1. Choose the questions a buyer asks
Not your category name, and not your product name — a buyer who already knows your product name is not the buyer you are losing. Write the question the way a customer writes it: "which 45-litre trekking backpack has a separate sleeping bag compartment", "a camera strap I can take off quickly between shots".
Five products, one question each, is enough to start.
2. Ask several models
At least four: the question goes to whichever assistant your buyer uses, and the share of useful answers varied from 52% to 83% between models in our own measurements (study). One model tells you about one model.
3. Record six things per answer
| Field | Why | | --- | --- | | provider and exact model name | behaviour changes under a stable name | | date | the same model answers differently next month | | was your product named | the first half of the outcome | | which URL was linked | the second half — and whether it is your page, your site, or someone else's | | any wrong fact | a wrong capacity or variant loses the sale as surely as absence | | was a web search used | an answer from memory and an answer from search are different measurements |
4. Repeat without changing anything
Run the same questions two weeks later, changing nothing on your site. The difference between the two runs is the noise. Without it you cannot tell a real change from the ordinary drift between answers.
5. Then change one thing
Fix one thing — identifiers in the HTML, a stable product URL, one missing fact — and run again. Changing several things at once makes the result unreadable: something moved, and you will not know what.
How to read the result
Count outcomes, not impressions: product named with your product page linked is the result worth having. A crawler visit, a mention without a link, or a link to your home page are steps toward it, and they are worth recording as what they are. The ladder we use for that is at /evidence.