A company's product catalogue should survive the replacement of its storefront, marketplace or integration vendor. That requires operational control over product identity, attributes and distribution rules.
Reuters reported on November 13, 2026 on World Product Graph, an internal Amazon proposal that the company says may no longer reflect its plans. No launch is confirmed. The engineering problem of distributing reliable product information is already real.
Model identity before connecting destinations
A master product catalogue is a logical system of records and rules governing product identity, provenance, versions and distribution. It can span multiple services rather than requiring one universal database.
Separate the product, variant and offer. A product groups a commercial family. A variant represents an identifiable configuration. An offer represents the opportunity to purchase a variant from a seller in a market under specific terms.
Consider an espresso machine with two voltage configurations. The product family shares a name and design. Each variant has its electrical specification. Offers vary by market, currency, seller and inventory. Storing voltage and price as unscoped properties in the same mutable record invites accidental overwrites.
Use a stable internal identifier. Retain legitimate manufacturer identifiers and GTINs when assigned to the relevant variant. Do not fabricate identifiers to satisfy a destination's validation rules. Marketplace identifiers belong in an external reference mapping, not as the sole business key.
Schema.org Product is useful for external representations. The internal model may need richer semantics than a public vocabulary provides, particularly for approval, history and regional scope.
Provenance and precedence belong at attribute level
A critical attribute needs more than a value. Preserve its source, supporting evidence, observation time, effective period, version and approval responsibility where appropriate. An approved value can coexist with pending proposals and conflicting observations.
Authority varies by field. A manufacturer validates electrical specifications. An inventory service reports available stock. A commercial service governs the active offer. A translation provider may suggest localized copy without gaining permission to change technical measurements.
Define conflict resolution explicitly. Depending on the field and evidence, the system may accept an update, retain the approved value or quarantine the conflict for review. Last-write-wins is rarely a suitable universal policy. Unknown values must remain distinguishable from zero, false and not applicable.
History makes it possible to reconstruct what the business believed was valid at a particular time. That matters when a customer challenges a specification shown before a correction. Retention policies should balance auditability, storage costs and applicable obligations.
Distribute changes through explicit contracts
The catalogue does not have to serve every destination directly. A distribution layer can translate approved records into channel-specific contracts, including those described in Google Merchant Center’s product data specification.
A reference workflow has five stages:
- Validate the proposed change against the schema, business rules and variant relationships.
- Persist the approved version and a distribution event atomically using a transactional outbox where the storage model supports it.
- Consume the event idempotently, identifying the variant, version and destination.
- Transform the record into the destination contract, validate the payload and deliver it with bounded retries.
- Record acceptance or rejection, then periodically reconcile the external representation with the desired state.
Retries must not create duplicate offers or restore obsolete versions. Idempotency keys and version checks reduce these risks, subject to the guarantees provided by each destination.
Incremental events avoid unnecessary work. Full snapshots provide a recovery mechanism when a destination misses updates. Both need explicit retirement semantics: removing a variant from an export is not sufficient unless the destination interprets absence as intended.
Isolate destination failures. Each adapter can maintain independent queues, concurrency limits and error records. Permanent validation errors should reach a review queue; transient failures can use exponential backoff with jitter. A destination outage should not prevent approval of a valid catalogue correction.
Reconciliation, security and acceptance criteria
A successful HTTP response does not establish that a public listing reflects the intended record. Reconciliation compares desired, delivered and, where observable, externally displayed versions. Keep those states separate in operational reporting.
Public specifications also need a clear boundary from private information. Supplier costs, margins, credentials and negotiated commercial terms should not leak into a public feed. Give each adapter access only to required fields. Assign editing privileges by role and consider independent approval for high-impact changes.
Tests should cover near-identical variants, duplicate events, out-of-order delivery, partial rejection, provider outages and product retirement. Verify that a technical correction can propagate without unintentionally changing the offer or stock.
Useful operational measures include critical attribute coverage, freshness by attribute class, distribution lag, discrepancy rate and resolution time. Set objectives according to business risk and the destination's actual capabilities rather than adopting arbitrary universal thresholds.
A product information management system may handle much of this work. An ERP, commerce platform and inventory service can remain authoritative for separate domains. The build-versus-integrate decision depends on which rules the business must govern, a distinction central to owning software decisions.
Frequently asked questions
Does a master catalogue require a graph database?
No. Product relationships can be represented in different storage engines. Query patterns, scale and operational constraints should guide the decision.
Should the PIM own inventory?
Not necessarily. It can consume availability from the authoritative inventory service without becoming the original source.
How should a distributed attribute be corrected?
Approve a new version, propagate it and verify destination state wherever observation is supported. Record unresolved discrepancies.
What if a destination has no read API?
Retain delivery evidence and use available checks, including manual inspection. Do not report confirmed synchronization without supporting evidence.
Catalogue ownership becomes tangible when a business can correct an attribute, reconstruct its history and replace a channel without losing product identity. The architecture should make those actions routine rather than exceptional.
SOURCES & REFERENCES
Explore the sources
- Product
Schema.org
The Blaxline Editorial Team brings together strategists, designers and engineers to examine how technology can make important business work clearer, more connected and easier to evolve.