Reference

Global Unit Data Loads — 3D Commerce, WMS & Merchant Feed Spec

Nov 13, 2026 · Yoonsun Lee

How g:id links to the Shopify GID namespace. The Shopify product GID gid://shopify/Product/8123456789 and variant GID gid://shopify/ProductVariant/4455667788 give their numbers to the Google offer ID shopify_KR_8123456789_4455667788, with item_group_id 8123456789. The Merchant API v1 names the product accounts/{merchantId}/products/ko~KR~offerId, so the full key is GID plus content language plus feed label. Back to Shopify: split the offer ID on underscores; the last part rebuilds the variant GID and the second last the product GID, with no lookup table. A custom offer ID such as fo1-KR-0142 has no numbers to parse, so the canonical record stores the GID and its offers. Rules: never change a live offer ID, derive it from the variant GID, one offer per variant per target country, up to 50 characters.

Related knowledge base: ECC direct to Google · ECC market prices and FX · ECC revenue and WMS risk register · SAP to Google EDI · SAP Fiori-style UI5 apps at the edge · An alternative to a full S/4HANA upgrade · From product.csv to Consented Conversions · Copper to Fiber: 2000s-Era Network Debt and Short-Term Paired Fixes

Live on fo1.io (the first use case): Furnish a home (floor plan, budget and a 3D room in meters) · Configure in 3D (made-to-measure pieces) · Portal (clients and architects, demo)

Overview

Every physical quantity is stored once, in SI units with an ISO/UN code, and every surface converts from that record at its own edge. The 3D viewport, the WMS, Shopify, the Google Merchant feed and every migrated app read the same canonical unit record; none of them stores its own copy of a measurement.

Canonical unit model

The canonical record stores the SI value plus the UN/CEFACT Rec 20 code; UCUM, CLDR and Shopify codes are lookups, not stored copies. A record is {"value": 1.8288, "unit": "MTR", "source_value": 72, "source_unit": "INH"}, so the original entry is never lost.

QuantityUnitUN/CEFACT Rec 20 (canonical)UCUMCLDR / IntlShopify metafield
LengthmeterMTRmmeterm
LengthcentimeterCMTcmcentimetercm
LengthmillimeterMMTmmmillimetermm
LengthinchINH[in_i]inchin
LengthfootFOT[ft_i]footft
Areasquare meterMTKm2——
Areasquare footFTK[sft_i]——
Volumecubic meterMTQm3—m3
VolumeliterLTRLliterl
VolumeUS gallonGLL[gal_us]gallonus_gal
WeightkilogramKGMkgkilogramkg
WeightgramGRMggramg
WeightpoundLBR[lb_av]poundlb
WeightounceONZ[oz_av]ounceoz
TemperaturekelvinKELK——
Temperaturedegree CelsiusCELCelcelsius—
Temperaturedegree FahrenheitFAH[degF]fahrenheit—

Conversion rules:

Linking units to global WMS

The WMS keeps its own unit-of-measure (UoM) master, so the link is a mapping table from each WMS UoM to the canonical UN/CEFACT code, plus GS1 identifiers on everything that ships. Packaging levels (each, case, pallet) stay in the WMS; physical measures resolve to the canonical record.

Physical measureGS1 Application IdentifierCanonical code
Net weight310n (kg) · 320n (lb)KGM
Length311n (m) · 321n (in)MTR
Width312n (m)MTR
Height / depth313n (m)MTR
Net volume315n (L) · 316n (m³)MTQ
Gross (logistics) weight330n (kg) · 340n (lb)KGM
Max temperature4330 (°F) · 4331 (°C)KEL
Min temperature4332 (°F) · 4333 (°C)KEL

The n digit in each AI is the number of decimal places. The temperature AIs were added in GS1 General Specifications v24 for freight temperature requirements and are expressed in hundredths of a degree; GS1 has no kelvin AI, so kelvin converts to °C or °F at the label.

Google Merchant feed schema

The feed is generated from the canonical record, never edited by hand, and pushed through the Merchant API v1. The Content API for Shopping reached its sunset on August 18, 2026, with intermittent 410 errors from September 1, 2026, so any direct integration must already be on the Merchant API. Shopify's Google channel handles that migration for its own sync.

ConcernFeed attributeFormat and unitsSource in canonical record
Currencyprice, sale_price, cost_of_goods_soldvalue + ISO 4217, e.g. 4500000 KRWprice per market
FXnone in the feedsubmit a price in the target country's currency rather than relying on conversionShopify Markets price + stored rate and rate date
Taxtax (US only), tax_categoryUS: country, region, rate; KR and EU prices are VAT-inclusivetax engine per market
ISO codesshipping.country, ships_from_country, content_languageISO 3166-1 alpha-2, ISO 639-1market + locale
Datesavailability_date, sale_price_effective_dateISO 8601 (interval for sale dates)order and promo calendar
Shipping weightshipping_weightlb, oz, g, kgKGM, converted
Shipping dimensionsshipping_length, _width, _heightcm or in, 1–400 cm, one unit across all threeMTR, converted
Product dimensionsproduct_length, _width, _height, product_weightcm / in; lb, oz, g, kgMTR, KGM
Unit pricingunit_pricing_measure, unit_pricing_base_measurevalue + unit, e.g. 750 ml / 100 mlvolume or weight record
Loyaltyloyalty_programprogram label, tier, member price, pointsloyalty migration (below)
Technical specsproduct_detailsection, name, value (e.g. Color temperature, 2700 K)Kelvin, area, materials

Rules:

Example item (XML feed):

<item>
  <g:id>fo1-KR-0142</g:id>
  <g:price>4500000 KRW</g:price>
  <g:availability>preorder</g:availability>
  <g:availability_date>2027-01-15T09:00+09:00</g:availability_date>
  <g:product_length>213 cm</g:product_length>
  <g:product_weight>48 kg</g:product_weight>
  <g:shipping_label>freight-white-glove</g:shipping_label>
  <g:ships_from_country>US</g:ships_from_country>
  <g:product_detail>
    <g:section_name>Lighting</g:section_name>
    <g:attribute_name>Color temperature</g:attribute_name>
    <g:attribute_value>2700 K</g:attribute_value>
  </g:product_detail>
</item>

The offer ID (g:id) is a plain string of up to 50 characters, not a GID. Shopify's Google channel builds it from the numbers at the end of the GIDs: gid://shopify/Product/8123456789 and gid://shopify/ProductVariant/4455667788 become g:id shopify_KR_8123456789_4455667788 and item_group_id 8123456789, one offer per variant per target country. The Merchant API v1 names the product accounts/{merchantId}/products/{contentLanguage}~{feedLabel}~{offerId}, so the full key is GID + language + feed label. Splitting the offer ID on _ gives both GIDs back with no lookup table (see the diagram at the top).

A custom offer ID such as fo1-KR-0142 has nothing to parse, so the canonical record stores the link: { "gid": "gid://shopify/ProductVariant/4455667788", "offers": [{ "offerId": "fo1-KR-0142", "feedLabel": "KR", "contentLanguage": "ko" }] }. Never change a live offer ID, since Google treats a new ID as a new product; derive it from the variant GID, never the SKU or title.

product.csv vs. Shopify Catalog data layer vs. Google Merchant

The GraphQL Admin API catalog is the data layer; product.csv is a flat, lossy export of it, and the Merchant feed is generated from GraphQL, never from the CSV. The CSV stores weight only in grams, has no dimensions, 3D media, translations or contextual pricing, and carries no explicit currency code, so it is used only for bulk legacy imports. Column names follow Shopify's current product CSV; legacy names are in parentheses.

Dataproduct.csv columnCatalog GraphQL (Admin API)Google Merchant attributeUnits, ISO and gaps
IdentityURL handle (Handle)Product.id, handle; ProductVariant.idid (variant), item_group_id (product), linkGID ↔ offer ID mapping kept in the canonical record
TitleTitletitletitleper-locale titles from translations
DescriptionDescription (Body (HTML))descriptionHtmldescriptionHTML stripped to text for the feed
BrandVendorvendorbrand—
CategoryProduct categorycategory (TaxonomyCategory)google_product_categoryShopify taxonomy → GPC ID mapping
TypeTypeproductTypeproduct_type—
TagsTagstagscustom_label_0–4 (by rule)—
StatusStatus, Published on online storestatus, publicationsexcluded unless active and published to the Google channel—
OptionsOption1–3 name / valueoptions, selectedOptionscolor, size, material, patternoption → attribute mapping per category
SKUSKU (Variant SKU)skuoften the offer id—
GTINBarcode (Variant Barcode)barcodegtinempty → identifier_exists = no for one-of-a-kind
MPN, gender, age group, conditionGoogle Shopping / … columnsmetafields (Google channel namespace)mpn, gender, age_group, condition—
PricePrice (Variant Price)price (MoneyV2: amount + currencyCode)priceCSV has no currency column; GraphQL carries ISO 4217
SaleCompare-at pricecompareAtPriceprice = compare-at, sale_price = priceplus sale_price_effective_date
Market pricePrice / [market]price lists, contextualPricing(context: {country})price per target countryKRW, JPY: zero decimals
CostCost per iteminventoryItem.unitCostcost_of_goods_sold—
TaxCharge tax, Tax codetaxable, tax codetax (US only), tax_category—
WeightWeight value (grams) (Variant Grams), Weight unit for displayinventoryItem.measurement.weight (value + unit)shipping_weight, product_weightCSV is grams only; GraphQL keeps g, kg, oz, lb
Dimensionsnonedimension metafieldsshipping_length/width/height, product_length/…not in CSV at all
Unit pricingUnit price total / base measure + unitunitPriceMeasurementunit_pricing_measure, unit_pricing_base_measure—
InventoryInventory quantity, Continue selling when out of stockinventory quantities, inventoryPolicyavailability (in_stock, out_of_stock, preorder, backorder)custom orders use preorder + availability_date
ShippingRequires shipping, Fulfillment serviceinventoryItem.requiresShippingshipping, shipping_labelfreight label for oversized pieces
ImagesProduct image URL, Image position, Image alt textmedia (MediaImage url, alt)image_link, additional_image_linkalt text per locale via translations
3D modelsnonemedia (Model3d, GLB / USDZ sources)none3D stays on Shopify, R2 and Webflow
Translationsnone (separate translation CSV)translations(locale), translationsRegisterlocalized title, description per feed languagem2m100 pipeline
Origin and customsnoneinventoryItem.countryCodeOfOrigin, harmonizedSystemCodenoneISO 3166-1; feeds WMS and customs
Custom specsMetafield columnsmetafields, metaobjectsproduct_detailkelvin, area, materials as display text

Cut over by loading the CSV once into GraphQL, then switching all writes to GraphQL mutations so the CSV is never a source again.

The SAP → e-commerce break

SAP already holds every quantity as value + unit with an ISO code, and the break happens when that data is flattened into the product.csv shape on its way to the storefront. Each app downstream then reads the flattened copy, keeps its own version, and has no path back to the ERP's units, currency or origin data, so the loss becomes permanent.

For ECC specifically, see ECC direct to Google (converting ECC's own data straight into Merchant Center's shape) and, for the screens, Fiori-style UI5 apps at the edge calling ECC's OData services through the gate.

SAP sourceWhat SAP holdsWhat the CSV shape keepsLossless target in the Catalog data layer
Material master weights (MARA-BRGEW, NTGEW, GEWEI)gross and net weight + unitone weight, grams onlyinventoryItem.measurement.weight + canonical KGM record with source unit
Volume (MARA-VOLUM, VOLEH)volume + unitnothingvolume metafield + canonical MTQ / LTR record
Dimensions per unit of measure (MARM-LAENG, BREIT, HOEHE, MEABM)L × W × H for each, case and palletnothingdimension metafields; packaging levels stay in WMS
Units table (T006 ISO code)internal unit → ISO/UN unit codenot applicablelookup to UN/CEFACT canonical codes
Currency and FX (TCURC, TCURR)ISO 4217 codes, rates by type and datea bare numberMoneyV2 with currencyCode; rate and rate date stored
Country of origin, commodity codeISO country, customs codenothingcountryCodeOfOrigin, harmonizedSystemCode
BOM and revisionsparent/child, quantity + unit, revisionnothingBOM records keyed by Digital Link URI

Where it breaks in practice:

The fix is to skip the flat shape: read SAP through its product APIs or MATMAS IDocs, map units through T006 ISO codes to UN/CEFACT, write directly to Shopify GraphQL and the canonical record, and let every app read from that record instead of storing its own copy.

Permissions and boundary events

Each field has exactly one system of record that may write it; every other system reads. Every event that crosses a system boundary is validated for unit codes, currency and consent before it is accepted, and rejected rather than guessed when a code is unknown.

Write ownership

Field groupSystem of record (writes)Readers
Dimensions, weight, volumeERP / WMS item masterShopify, feed, 3D scene
Temperature specs and telemetryWMS + IoT ingestERP, ops dashboards
Price and currency per marketShopify (Markets)feed, Webflow, ERP
FX rate and rate dateERP / financeShopify Markets, PO costing
Tax ratestax engine per marketShopify, feed (US only)
Reviews and ratingsreview platform → Merchant API ReviewsWebflow, feed
Loyalty tiers and pointsloyalty platformShopify, feed loyalty_program
Marketing consent and identityconsent layer (CRM Sync)CRM, loyalty, ad platforms

Access scopes: Shopify Admin API scopes limited per integration (e.g. read_products for the feed builder, write_inventory only for the WMS connector); the Merchant API through a dedicated Google Cloud project with the content OAuth scope; Merchant Center user roles kept to named people.

Boundary events

BoundaryEventChecks at the boundary
ERP → Shopifyitem master changeknown UN/CEFACT codes, SI conversion, source_value kept
Supplier → WMSPO (EDI 850), ASN (EDI 856)GS1 AIs parsed to canonical units, currency + FX stamp on cost
IoT → WMStelemetry reading, threshold breachunit mapped, thermodynamic K vs color temperature tagged
Shopify → Merchant APIproduct create/updateprice currency matches target country, units within Google limits
Merchant API → opsproduct status change notificationdisapprovals routed to field owner
Shopify → CRM / loyalty / adsorder, customer updateconsent checked per subject at send time, not from a batch export
Any → archiveconversion or rejectionraw payload and rule version logged for audit

Migration from apps

Each app is migrated by extracting what it holds, normalizing every unit, currency, country and locale to the canonical codes, and writing it back to its system of record. Apps then read from the canonical data instead of keeping their own copies.

App typeUnit and ISO data it holdsNormalizationLands in
ERP (NetSuite, SAP, Acumatica, Odoo)UoM tables, item dimensions and weights, currency, FX tables, tax codesmap each ERP UoM to its UN/CEFACT code; keep FX rates with their dates; tax codes per ISO countryitem master (canonical), FX table
CRM (Klaviyo, HubSpot, Salesforce)country, locale, currency, custom-order measurements, consentISO 3166-1, BCP 47, ISO 4217; add preferred unit system; consent moved as per-subject records, not list flagscustomer profile + consent layer
Loyalty (Smile.io, LoyaltyLion, Yotpo)points, tiers, reward values in one store currencyreward values per market in ISO 4217; points stay unitless integers; member prices per marketcustomer metafields, feed loyalty_program
Reviews (Judge.me, Okendo, Yotpo)ratings, product IDs, fit and size answers (often imperial text)map products to SKU/GTIN; parse measurements into canonical records; strip reviewer PII from exportsreview platform → Merchant API Reviews, Webflow
Comparison (spec tables, comparison apps, shopping engines)free-text specs such as "72 in" or "approx. 1.8 m"parse to value + code; unparseable values go to a review queuecanonical record; tables render by locale

Migration order: ERP item master first, since every other app references its dimensions and currency; then CRM and consent; then loyalty, reviews and comparison, which depend on both. Run each app in read-only parallel until its counts and totals reconcile, then cut over.

Where each app now lands in Google

Apps that kept their own data can now publish it to Google through Merchant Center surfaces, fed from the canonical record rather than from each app's copy. Each surface carries ISO currency and country codes natively, so the unit and currency spine reaches Google intact.

App typeGoogle surfaceWhat it carriesLimits to plan for
RMA / returnsOnline return policies (accounts.onlineReturnPolicies get, list, create)countries (ISO 3166-1), return window in days, return methods (by mail, in store, kiosk), item conditions, restocking and return shipping fees with currencyCode, refund processing days, exchangespublishes the policy, not individual RMAs; policies made in the Merchant Center UI take up to 10 business days to review; must match the policy on the website
LoyaltyMerchant loyalty program + loyalty_program attribute; Loyalty Customer Match via Merchant APItiers, member prices, redeemable points, member free shippingone supported country per Merchant Center account (KR and US are both supported, so each needs its own account); member price shows in ads only at 5% or 5 currency units off or more; uploaded customer identifiers need a per-subject consent check
Reviews (store)Google store ratingsstar rating, review count, delivery-time qualifierabout 100 eligible post-purchase reviews per country; ads need 3.5 stars or more; calculated per country, not globally; sourced from Google Customer Reviews or listed partners (Okendo, Yotpo, Stamped and others)
Reviews (product)Merchant API product reviewsrating, review text, product identifiers (GTIN, MPN, SKU)reviews must map to the same identifiers as the feed; reviewer PII stays out
ComparisonComparison Shopping Services (CSS) in the Merchant APIthe same product data, listed through a CSS partnerCSS applies in Europe; products come from the canonical feed, not a separate spec table

The consent layer sits on the loyalty and review paths: Loyalty Customer Match and review requests both send customer identity to Google, so each send is checked per subject at the moment it happens.

Regional model loading: APAC, EU, the Americas

Each region loads two kinds of models from the same canonical record: translation models for content, and 3D models for the viewport. Translation runs through PIM Sync: Meta m2m100-1.2B on Cloudflare Workers AI, pushed to Shopify with the GraphQL translationsRegister mutation and tracked from pending through translated, reviewed and published. 3D files (GLB, USDZ, STL, OBJ, 3MF) are hosted on Cloudflare R2 and loaded into Three.js / A-Frame scenes in meters.

RegionExample marketsCurrency (ISO 4217)Display unitsTranslation3D asset storagePrice and tax notes
The AmericasUS, CA, MX, BRUSD, CAD, MXN, BRLUS customary in US; metric elsewherenone for US and CA English (Shopify auto currency conversion); m2m100 for es, pt, frR2 location hint for North AmericaUS prices pre-tax with Merchant tax; BR VAT-style inclusive
EU + UKDE, FR, NL, ES, IT, PT, PL, SE, UKEUR, PLN, SEK, GBPmetricm2m100 for de, fr, nl, es, it, pt, pl, sv; UK English uses auto conversionR2 bucket in the EU jurisdiction for residencyVAT-inclusive prices
APACKR, JP, AU, NZKRW, JPY, AUD, NZDmetricm2m100 for ko, ja; AU and NZ English use auto conversionR2 location hint for Asia-Pacific / OceaniaKR 10% VAT included; KRW and JPY have zero decimal places

Loading rules:

Non-commerce specs: LiDAR, BOM, QA, agents and events

Every physical item has one identity key that scans, parts, QA results, agent actions and events all attach to: a GS1 Digital Link URI built from its GTIN and serial, e.g. https://id.gs1.org/01/09506000134352/21/FO1-0142. One-of-a-kind and antique pieces with no GTIN use an internal ID plus MPN, and the Merchant feed sets identifier_exists to no.

Identifiers

IdentifierUseApplies to
GTIN (AI 01)trade itemcatalog products, purchased parts
Serial (AI 21)individual uniteach custom order, each antique
MPNmanufacturer part numberparts and products without GTIN
GMN (AI 8013)model numbermade-to-order furniture lines
CPID (AI 8010) + serial (AI 8011)component or partBOM components, replacement hardware
GIAI (AI 8004)fixed assetscanners, warehouse machines, IoT sensors

LiDAR and 3D capture

Machine part BOM

BOM fieldFormat
Parent and child itemDigital Link URI, GTIN or MPN
Quantity per parentdecimal + UN/CEFACT unit code (C62 = each, MTR, KGM)
Revisionrevision letter or number + effective date (ISO 8601)
Material and finishcontrolled list per category
Suppliermanufacturer name, MPN, country of origin (ISO 3166-1)
Engineering geometrySTEP AP242 or the source CAD file, linked by revision
Substitutesapproved alternates by MPN, each with its own QA status

QA

Agent automation and permissions

Event management

Events use the CloudEvents envelope (id, source, type, time, subject), with subject set to the item's Digital Link URI so every event traces back to one physical item.

Event typeRaised byEffect
scan.capturedscanner appcreates measured dimensions, triggers QA
qa.passed / qa.failedinspector or QA ruleunblocks or blocks publish and ship
bom.revisedengineeringnew revision; open orders flagged for review
part.substitutedpurchasingrequires QA status on the substitute
agent.proposedagentqueued for review, nothing published
agent.approved / agent.rejectednamed personapplies or discards the proposal
item.publishedpublish jobpushes to Shopify, Merchant API, Webflow

Geospatial data: KML for land, ocean and space

Locations are exported as KML (OGC KML 2.3), with every placemark carrying the same Digital Link URI and unit codes as the canonical record. KML coordinates are WGS84 longitude, latitude, altitude (in that order), with altitude in meters, so they line up with the meter-based 3D scene. LiDAR scans in a local or projected CRS are reprojected to WGS84 before export.

DomainWhat is mappedKML altitude handlingLinked data
Landwarehouses, install sites, delivery addresses, supplier factoriesclampToGround, relativeToGround (e.g. floor height in a building)WMS location, order, scan, ISO 3166-1 country
Oceanfreight routes US → KR, ports, container positions, sea-floor assetsgx:clampToSeaFloor, gx:relativeToSeaFloor; depth as negative metersUN/LOCODE ports, SSCC containers, IoT temperature readings
Air and spaceair-freight legs, GNSS fixes, satellite IoT ground tracks, sky positionsabsolute altitude in meters; sky data with hint="target=sky" (longitude = right ascension − 180°, latitude = declination)GIAI tracker assets, telemetry events

Rules:

Example placemark:

<Placemark>
  <name>FO1-0142 at Busan</name>
  <TimeStamp><when>2027-01-08T14:20:00+09:00</when></TimeStamp>
  <ExtendedData>
    <Data name="subject"><value>https://id.gs1.org/01/09506000134352/21/FO1-0142</value></Data>
    <Data name="unlocode"><value>KRPUS</value></Data>
    <Data name="temp_max"><value>298.15 KEL</value></Data>
  </ExtendedData>
  <Point>
    <altitudeMode>clampToGround</altitudeMode>
    <coordinates>129.04,35.10,0</coordinates>
  </Point>
</Placemark>

ISO data in UTF-8: space, sea, air and LiDAR

KML is the display layer. Behind it, each domain has its own identifiers and source formats, and the canonical record keeps those as text in UTF-8 (RFC 3629, the encoding form of ISO/IEC 10646), with numbers stored as numbers. The table lists the codes each domain carries into the record. The same Digital Link URI is the subject throughout.

DomainIdentityPosition and reference systemTimeSource format (kept as is)Display
SpaceCOSPAR ID (1998-067A), NORAD catalog number, GIAI for owned trackersorbit in an inertial or Earth-fixed frame (GCRF, ITRF); ground track in WGS84UTC in ISO 8601; source time scale (UTC, TAI, GPS) recordedCCSDS Orbit Data Messages (CCSDS 502.0, adopted as ISO 26900), OEM or OPMderived ground track or position in KML (absolute)
SeaIMO ship number (7 digits), MMSI (9 digits; Korea uses MID 440 and 441), ISO 6346 container code (CSQU3054383, check digit), SSCC, UN/LOCODE portsWGS84 lon/lat; depth as negative meters with its vertical datumAIS report time in UTC, ISO 8601AIS messages (ITU-R M.1371), ISO 19848 shipboard sensor data, IHO S-100 for chartsKML gx:clampToSeaFloor / gx:relativeToSeaFloor, or clampToGround at the surface
AirICAO 24-bit aircraft address, ICAO and IATA airport codes (RKSI / ICN), IATA air waybill (3-digit prefix + 8 digits, mod-7 check digit)WGS84 lon/lat; barometric altitude (feet, flight level) and GNSS geometric height kept apartUTC, ISO 8601ADS-B, cargo messages from the carrier; at least two feed providers (see the legacy network note below)KML absolute, converted to meters with source_unit FOT kept
LiDARDigital Link URI of the scanned item or site, GIAI of the scannerCRS as WKT2 (ISO 19162) or an EPSG code, horizontal plus vertical (e.g. EPSG 5186 Korea 2000 Central Belt 2010 + a height datum); reprojected to WGS84 only for displaycapture time, ISO 8601LAS 1.4 / LAZ (CRS in the WKT VLR), COPC for streaming, E57 (ASTM E2807)derived mesh as GLB in the meter scene; footprint or point in KML

Rules for every domain:

UTF-8 rules:

Example canonical record for one sea leg:

{
  "subject": "https://id.gs1.org/00/095060001343520011",
  "container": "CSQU3054383",
  "vessel": { "imo": "9074729", "mmsi": "440123456", "name": "한바다호" },
  "from": "USNYC",
  "to": "KRPUS",
  "position": { "lat": 35.1000000, "lon": 129.0400000, "h": 0, "h_ref": "EPSG:3855", "accuracy_m": 10 },
  "time": "2027-01-08T05:20:00Z",
  "readings": [{ "n": "temp", "u": "Cel", "v": 21.5 }],
  "source": { "format": "AIS", "spec": "ITU-R M.1371", "raw_ref": "r2://raw/ais/2027-01-08/…" }
}

All identifiers above (SSCC, container, IMO, MMSI, vessel name) and the storage path are placeholders. readings uses SenML (RFC 8428) names and units.

AI scaling and security for geospatial data

AI reads this data to summarize routes, flag QA failures, match scans to products and answer questions such as "where is order 0142?". It doesn't get the raw data. These are the requirements that keep it affordable and safe.

Scaling

NeedRule
Point clouds are too large for a modelA survey scan is hundreds of millions of points (gigabytes), and even a phone room scan is megabytes. The model gets derived values: dimensions, accuracy, QA result and a thumbnail. Raw LAZ/COPC stays in R2 and is read by range, tile by tile, by code.
Tracks are too long for a context windowA container's AIS track can be thousands of fixes. A Worker reduces it to legs (port to port, with times and the gaps), and the model reads the legs.
Conversions must be exactUnit, datum and coordinate conversions run in code with exact factors and the geoid model. The model never computes a conversion; it reads the converted record.
Compact, coded inputRecords go to the model as JSON with ISO and UN codes (KRPUS, KGM, Cel), not prose. That uses fewer tokens and gives less room to misread.
Fan-outEach scan, AIS batch or telemetry window is one Queue message. AI calls run per message with retries and a dead-letter queue, so a spike in scans doesn't become a spike of failed calls.
RetrievalEmbeddings index the metadata (names, notes, QA text) keyed by Digital Link URI. Coordinates and measurements are queried exactly in Xano and never matched by similarity.

Security

RiskRule
Location is personal dataDelivery addresses and live positions of a person (a driver, an installer, a customer's home) are personal data. In Korea, personal location information also falls under the Location Information Act as well as PIPA. Positions tied to a person are blurred (for example, to postal-code or 3-decimal precision) before any model sees them, and pass the per-subject consent check like any other personal field.
Asset location is sensitiveA live container or tracker position is cargo-theft data. Only delayed or leg-level positions go to AI or to a public map. The live feed stays behind authenticated endpoints.
Residency and exportKorean site scans and positions stay in APAC storage. Korea restricts taking high-precision survey map data out of the country, so check before a site-level LiDAR scan or a survey-grade model of a Korean site leaves the region.
Text inside data is untrustedShip names, KML description and ExtendedData, scan notes and AIS free-text fields can contain instructions aimed at a model. They are passed to the model as quoted data, and model output can only propose (agent.proposed), never publish.
Unicode tricksBidirectional control characters (the Trojan Source class, CVE-2021-42574) and confusable letters (Cyrillic а in place of Latin a) are stripped or rejected in identifiers, using the Unicode Security Mechanisms (UTS #39). They're allowed in display names only after the bidi controls are removed.
Parser attacksKML and GML are XML, so they're parsed with external entities and DTDs disabled (XXE). NetworkLink and remote href values are stripped on import, so a file can't make the server fetch a URL (SSRF). KMZ and LAZ are decompressed with a size limit to stop zip bombs.
Tampered recordsEach canonical row is signed (HMAC over the stored fields). A batch of position or QA changes that fails signature checks freezes sync instead of publishing, as with the integrity tripwire on the Xano side.
TrainingCustomer positions, site scans and route history are not used to train or fine-tune shared models. Each agent call logs the agent ID, model, version and the input record IDs, not the raw coordinates.

Non-browser data chains

The helm: one manifest per flow steers four duties. Sources on the left (Shopify webhooks; ERP and WMS with EDI and batch files; machines such as radar, VMS, IoT, LiDAR and trackers; shoppers and agents; design surfaces, outbound only unless allowed) flow into the helm, a control layer with one manifest per flow, versioned in git, validated, canaried and rolled back by pointer. It sets four duties. Data direction: inbound, outbound or both per flow, undeclared directions refused, nothing sent out read back. Orchestration: request, event and scheduled paths; load, validate, canary, switch; idempotent jobs; reads serve the last good copy. Security: one key per flow and device, signed messages with size and type limits, consent checked before personal data moves, agents propose and people approve. Scaling: a partition key per flow, bounded queues with backpressure, a memory budget per task, more instances rather than a bigger machine. Outputs on the right: Xano as the system of record, the edge cache with preloaded reads and A/B slots, channels (Merchant API, Meta, Amazon, Shopify), operations screens and the R2 archive. The helm steers three paths: the request path with Node preload at the edge, the event path with a queue feeding Rust and Tokio tasks, and the scheduled path with Xano background tasks.

Every chain below runs under the helm: one manifest per flow sets its direction, orchestration, security and scaling. The full model, including an example manifest, is in Copper to Fiber: Orchestration, the helm.

Most of this data never touches a browser: a scanner, an AIS receiver, a satellite IoT tracker, an ERP or a carrier's EDI. There's no cookie, no consent banner and no user session. So identity, consent and delivery guarantees have to be built into the chain itself.

ChainTransportDevice or sender identityPayloadWhat goes wrong
Scanner → edgeHTTPS upload, resumablemTLS client certificate per scanner (GIAI), checked at CloudflareLAZ/E57 or USDZ plus a JSON sidecarpartial upload; the scan is attached to the wrong item
Tracker or sensor → edgeMQTT 5 over TLS, or CoAP over DTLSper-device certificate or key; device ID = GIAISenML in CBOR (RFC 8949), batchedoffline at sea for days, then a burst; clock drift
AIS / ADS-B feed → edgeprovider API or streamprovider API key plus an IP allowlistdecoded messages with UTC timesduplicate and out-of-order reports; a single provider going quiet during an upstream network outage
Space data → edgeprovider APIprovider keyCCSDS OEM/OPMwrong frame or time scale assumed
ERP / carrier → edgeAS2 or SFTP for EDI, or a signed webhookAS2 certificates, an SSH key, or an HMAC signatureEDIFACT (UNOW), X12 856/214, JSONencoding not declared; a resent file is processed twice
Edge → system of recordQueue → XanoWorker service credential, scoped per workspacecanonical recordat-least-once delivery writes twice without a key

Rules for every chain:

Load pipeline and validation

Every load, from an app migration or a daily feed, runs the same six steps, and a record that fails a check is quarantined rather than written.

  1. Extract the raw payload as UTF-8 and archive it unchanged.
  2. Parse each measurement into value + source unit; each price into amount + ISO 4217; each place into ISO 3166-1.
  3. Map source units to UN/CEFACT codes through the lookup table; unknown codes stop the record.
  4. Convert to SI with exact factors, keeping source_value and source_unit.
  5. Validate against the rules below.
  6. Publish to each reader: Shopify metafields, WMS, Merchant API, Webflow JSON-LD and the 3D scene.

Validation rules:

Open questions

Sources