Authorized Retailers Specification v0.8
Draft v0.8 · 23 October 2026
Editor: Authorized Retailers LLC (authorizedretailers.ai)
Licensed under CC BY 4.0. JSON Schemas and reference code: Apache-2.0.
1. Overview and scope
This spec defines how a brand records which retailers it has authorized to sell its products, and how an AI shopping agent verifies that status before recommending or buying. It is an open format with a single registry: anyone may read, cache or relay what the registry publishes, and anyone may build software on the format, but authorizedretailers.ai is the only registry. A list or answer counts only if it verifies against the registry’s published key set (section 8).
The registry works like a notary. It checks that a brand controls its domain and is independently linked to it (section 4), records the brand’s own decisions and signs them. The signature says who published a list, not who is authorized. Only the brand decides that (section 3).
Status: draft v0.8, for discussion with brands, retailers and agent developers. Breaking changes are expected before v1.0. Changes are listed in section 14.
The spec asserts one thing: whether a brand has authorized a given seller, on a given channel, in a given territory, for a given product, as of a given date.
Scope: the spec covers consumer-facing online channels: marketplaces and web stores. Every channel is identified by a domain, because agents buy from domains. Physical storefronts are not channels: an omnichannel retailer is listed by its web domain. Trade, wholesale and practitioner-only channels are out of scope. Sellers there are not listed, so their resale on consumer channels verifies as unlisted. So are sellers on a platform that hosts many independent sellers under one domain (a practitioner dispensary platform, say): the domain does not identify the seller, so authorizing it would authorize every seller on it, and a brand that does that has authorized them all. Purchases across borders (for example on amazon.com for delivery to Canada) are out of scope for v0.6.
It does not assert:
- Authenticity. An authorized retailer can still sell counterfeit goods or have its account compromised. Authorization and authenticity are separate questions.
- Pricing. The format carries no price, MAP or discount terms.
- Legal status. A seller not listed is unlisted: the brand hasn’t authorized it, and nothing more is implied. A seller the brand declares unauthorized (section 3.2) is one the brand has found selling its products without authorization. The registry records and signs that declaration; it does not investigate the seller. An agent SHOULD NOT send a shopper to a declared seller: goods from an unauthorized seller may be materially different from what the brand sells, or counterfeit, and the brand’s warranty may not apply.
The key words MUST, SHOULD and MAY are used as defined in RFC 2119.
2. Terminology
| Term | Meaning |
|---|---|
| Brand | The owner of a trademark whose products are sold, identified by one or more verified domains. |
| Seller | Anyone who sells a brand’s products on a channel, whether or not the brand has authorized it: a marketplace seller or a web store. |
| Retailer | A seller the brand has authorized to sell its products (an authorized retailer). |
| Authorization | A brand’s statement that a seller may sell its products within a stated scope, until the brand revokes it or until an end date the brand sets. By that act the seller becomes an authorized retailer. |
| Declaration | A brand’s statement that a seller on a channel is not authorized to sell any of its products (section 3.2). |
| Channel | A place a seller sells, identified by a domain: a marketplace domain and seller ID, or a web store’s domain. |
| Catalog | The brand’s list of products on the registry. It defines what an authorization covers and lets agents find the brand from a product (sections 3.1, 9.5). |
| Product | One entry in a catalog: a brand-chosen ID, a name and at least one GTIN, ASIN, Walmart item ID or WPID. |
| GTIN | A GS1 Global Trade Item Number: a UPC (12 digits), EAN (13 or 8 digits) or GTIN-14. A UPC is a GTIN. |
| Product line | A named group of products, such as “Classic Tee”. |
| Observation | Evidence from outside the brand’s list that an authorized retailer is actively selling on a channel. |
| Registry | authorizedretailers.ai: the service that verifies brands, records their decisions, signs lists and answers and responds to verify requests. |
| List | The registry’s signed copy of one brand’s authorizations and catalog (section 5.2). |
| Pointer | An optional file on a brand’s domain that names the brand’s list and repeats its verification token (section 5.1). |
| Mirror | A copy of the registry’s published keys, public lists and revocation feed, run on independent infrastructure (section 10.2). |
| Agent | Any automated system that queries authorization status before recommending or buying. |
3. Authorization model
Every authorization MUST come from the brand. No other party can authorize a seller, including the registry itself.
Sellers can request authorization. Only the brand’s approval creates one. Anything the brand has not approved verifies as unlisted.
The registry MUST NOT add, remove or change an authorization except on the brand’s instruction, or when the brand approves a seller’s request, or, as its terms of use reserve, to remove an entry it reasonably believes is unlawful or abusive. The registry MUST keep a change log for each brand, recording what changed, who changed it and when (and, for a removal the registry makes, why), and MUST show that log to the brand.
A request to a brand that is not yet registered MAY be held by the registry and delivered once the brand registers. A held request has no effect on verify answers.
A registry MAY ask a seller that requests authorization for evidence of where it buys the brand’s products, such as an invoice from a distributor or from the brand, and MAY show that evidence to the brand with the request. The evidence is for the brand’s decision only: it MUST NOT be published, MUST NOT change verify answers, and the registry SHOULD delete it within a set time after the brand decides.
Each authorization MUST carry:
- Channels: at least one channel identifier (section 6).
- Product scope: product lines (
product_lines: named lines, or["all"]), individual products (products: catalog product IDs), or both. Exclusions MAY narrow it (section 3.1).
Each authorization MAY carry:
- Territories: ISO 3166-1 alpha-2 country codes, or
*for worldwide. Most channels take their territory from their domain, so territories are needed only for web stores on generic domains (section 6.2). - End date:
expires, set by the brand. Without one, the authorization lasts until the brand revokes it (section 10).
3.1 Products and catalog
A brand MAY keep a catalog on the registry: a list of products. The catalog does two things. It defines what an authorization covers, and it is what makes a brand findable from a product page.
An agent on a product page sees a product identifier and a seller, not the brand’s domain. The registry finds the brand from the product through the catalogs of registered brands (section 9.5). A brand can register and authorize retailers without a catalog, but agents can’t reach it from a listing until it adds products.
- GTIN is the baseline. A UPC or EAN column is enough to start. Most product feeds and listings outside Amazon expose a GTIN.
- ASINs for Amazon. A brand that sells on Amazon SHOULD also list its ASINs. Amazon listings expose ASINs more reliably than barcodes.
- Walmart item IDs and WPIDs for Walmart. A brand that sells on Walmart SHOULD also list its Walmart item IDs, the number in a walmart.com product URL, which is what an agent on the page can read. It MAY add WPIDs, the product IDs in Walmart Seller Center.
The catalog is not a general product database. It holds only what the brand chooses to add. Finding the brand from a product names the brand and nothing else from the catalog (section 9.5).
{
"id": "EB-TEE-BLK-M",
"name": "Classic Tee, Black, M",
"product_line": "Classic Tee",
"attributes": { "color": "Black", "size": "M" },
"gtin": "00012345678905",
"asins": ["B0EXAMPLE1"],
"walmart_item_ids": ["123456789"],
"wpids": ["1A2B3C4D5E6F"],
"mpns": ["EB-100-BLK"],
"style_code": "EB-TEE",
"regions": ["US", "CA"]
}
| Field | Rules |
|---|---|
id | Required. Chosen by the brand (a SKU works) and unique within the catalog. |
name | Required. |
product_line | Optional. The same product line named in scope.product_lines, so a brand can authorize a whole line in one step. |
attributes | Optional flat string key/value pairs, for display and filtering. Never used for matching. |
gtin | Optional. 8, 12, 13 or 14 digits with a valid GS1 check digit (a UPC, EAN or GTIN). Compared as GTIN-14: left-padded with zeros to 14 digits. |
asins | Optional. Amazon ASINs: 10 uppercase letters or digits. An ASIN applies on every Amazon marketplace domain. |
walmart_item_ids | Optional. Walmart item IDs: 5 to 15 digits, as in a walmart.com product URL (walmart.com/ip/<name>/<item ID>). |
wpids | Optional. Walmart product IDs (WPIDs): 8 to 16 uppercase letters or digits. |
mpns | Optional. Up to 10 manufacturer part or model numbers, such as regional variants of one model. |
style_code | Optional. The brand’s style code, often shared by every size of a style or colorway. |
status | Optional. active (the default), discontinued or recalled. |
status_since | Optional. The ISO 8601 date the status took effect. |
recall_url | Required when status is recalled: the official or brand recall notice, over HTTPS. |
regions | Optional. Where the product is made to be sold: ISO 3166-1 alpha-2 codes or named groups. Without it, the product is made for every territory. |
Each product MUST have a gtin or at least one ASIN, Walmart item ID or WPID. A product identifier (a GTIN, ASIN, Walmart item ID or WPID) MUST NOT appear on two products in the same catalog. Walmart item IDs and WPIDs apply on walmart.com, the marketplace that issues them.
Model numbers and style codes are unique only within a brand: many brands use “X100”. They MUST NOT find a brand or a product on their own (section 4.2). An agent MAY send them with a product identifier, and the answer says whether they match (section 9).
Status. A recall is a fact about the product, not the seller, so it doesn’t change whether a seller is authorized. It is carried on every answer about the product (section 9). The registry SHOULD tell the retailers the brand has authorized for a recalled product, when they have registered. A discontinued product can be legitimate clearance stock. Counterfeits of discontinued products are also common, so brands see how often agents ask about them.
Regions let one catalog hold products made for different markets, such as a US and an EU formula with different GTINs. A product is covered only in the territories its regions include, whatever an authorization says, as with exclusions. A region is a country code or one of these groups:
| Group | Countries |
|---|---|
EU | The 27 member states of the European Union. The United Kingdom (GB) is not included. |
Each market’s version of a product has its own GTIN, as GS1 rules require when a product’s formula changes, so the registry can tell them apart.
An authorization’s scope can be narrowed with exclusions, and a brand can exclude products from every authorization:
scope.exclude_product_linesandscope.exclude_products: products this retailer is not authorized for, such as every line except one, or a line except its newest release.exclusions.product_linesandexclusions.productson the list: products no retailer is authorized for, such as products sold only direct to consumers.
A product is covered by an authorization when product_lines is ["all"], or includes the product’s line, or products includes its id, and it is not excluded by the authorization or by the brand. Exclusions always win over inclusions.
Every product line named in a scope or an exclusion MUST exist in the catalog when the brand has a catalog. Every product ID named in products, exclude_products or exclusions.products MUST exist in the catalog. Named products and exclusions therefore need a catalog.
Matching a verify request against the catalog is described in section 9.2.
3.2 Declared unauthorized sellers
Unlisted means only that the brand hasn’t authorized a seller. A brand that has found a seller selling its products without authorization, a diverter say, MAY say so with a declaration. Verify answers about that seller are then unauthorized instead of unlisted (section 9).
- A declaration names one channel identifier (section 6): a marketplace domain and seller ID, or a web store’s domain. It is matched as in section 6.1 and covers every territory on that channel identifier.
- A declaration MAY carry a product scope,
scope, with the same shape as an authorization’s:product_lines(named lines, or["all"]),products(catalog product IDs), or both (section 3). Without one, or withproduct_lines: ["all"], it covers every product. A scoped declaration covers a catalog product whose line is inproduct_linesor whose ID is inproducts, a request’sproduct_linewhenproduct_linesnames it, and, seller first, any request that names no product or product line. A product outside the scope, or not in the catalog, is answered as if the declaration didn’t exist, with the signalunauthorized_for_other_products(section 9.1). Exclusions don’t apply: a declaration names what it covers. - A declaration and an active authorization MUST NOT cover the same channel identifier, whatever their scopes. The registry MUST refuse a declaration for a channel identifier that an active authorization names, and MUST refuse an authorization for a declared one until the declaration is withdrawn.
- Declarations are not published. A declaration MUST NOT appear in a signed list, a list endpoint or the revocation feed, and the registry MUST NOT reveal one except in a verify answer about that channel identifier. This holds whether or not the brand’s list is public.
- Only the brand makes or withdraws a declaration, under the same rules as an authorization: the registry MUST NOT add, remove or change one on its own, except to remove one it reasonably believes is unlawful or abusive, as its terms of use reserve, and every change is in the brand’s change log (section 3). A withdrawn declaration leaves the seller
unlisted. - A seller can dispute a declaration (section 11).
The registry records and signs the brand’s declaration. It does not investigate the seller.
4. Brand identity and domain verification
A brand is identified by its web domain. The registry MUST verify control of a domain before accepting any authorization for it.
The registry issues a verification token for each domain. The brand proves control with a DNS TXT record, authorizedretailers-verify=<token>, on the brand’s domain. This is the only method. It works on every site platform, because every domain’s DNS can hold a TXT record. A file on the brand’s site, including the pointer (section 5.1), never proves control of the domain.
The token is public. It proves control only of the domain it was issued for, found on that domain, and it grants no access to the brand’s account.
Domain control does not prove brand ownership, because a counterfeiter can register a lookalike domain and verify it. The registry MUST therefore review each brand before its list counts, and MAY decline or end a registration. The registry SHOULD publish how it reviews.
The registry SHOULD recheck each domain at least weekly. If the DNS record is not found, the registry MUST treat the domain as verification_lapsed and notify the brand. It MUST NOT keep returning authorized on a lapsed domain.
A brand is registered for a domain when it has an account on the registry, the domain is verified by its DNS record and the registry has approved the brand. For any domain that is not registered, verify answers are brand_not_registered with a reason (section 9):
reason | Meaning |
|---|---|
not_registered | The registry holds no account for this domain. A file on the brand’s domain alone doesn’t count |
domain_unlinked | Domain control is verified, but the registry has not approved the brand yet |
verification_lapsed | The DNS record is not in place any more |
4.1 Several domains per brand
A brand account MAY hold several domains (for example brand.com, brand.co.uk and brandstore.com). Each is verified and linked separately. For each domain it adds, the brand chooses to share an existing domain’s list (one list, kept in sync), copy it (a separate list that starts identical and is then edited on its own) or start a new list.
A domain is identity, not territory: a brand whose domain is brand.co.uk can authorize US channels.
Verification never carries over from one domain to another. Each domain a brand adds proves control with its own DNS record (section 4), and the registry MUST NOT treat control of one domain, or a relationship between two domains or companies, as proof of control of a third. A parent company that owns several brands gets no shortcut: brands under one parent are often run independently, so each brand verifies its own domains and keeps its own list.
4.2 Brand names and aliases
A brand’s name, and any aliases it adds (such as a former name or a common short form), are for display and search. A list MAY carry them as brand.name and brand.aliases. A name or alias MUST NOT resolve a brand for a verify answer: names are not unique and anyone can type one. The same holds for model numbers and style codes: they confirm a product found by its product identifier, and never find a brand or product alone. Only a verified domain or a product in a registered brand’s catalog finds a brand (section 9.5).
5. The pointer file and the signed list
5.1 The pointer (optional)
A brand MAY serve a pointer at https://<brand-domain>/.well-known/authorized-retailers.json, as JSON over HTTPS. The pointer is optional and is offered only once the domain is verified by DNS (section 4). It helps software that finds brands by domain. It does not verify the domain, and a brand’s list counts without it.
{
"spec": "authorized-retailers/0.8",
"form": "pointer",
"brand": { "name": "Example Brand", "domain": "brand.example" },
"verify_token": "<token from the registry>",
"list": "https://authorizedretailers.ai/brands/brand.example/list"
}
specMAY beauthorized-retailers/0.2,/0.3,/0.4,/0.5,/0.6,/0.7or/0.8. The pointer format has not changed since v0.2, so a pointer published then stays current.formispointer. It is the only form. The field stays so that later versions can add forms. Readers MUST treat a file with any otherform, including the v0.1fullandprivateforms, as no file.brand.domainMUST be the domain the pointer was fetched from. A pointer vouches only for its own domain.listMUST be anhttps://authorizedretailers.ai/URL. A pointer naming any other host is treated as no file. The registry issues the unversioned path shown above; a pointer that names the list under/v0/(every pointer issued before 2026-10-22) stays current, since the registry serves both (section 9).verify_tokenis the domain’s token from the registry (section 4). It lets the registry tell whether a pointer is current. It is not proof of control.
The pointer carries no authorization data, so it does not change when the brand’s list changes, and it is not signed. It needs replacing only when its own fields change: a new verification token, domain or spec version. The registry SHOULD tell a brand when a pointer it finds on the brand’s domain is out of date. The list it names is signed, and so is every verify answer.
Readers MAY follow up to 3 HTTPS redirects when fetching the pointer (for example from the bare domain to www), but brand.domain MUST still match the domain the reader first asked for, ignoring a leading www.. Readers SHOULD accept the file when it is served as application/json or text/plain.
authorized-retailers.json is being registered in the IANA Well-Known URIs registry. This section will link to the entry once it is registered.
5.2 The signed list
The registry publishes each brand’s list at https://authorizedretailers.ai/brands/<domain>/list (and, for existing readers, under /v0/, section 9). Brands never build or sign their own list: they make authorization decisions on the registry, and the registry builds and signs the list.
{
"spec": "authorized-retailers/0.8",
"brand": { "name": "Example Brand", "domains": ["brand.example", "brand.example.co.uk"], "aliases": ["EB"] },
"public": true,
"issued": "2026-09-28T14:00:00Z",
"valid_until": "2026-09-28T14:15:00Z",
"stale_until": "2026-09-29T14:00:00Z",
"catalog": [
{ "id": "EB-TEE-BLK-M", "name": "Classic Tee, Black, M", "product_line": "Classic Tee", "gtin": "00012345678905", "asins": ["B0EXAMPLE1"] }
],
"exclusions": { "product_lines": [], "products": [] },
"authorizations": [
{
"id": "auth_01J9X2",
"revocation_id": "rv_Q2l8dX4bT0aN5kPw",
"retailer": { "name": "Example Retail LLC", "entity_id": "ent_7Q4M" },
"channels": [
{ "type": "amazon", "domain": "amazon.com", "seller_id": "A1B2C3D4E5F6G7" },
{ "type": "walmart", "domain": "walmart.com", "seller_id": "101234567" },
{ "type": "web", "domain": "retailer.example" }
],
"scope": { "territories": ["US"], "product_lines": ["all"] }
}
],
"signature": { "kid": "ar-20260928-6e37", "jws": "<detached JWS>" }
}
Earlier versions carried a proposed_by field on each authorization. It is removed in v0.5: a v0.5 list MUST NOT include it, and readers MUST ignore it when it appears in a list published under an earlier version.
brand.domainsnames every domain the list applies to, so an agent can see that two domains share one list.brand.aliasesis optional (section 4.2).issued,valid_untilandstale_untilgovern freshness of this signed copy, not the authorizations (section 10).catalogandexclusionsare optional.revocation_idis a random value that changes whenever the authorization’s coverage shrinks (section 10.3).
A list counts only if its signature verifies against the registry’s key set and brand.domains includes the domain the reader asked about. Readers MUST treat an unsigned list, or one whose signature does not verify, as no list.
Whether a brand’s list is public is a setting on the registry. The catalog follows the same setting. For a brand whose list is not public, the list URL returns a signed response with "public": false and no catalog or authorizations:
{
"spec": "authorized-retailers/0.8",
"brand": { "name": "Example Brand", "domains": ["brand.example"] },
"public": false,
"issued": "2026-09-28T14:00:00Z",
"valid_until": "2026-09-28T14:15:00Z",
"stale_until": "2026-09-29T14:00:00Z",
"signature": { "kid": "ar-20260928-6e37", "jws": "<detached JWS>" }
}
Agents then use the verify endpoint, which answers one question at a time (section 9). A domain that is not registered returns HTTP 404 from the list URL.
6. Seller identity and channel identifiers
Agents see sellers as channel identifiers, not company names, so every authorization MUST name at least one channel identifier. A name alone is not verifiable.
| Channel type | Required fields | Identifier |
|---|---|---|
amazon | domain, seller_id | Amazon merchant ID on that marketplace domain |
walmart | domain, seller_id | Walmart Marketplace partner ID |
ebay | domain, seller_id | eBay user ID |
web | domain | The seller’s store domain |
Every domain is lowercase, with no scheme, port, path, trailing dot or leading www.. For amazon, walmart and ebay, domain MUST be one of that type’s marketplace domains in the table below. A channel whose type and domain don’t match (for example "type": "amazon" with walmart.com) is invalid, and a web channel MUST NOT use a marketplace domain. The v0.1 marketplace field and physical channel type are removed, and readers MUST reject them.
| Marketplace domain | Type | Country |
|---|---|---|
| amazon.com | amazon | US |
| amazon.ca | amazon | CA |
| amazon.co.uk | amazon | GB |
| amazon.de | amazon | DE |
| amazon.fr | amazon | FR |
| amazon.it | amazon | IT |
| amazon.es | amazon | ES |
| amazon.com.au | amazon | AU |
| walmart.com | walmart | US |
| walmart.ca | walmart | CA |
| ebay.com | ebay | US |
| ebay.co.uk | ebay | GB |
| ebay.ca | ebay | CA |
| ebay.de | ebay | DE |
| ebay.com.au | ebay | AU |
New channel types and marketplace domains are added by the registry and published in the changelog.
6.1 Matching identifiers
Agents MUST match on the channel identifier they see, never on the seller name. Two identifiers are the same when their channel keys are equal:
| Channel type | Channel key |
|---|---|
amazon | amazon:<domain>:<seller_id>, seller ID trimmed and uppercased |
walmart | walmart:<domain>:<seller_id>, seller ID trimmed, case kept |
ebay | ebay:<domain>:<seller_id>, seller ID trimmed and lowercased |
web | web:<domain> |
Domains in keys are normalized as above. A seller ID is 1 to 128 characters with no whitespace.
The registry MAY group a retailer’s identifiers under one entity_id, so one retailer selling on several channels is one record. Registry-proposed links between identifiers MUST be confirmed by the brand before they count as authorized.
6.2 Territory
Each channel’s territory comes from its domain:
- Marketplace domain: the country in the table above. It cannot be changed. If an authorization lists
scope.territoriesand they leave out a marketplace channel’s country, the authorization is invalid (territory_mismatch). - Web store on a country-code domain (
.ca,.co.uk,.deand so on): that country by default. An authorization MAY list other territories. The validator warns (territory_differs_from_domain) when they leave out the domain’s country. - Web store on a generic domain (
.com,.net,.shopand so on): no default. The authorization MUST listscope.territories, which MAY be["*"]for worldwide.
In a verify request, the territory is where the sale happens. A marketplace’s domain gives it, so the agent doesn’t need to. For a web store on a generic domain, the agent sends the buyer’s country, which it knows from the order. The registry never infers a territory from the seller’s authorization: a store authorized only for the US would then look authorized to a buyer anywhere.
A channel’s effective territories are its domain’s country for a marketplace, and scope.territories when given, otherwise the domain’s country, for a web store.
A country-code domain is one whose last label is a two-letter ISO 3166-1 code, with uk read as GB. These two-letter endings are widely used as generic domains and are treated as generic: ac, ai, cc, co, fm, gg, io, la, ly, me, sh, so, to, tv, vc and ws. Adding to this list is a registry change published in the changelog.
Country codes are ISO 3166-1 alpha-2: the United Kingdom is GB, not UK.
6.3 Proving channel identifiers
The registry MAY let sellers prove their own channel identifiers before they request authorization: a web domain by an account signed in with an email address at that domain, without publishing anything on it, and a marketplace seller ID by confirming with the marketplace that it exists on that domain. Confirming that a seller ID exists does not prove that the seller controls it, and the registry MUST NOT present it as proof of control.
The registry MAY decline to register a marketplace seller that does not publicly list a seller name, business name and address on its marketplace.
7. Evidence and observation
Every answer about a seller rests on the same evidence: a verified domain whose brand the registry has approved (section 4), the brand’s own approval (section 3), and the registry’s signature (section 8). There are no weaker grades of answer. Where that evidence is missing, the answer is brand_not_registered, whether or not a file exists on the brand’s domain.
The registry MAY also observe sellers: evidence from outside the brand’s list that an authorized retailer is actively selling the brand’s products on the stated channel. Every answer MUST carry observed. On an authorized answer it is { "last_seen": <timestamp> } when the registry has seen this seller offering one of the brand’s products on the stated channel within the last 30 days, and otherwise null. On every other status it MUST be null. Observation never changes the status.
What counts as a sighting depends on the channel type. On a marketplace (amazon, walmart, ebay) it is a current offer from the seller ID on a product the brand lists, or one the marketplace shows under the brand’s name. On a web channel it is the brand’s product offered for sale at the stated domain. A registry MAY observe some channel types and not others, and MUST say which ones, and for which territories, in its documentation of the verify API.
observed: null means only that the registry has no sighting. It is never evidence that a seller has stopped selling or is not authorized, because the registry may not observe that channel at all. Agents MUST NOT treat null as a negative signal, and MUST NOT withhold or downgrade an authorized answer because observed is null.
Observation concerns authorized retailers. A registry MAY also search for sellers carrying a brand’s products that the brand has not listed, and show them to the brand as leads. That is outside this spec: a lead is not a status, never appears in an answer, a list or the revocation feed, and a seller stays unlisted until the brand authorizes or declares it (sections 3 and 3.2).
Observation can also raise flags on an authorized retailer: sudden changes in account name, address, catalog or volume that suggest a compromised account. The registry MAY return these as signals alongside the answer, without changing the authorization status.
8. Signing and key management
Lists, verify answers and the revocation feed are signed with a detached JWS (RFC 7515) over the JSON canonicalized per RFC 8785, using Ed25519 (EdDSA). Anyone can check a signature without trusting the registry’s servers.
- Public keys: published at
https://authorizedretailers.ai/.well-known/jwks.json. Every signature names the key that made it (kid). - Agent behavior: agents SHOULD cache the key set, for no longer than the
Cache-Controlon the key set response allows (the registry serves it with a 5-minute lifetime, so a removed key is out of every cache that honors it within 5 minutes), and refetch it when they see an unknownkid. They MUST reject signatures from keys no longer in the set. Mirrors serve a copy of the key set, but an agent MUST NOT trust a key it has not also fetched from authorizedretailers.ai. An agent that passes an answer on without checking its signature SHOULD say that the answer is unverified. - Scheduled rotation: every 90 days. The new key is published at least 7 days before first use. The old key stays published until everything it signed is past its
stale_until(section 10). - Emergency rotation: a compromised key is removed from the set immediately, and every current list is re-signed with a new key.
- Storage: the registry SHOULD hold private keys in a managed key service, sign server-side, and never expose keys to application code or staff.
Brands do nothing during rotation.
9. Verify API and MCP interface
An agent asks one question: is this seller authorized for this brand, on this channel, in this territory, for this product? The agent MAY name the brand by its domain, or let the registry find it from the product (section 9.5). The answer is signed and is the same whether the brand’s list is public or not.
Request: POST /verify
The verify API’s paths carry no version: a request without one gets the current specification. A registry MUST also serve each of them under /v0/ (/v0/verify, /v0/resolve, /v0/revocations, /v0/brands/<domain>/list) for clients that already use that form, with the same answers.
A registry MAY require an API key on this endpoint, on POST /resolve and on the MCP server’s tools (section 9.6). authorizedretailers.ai does. The key set, public lists and the revocation feed need no key.
{
"channel": { "type": "amazon", "domain": "amazon.com", "seller_id": "A1B2C3D4E5F6G7" },
"product": { "asin": "B0EXAMPLE1" }
}
The same request with the brand’s domain, a territory and a date:
{
"brand_domain": "brand.example",
"channel": { "type": "amazon", "domain": "amazon.com", "seller_id": "A1B2C3D4E5F6G7" },
"territory": "US",
"product": { "gtin": "012345678905", "asin": "B0EXAMPLE1" },
"as_of": "2026-09-28"
}
| Field | Rules |
|---|---|
brand_domain | Optional when product is given. Any registered domain of the brand. The answer uses that domain’s list. |
channel | Required. A channel identifier (section 6). type is optional in a request: a marketplace domain implies its type (amazon.com is amazon), and any other domain is web. A type that is sent MUST agree with the domain. |
territory | Optional. Defaults to the channel domain’s country. Required for a web store on a generic domain. |
product_line | Optional. Asks about a whole product line. |
product | Optional when brand_domain is given. One or more of gtin, asin, walmart_item_id and wpid. A GTIN is a UPC-A (12 digits), EAN-13, GTIN-8 or GTIN-14 with a valid check digit, compared as GTIN-14. An ASIN is read on the channel’s Amazon marketplace, and a Walmart item ID or WPID on walmart.com. mpn and style_code MAY be added to a product identifier to confirm the match, never on their own. |
as_of | Optional. An ISO 8601 date or UTC date-time. The answer reflects status at that time. Defaults to now. |
price | Optional, and read only when product is given. The price the agent is looking at for this seller and product: { "amount": "34.50", "currency": "USD" }, amount a decimal string, currency an ISO 4217 code. It never changes the answer (below). |
A request MUST name a brand_domain, a product or both. A request that finds the brand by ASIN alone MUST use an amazon channel, and one that finds it by Walmart item ID or WPID alone MUST use a walmart channel on walmart.com. The registry MUST reject a GTIN with a bad check digit rather than resolve it.
price is for the brand’s eyes, in aggregate, not for the answer. The registry MUST NOT use it to decide status or anything else in the answer, and MUST NOT reveal a brand’s minimum advertised price (MAP) or whether a price is below it: MAP is a brand’s own policy, not an authorization fact, and an answer that depended on a sent price would invite spoofing. The registry MAY keep the price with the query log (section 12) and show the brand how often agents named a seller and product at a price below its MAP, and which agents did (section 9.6). An agent gains nothing from sending it, so a brand reading those numbers must know they cover only the agents that do.
Response:
{
"status": "authorized",
"brand": { "domain": "brand.example", "name": "Example Brand" },
"resolved_by": "asin",
"authorization_id": "auth_01J9X2",
"revocation_id": "rv_Q2l8dX4bT0aN5kPw",
"checked": "2026-09-28T14:02:11Z",
"valid_until": "2026-09-28T14:17:11Z",
"stale_until": "2026-09-29T14:02:11Z",
"product": { "in_catalog": true, "id": "EB-TEE-BLK-M", "product_line": "Classic Tee" },
"observed": null,
"signals": [],
"signature": { "kid": "ar-20260928-6e37", "jws": "<detached JWS>" }
}
status | Meaning |
|---|---|
authorized | An authorization covers this seller, channel, territory and product |
unlisted | The brand is registered and no authorization covers this request |
unauthorized | The brand has declared this seller unauthorized (section 3.2) |
expired | An authorization with an end date covered this seller, and the end date has passed |
brand_not_registered | The brand domain is not registered (section 4) |
disputed | A dispute is open on this authorization, and the brand has flagged it for review, or on a declaration the brand hasn’t yet reconfirmed (section 11) |
product_not_found | The request gave no brand_domain, and no registered brand’s catalog contains the product (section 9.5) |
brand_ambiguous | The request gave no brand_domain, and more than one registered brand claims the product (section 9.5) |
brand_mismatch | The product is not in the catalog of the brand named by brand_domain, and other registered brands claim it (section 9.5) |
brandandresolved_byappear on every answer exceptproduct_not_foundandbrand_ambiguous.brandhas the brand’sdomainand, when registered, itsname.resolved_byisbrand_domain,gtin,asin,walmart_item_idorwpid: how the registry found the brand.candidatesappears only onbrand_ambiguousandbrand_mismatch: the registered brands that claim the product, each as{ "domain", "name" }.authorization_idappears onauthorized,expiredand adisputedauthorization.expiresappears with it when the authorization has an end date.declaration_idanddeclared(when the brand made the declaration) appear onunauthorizedand adisputeddeclaration, and nowhere else. Adisputedanswer carries eitherauthorization_idordeclaration_id, never both.revocation_idappears only onauthorized(section 10.3).as_ofappears when the request gave one.- Every signed answer MUST carry
valid_until, no later than 15 minutes afterchecked, andstale_until, no later than 24 hours afterchecked(section 10). productappears when the request gave one. It always hasin_catalog. For a product in the catalog it also has:status(when notactive),status_since,recall_url(when recalled) andregions, for every brand;confirmedandmismatched: which of a sentmpnandstyle_codematch the catalog. A mismatch adds the signalproduct_details_mismatch, for example an ASIN whose listing shows another model, and never changes the status;id,product_line,name,mpnsandstyle_code, only for a brand whose list is public.
- When
product.statusisrecalled, an agent MUST tell its user and SHOULD NOT complete the purchase, whatever the seller’s status.discontinuedis information for the user.
New statuses can be added, as unauthorized was in v0.4. An agent MUST NOT treat a status it doesn’t recognise as authorized, and SHOULD treat the seller as unconfirmed, as for an unreachable registry (section 10.2).
reason is optional on brand_not_registered (section 4) and unlisted, and MUST NOT appear on any other status:
unlisted reason | Meaning |
|---|---|
territory_mismatch | The seller is authorized on this marketplace domain, but the requested territory is not that domain’s country |
product_excluded | The product is in the catalog and excluded, by the authorization or by the brand |
product_not_in_scope | The product is in the catalog but the authorization doesn’t include it |
product_not_in_catalog | The product is not in the catalog and the authorization is limited to named lines or products |
product_not_for_region | The product is in the catalog, and its regions don’t include the territory (section 3.1) |
signals is an array on every answer, empty when there is nothing to say. Each entry has a type, when it was observed, and an optional detail. A signal never changes the status. A registry MAY add its own; an agent MUST ignore a type it doesn’t recognise. This specification defines:
signals[].type | On | Meaning |
|---|---|---|
product_not_in_catalog | authorized | The product isn’t in the catalog, and the authorization covers all products (section 9.2) |
product_details_mismatch | any answer with a product | A sent mpn or style_code doesn’t match the catalog (section 3.1) |
authorization_scoped | authorized | The request named no product, and the authorization is limited to named lines or products or has exclusions: ask again with the product (section 9.1) |
unauthorized_for_other_products | unlisted | The brand has declared this seller unauthorized for other products; the one asked about is outside that declaration’s scope (section 3.2) |
9.1 Seller first
Authorization is decided for the seller first: channel, seller ID, territory and date. Only authorizations that cover the seller are then checked against a product line or product. A product never makes an uncovered seller authorized. In particular, an unauthorized seller that sets up a duplicate listing under a new ASIN stays unlisted.
A declaration is decided at the same step. Once the brand is found, a declaration covering the seller answers unauthorized (or disputed, section 11) whatever the territory asked. A declaration without a product scope covers every product and product line. One with a scope (section 3.2) answers unauthorized to a request that names no product or product line (the question is about the seller, and the brand has declared it), to one naming a catalog product inside the scope, and to one naming a product_line it lists. A request naming a product outside the scope, or one not in the catalog, is answered as if there were no declaration, from the brand’s authorizations, which for a declared seller means unlisted; that answer carries the signal unauthorized_for_other_products, so the agent knows the brand has declared this seller unauthorized for other products. With as_of, a declaration counts if the brand made it by then and hadn’t withdrawn it by then, with its current scope: the registry keeps no history of scope edits.
The same holds the other way for an authorization limited to named lines or products, or with exclusions: a request that names no product is answered authorized for the seller, with the signal authorization_scoped. An agent that gets it SHOULD ask again with the product before relying on the answer for a purchase.
9.2 Matching a product
When a request names a product, the registry looks it up in the catalog by GTIN-14, ASIN, Walmart item ID or WPID, and for the authorizations that cover the seller:
- Product in the catalog, and the territory is outside its
regions:unlisted, reasonproduct_not_for_region. - Product in the catalog and covered by an authorization:
authorized. - Product in the catalog, covered by none of them, and excluded by the brand or by at least one of them:
unlisted, reasonproduct_excluded. - Product in the catalog and not included by any of them:
unlisted, reasonproduct_not_in_scope. - Product not in the catalog, and every covering authorization is limited to named lines or products:
unlisted, reasonproduct_not_in_catalog. - Product not in the catalog, and a covering authorization has
product_lines: ["all"]:authorized, with the signalproduct_not_in_catalog.
A product that isn’t in the catalog may be a duplicate listing, which is why every answer about a product says whether it is in_catalog.
A request with product_line is covered when an authorization’s product_lines is ["all"] or includes that line, and the line is not excluded.
9.3 Lists that are not public
For a brand whose list is not public, the verify API answers only the question asked. It MUST NOT return the brand’s other retailers or catalog, beyond the asked product’s status fields and whether sent details match (section 9). Status is safety information, so it shows whether or not the list is public. The API SHOULD rate-limit queries per caller (per key, where a registry requires keys, section 9.6) so the list can’t be reconstructed by enumeration.
9.4 MCP
The same capability is exposed as MCP tools: verify_seller (the request above, with or without brand_domain), resolve_brand (a product in, the brand or the candidates out, signed as in section 9.5), get_brand (registration status, whether the list is public, last update) and list_authorized_retailers (for brands whose list is public). A registry that requires keys (section 9.6) takes the key as the same Authorization: Bearer header on the MCP transport, and MAY serve initialize and the tool list without one.
9.5 Finding the brand from a product
When a request gives no brand_domain, the registry finds the brand from the product:
- Only registered brands take part (section 4): the domain is verified and the registry has approved the brand. A catalog on a domain that is not registered MUST NOT resolve, so a lookalike brand can’t take over a product by uploading someone else’s barcodes.
- The registry looks up every identifier given (the GTIN as GTIN-14, the ASIN, the Walmart item ID and the WPID) in every registered brand’s catalog. Domains that share one list count as one brand.
- No brand claims the product:
product_not_found. - Exactly one brand claims it, by every identifier given: the registry answers from that brand’s list, as it would for its
brand_domain, withresolved_byset to the first identifier given in this order:gtin,asin,walmart_item_id,wpid. - More than one brand claims it, or two of the identifiers given point to different brands:
brand_ambiguouswith thecandidates. The registry MUST NOT answerauthorizedand SHOULD flag the product for review.
When a request gives a brand_domain and a product, the brand comes from the domain (resolved_by is brand_domain). If the product is not in that brand’s catalog and other registered brands claim it, the registry MUST answer brand_mismatch with those brands as candidates, rather than trust either the domain or the catalog. If the product is in that brand’s catalog, the registry answers as usual, whether or not another brand also claims it.
Finding the brand reveals only which registered brand a product belongs to. It MUST NOT reveal a brand’s retailers, and for a brand whose list is not public it MUST NOT reveal anything from the catalog beyond in_catalog (section 9.3).
POST /resolve (and the MCP tool resolve_brand) finds the brand without asking about a seller:
{ "product": { "asin": "B0EXAMPLE1" }, "domain": "amazon.com" }
domain is optional: the marketplace the product was seen on. With an ASIN alone it MUST be an Amazon marketplace domain, and with a Walmart item ID or WPID alone it MUST be walmart.com. The signed answer has status (resolved, product_not_found or brand_ambiguous), brand and resolved_by (gtin, asin, walmart_item_id or wpid) when resolved, candidates when ambiguous, and checked, valid_until, stale_until and signature as in section 9.
{
"status": "resolved",
"brand": { "domain": "brand.example", "name": "Example Brand" },
"resolved_by": "asin",
"checked": "2026-09-29T14:02:11Z",
"valid_until": "2026-09-29T14:17:11Z",
"stale_until": "2026-09-30T14:02:11Z",
"signature": { "kid": "ar-20260928-6e37", "jws": "<detached JWS>" }
}
For now an ASIN resolves on every Amazon marketplace, as it does in a catalog (section 3.1).
9.6 API keys and identifying the agent
A registry MAY require an API key on POST /verify, POST /resolve and the MCP server’s tools. authorizedretailers.ai does. A key tells the registry which operator is asking, so it can rate-limit each operator on its own, contact it about the service and see which agents use it. It is not part of an answer: the registry signs every answer, and anyone with the registry’s public keys can check one.
- Sending a key. An agent sends its key as
Authorization: Bearer <key>. An agent MUST NOT send it in a URL, query string or request body, and a registry MUST NOT accept it there, so keys stay out of logs and referrers. - No key, or a bad one. A registry that requires keys answers a request with no key, a malformed key, an unknown key or a revoked key with HTTP 401 and a JSON body,
{ "error": "api_key_required" }or{ "error": "invalid_api_key" }. It doesn’t answer the question. The agent treats the seller as unconfirmed (section 10.2). A 401 means the agent’s setup is wrong, not that the registry is down. - Limits. Rate limits apply per key, and per key and brand (section 9.3). A request over a limit gets HTTP 429 with
Retry-After, and is unconfirmed too. - A key never changes an answer. Two requests that differ only in the key get the same answer. A registry MUST NOT give one operator a different status, signal or product detail from another’s, and MAY give different rate limits.
- Keys. A registry SHOULD give its keys a recognisable prefix, so a leaked key can be found by secret scanning, MUST store only a hash of a key, and shows a key to the operator once. An operator can hold several keys and revoke any of them, and a revoked key stops working within a minute. How an operator gets a key, and what the registry asks of it, is the registry’s own process and terms, not this specification.
- No key needed. The key set (section 8), public lists (section 5) and the revocation feed (section 10.3) are public, signed data. They need no key, from the registry or from a mirror.
Agents SHOULD send a User-Agent header that names the agent and its operator and gives a way to reach them, for example ExampleShopper/2.1 (+https://example.com/agent; agents@example.com). The registry MAY record it, with the name of the organization that owns the caller’s network, to learn which agents use it and to contact their operators about the service (section 12). The header never changes an answer: an agent that sends a generic one gets the same signed answers.
9.7 With agentic checkout protocols (informative)
Protocols for agents that buy, such as the Universal Commerce Protocol (UCP) and the Agentic Commerce Protocol (ACP), describe how an agent finds a merchant’s products and checks out with it. They say nothing about whether a brand authorized that merchant, and a merchant can’t vouch for itself: only a signed registry answer counts (section 8). The check belongs between choosing a seller and starting checkout. The agent asks the question in section 9 about the seller the merchant’s channel identifies (its web domain, or its marketplace and seller ID), and starts checkout only for a seller it may describe as authorized, or one the shopper accepts as unconfirmed. Nothing here depends on those protocols, and they don’t depend on this one.
9.8 Buying from a seller the brand hasn’t authorized
An agent that buys on a person’s behalf has to decide what to do with an answer that isn’t authorized. This section says what conforms. It is about the agent’s conduct: the answers don’t change, a person’s decision never reaches the registry, and the registry still makes no finding about any seller (section 1).
unauthorized: the brand has declared the seller unauthorized (section 3.2). An agent MUST NOT complete a purchase from that seller on its own. It MAY complete one only after the person it acts for has confirmed that purchase, having been told that the brand has declared the seller unauthorized, that the goods may be materially different or counterfeit, that the brand’s warranty and support may not apply, and that the registry cannot say whether the goods are genuine. A standing instruction from the person to buy from such sellers, for example for enforcement test purchases, counts as that confirmation. The agent SHOULD keep the signed answer with the order’s records.unlistedandexpired: the brand is registered and no authorization covers this request, and nothing more is implied (section 1). An agent SHOULD prefer a seller the brand has authorized when one has the product, and SHOULD tell the person when it chooses an unlisted seller instead, for example because the authorized ones are out of stock, and what that means. No confirmation is required.- Every other answer, a status the agent doesn’t recognise, or no usable answer: the seller is unconfirmed (sections 9 and 10.2). The agent SHOULD say so when it describes the seller. No confirmation is required.
The conformance decision is unchanged: every status but authorized is still “not authorized”. This section adds what an agent does next.
10. Expiry, revocation and freshness
An authorization lasts until the brand revokes it, unless the brand gives it an end date (expires). There is no default end date and no maximum.
- Reminders: for an authorization with an end date, the registry reminds the brand 30 and 7 days before it. One confirmation renews every authorization the brand selects.
- Revocation: a brand can revoke an authorization at any time. The verify API MUST reflect a revocation within 1 minute.
- Retailer notice: the registry SHOULD notify a retailer when its authorization is revoked or reaches its end date, if the retailer has registered.
10.1 Freshness
Every signed list and answer carries valid_until, at most 15 minutes after it was issued, and stale_until, at most 24 hours after it was issued. These describe the signed copy, not the authorization. An agent that shows valid_until to a person SHOULD present it as when the answer needs re-checking, not as when the authorization ends: an authorization lasts until the brand revokes it or until its own expires, when it has one. The registry re-signs each list at least every 15 minutes, and immediately on any change. A revoked seller can therefore look authorized in a cached copy for about 16 minutes at most.
Agents SHOULD decide from the verify endpoint and MUST NOT rely on a cached list or answer past its valid_until, except as section 10.2 allows.
An answer that carries expires stops being usable as authorized at that instant, even if its valid_until is later: an agent MUST NOT treat a seller as authorized after the expires on the answer, and the registry SHOULD NOT set valid_until later than expires. Reaching an end date is not a revocation, so it does not appear in the revocation feed (section 10.3).
10.2 When the registry is unreachable
A registry outage never makes a seller unauthorized and never blocks a purchase on its own.
- Fail to unconfirmed, never to authorized. An agent that can’t get a current answer, or a stale one allowed below, treats the seller as unconfirmed. It MAY carry on as it would without this standard, but MUST NOT describe the seller as authorized. A 401 or a 429 (section 9.6) means the registry is reachable, so the stale-answer rule below doesn’t apply: the agent fixes its key or waits out the limit, and treats the seller as unconfirmed meanwhile.
- Stale answers. Between
valid_untilandstale_until, an agent MAY use a list or answer only if the verify API is unreachable, and only with a revocation feed (section 10.3) that was issued at or after the answer’schecked(or the list’sissued) and is itself still usable, up to its ownstale_until. A feed issued before the answer says nothing about revocations since. The feed comes from a mirror, or from the registry if only its verify API is down; with no such feed the answer is unusable. The agent MUST mark the result as stale to its user or in its logs. Afterstale_untilthe list or answer is unusable. This follows the pattern of HTTPstale-if-error(RFC 5861). - Mirrors. Mirrors run on a different provider and DNS from the registry. They serve the key set, every public list and the revocation feed, all signed by the registry. They never answer verify requests. Reading a mirror needs no account or key, because everything on it is already public and signed; the operator MAY rate-limit at the provider to control cost. The registry’s first mirror will run on Google Cloud. Mirror base URLs are listed here when mirrors launch. There are none yet.
- Order. An agent tries the live verify API, then a mirror (a public list, or a stale answer checked against the revocation feed), then treats the seller as unconfirmed.
10.3 Revocation feed
Every authorization carries a revocation_id: a random value that reveals nothing about the brand or seller. It appears in the list and on authorized answers only. When an authorization stops covering something it covered before (it is revoked, a channel is removed, its territories or products shrink, or its end date moves earlier), the registry publishes its revocation_id in the revocation feed and gives the authorization a new one.
The feed is a signed, append-only record of the last 24 hours, published every minute at https://authorizedretailers.ai/revocations (also under /v0/, section 9) and on every mirror:
{
"spec": "authorized-retailers/0.8",
"feed": "revocations",
"issued": "2026-09-28T14:02:00Z",
"valid_until": "2026-09-28T14:17:00Z",
"stale_until": "2026-09-29T14:02:00Z",
"entries": [{ "revocation_id": "rv_Q2l8dX4bT0aN5kPw", "revoked": "2026-09-28T13:40:12Z" }],
"signature": { "kid": "ar-20260928-6e37", "jws": "<detached JWS>" }
}
An agent holding a stale answer or list MUST check it against a revocation feed issued at or after the answer’s checked (or the list’s issued), which MUST itself be before its own stale_until. It MUST drop any authorization whose revocation_id is in the feed. Only someone who was told a seller is authorized holds its revocation_id, so the feed can be public without revealing who was revoked.
11. Disputes and unclaimed listings
Disputes. A retailer can see its own status for any brand and can open a dispute if it believes a listing is wrong, for example a revocation it says was issued in error or an identifier mapped to the wrong entity.
- The retailer submits the dispute with evidence.
- The brand is notified and has 30 days to confirm, correct or reject.
- While a dispute is open, verify answers return
disputedonly if the brand has flagged the authorization for review. Otherwise the brand’s current decision stands. - The registry does not rule on commercial relationships. The brand’s decision is final for authorization status. Identifier errors are corrected by the registry.
- If the brand does not answer within 30 days, the dispute closes and the brand’s current decision stands. The person who opened the dispute is told how it closed.
Disputing a declaration. A seller declared unauthorized (section 3.2) can dispute the declaration. Marketplace sellers can’t prove control of a seller ID (section 6.3), so the registry MUST NOT require that proof: the dispute gives a contact email address and evidence. For a web store, the registry MAY require an address at the store’s domain, as it does to confirm retailer web stores (section 6).
- While the dispute is open, verify answers about the seller are
disputed, with thedeclaration_id. - The brand is notified and MUST either reconfirm the declaration, which answers
unauthorizedagain, or withdraw it, which answersunlisted. - If the brand does neither within 30 days, the dispute closes and the answer goes back to
unauthorized. The brand’s declaration stands until the brand withdraws it. - A declaration has at most one open dispute. The registry MAY limit how often disputes on the same declaration can be opened.
- The person who opened the dispute is told how it closed.
The registry’s liability position and data handling are set out in its terms of use, not in this spec. The dispute timelines above are part of the standard, and the terms of use repeat them.
Unclaimed listings. The registry MAY index authorized-retailer or dealer-locator pages that a brand already publishes, as unclaimed entries. These MUST be labeled as unverified, MUST NOT verify as authorized, and MUST link to the source page. A brand claims its listing by registering (section 4), after which the entries become editable and signable.
12. Security and privacy considerations
- Brand impersonation: mitigated by domain verification plus the registry’s review of each brand (section 4). This is the highest-impact attack, because a false brand could authorize diverters.
- Account takeover at the registry: brand accounts MUST use multi-factor authentication. Revocations and new authorizations SHOULD trigger an email to every brand admin.
- Registry changes: the registry changes a brand’s list only on the brand’s instruction, except to remove an entry it reasonably believes is unlawful or abusive, and every change, with the reason for any such removal, is in a log the brand can see (section 3).
- Compromised retailer accounts: an authorized retailer’s marketplace account can be taken over. Observation signals (section 7) help, but authorization does not guarantee the retailer’s current conduct.
- Declarations: a brand could declare a seller it did in fact authorize, or one it has no relationship with. The answer reports the brand’s declaration, not a finding by the registry (section 1), the seller can dispute it (section 11), and declarations are never published as a list (section 3.2), so a declaration can’t be read as a public blocklist.
- Dispute abuse: anyone can dispute a declaration, which makes the answer
disputedfor up to 30 days. The brand can end it at once by reconfirming, and the registry can limit repeat disputes (section 11). - Mailbox addresses: a seller whose listed address is a PO box or a commercial mailbox can be hard to trace. The registry MAY warn a brand before it authorizes such a seller. The warning is for the brand only and MUST NOT change verify answers.
- Product takeover: a lookalike brand could upload another brand’s barcodes to capture product lookups. Only registered brands take part in finding a brand, and a product claimed by two brands is never authorized (section 9.5). The registry SHOULD warn a brand when it adds a product another brand already claims.
- List enumeration: lists that are not public are protected by answering single questions only, with per-caller rate limits and query logging.
- Query logging: the registry logs each verify and resolve query to rate-limit, to detect enumeration and to handle abuse. The log MUST NOT hold a caller’s IP address. It MAY hold a keyed hash of the address, the ID of the API key the request used (never the key’s secret), the name of the organization that owns the caller’s network, and the User-Agent the client sent, so the operator can tell which platforms and agents use the registry, and the
pricea verify request carried (section 9), so a brand can see how often agents named a seller and product below its minimum advertised price. The registry MUST NOT tell a brand which caller asked about it, whether the caller is known by an address, a key or a name; a brand may see how often it was checked and which sellers were asked about. The registry SHOULD state in its privacy policy what its query log holds, who can see it and for how long. - API keys: a leaked key lets someone use an operator’s rate limits and appear as that operator in the registry’s logs. It can’t forge an answer, because the registry signs every answer. Keys are shown once, stored only as a hash, carry a recognisable prefix so they can be found by secret scanning, and can be revoked at once; an operator whose key leaks revokes it and makes another (section 9.6).
- Revocation feed: entries are random
revocation_idvalues, not hashes of brand and seller, so nobody can test guesses against the feed. - Personal data: channel identifiers and business names are business data. The registry MUST NOT publish personal addresses of sole-trader retailers.
- Request evidence: an invoice a seller attaches to a request (section 3) names the seller’s supplier and its prices. The registry MUST show it only to the brand asked and to the registry’s own operators, MUST NOT publish it or any part of it, and SHOULD state in its privacy policy how long it keeps it.
- Replay: agents SHOULD check the
checkedtimestamp on a signed answer and MUST reject answers past theirvalid_until, except as section 10.2 allows.
13. Open questions for v0.9
- Territory splits where the trademark has different owners by country: one brand record or several?
- Multi-seller platforms and channel audience: sellers that share one platform domain (such as practitioner dispensaries), and whether a channel needs an
audiencefield. Deferred. They stay out of scope (section 1) and are assessed case by case, when a brand needs one, and only if compatible with the format. The likely shape is a platform channel type with a seller handle, like a marketplace’s seller ID. - Cross-border purchases: buying on one country’s marketplace for delivery to another.
- Per-marketplace ASINs: products whose ASIN differs between Amazon marketplaces.
- GS1 company prefix: finding a brand without a catalog from the company prefix in its GTINs, when the registry can confirm the brand owns that prefix.
14. Changelog
- 2026-09-23
- Verify answers carry
valid_until(at most 24 hours afterchecked). Key retirement is based on it (sections 8, 9, 10, 12). brand_unverifiedanswers MAY carry areason(not_registered,domain_unlinked,verification_lapsed) (section 9).- List endpoints respond to private-form brands exactly as to unknown brands (section 9).
- Verify answers carry
- 2026-09-24
- Key storage in a managed key service is a SHOULD for every registry, replacing the description of one registry’s setup (section 8).
- Evidence tiers are removed. Every answer about a seller needs a verified, independently linked brand domain, the brand’s approval and a registry signature; without them the answer is
brand_unverified(sections 5, 7, 9). - Only registry-signed files count. An unsigned file, or one whose signature doesn’t verify, is treated as no file, and a file vouches only for the domain it’s fetched from (section 5).
- Observation is a separate
observedfield ({ "last_seen" }ornull) on every answer, instead of a tier (sections 7, 9).
- 2026-09-25
- Distributors: a registry SHOULD accept proposals only from distributors the brand has designated and that have proven control of their domain. Brands MAY limit what a distributor can propose (sections 2, 3).
- A registry MAY hold a request to a brand that isn’t registered yet and deliver it once the brand verifies (section 3).
- Retailer-side verification of channel identifiers, an open question, is now a MAY. Confirming a marketplace seller ID exists MUST NOT be presented as proof of control (sections 6, 13).
- A registry MAY decline marketplace sellers that don’t list a seller name, business name and address (section 6).
- A registry MAY warn brands about sellers at PO boxes or commercial mailboxes. Verify answers don’t change (section 12).
- Corrected: the reference registry doesn’t observe sellers yet (section 7).
- Examples use reserved
.exampledomains.
- 2026-09-26
- Scope: the spec covers consumer-facing online channels, including omnichannel retailers identified by their web domain. Trade, wholesale and practitioner-only channels are out of scope (section 1).
- Every authorization MUST name at least one online channel identifier (
amazon,walmart,ebayorweb). Aphysicalidentifier alone no longer authorizes anything (section 6). - Two open questions: dispensary and multi-seller platforms, and a channel
audiencefield (section 13).
- 2026-09-28 (v0.2, breaking)
- One registry. authorizedretailers.ai is the only registry, and its key set is the trust anchor. The format stays open (section 1).
- The registry MUST NOT change a brand’s list on its own, and MUST keep a change log the brand can see (section 3).
- Domain verification: a DNS TXT record (recommended) or the
verify_tokenin a pointer. The hostedauthorized-retailers-verify.txtfile is removed. The token is public (section 4). brand_unverifiedis renamedbrand_not_registered, with the same reasons (sections 4, 9).- A brand account can hold several domains, each verified separately, sharing or copying a list (section 4.1).
- The
fullandprivatefile forms are removed. The only file is an optional, unsigned pointer whoselistMUST be on authorizedretailers.ai. Readers MAY follow up to 3 redirects (section 5.1). - The signed list lives on the registry. It names every brand domain it applies to, carries
issued,valid_untilandstale_untilin place ofexpires, and adds an optional catalog and brand-level exclusions. Whether it is public is a registry setting; a list that isn’t public returns a signed"public": falseresponse instead of a 404 (section 5.2). - Marketplace channels are identified by
domain(for exampleamazon.com) instead ofmarketplace(US). Thephysicalchannel type is removed. A normative table lists the marketplace domains and their countries (section 6). - Channel keys and seller ID normalization are written out (section 6.1).
- Territory follows each channel’s domain.
scope.territoriesis optional, needed only for web stores on generic domains, and must not contradict a marketplace domain’s country (section 6.2). - Brand catalog: authorizations can be scoped by product as well as by product line, with exclusions per authorization and per brand (section 3.1).
- Verify requests take an optional
product(GTIN or ASIN) andas_of, andterritoryis optional. Answers carryproduct,stale_untiland, when authorized,revocation_id.unlistedanswers MAY carry areason(section 9). - Authorization is decided for the seller first, then for the product (sections 9.1, 9.2).
- No default or maximum expiry: an authorization lasts until revoked unless the brand sets an end date. Reminders apply only to authorizations with one (section 10).
- The verify API MUST reflect a revocation within 1 minute (was 15).
valid_untilis at most 15 minutes (was 24 hours) (sections 9, 10). - Availability:
stale_until, mirrors, a revocation feed of random IDs, and the rule that an outage leaves a seller unconfirmed, never authorized or unauthorized (section 10.2, 10.3). - Bring-your-own-key signing and SKU-level scope are removed from the open questions. Cross-border purchases, per-marketplace ASINs and finding the brand domain are added (section 13).
- 2026-10-07 (v0.3, breaking; published together with v0.4)
- Why DNS only: a file on a brand’s site is weaker proof than DNS and many site platforms can’t serve it, and distributors and retailers shouldn’t have to publish anything when the brand’s list is what authorizes them.
- Brand domain verification is by DNS TXT record only. The pointer no longer verifies a domain; it is offered once DNS verification is done. A lapsed domain is one whose DNS record is gone (sections 4, 5.1).
- Distributors and retailer web stores publish nothing on their domain. The registry confirms them through an account signed in with an email address at that domain (sections 3, 6).
- The pointer’s
verify_tokenlets the registry tell whether a pointer is current. The registry SHOULD tell a brand when its pointer is out of date (section 5.1). - Why product-first: an agent on a product page sees a product identifier and a seller, not the brand’s domain. Requiring
brand_domainmeant most agents couldn’t ask. - Verify requests MAY omit
brand_domainwhen they give aproduct. A product MAY give both agtinand anasin(section 9). - The registry finds the brand from the product through registered brands’ catalogs. Answers carry
brandandresolved_by. New statusesproduct_not_found,brand_ambiguousandbrand_mismatch(section 9.5). POST /v0/resolveand the MCP toolresolve_brandfind the brand without asking about a seller (sections 9.4, 9.5).- The catalog is what makes a brand findable from a product page. GTIN is the baseline; brands on Amazon SHOULD list ASINs (section 3.1).
- Brand names and aliases are for display and search, and never resolve a brand (section 4.2).
- Finding the brand is removed from the open questions. GS1 company prefix is added (section 13).
- A pointer MAY name v0.2 or v0.3: its format is unchanged, so brands don’t need to replace it (section 5.1).
- Why product status and regions: a recall is the most important thing an agent can learn about a product, and brands selling different formulas per market shouldn’t need a list per domain.
- Products MAY carry
status(active,discontinued,recalled),status_sinceandrecall_url. The registry SHOULD tell registered retailers authorized for a recalled product. Answers show them for every brand, public list or not. Agents MUST tell the user about a recall and SHOULD NOT complete the purchase (sections 3.1, 9, 9.3). - Products MAY carry
regions. A product asked about outside them isunlisted, reasonproduct_not_for_region. Region groupEU(sections 3.1, 9.2). - Products MAY carry
mpnsandstyle_code. Requests MAY addmpnandstyle_codeto a GTIN or ASIN; answers say which areconfirmedormismatched, and a mismatch adds the signalproduct_details_mismatch. They never find a brand or product alone (sections 3.1, 4.2, 9). - Public lists’ answers also give the product’s
name,mpnsandstyle_code(section 9). - Why Walmart identifiers: agents on walmart.com see a Walmart item ID, rarely a barcode, so a brand selling there needs a Walmart identifier to be findable, as ASINs do for Amazon.
- Products MAY carry
walmart_item_idsandwpids. Requests MAY givewalmart_item_idorwpid, alone only on awalmartchannel on walmart.com;resolved_bymay bewalmart_item_idorwpid(sections 3.1, 9, 9.5). channel.typeis optional in a request: the domain implies it (section 9).- For a web store on a generic domain, the territory is the buyer’s country, sent by the agent; the registry never infers it from an authorization (section 6.2).
- 2026-10-07 (v0.4)
- Why declarations:
unlistedcan’t tell a seller the brand hasn’t got round to listing from one it knows is diverting. Brands asked to say so, and agents asked to know. - A brand MAY declare a seller on a channel identifier not authorized. Verify answers about it are
unauthorized, withdeclaration_idanddeclared, whatever the product. A declaration and an active authorization never cover the same channel identifier. Declarations are never published in lists or the feed (sections 2, 3.2, 9, 9.1). - A seller can dispute a declaration without proving control of the seller ID; a web store MAY be asked for an address at its domain. While the dispute is open the answer is
disputedwith thedeclaration_id. The brand reconfirms or withdraws; after 30 days with neither, the answer goes back tounauthorized(section 11). - Why agents must handle unknown statuses: the spec can add statuses, as this version does, and requests carry no version to opt out of them. A status an agent can’t parse is safe if it’s treated as not authorized.
- Agents MUST treat a
statusthey don’t recognise as not authorized, and SHOULD treat the seller as unconfirmed (section 9). - A declaration is the brand’s statement, not a finding that the seller acts unlawfully (sections 1, 12).
- A pointer MAY name v0.2, v0.3 or v0.4 (section 5.1).
- Brand-declared
unauthorizedis removed from the open questions (section 13).
- Why declarations:
- 2026-10-15 (v0.5, breaking)
- Distributors are removed. The registry has two parties: the brand that decides and the seller that sells to consumers. The terminology row, the proposal flow and the designation rules are gone (sections 2, 3).
proposed_byis removed from the authorization object and fromcommon.json. Readers MUST ignore it if present in a list published under an earlier version (section 5.2).- Terms: a seller is anyone who sells a brand’s products on a channel, whether or not the brand has authorized it, and a retailer is a seller the brand has authorized. Authorizing makes a seller an authorized retailer. §2 now defines both. The diagram and sections 6 and 12 follow it: a party whose status is not yet known is a seller, an authorized one is a retailer (sections 2, 6, 7, 12).
- A pointer MAY also name v0.5 (section 5.1).
- 2026-10-19 (v0.6)
- Observation is defined for every channel type, not only Amazon: what counts as a sighting, that a registry says which channels and territories it observes, and that
observed: nullis never a negative signal (section 7). Observation on channels beyond Amazon and multi-brand owners are removed from the open questions. - Finding sellers a brand hasn’t listed is declared outside the spec: a lead is not a status and never appears in an answer, list or feed (section 7).
- Each domain verifies with its own DNS record, and nothing carries over between domains or companies, so a parent company has no shortcut (section 4.1).
- Agents SHOULD send a descriptive
User-Agentwith a way to reach the operator; it never changes an answer (section 9.6). - Fix: a stale answer is usable when the verify API is unreachable, with a revocation feed issued at or after the answer. The old condition (registry and every mirror unreachable) could never be met, because the feed comes from one of them (sections 10.2, 10.3).
- An answer is unusable as
authorizedafter itsexpires, even before itsvalid_until(section 10.1). - Section 12 says what a registry’s query log may hold.
- Mirrors are openly readable, with no account or key; operators MAY rate-limit. The first mirror will run on Google Cloud (section 10.2).
- Sellers that share one platform domain stay out of scope, and say why: authorizing the domain authorizes every seller on it (section 1). The platform and audience questions are deferred to a case-by-case basis (section 13).
- A short informative section places the check in agentic checkout flows such as UCP and ACP (section 9.7). The extension-field question is closed.
- Observation is defined for every channel type, not only Amazon: what counts as a sighting, that a registry says which channels and territories it observes, and that
- 2026-10-20 (0.6.1)
- Agents SHOULD cache the key set no longer than its
Cache-Controlallows (5 minutes at the registry), so an emergency rotation reaches them, and SHOULD say when an answer’s signature was not checked (section 8). valid_untilis presented as when an answer needs re-checking, never as when an authorization ends; the authorization’s own end isexpires(section 10.1).
- Agents SHOULD cache the key set no longer than its
- 2026-10-22 (v0.7)
- Why registry review: anyone can look up a trademark filing and submit it, so the independent link proved little, and the registry reviews every brand by hand anyway. Domain verification stays; the registry MUST review each brand before its list counts, MAY decline or end a registration, and SHOULD publish how it reviews.
domain_unlinkedkeeps its value and now means the registry has not approved the brand yet (sections 4, 7, 9.5, 12). Residual risk: a lookalike that review misses can claim an unregistered brand’s barcodes; once the real brand registers, section 9.5 answersbrand_ambiguousand neverauthorized. - Verify requests MAY carry a
price(amount,currency), the price the agent sees for the seller and product. It never changes the answer and never reveals a brand’s minimum advertised price; the registry MAY keep it with the query log and show the brand, in aggregate, how often agents named a seller and product below its MAP, and which agents did (sections 9, 12). Reason: brands now give the registry a MAP per product, and the registry can otherwise compare only the prices its own seller scans saw. For the 0.7 release:priceis added toverify-request.json(the 0.6 schema refuses unknown fields). - A declaration MAY carry a product
scope(product_lines,products, or both), with the same shape as an authorization’s. A scoped declaration answersunauthorizedto a request with no product (seller first) and to a catalog product orproduct_lineinside its scope; a product outside it is answered as if there were no declaration, with the signalunauthorized_for_other_products. Without a scope a declaration covers every product, as before. A declaration and an active authorization still never share a channel identifier (sections 3.2, 9.1). - Signals are listed in section 9, with two new ones:
authorization_scopedon anauthorizedanswer to a request with no product when the authorization is limited to named lines or products or has exclusions (ask again with the product), andunauthorized_for_other_productsabove (sections 9, 9.1). - An authorization dispute the brand doesn’t answer within 14 days closes with the brand’s current decision standing, as a declaration dispute does after 30 days (section 11).
- Housekeeping: a pointer MAY name any version from 0.2 to 0.7 (section 5.1; the 0.6 schema had left 0.5 out of
pointerSpecVersion, fixed in 0.7), and the examples show the version of the document (sections 5.1, 5.2, 10.3). - Schemas:
schemas/0.7/addspricetoverify-request.jsonand the pointer version list; nothing else changes shape. Documents nameauthorized-retailers/0.7. - Wording: a declared seller is one the brand has found selling without authorization, and an agent SHOULD NOT send a shopper to it, since the goods may be materially different or counterfeit. The registry still makes no finding of its own (sections 1, 3.2, 12). An unrecognised
statusMUST NOT be treated as authorized, replacing “as not authorized” (section 9). - The verify API’s paths carry no version (
POST /verify,POST /resolve): a request without one gets the current specification. A registry MUST also serve them under/v0/for existing clients (sections 5.2, 9, 10.3). The README and the conformance runner use the unversioned paths; the pointer example names the list at the unversioned path, and a pointer naming it under/v0/stays current (section 5.1). - New section 9.8, what an agent does with an answer that isn’t
authorized: it MUST NOT complete a purchase from a seller the brand has declaredunauthorizedon its own, only after the person has confirmed that purchase knowing what the declaration means (or has given a standing instruction, such as for enforcement test purchases); forunlistedit SHOULD prefer an authorized seller and say when it picks another; anything else is unconfirmed and said so. Reason: a shopper whose authorized channels are sold out, or a brand’s own test buy, can still go ahead, but knowingly and never by the agent alone. Answers, schemas and the conformance decision are unchanged.
- Why registry review: anyone can look up a trademark filing and submit it, so the independent link proved little, and the registry reviews every brand by hand anyway. Domain verification stays; the registry MUST review each brand before its list counts, MAY decline or end a registration, and SHOULD publish how it reviews.
- 2026-10-23 (0.7.1)
- Text only: no schema or version-string change. Documents still name
authorized-retailers/0.7. - An authorization dispute the brand doesn’t answer now closes after 30 days, not 14, the same period as a declaration dispute. A retailer disputing a listing and a seller disputing a declaration get one clock to reason about, and the brand has the same time to review evidence in both (section 11).
- Section 11 no longer says dispute timelines are set out in the registry’s terms of use: the spec states them, and the terms repeat them.
- The registry reserves the right, in its terms of use, to remove an authorization or declaration it reasonably believes is unlawful or abusive. The removal is recorded in the brand’s change log with its reason; every other change still needs the brand’s instruction or approval (sections 3, 3.2, 12).
- Text only: no schema or version-string change. Documents still name
- 2026-10-25 (v0.8)
- API keys (section 9.6): a registry MAY require a key on verify, resolve and the MCP tools. The key goes in
Authorization: Bearer, never in a URL or body. No key or a bad one is HTTP 401 (api_key_required,invalid_api_key) and is unconfirmed; a key never changes an answer; a registry stores only a hash, shows the key once, and revokes within a minute. The key set, public lists and the revocation feed stay open. Reason: the registry wants to know who is testing and using it, and to rate-limit each operator on its own instead of by network address.User-Agentstays as identification, and is no longer described as never a condition of service (sections 9, 9.3, 9.4, 9.6). - A 401 or a 429 means the registry is reachable, so the stale-answer rule in section 10.2 doesn’t apply to it; the seller is unconfirmed meanwhile.
- The query log MAY hold the ID of the key a request used, never its secret, and a brand still never learns which caller asked, whether by address, key or name (section 12). New risk entry for leaked keys.
- Schemas:
schemas/0.8/has the same shapes as 0.7. Documents nameauthorized-retailers/0.8, and a pointer MAY name any version from 0.2 to 0.8 (section 5.1). The conformance runner takes--key. - Open questions are now for v0.9 (section 13).
- API keys (section 9.6): a registry MAY require a key on verify, resolve and the MCP tools. The key goes in
- 2026-10-31 (0.8.1)
- Text only: no schema or version-string change. Documents still name
authorized-retailers/0.8. - A registry MAY ask a seller requesting authorization for evidence of where it buys the brand’s products, such as a supplier invoice, and show it to the brand with the request. The evidence is for the brand’s decision only: never published, never a factor in verify answers, and deleted within a set time after the decision (sections 3, 12). Reason: authorizedretailers.ai now requires an invoice on every listing request, so a brand can see whether its products reach the seller through its authorized channels rather than retail arbitrage. No schema or version-string change.
- Text only: no schema or version-string change. Documents still name