How a brand becomes an ownership call

Tariffed publishes two datasets. This page says how each one decides what it knows, what it refuses to say, and where it would break.

Written 2 September 2026, against the brand build of 28 August 2026. The short version is on How we check.

Two datasets, two questions

The ownership dataset answers one question. Given a brand that paid for an ad, which company owns it, and what country do public records tie that company to? 710 brands are seeded and 415 clear the bar to be hidden. Every call and the records behind it sit on the brand list.

The Canadian product registry answers a different question. Given a Canadian business, does the product it makes appear on a published United States tariff list? 1,740 listings, on the registry page.

They share one rule. A claim is published when public records carry it, and the record is linked on the row so a reader can check the work. The rest of the two methods differs, so they are described apart.

The chain from an ad to a country

An ad on a page is a slot. The filter walks it in one direction.

Eight rules try to name the brand, in a fixed order, first answer wins. The strongest reads the click destination, because a paid link says where the money goes. Weaker rules read the advertiser name node, a verified byline, the words inside a paid label, and last the product named in the ad copy. A rule that fires late carries less weight, and the order is the weighting.

Name matching reads from the head of a phrase, up to three words, longest match first. A country suffix on the end is ignored, so "Nike Canada" resolves to Nike. A single everyday word is refused unless the slot has already proved it is paid. Without that guard the word "target" hides a sentence about archery.

Three registries, asked the same question

The country of the ultimate parent is asked of three separate sources.

The three are compiled by different people from different inputs. Wikidata is edited by volunteers. GLEIF is fed by the banks that issue legal entity identifiers. EDGAR is fed by the filings themselves. Agreement between two of them is stronger evidence than one record on its own, and it is not proof. Nothing in the code checks whether a Wikidata editor read the LEI record before writing the page.

They also do not report the same thing, and this page should not pretend they do. Wikidata reports the country on the parent's own entry, property P17, falling back to the country of the headquarters and then to the country of origin. On this build that is 556 brands by P17, 8 by headquarters and 3 by country of origin. GLEIF reports the country of the legal address on the LEI record, which is not the jurisdiction of incorporation. EDGAR reports the state of incorporation, and is the only one of the three that reports incorporation at all.

Ownership follows P749 alone. P127, "owned by", is ignored, because index funds hold a slice of everything and following that property walks half the S&P 500 to BlackRock.

The walk is best effort. It follows the first P749 claim on each company and stops after five steps, so a chain that branches or runs deeper is cut short. "Ultimate parent" means the end of that walk, and not a proven top of the tree.

A source that cannot answer has to abstain

Every registry is allowed to say nothing, and saying nothing is treated as silence rather than as evidence.

GLEIF searches by legal name, and legal names collide. "NIKE" is registered in Luxembourg and "NIKE, INC." is registered in Oregon. Both normalise to the same string, so picking the first record picked Luxembourg. When the matched records disagree on a country, GLEIF abstains. "Abbott Laboratories" matches 8 records across Russia, the Philippines, the United States, Germany, Switzerland, Britain, Spain and Canada. "AbbVie" matches 14 records across 14 countries. Both abstain.

Name normalisation upper-cases the name, turns punctuation into spaces, and deletes a fixed list of corporate words wherever they appear. INC, LTD, GMBH, THE and about twenty others are on that list. GROUP, HOLDINGS, INTERNATIONAL and WORLDWIDE are not, because they are parts of a company's name. Adding them merged "Shopify International Limited" in Dublin into "Shopify Inc." in Ottawa, and two different companies reading as one is how a wrong country gets confirmed. Deleting a word wherever it appears is blunt, and a company whose real name contains one of those words normalises to something shorter than its name.

A near miss is recorded as no match at all, and never as a conflict.

The rule that decides

There are two ways a brand becomes hidden, and the 415 hides split 317 and 98 between them.

By registry vote, 317 brands

All five of these have to hold.

1. At least two of the three registries answered.
2. Every registry that answered gave the same country.
3. That country is the United States.
4. The curator's prior, written down before the run, does not disagree. A brand with no prior passes, because silence is not assent.
5. The brand's own country does not contradict the vote. A brand with no country of its own passes, because silence is not a Canadian answer. On this build no brand was held back by this test alone.

A brand that fails the vote lands in unclear and is never hidden. A brand that passes the vote with a country other than the United States is confirmed and not hidden, and 110 brands sit there.

By curator override, 98 brands

A person read the filings and wrote the country by hand. An override sets the country directly and skips the five tests above, so neither the two-registry floor nor the contradiction test applies to those 98 rows. They rest on the two source links in the override row and on nothing else. That is 24 percent of everything the filter hides, and it is the largest hole in the method.

The two errors do not cost the same. Missing a hide costs one ad on one page. Hiding an ad that should have stayed costs the product, so the rule is built to fail towards missing.

The curator tier

The registries answer with the country a company registers in. That is Bermuda for a jeweller run out of Ohio. Registration is a filing choice, and for this question the filing choice is the wrong answer.

112 rows hold a verdict a person wrote after reading the filings. Each row carries a country, a reason in one sentence, and two source addresses, so the claim can be checked rather than trusted.

The override is applied after the registries have been asked, never instead of them. Alo Yoga keeps its Wikidata identifier and its GLEIF record in the source list alongside the override, so a reader sees the machine answer and the human one together. An override sets the tier to curated and leaves the disagreement on the row.

What the numbers say

710 brands. 415 hidden, 295 not hidden. The 295 split four ways, and two of those splits are refusals.

So 171 brands out of 710, a shade under a quarter, are refusals. 124 more are answered and are not American.

Source links by origin: Wikidata 680, GLEIF 456, SEC EDGAR 244, curator 224. The curator figure is 112 override rows carrying two links each, and not 224 separate readings.

Coverage, counted by how many of the four kinds of source a brand carries: 229 brands carry three, 331 carry two, 143 carry one, and 7 carry none. Counted by raw links the same file reads 31, 274, 260, 138 and 7, which is the larger-sounding number and the less useful one.

Ownership, top countries: United States 417, Canada 75, Japan 31, Germany 23, Britain 22, China 19, France 14. No brand is hidden with a non-American owner.

Why a brand is refused

Every entry carries a review note. The engine writes down why it hesitated, and the reason travels with the row into the published file.

A brand can carry more than one reason, and a reason can sit on a brand that an override later resolved, so the column does not sum to the refusal count.

Six of the 32 conflicts name Cayman on one side. Wikidata reports where the company operates and GLEIF reports the legal address it filed, so the two split. Three of the six are settled by an override, being Herbalife, Razer and Webull. ByteDance, NetEase and Trip.com are still unclear and still visible. Whether all six are flags of convenience is a reading of the companies, and the data file does not carry that judgement.

Refusal one, Pet Valu

Wikidata answered United States. GLEIF answered Canada on the record for Pet Valu Holdings Ltd., jurisdiction British Columbia. Two sources spoke and they disagreed, so the tier is unclear and the ad stays. A Canadian retailer listed on the TSX was kept visible by the disagreement rule. Under a one-source method it would have been hidden on Wikidata's word.

Refusal two, Scopely

Only Wikidata answered, and it said United States. The curator's prior said Saudi Arabia, because Savvy Games Group, the Saudi sovereign fund's gaming vehicle, bought Scopely in 2023 and the Wikidata parent chain has not caught up. The tier is unclear, so nothing is hidden. This is the case the two-source rule was written for. A stale encyclopedia entry cannot hide a company on its own.

A brand that passes, American Airlines

The seed file pins the Wikidata entity for this brand to Q32396 rather than leaving the domain to resolve it. From there Wikidata walks the parent chain up to American Airlines Group and reports United States. GLEIF matches "American Airlines Group Inc." to LEI IWUQB36BXD6OWD6X4T14, legal address United States, jurisdiction Delaware. EDGAR matches the same name to CIK 0000006201, state of incorporation Delaware. Three registries, one answer, an empty review note.

The Canadian product registry

The registry asks a different question and holds to the same standard of proof. Candidates come from public directories, and each one has to clear four gates.

1. The host answers over https at its own root. A 403 or a 503 counts as alive, because a bot wall and a busy server both mean the business is still there. Silence or a 404 counts as gone. 20 dead hosts were dropped.
2. The website host is not the directory that supplied it. The checker repeats this test after every merge, so a directory address cannot quietly become a business website.
3. The host is new to the registry and new to the run. Deduplication is by normalised hostname, and www, m, shop and store fold in however deep they are stacked. 21 duplicates were dropped.
4. The source, the business name and the words in the domain together support a tariff heading. 29 apparel rows failed this and were dropped, and 6 survived a hand check.

The bulk merge of 24 August 2026 put 1,378 candidates in and admitted 1,308. The file holds 1,740 listings today.

The reading stops at six digits

The Section 338 list holds 554 eight-digit codes across 58 chapters. Those collapse to 373 six-digit headings, of which 217 are fully covered and 156 are covered in part. A partly covered heading returns Needs review. Six digits cannot settle the duty on a product, and picking the seventh and eighth digit for somebody else's goods is customs business under 19 CFR 111. A licensed broker settles it. 43 of the 558 published listings carry that reading today.

The reading is never read from the file. The checker recomputes it on every run and either writes it back or reports the file as wrong, so nobody can hand-write a severity. Antidumping duties and patented pharmaceuticals are out of scope, because heading 9903.04.60 follows patent status rather than a code.

Who said what

Three fields carry the evidence, and the card prints which one applies.

• Who supplied the listing. All 1,740 rows today came from a public source, and a public row cannot be published without the page it came from.
• Whether anyone has stated that the goods go to the United States. 12 rows say yes. 1,728 say maybe, which means the business makes a covered good and nobody has said whether it crosses the border.
• Who said it. A business speaking about its own sales stands on that line alone. Anyone speaking about somebody else's business has to hand over a page a reader can open.

A share-of-sales band is refused on a listing where nobody has stated that the sales happen. A percentage attached to an unproven claim is decoration.

What held back means here

1,182 of the 1,740 listings are held back, and every one of them is a drink maker. That is an editorial hold on an over-represented category. Those rows are checked and correct, and they wait off the page until a business asks to be drawn. Keeping them in the file is what stops the merge tool adding them back.

The registry's real refusals are the 43 readings that need review and the 1,728 rows that carry no stated United States sales.

Provincial coverage is uneven. British Columbia has 665 listings and Saskatchewan has 30. That is a gap in the source directories, and it is recorded rather than filled with guesses.

Where this breaks

Five weaknesses, in the order they worry me.

• Wikidata is the only source that resolves a domain to a company. GLEIF and EDGAR both search by legal name, so a brand with no Wikidata page needs a curator to supply a legal name before the other two can be asked. 30 brands hit this. The fix is a second domain-to-entity source, and there is no free one.
• One volunteer edit can move a country. Wikidata answered 680 of 710 brands, far more than either other source. The two-source rule contains the damage. 186 brands were left with Wikidata as the only voice, 127 of those stayed unclear, and a person read the filings on the other 59. The rule limits the exposure and does not remove it.
• The curator tier is one person's reading, and it skips the safety rules. 112 overrides carry 98 hides, which is 24 percent of everything the filter hides. Those rows do not need two registries and are not tested against the brand's own country. Each one has two sources, so the reasoning is checkable, and none of it is peer reviewed.
• Ownership goes stale faster than the records do. Scopely changed hands in 2023 and Wikidata still walks the old chain. Every source row carries the date it was read, and nothing re-checks on a schedule.
• The registry proves a business exists and makes a covered good. It does not prove the goods cross the border. 1,728 of 1,740 rows say maybe for that reason, and it is the largest open question in the file.

Three things would break the method outright. A licence change at GLEIF or EDGAR, because the data is publishable only while all three sources are CC0 or public domain. A registry going quiet, because two sources is the floor and losing one would turn every abstention into a refusal. A parent chain running through a jurisdiction with no public register, because the Cayman rows are already resolved by hand and hand resolution does not scale.

Checking the figures

Every count on this page was read on 2 September 2026, against the brand build of 28 August 2026. The brand figures come from the file the extension ships, and the same file draws the brand list, where each row names its sources and links them. The registry figures come from the file behind the registry page, checked by a script that exits non-zero when a listing breaks a rule. Two sets of numbers come from somewhere else. The tariff-list figures, 554 codes down to 373 headings, are read from the Section 338 list itself. The merge figures of August 2026 are read from the build log, because the registry file keeps no history.

The description of the ownership engine on this page was read back against the engine's own code by a second model on 2 September 2026. It found eight places where the prose claimed more than the code does. All eight are corrected above.

If a figure here disagrees with a page, the data file is the one to believe, and the disagreement is a bug worth an email.