Short answer
Check product facts for a specific variant. Compare visible information, structured data and the AI service's answer. Record sources and check times for price, stock and suitability. Keep missing evidence visible as an open question.
1. Define a specific purchase decision.
Choose a product sold in several variants. Record the selected size, colour or version, delivery country and buyer requirements. Test that exact variant. Product-family information may not be enough.
A useful question is: Does this version fit my device, is it available, and what extra costs apply? Add the exact device model and version. Establish the expected answer from shop and manufacturer information before testing. If those sources do not provide evidence, leave the question open.
2. Collect evidence for this variant.
Create a short checklist. Record the source, affected variant and time checked for each fact. Product specifications, shipping rules and return terms may live on separate pages. Check that they describe the same purchase.
- Identity: the product name, model and SKU match the selected variant.
- Suitability: dimensions, materials, connectors or other required properties are stated explicitly, with their limits.
- Offer: price, currency and stock apply to this version. A starting price is not a confirmed variant price.
- Delivery: costs, destination and stated timing match the delivery country. A general promotional claim cannot replace these details.
- Returns: linked terms state the period, costs and exceptions where they apply to the product.
3. Check what happens when the variant changes.
Open the selected variant's link in a new browser window. Is the same version selected? Do the image, SKU, price and stock still match? Switch size or colour and check again. This exposes information that remains unchanged when it should update.
Compare visible text with structured product data. Google documents ProductGroup with hasVariant, variesBy and productGroupID for grouping variants. This markup helps identify relationships; it does not establish suitability for a buyer or a recommendation by an AI service. The implementation also depends on whether variants share a page or use separate pages.
4. Ask questions that expose mistakes.
Use the same questions for the manual check and for the AI service you want to assess. Save its answer and linked sources. Record a correct statement without matching evidence separately from an evidenced answer.
- What is the selected version's SKU? Expected: that variant's exact identifier, not a similar model's.
- Does it fit the device I named? Expected: an explicit compatibility statement, or a clear statement that compatibility is unverified.
- Can I order this size now? Expected: the stock checked, with no silent substitution of another size.
- What shipping costs apply to my country? Expected: the matching rule, or an open answer if an address or cart is needed.
- What limitation could rule out this purchase? Expected: known restrictions, not just a list of benefits.
5. Separate missing facts from misread facts.
If the shop lacks a needed fact, add it from a reliable source. If the page and markup disagree, fix their shared data source. If the fact exists but the AI service assigns it to the wrong product, check its placement and variant label.
Example test case: two sizes have different prices, but the answer gives only the lowest price. Check the variant in the link, then the visible price, then the offer in the markup. This narrows down the error.
6. Keep a record others can check.
Each record contains the date, product URL, variant, question, expected answer, evidence location, actual answer and finding. For AI tests, also record the provider, displayed model and whether you used a new session. Save prices and stock with the check time so later changes do not look like reading errors.
Use four findings: correctly evidenced, incorrect, unsupported or not testable. Count the cases in each group and state the scope. Repeat the same tasks after a change. A few correct answers do not establish general reliability or increased visibility.
Then check the handoff to the cart.
Start with a frequently purchased product and a variant with an important restriction. This keeps the first check manageable while covering more than a favourable standard case.
Once the product facts agree, check that variant, quantity and price remain correct in the cart. Research ends with an evidenced choice; a purchase also needs clear approvals and reliable order steps.
FAQ
Is structured product data enough?
No. Also check visible information and the selected variant. Technically valid markup can contain outdated prices or facts for another version.
How should I handle missing compatibility information?
Mark it as unresolved. A similar model number or general marketing claim does not establish compatibility. Add information only from a matching manufacturer or shop source.
Does a correct AI answer prove the change worked?
It supports that test case only. Repeat the same tasks, record errors and sources, and log other changes. Search visibility and revenue effects remain separate measurements.
Want to know where the purchase gets stuck?
Book a 30-minute diagnostic. We check one buying path and name the first fix worth reviewing.
Next practical step
Sources and research basis
Related guides
Checkout Readiness
Checkout readiness: rules agents can complete
Test variants, totals, delivery, approval and order status. A practical checklist with failure cases and repeatable tests for AI agents.
Measurement
Measuring Agentic UX before AI traffic shifts
A practical framework for measuring Agentic UX in ecommerce: journey tests, answer coverage, evidence quality, checkout completion, QA routines, and governance.