Product Schema in 2026: From Required Fields to Citable Facts

Monday morning at a DTC camping-furniture brand shipping into Germany: an ops lead asks ChatGPT how much weight the folding table holds and how fast it reaches Munich. The answer cites a German gear-review blog and a marketplace listing. The brand’s own page is missing. The developer pulls the source: JSON-LD is there, the validator is green, but gtin is empty, price is typed into the template, and availability reads InStock on every variant.

Product schema earns its keep not by passing validation but by letting a model confirm four things — which product this is, what it costs now, whether it ships, and how it was reviewed — and missing any one drops the page back to “unverified”.

A green checkmark says nothing about truth

Validators check syntax, not truth: a page declaring a price of 0.01 gets the same checkmark as a correct one. “Required” is also a low floor — snippet eligibility and merchant-listing eligibility are different bars; clearing only the first leaves the page mute on what buyers ask.

A model does not read fields; it confirms facts. When JSON-LD, visible text, the feed, and a third-party page disagree about one price, the rational move is to stop trusting you. No error is raised; the citation is simply swapped.

Hard facts — price, stock, dimensions, material — compare character by character. Soft facts — positioning, use case — are only read whole. Schema carries hard facts, so one wrong one cannot be rescued by better copy; models match entities and relationships, not isolated fields, which is the work in the product graph behind a SKU.

Three layers, in a fixed order

Layer	Fields	Answers	Cost when missing
Identity	brand, sku/mpn, gtin	Which product is this	Variants merged; nothing to join on
Transaction	price, priceCurrency, availability, priceValidUntil, shippingDetails, hasMerchantReturnPolicy	Buyable now, at what landed cost	Price answers pull a third party
Trust	aggregateRating, review	How has it been rated	Filed as a new product, unreviewed

The sequence cannot be reversed. Get identity wrong and the other two are wrong too, quietly: six colors collapse into one product, reviews pool, and “4.8 stars, three colors” is false in every word.

Trust is the only layer where the rule is if it is not real, leave it empty. A series average pasted onto every variant buys a prettier page and spends the batch’s credibility; rating markup is advertising too, governed by platform policy. In a second market, GTIN and currency break before the translation does: one markup set across locales, prices pinned per template, two markets contaminating each other.

Five mistakes that survive validation

Mistake	Looks like	Consequence	Fix
Hardcoded price	Template constant, or cached for days	Disagrees with the page during a promotion	Inject from the price service at render
Stock always available	Out-of-stock pages keep InStock	Buyer lands on a dead page	Generate from inventory; use OutOfStock
GTIN missing, or an internal SKU standing in	Field is “filled” but joins nothing	Cannot merge with feeds or the graph	Use the manufacturer GTIN, or leave it blank
Rating borrowed across variants	One series score on every color	Mismatch with the page; policy risk	Mark at the real review granularity
JSON-LD only after JavaScript	Absent in source, added at runtime	Some fetch paths read an empty shell	Server-render or pre-generate

GTIN attracts the “just fill it in” reflex. It is the identity card issued by the outside world; inventing one is worse than a blank, because it binds your product to somebody else’s. The price row travels with the last one: compute in the browser, write into JSON-LD on the way past, and consistency has been outsourced to the client.

Verification is a regression test, not an audit

Step	Action	Owner	Passes when
1 Inventory	Export JSON-LD from every PDP; group by template, market, variant	Engineering	Ten URLs per template
2 Single source	Price, stock, rating keep one upstream, rendered with the HTML	Engineering	Change once, both copies move
3 Regression	Validation and consistency diff in the release pipeline	Engineering	Conflicts block the deploy
4 Retest	Re-run the prompt set; did the cited source change?	Growth	An explainable before-and-after

Most teams do steps 1 and 3 and skip 2, so every release fixes the same defect again. Consistency is not validated into existence; it comes from rendering both copies from one source. Step 4 gets skipped because it returns no red-or-green result, yet without it you prove the markup is legal, not that it was used. An AI visibility baseline shows how to build the prompt set.

Example: one SKU page, five changes

Example — a hypothetical folding table, three colors and two sizes, no real brand and no figures. Its JSON-LD carries a name, images, a price that never moves with promotions, an empty gtin, and an aggregateRating copied from the whole series.

Five changes: each of the six variants gets its own URL and its own Offer; price is injected from the price service with a priceValidUntil; availability follows inventory; GTIN comes from supplier data, and variants without one stay blank; the series rating is deleted, leaving only that variant’s real reviews.

Acceptance is three checks: JSON-LD appears in HTML fetched without JavaScript; one price change moves both copies; on retest, your domain shows up in price answers.

Where this breaks

First, categories that convert inside marketplaces: on Amazon or Shopee, answers lean on listings, reviews, and fulfillment, so cleaner PDP schema changes little. Fix the feed.

Second, categories with no GTIN — handmade, custom, bundles. Substitute brand plus MPN plus consistent titles, and accept poor variant merging.

Third, pages with no readable attributes or use cases. Schema translates existing facts; it cannot translate slogans.

Fourth, white-label goods and same-GTIN multi-seller listings. Complete markup makes you easier to slot into a price comparison — not always what you want.

Fifth, schema as a ranking switch. It changes the probability of being understood, not your weight; teams expecting citations from a few fields usually quit altogether.

This sits on the technical foundation line. If a page still has no sentence a model could lift, start with what a product looks like inside an AI answer; if you doubt AI entry points matter, read the four entry points first.

This month, do one thing: export the JSON-LD from your 20 best-selling SKU pages, point price and stock at a single server-side source, delete rating fields without real reviews, and add those three checks to the release pipeline.