Product Data Contracts for AI Buying Agents: Build, Validate, and Monitor

Start with identifiers. Give the product family a stable internal ID and give every purchasable variant its own stable variant ID. Keep merchant SKU, GTIN when legitimately assigned, URL, currency, locale, and channel identifiers as mapped fields rather than interchangeable identity keys.

Do not invent identifiers to satisfy a validator. If a legitimate GTIN is absent, represent that absence according to the destination's rules and send the row to review when necessary. A syntactically valid invented identifier creates a worse failure: it can merge unrelated products and make later corrections hard to trace.

Variant selection must be deterministic. Color, size, material, pack quantity, or region can change price, inventory, images, or shipping. An agent should never receive a family-level offer that silently borrows fields from whichever variant happened to render first.

A commercial record should answer one concrete offer question: for this variant, currency, market, customer state, and time, what can be purchased? Price without currency is incomplete. Availability without a warehouse or market scope may be misleading. A sale price without its validity window can outlive the promotion.

Normalize values but do not erase meaning. Preserve the raw source, normalized value, currency, tax or shipping scope when relevant, and effective timestamp. If the page and feed disagree, report both values and their sources instead of selecting the lower or newer value by intuition.

Merchant forums repeatedly surface duplicate product JSON-LD generated by themes and apps. That discussion is useful operational context: the failure often comes from multiple emitters, not from missing markup. The repair is to select one owner per field and remove or reconcile duplicate projections.

For high-impact fields, store where the value came from and when it was observed. A useful receipt contains the canonical record ID, source system, source path or endpoint, raw value, normalized value, observation time, validator version, and result. This makes a mismatch reproducible instead of anecdotal.

Receipts also separate platform rules from community experience. Official product documentation defines accepted fields and formats. Independent citation research can reveal which content structures appear in answer surfaces. Merchant discussions can expose implementation traps. Only the merchant's own current record can prove the price and inventory it intends to offer.

Set freshness by business risk rather than one global timer. Product descriptions can tolerate slower review than availability. A receipt that is older than its field policy becomes stale; stale is not the same as false, but it is not eligible for an agent-ready claim.

A validator that only checks happy-path syntax is not enough. Seed controlled conflicts: swap two variant URLs, expire a sale price, duplicate a JSON-LD Product node, remove currency, supply a malformed identifier, and make checkout disagree with the feed. The test passes only when the system catches the conflict and identifies its sources.

Define precedence only where the business owns the rule. For example, checkout may be the final purchasability observation while the catalog remains the canonical merchandising record. If that relationship is not documented, the validator should not choose a winner. It should return conflicting values, evidence receipts, and a remediation owner.

The safe machine response has four states: valid, invalid, stale, and unknown. Unknown is essential. It prevents a confident-looking agent response from filling a gap with a guess and lets the operator repair the source rather than patching one downstream export.

Run contract checks whenever the canonical record changes and on a scheduled sample of live products. Change-triggered validation catches planned edits; scheduled validation catches theme updates, app injections, cache problems, broken templates, and endpoint drift that never touched the catalog record.

Prioritize by commercial impact. A missing marketing description is not equivalent to a wrong price or an unavailable variant presented as purchasable. Route high-impact conflicts to an immediate block or rollback, and route low-impact presentation issues to a repair queue with an owner and deadline.

End each run with an artifact: the affected variants, sources compared, exact field-level diff, validation state, and safe next action. That artifact is useful to the merchant and to any downstream agent integration because it explains not only that a row failed, but why it was withheld.