Valid schema can still leave an AI shopping agent guessing
I kept running into a weird problem while checking ecommerce product pages. The information was technically there, but an agent still could not use it with much confidence.
After a few audits, I started checking every product fact in three ways:
- Can a customer see it?
- Is it exposed in machine-readable product data?
- Does it agree with the other sources?
That catches things a normal schema check misses:
- ratings shown on the page, but no AggregateRating
- €19 on the product page and €20 in the feed
- InStock in JSON-LD while inventory says sold out
- a 30-day return policy on the site, but a final-sale rule on the product
- material shown in an image and nowhere else
The schema can be valid while the data is stale. The feed can be right while the page is wrong. Both can exist and still leave the agent guessing.
A single score hides too much. The useful part is the audit trail:
source -> fact -> conflict -> fix
For AI shopping, more copy is not the answer to this problem. The page, schema, feed, catalog, inventory and policies have to agree.
I am still working out how to prioritize these conflicts. Right now I treat price and stock as blockers, missing attributes as discovery issues, and policy conflicts as trust issues.
How would you rank them?