the tool and parameters, the result, and what it means.

Loading…

Search all calls

Every Sayari and Tavily call, including re-checks. Type to search; results appear below.

Show the full log
IDServerToolQuestion and resultScoreTime

Results are our own summaries. We don't publish raw Sayari records; the data is licensed. Tavily summaries are paraphrased. Tavily timestamps are approximate.