Answer in brief
A practical, evidence-led worksheet for fashion stores: compare the visible product page, variant markup, merchant feed and returns policy before testing AI shopping answers.
The question is answerable only for a particular product
Can an AI shopper tell someone whether a jacket is available in medium and whether it can be returned? A fashion store cannot answer that responsibly with a generic claim that its catalogue is AI-ready. The answer depends on the exact size and colour, the destination market, the time of the stock check and the applicable return terms. This guide is a method for store owners, merchandisers and developers to test those facts before an assistant is invited into the buying journey. It is not a report of an audit of VITON13 or any other retailer. No SKU was sampled for this article, and no AI platform was measured.
The audit has four evidence surfaces: the visible product detail page, the page's rendered structured data, the exported merchant-feed row and the policy a buyer can actually read. Its output is a discrepancy register, not a single visibility score. Shopify describes its own agentic-commerce infrastructure as moving product information to connected AI surfaces, then verifying options, inventory and pricing on the way to checkout. That is a description of Shopify's ecosystem, not evidence that every assistant uses the same feed or answers from the latest version. A useful test therefore starts with the retailer's own controllable facts and treats a chatbot response as a separate observation.
Build a sample around variants, not parent products
Choose a small, dated sample that can fail in different ways: an in-stock core size, a sold-out size, a colour with a different image, a sale-price variant and an item with a returns exception. Record the date, time zone, country, currency, channel and the precise size-colour combination for each row. Keep the sample selection rule before looking at results so the audit does not quietly omit awkward items. Five variants are enough to demonstrate the worksheet, not enough to claim a catalogue-wide error rate. If the store has many warehouses or country-specific catalogues, treat those as distinct contexts instead of averaging them into one convenient answer.
A hypothetical record might be DEMO-COAT / black / M / destination France. This is a fabricated training identifier, not a real product or finding. Give that row one canonical internal variant ID, the shopper-facing URL after the selections are made and the feed ID sent to each channel. If a parent listing says the coat exists but medium-black is unavailable, a recommendation for the parent is not yet a correct answer to the shopper's question. Put the exact question in the worksheet: ‘Can I order black M for delivery to France today, and what would a return cost?’ The rest of the audit tests each clause of that question independently.
Record what the product page actually tells a buyer
Open the product detail page in a fresh browser session, choose the exact variant and wait for the page to finish updating. Capture the visible size label, measurements and units, fit guidance, selected image, active price with currency, stock message, add-to-cart state, delivery destination and returns link. Repeat on a phone-sized screen if a selector or policy link disappears at a breakpoint. Save the URL and a time-stamped screenshot; an internal catalogue field is not a substitute for what a buyer sees. A size labelled M may be a garment measurement, a body measurement or just a code. If the chart does not say which, the ambiguity is a product-content defect even if the feed faithfully says M.
Follow the purchase path far enough to detect changed claims, without submitting an order. Does the cart retain the chosen colour and size? Does a sold-out variant remain selectable after a refresh? Is the returns link attached to the product or hidden in a generic footer? Note whether a ‘free returns’ badge is scoped to the selected country and category. These are observations to collect in a real audit, not assertions about any store here. The visible page is the human reference point; it also helps identify when a technically valid data layer contradicts the offer the customer could actually buy.
Match each variant to the rendered structured data
Inspect the HTML a crawler receives, not only the source component in a repository. Google's Product variant documentation describes ProductGroup with variesBy, hasVariant and productGroupID as one way to connect size and colour variants; it also documents single-page and multi-page implementations. Locate the Product or ProductGroup node for the selected item and map its SKU, size, colour, image and Offer URL to the visible selection. Check the Offer's priceCurrency, price and availability against the product page. Record whether a variant URL selects the correct combination when opened directly. A generic parent Offer can make a page appear complete while leaving the exact black-M question unanswered.
Run Google's Rich Results Test as a syntax and eligibility check, then still perform the comparison by hand. A passed test does not establish that the displayed size chart is true, that inventory is current or that a third-party AI assistant will retrieve this markup. Google also says structured data should represent page content and documents return-policy data at organisation or offer level. If a site stores a global policy under Organization, the product's Offer should not invent a different window unless an actual exception applies. Save the rendered JSON-LD and the tool result with the audit timestamp; a developer can then reproduce a discrepancy instead of debating a screenshot.
Reconcile Merchant Center rows with the page
Export the merchant-feed rows for the same variant IDs and market. Google's Product data specification distinguishes the product's item ID from item_group_id used to group variants; it describes size, size_type and size_system where applicable, plus price, availability and landing-page link. For each sampled combination, compare the feed's size and colour with the selector, and the group ID with the parent product. Check whether the link lands on that variant or merely on a parent page that defaults to another size. For frequently changing stock and prices, record the feed generation and processing times as well as the page capture time. A difference caused by refresh delay is still a buyer-facing risk, but its remedy may be synchronization rather than copy editing.
Do not force all channels into an imaginary universal schema. Merchant Center requirements depend on product category and destination; other platforms can have different fields and update paths. The Google specification says apparel sizing is required in defined cases and availability should match landing and checkout pages and structured data. Shopify's article describes how its own catalogue supplies connected AI experiences; it does not prove that a standalone Google feed is read by every conversational agent. The worksheet should have separate columns for source system, exported value and observed value. That separation keeps a tidy feed from being mistaken for the entire commercial truth.
Read returns as a policy, not a badge
Open the full returns page from the sampled product, with the destination market recorded. Extract the return window and the event that starts it, the condition of the item, exclusions, shipping fees, refund method and the steps a customer must take. Ask whether a sale item, personalised item or hygiene-sensitive garment has different terms. Do not generalise from a homepage slogan: a ‘30-day returns’ badge without country, exception or fee context is not a complete answer. The audit is a content-consistency check, not legal advice; the merchant's legal or policy owner must approve the operative wording for each market.
Compare that wording with the structured return data and any Merchant Center return settings. Google Search Central documents MerchantReturnPolicy on Organization for a standard policy and offer-level markup for product-specific overrides, with fields such as applicableCountry, merchantReturnDays and returnFees. Those fields can help a supported search experience interpret declared terms, but they cannot explain every edge case in a long policy or certify that a warehouse will honour it. In the hypothetical DEMO-COAT row, suppose the page says ‘returns within 30 days of delivery’ while a feed setting says 14 days. Mark it as a hypothetical conflict requiring policy-owner review; do not choose the more attractive number for an AI-facing answer.
Turn discrepancies into fixes and a repeatable test
Use a register with one row per claim: variant ID, market, claim type, visible value, structured-data value, feed value, policy value, capture time, source URL, owner and resolution status. Classify a mismatch by customer impact rather than by which department owns the field. Wrong purchasability or return cost for a selected size deserves immediate attention; a cosmetic colour-name variation can wait if it does not change the item a customer orders. For the hypothetical black-M coat, the safe interim response is ‘availability or return terms could not be verified’ until the conflicting sources agree. The article makes no claim that any real merchant has this defect.
After a fix, refresh the page and feed, re-open the exact variant URL and repeat the shopper question in each AI surface you actually support. Capture the prompt, response, date, location or market settings and linked evidence. A correct response once is not a stable accuracy rate; assistants may retrieve different material, cache older data or decline to answer. Establish a small acceptance rule before testing: every sampled variant must display the same purchasability and price across the merchant's own surfaces, every applicable returns statement must have an approved source, and unresolved conflicts must be visible to a human owner. The practical result of this product-truth audit is a corrected source of record and a repeatable check. It improves the information an AI shopper can use, without promising that any particular model will quote, rank or recommend the product.
Practical checklist
- Select a dated, representative sample of exact size-colour variants and record their market and URLs.
- Capture visible size guidance, price, stock and the return-policy link after choosing each variant.
- Compare the rendered Product or ProductGroup markup with the selected variant and its offer.
- Export the corresponding merchant-feed rows and compare identifiers, size, price, availability and landing URLs.
- Read the full return terms for the test market, then record and retest every mismatch with an owner.
Questions and answers
Does valid Product schema prove that an AI shopper will quote the right size?
No. Validation checks whether markup is interpretable under a particular tool's rules. It does not verify garment measurements, the accuracy of a size chart, a live inventory change or what an unrelated assistant retrieved at answer time.
Should every colour and size have a separate product page?
Not necessarily. Google documents both single-page and multi-page variant patterns. The audit asks whether each selectable combination has a stable identity and consistent URL, offer and availability wherever the merchant exposes it.
Which return policy should the audit compare?
Use the policy that actually applies to the selected item, destination country and purchase channel. Check the shopper-facing wording, any organisation-level or offer-level structured data and Merchant Center settings separately; do not assume one automatically updates the others.
Can a merchant feed fix a wrong size chart?
No. A feed can transmit a size attribute, but it does not make a garment fit according to an inaccurate chart. Correct the underlying product record and human-facing guidance, then reconcile all exported surfaces.
What counts as a pass for the product-truth audit?
A pass means the sampled claims agree at a recorded time, exceptions are resolved by named owners and a buyer can find the same terms through the visible route. It is not a promise of placement, citation or correct answers on every AI platform.