CASE STUDY · E-COMMERCE BRANDS
A catalog graded the way an agent reads it, fixed revenue-first
A DTC brand with a large SKU count had a catalog written for human browsers. We scored every sellable SKU against the Agent Readiness Ladder and fixed attribute coverage on the products carrying the revenue first.
How to read these numbers
This client asked not to be named. The engagement shape is real; the figures below are directional estimates pending client sign-off, not audited results. (Directional estimate · client not named)
Top 10%
Of SKUs remediated first
Ranked by revenue, not alphabetically
~3x
Increase in structured attribute coverage
Directional; pending client sign-off
1 feed
Machine-readable product feed shipped
Validated against the schema contract
We had been writing product copy for shoppers for a decade. This was the first time anyone showed us what the catalog looks like to a machine.
Paraphrased from the engagement and shown without a named person while we wait on client sign-off. We will not attach a name or a title to a quote the client has not approved.
Context
- A mid-market direct-to-consumer brand selling across its own storefront and two marketplaces.
- Product data was maintained by a merchandising team optimising for on-page copy, not machine readability.
The problem
- Attributes an agent needs to answer a comparison question, material, fit, compatibility, dimensions, were inconsistent or buried in prose.
- No structured feed existed, so agent-mediated discovery had nothing reliable to read.
What we did
01 · Define
Scored the full catalog on the Agent Readiness Ladder and ranked gaps by revenue exposed.
02 · Design
A canonical attribute schema per category, with a validation contract at the source of truth.
03 · Build
Agent-assisted attribute extraction from existing copy and supplier data, with human review on every low-confidence field.
04 · Measure
Re-scored the catalog and tracked structured coverage on the top revenue decile.
Tooling shown is what was used in this engagement, not a required stack. We build to the platforms and systems you already run.
What we would do differently
We would agree the category schema with merchandising before extraction. Retrofitting two categories cost more than defining all of them up front would have.
NEXT STEP
Recognise the pattern?
Tell us where your signals live and we will tell you whether this shape fits.