How often we match CBP

Results

Exact 10-digit

44.2%

106 of 240

Same 8-digit subheading

50.8%

122 of 240

Same 6-digit HS code

57.1%

137 of 240

Same 4-digit heading

69.6%

167 of 240

Read it plainly: on these rulings, the code we suggest is fully right about 44 times in 100, and right at the heading level about 70 times in 100. Confidence helps: when the classifier said high confidence (112 rulings), the 6-digit code matched 75% of the time; at medium (113) it matched 44.2%, and at low (14) 21.4%. Treat a medium or low answer as a starting point and check the alternatives and questions it shows.

Method

1. Source: CBP Customs Rulings Online Search System (CROSS), New York (NY) ruling letters. We listed 4,170 NY classification rulings dated 2025 or later.
2. Filters: Category Classification, NY collection, ruling date in 2025 or 2026; Exactly one tariff number cited (one product per ruling); Not revoked, modified or operationally revoked; Ruled code exists in the current HTS data; chapters 98 and 99 excluded. That left 1,231 rulings.
3. Sample: 240 rulings dated January 3, 2025 to August 14, 2026, drawn at random with a fixed seed (2026) and spread across 70 chapters. The sample was fixed before the classifier ran and was not changed after.
4. Input: Product facts from the ruling letter only: the item named in the subject line plus the description paragraphs before CBP's analysis, with every tariff number and every sentence that mentions a heading, HTSUS, GRI or a suggested code removed.
5. Classifier: the same two-step pipeline the site and API use, with the tariff data downloaded October 22, 2026. No ruling text or answers were added to its prompts.
6. Scoring: exact 10-digit agreement, then whether the first 8, 6 and 4 digits match. Exact 10-digit is scored on the 240 of 240 rulings that cite a full 10-digit statistical number.

Limitations

• Ruling requests are not typical products. People ask CBP when the answer is unclear, so these are harder than an average catalog. The sample is also spread across 70 chapters on purpose, which gives chemicals, metals and machinery more weight than a typical online store has.
• Ruling letters describe products in more detail than most sellers do. A short product title would usually give a lower score.
• We removed tariff numbers and sentences about headings or the requester's suggested code, but some wording in the facts can still hint at the answer.
• Only rulings that cite a single tariff number were used, so multi-product rulings are not tested. Rulings whose code no longer exists in the current schedule were left out.
• 2 of 240 runs did not return an AI answer (an error, or the text was judged not to be a product). They count as misses.
• One run per ruling. Model answers can vary a little from run to run.
• A CBP ruling binds only for the goods it describes. Our result is guidance; the importer of record is responsible for the final classification.