I'm building the data layer AI shopping agents actually read — and nobody's measuring it yet

Background: I'm a physician and accountant by training, now building infrastructure for ecommerce catalogs. Odd path, but the habit of measuring before treating turned out to be the whole thesis.

The problem I stumbled into: I was working with a shapewear brand — good products, detailed pages, everything a human shopper needs. Then I looked at what an AI shopping agent could actually read from those pages. Most of it wasn't there. The category resolved to something generic. The attributes a machine compares against were empty. The meaning was locked in prose no agent can verify.

That's not a copywriting problem. It sits in the structured data underneath the store, and almost nobody measures it.

So I built an audit that scores it: what a machine reads from a product page today, across three dimensions — are the fields present, does the meaning resolve, can it be found and matched. Free, no signup, read-only on public pages. Then an optimizer that closes the gap by writing verified data back into the store.

One rule throughout: if the page doesn't say it, we don't write it. The field stays empty and we report which one. Generated product content that sounds plausible and contradicts the actual product is worse than a gap.

Where I actually am: paying customers, one documented case study, and a paper on the framing deposited on Zenodo. Two independent reviewers have gone at the instrument and found real bugs — one caught our enrichment escalating "protects against drops" into "drop-tested," which broke our own first rule. Fixed it, shipped it, and the review made the product better.

What I'm still figuring out: the honest evidence problem. Every lift number we report is a projection, not a post-deployment measurement. The experiment that would prove or kill the whole thesis — agent task-success on high- versus low-scoring catalogs, everything else held constant — is expensive and slow, precisely because the public surfaces are non-deterministic. I'd rather publish a small honest number than a big projected one, but that means going to market with less proof than I'd like.

Curious if anyone here has dealt with that: selling something you believe in while the rigorous evidence is still years out.