Start without the product name.

Imagine the question: “Which jacket would suit a 30-minute run in the rain without getting too hot?” We would record the products returned, how they are described and any sources cited. If your jacket appears, that is an observation from this test—not a guaranteed position.

Then check the description.

A second question might name your jacket and ask about rain protection and comfort during a run. That helps us inspect its description. It cannot show that the system would have discovered the jacket unaided.

Find the detail worth checking.

Suppose an answer calls the jacket waterproof. We would check the brand’s specification and supporting evidence. If the page only claims water resistance, there is a discrepancy to investigate. We would not simply strengthen the page’s claim to match the answer.

Or perhaps the answer misses packed dimensions. First establish whether those dimensions are useful, verified and clearly published. Then identify a specific change worth testing.

Keep a record you can compare.

Our test record includes the exact prompt, date, platform and available model/context information, response, cited sources and product information at the time. We repeat the agreed questions and compare observations. Different answers alone do not prove that a page edit caused the change.

The jacket and questions here illustrate our method. They are not a published client test or a visibility benchmark.