A second store we operate — a German online shop for garden saunas, not yet open — run through the whole loop: questions written from its own catalogue, a baseline on five assistants, and pages drafted from the questions it lost. It is labelled as ours wherever it appears.
Figures from one product run on 9 October 2026; the store is not named because it is a separate company.
Answers citing the shop. Expected for a site that is new and not yet open to search engines — the report states the ceiling instead of a shrug.
What the baseline cost us in API fees across ChatGPT, Gemini, Perplexity, Google AI Overviews and AI Mode.
Facts the four drafts marked as unverified. Publishing is refused while any marker remains.
Pages published. The other seven questions were already covered, local, about products the shop does not sell, or health-adjacent.
The tool read the shop first and wrote twelve German buying questions from its own products. One was medical and was dropped by hand; eleven were measured, each asked repeatedly on every assistant.
Two large garden-building retailers, a gardening forum, YouTube, a sauna manufacturer and a classifieds site — 331 cited pages, all read. That list is the brief: the shape and facts of the pages the assistants already trust.
Seven of eleven lost questions were rejected as subjects. A tool that drafts whatever it lost would have published pages about products the shop cannot deliver.
Of 75 flagged facts, 45 were filled from the shop's own data and 30 were deleted — prices for heaters it does not sell, and placeholders. That is why nothing publishes with a marker left.
A zero with a ceiling is a baseline. Without it, the first good month after launch proves nothing.
The cited pages tell you what to build — length, comparisons, the facts buyers check — better than any keyword list.
Drop lost questions you cannot answer honestly. An assistant that sends a buyer to a product you cannot deliver costs you twice.
After launch the same eleven questions run again, and the report will say "changed", "within noise" or "new baseline" — whichever the test returns.