What the static scan checks
Surfaced begins with a public, repeatable scan of a Shopify storefront. The scan checks whether meaningful pages can be reached without a login, whether basic crawler directives are readable, and whether key product facts are expressed in the page and its structured data. It looks for obvious conflicts between canonical URLs, redirects, and the public destination a shopper would use. It can also flag the presence of agent-facing guidance and link to the pages that guidance names.
The scan is intentionally narrow. It does not crawl private account pages, infer inventory that is not public, or claim that a product is eligible for a particular answer experience. A pass means a signal was observed at scan time. A fail means the signal was absent or inaccessible under the stated check. A warning means the signal needs a closer human review.
How AI sampling works
Answer-engine monitoring is a separate measurement. We define a set of retrieval-backed prompts for a category, run them multiple times each week, and record whether a merchant’s store appears as a retrieved source, citation, or named result where that surface is available. We retain the prompt, timestamp, model or endpoint identifier, and the observable source evidence. Results are aggregated as an appearance rate across the sampled runs.
This approach is designed to show directional change and compare a store against its own baseline. It is not a census of every answer produced by every model. Prompt wording, location, model updates, retrieval indexes, personalization, and product availability can all change a result. We therefore avoid language such as “what users actually see” or “guaranteed ranking.”
How to read a result
Treat a result as a starting point for diagnosis. A low access score suggests checking directives, redirects, and canonical pages. A structure issue suggests checking whether visible product facts and JSON-LD agree. A lower appearance rate may suggest that the relevant evidence is missing, weak, or less useful than another source for that prompt. It does not prove a causal ranking factor.
What we publish in reports
Every public report will list the sample size, collection period, prompt scope, observed surfaces, and material limitations. We will not publish fabricated benchmarks or imply a larger sample than we collected. The useful claim is modest: Surfaced reports sampled monitoring, not a universal view of private AI experiences.