A field-level product-truth reconciliation workflow for ChatGPT feeds, Google Merchant Center, Product structured data and the visible product page.
Documented fact: OpenAI says product feeds supply current pricing, availability and seller context for ChatGPT, while Google recommends aligning page structured data and Merchant Center feeds and documents mismatches as a source of display or eligibility problems.
CDM inference: The Product Truth Reconciler, field ownership model and release gate are CDM operations design, not a joint OpenAI–Google specification.
Four surfaces can describe one product differently
A visible product page serves people and crawlers. Product structured data expresses selected facts in machine-readable markup. Google Merchant Center receives catalogue data for Google shopping experiences. An OpenAI product feed supplies structured catalogue data for discovery and commerce in ChatGPT.
OpenAI’s current product-feed documentation says feeds help ChatGPT surface products with accurate pricing, availability and seller context, and instructs merchants to keep price and inventory updates fresh. Google’s Merchant Center product-data specification says inaccurate, missing or conflicting information can cause disapprovals, limited eligibility or incorrect displays.
Google also says providing both on-page Product structured data and a Merchant Center feed can maximise eligibility and help it understand and verify product information. Its Product structured-data guide notes that some experiences combine feed and page data.
These systems have different schemas and update paths. Consistency does not mean identical files; it means that equivalent claims resolve to the same governed fact for the same product, market and moment.
Define the governed product record
Choose the system that owns each fact. A product information management system may own title, brand, identifiers and variant relationships. Commerce may own current price and availability. Logistics may own shipping. Legal or customer operations may own returns. The web page and feeds are published views, not independent truth stores.
Identity comes first. Stable product and variant identifiers prevent a price or stock value from attaching to the wrong item. Google advises merchants to use a unique product ID, keep it stable and use SKU where possible. Map retailer SKU, Global Trade Item Number, manufacturer part number, parent group and platform-specific IDs explicitly.
Define market and currency in every record. “£49” is incomplete without the applicable country, tax treatment, variant and effective time. Availability should describe the same purchasable unit shown on the landing and checkout paths.
Build the Product Truth Reconciler
| Field group | Governed source | Cross-surface test | Severity if mismatched |
|---|---|---|---|
| Identity | Catalogue/PIM | ID, brand, title, GTIN/MPN and variant map refer to the same item | Critical |
| Price | Commerce pricing | Amount, currency, sale period, membership and checkout total agree | Critical |
| Availability | Inventory/commerce | Feed, schema, page and buy control reflect the same state | Critical |
| Description | PIM/editorial | Core facts and included items agree; channel-specific copy does not contradict | High |
| Image | Asset system | Correct product and variant, crawlable URL, rights and quality | High |
| Shipping and returns | Operations/policy | Market, cost, timing, exceptions and policy URLs agree | High |
| Seller | Merchant master | Legal/display identity and support details are consistent | High |
CDM asset: Product Truth Reconciler row
For every field, store the canonical value, system owner, transformation rule, ChatGPT mapping, Merchant Center mapping, schema path, visible-page selector, freshness target, last update, test result, severity and remediation owner.
Transformations should be explicit. A feed title may have a length constraint; the page title may include editorial context. Both can be valid if they identify the same product and do not contradict. Record the transformation rather than judging differences by string equality alone.
Model temporal truth. Sale price has an activation and expiry. Inventory can change faster than editorial copy. Use event or scheduled updates appropriate to volatility, and make stale-state handling explicit.
Create a field-level release gate
- Select a representative set of products, including variants, sale items and out-of-stock states.
- Extract the governed record and every rendered or delivered representation.
- Normalise currency, identifiers, whitespace and enumerations without erasing meaning.
- Compare critical fields and transformation rules.
- Validate page structured data with supported testing tools.
- Check landing and checkout truth manually in each target market.
- Block critical mismatches; assign high and medium issues with deadlines.
- Store feed version, page build, schema snapshot and test result.
Use two gates: schema validity and business truth. A JSON-LD document can be syntactically valid while stating an old price. A feed can satisfy a file schema while mapping a child variant to the parent image. Technical success does not replace factual reconciliation.
Test update latency. Change one controlled field in staging or a safe test product and measure when each surface receives it. That observation informs freshness targets and incident expectations.
Handle mismatches as customer and trust incidents
Prioritise identity, price and availability. Stop promotion or purchase paths where the mismatch could mislead or frustrate. Correct the owning source rather than patching four outputs independently.
Capture discovery time, affected products and markets, source version, downstream surfaces, customer impact, correction, propagation and prevention. A feed re-upload is not closure until the page and relevant destinations are verified.
Keep a fallback for stale dynamic fields. If real-time stock cannot be trusted, show a truthful availability check rather than a precise false claim. Review platform rules before omitting required data.
Monitor recurring drift by field and owner. Repeated title mismatch may indicate transformation logic; repeated availability mismatch may indicate latency or variant mapping. Fix the system, not only the row.
Limits and risks
OpenAI and Google specifications change, differ by programme and market, and may include fields not covered here. Confirm stable production documentation, access and contractual terms before implementation.
External platforms determine eligibility and presentation. Correct data improves reliability but does not guarantee discovery, ranking, citation or sales. Pricing, tax, product safety and consumer terms require qualified owners.
CDM verdict
Product discoverability starts with product truth. Assign each fact an owner, map it field by field, test both syntax and meaning, and block material contradictions before release. Feeds, schema and pages are distribution surfaces; none should silently become the master record.
Related guides
Frequently asked questions
Should the product page or merchant feed be the source of truth?This article is editorial guidance. Apply the principles in proportion to your market, evidence, and responsibilities.
