What we read
Four public sources, in one pass: the store’s /products.json feed (up to 250 products), /robots.txt, /sitemap.xml, and the rendered HTML of a single product page for its structured data.
Nothing is authenticated. We do not touch the admin, orders or customers, and there is nothing to disconnect afterwards because nothing was connected. The feed is a documented part of every Shopify storefront.
How the score is built
19 checks sit in four layers. Each check returns a ratio between zero and one rather than a pass or fail, so a store with 88% of variants carrying a SKU is not scored the same as one with 12%. Checks are weighted within their layer, layers are weighted against each other, and the result is rounded to a single number.
Layer weights are retrieval 30, structure 28, substance 26, trust 16. The ordering is deliberate. A store an agent cannot reach gets no benefit from immaculate product copy, so retrieval dominates. Trust signals matter but are the easiest to improve, so they carry least.
Known limits
It is a sample. The feed returns 250 products per request. Stores larger than that are scored on the first 250, which is usually recency-ordered rather than representative.
Structured data is checked on one page. Themes are consistent, so one product page is a fair proxy, but a store with hand-edited templates on some products can score better or worse than the sample suggests.
Some fields are invisible from the feed. Image alt text is the clearest example: Shopify omits it from /products.json entirely, so it cannot be measured this way and is deliberately not scored. A check that reports zero for every store is not a finding, it is a bug.
Spec detection is category-blind. The substance check looks for measurable figures such as sizes, weights and volumes. Some categories genuinely carry fewer of these, and will score lower without being worse.
No agent publishes its ranking function.These checks are built from schema.org, Google’s merchant listing requirements and published crawler documentation. They are a well-grounded proxy for machine legibility, not a reconstruction of any specific system’s scoring.