A product can be easy to find and still be difficult to buy with an AI assistant. Practical Ecommerce's five-test guide separates the problem into identification, proof, verification, supplying evidence and shopping. The useful outcome is a list of facts a merchant can correct and test again.
Identification asks whether the assistant understands the product and its intended use. Proof asks what supports a claim about it. Verification checks the answer against the product's actual specifications and terms. Supplying evidence asks whether a reader can reach the supporting source. Shopping checks whether the recommendation leads to the right purchasable item. Record the assistant's answer and links at each step so a plausible explanation does not substitute for a working purchase path.
Consider an illustrative test for a 750 ml insulated bottle. Ask for a bottle that fits a cup holder with a 7 cm opening. If the product page gives capacity but no base diameter, the assistant cannot verify the fit from that page. An answer that says it fits has identified the product but has not established suitability. Add the measured base diameter to the page and relevant product data, then repeat the question.
The evidence check goes further: open the cited page and confirm it supports the particular claim. A link to a general collection page may identify the brand while saying nothing about the bottle's dimensions. The shopping check then follows the exact variant through to the cart, checking its price, availability and delivery terms. A correct recommendation for an unavailable size still leaves the buyer without a usable answer.
Use the same prompt, assistant mode and location for the follow-up, and keep several observations rather than treating one improved answer as proof of a lasting change. A merchant's own page is one possible source among several, so record any external page the assistant used as well. The product-page checklist explains how to keep visible facts, structured data and feed information consistent.
The practical implication is to fix the first unsupported step in the buyer's task. “The assistant named us” and “the buyer could verify and purchase the right item” are different results, and the five checks make that difference observable.
Common questions
What should a merchant record for each test?
Keep the exact prompt, assistant and mode, date, location where relevant, answer, cited URLs and product variant. Record whether each check passed and the specific evidence for that decision.
Does a missing recommendation mean the product page is unreadable?
No. It may reflect the request, availability, competing products or the sources the assistant selected. Inspect access and product facts, then compare several equivalent observations before attributing the omission to one cause.
How can a team tell whether a fix helped?
First verify that the corrected fact is visible and consistent across the page, markup and feed. Then repeat the relevant buyer task under comparable conditions and report the observed results with their sample size.
Is appearing in an answer enough to pass?
No. The answer also needs to support the buyer's constraints and lead to the correct offer. A recommendation with an unsupported fit claim, wrong variant or unavailable item should remain a failed task.