Dated, sourced and never paid for.
What 162 shopper searches returned from Shopify's agent catalog on October 26, 2026: 1,620 listings, which facts were present or blank, worked examples and a mistake we made.
Reconstruct our minus-30 normalization failure with raw passages, expected matches and regression cases before labeling a product fact absent.
Ten results can be one seller: a recount of dog-boot, backpack and tent shelves in Shopify's agent catalog by product ID, variant and seller.
Rebuild the failure of our first gap score using missingness, normalization errors and component contribution checks rather than a replacement grade.
Plain titles were cited almost every time, clever ones mostly weren't: a simulation of which article titles Claude, GPT and Gemini models cite and choose to read.
Audit absent, null, explicit-zero and positive review states with field paths and denominator rules before calculating summaries.
Use a proxy-to-outcome checklist to distinguish returned Catalog facts from live assistant retrieval, recommendations, seller choice and orders.
Recount first-variant checkout flags in the tested unpersonalized context and separate a false flag from all live purchase routes.
Vague listings were almost never picked; factual ones won when they answered the shopper's brief. Pick counts across Claude, GPT and Gemini models, by brief and product.
Every noise statement on a humidifier shelf in Shopify's agent catalog, traced to the field it came from: which listings gave a decibel figure and which only said quiet.