Global Trade Signals, November 2026: A $336bn Import Month, Single-Source Exposure and the AI Citation Gap

1 — August imports: $336.17bn, and where single-source risk sits

The US Census international-trade series for August 2026 carries 582,823 HS10 rows covering 18,920 distinct ten-digit commodity lines and 232 supplier countries. Summed over real country rows it is $336.17bn of general imports — and it reconciles to the all-country total rows to the dollar (delta $0), which is the check that lets every share below be read as a share of the month.

Concentration is the headline: 2,592 of 5,470 HS6 groups (47.4%) already draw more than half their value from one partner, and those groups carry 34.0% of the month’s import value ($114.36bn).

Source: data/census-intl-trade/monthly/2026-08.jsonl, US Census Bureau International Trade API (api.census.gov HS endpoint, public domain), fetched 2026-10-21T13:53:44Z .. 2026-10-24T05:02:28Z; register C1–C4.

2 — Two chapters carry 44.0% of the import bill

Ranked by chapter, the month is not a spread — it is machinery and electronics. Chapter 84 (machinery, including the HS 8471 computing lines) holds 26.6% and chapter 85 (electrical machinery) 17.4%. Vehicles (87), mineral fuels (27), pharmaceuticals (30) and optical instruments (90) together add 22.0%.

The concentration matters for classification work rather than for procurement: a ten-digit error on a chapter-84 line moves a duty rate on a large base, and those are the lines where HTS rulings and Section 301 lists overlap most densely.

Source: same artifact as §1 (api.census.gov HS endpoint), chapter rollup of the HS6 sums, observed 2026-11-02T17:40:00Z; register C3.

3 — Partner breadth: the median HS6 line already has 20 suppliers

Diversification is the norm, not the exception. Across 5,470 HS6 groups the partner count has a median of 20, a mean of 26.9, a 90th percentile of 60 and a maximum of 228. Only 196 groups (0.1% of value) have exactly one supplier.

Read the same month by partner instead: MEXICO leads at 18.0%, and the top six origins supply 56.5% of the month.

Source: same artifact as §1, partner-count distribution and country rollup, observed 2026-11-02T17:40:00Z; register C4.

4 — AI answer engines cite this site on 7.0% of monitored questions

We ask the same 10 questions of an answer engine on a fixed schedule and record whether the response cites exportsonar.com. Across 11 runs (2026-10-20 to 2026-11-02) that is 143 answered legs and 10 citations — 7.0%, with 0 errors and 0 quota skips, so the denominator is real responses rather than attempted ones.

The average hides the whole story: every citation comes from the new cohort, and inside it from the questions phrased as a direct factual ask. The legacy cohort — long-tail questions about stacked tariffs, de minimis and HTS digits — was cited zero times in 88 legs. Negative controls on the same runs (simplyduty.com, zonos.com) returned 0 citations, which is what makes the 7.0% interpretable: the engine does cite sources.

Source: data/ai-citation-shifts.jsonl, our own leg log (channel pollinations_openai), runs 2026-10-20 .. 2026-11-02, filtered to the observation stamp 2026-11-02T17:40:00Z; register C5–C6.

5 — Crawler mix: 68 answer-engine hits in 23 days, and 1,195 that are us

The funnel ledger tags every hit with its user-agent family. Over 23 days it recorded 68 answer-engine crawler hits (meta-webindexer 39, OAI-SearchBot 21, meta-externalagent 7, GPTBot 1) — and 1,195 headless-Chromium hits that are our own rendering and screenshot passes. Counting those as traffic would inflate the crawler total by 17.6×, which is why the report states the exclusion instead of quietly netting it out.

For a data site the distinction is operational, not cosmetic: an answer-engine crawler is fetching a page to quote it, so the page it lands on is the page whose sentence gets quoted. That is the same mechanism §4 measures from the other end.

Source: data/funnel-ledger.jsonl, first-party beacon ledger, rows at or before 2026-11-02T17:40:00Z; register C7.

6 — Funnel instrumentation: 2,715 human events, one step that cannot be right

Stripping bot and probe rows leaves 2,715 human events across 602 sessions. The counts are published with their caliber rather than as a conversion rate, because one step fails an obvious plausibility test: 548 checkout views appear across 602 sessions — 91.0% of them — while only 1,304 landing views were recorded. A funnel step cannot fire on more sessions than its own entry point.

The repair is a beacon question, not a reporting one, so the honest output this month is the count table plus the flag. What the ledger does support: 35 completed signups, 104 USDT invoices created and 36 CTA clicks in the window, each counted from a single event type.

Source: data/funnel-ledger.jsonl, human rows only (bot and probe rows excluded, both flags tested), rows at or before 2026-11-02T17:40:00Z; register C8.

7 — Competitive board: 6 live tools, one 404, and an 8.8× index gap

Batch 2026-10 of our competitor board measured 6 live tools between 2026-10-11 and 2026-11-01. One challenger is gone: TariffBench returned HTTP n/a at its own domain on n/a — a n/a-byte Vercel “deployment not found” body, confirmed under a second content negotiation.

Two gaps are worth naming. Index size: 7,052 sitemap URLs against 804 for the largest live rival, 8.8×. Machine-readability: 3 of 6 live tools publish an llms.txt. Response time: 0.209s against 1.397s for the fastest rival measured — a single-sample homepage fetch, so read it as an order of magnitude, not a benchmark.

Source: data/competitors.json (ExportSonar competitor board, batch 2026-10, cells scraped 2026-10-11 .. 2026-11-02); dead-tool probe recorded in the same file under status_evidence; register C9–C10.

Claim register

Every figure in this report resolves to one row here: claim id, the figure as published, the artefact it was computed from, the upstream source, and the observation stamp. A figure without a row does not ship.

Frequently asked questions