Services / Catalogue feed diagnostic

Catalogue feed diagnostic

Which setup can you safely use to enrich your product feed, across your catalogue, and what does it cost per 1,000 products? In about three weeks you get a ship, revise or reject decision, the evidence behind it, and a test suite you keep.

Talk to usSee CatalogBench

Who it is for

  • Retailers and brands using AI to write titles, descriptions and attributes, with people still checking the output before it goes live
  • Product-data software companies (PIM, feed management, marketplace tools) shipping AI enrichment to their own customers

A good fit if

  • AI enrichment is live, or in a funded pilot
  • Mistakes are costly: thousands of products a year, or channels that penalise bad data
  • A decision is coming: switching model, changing the prompt, a new category, language or channel, or cutting human review
How it works

Three weeks from receiving your inputs.

  1. Days 1–3

    Define

    Agree the decision, the categories and channels in scope, and what counts as a critical failure, such as an invented material or safety claim.

  2. Days 3–8

    Build the truth set

    Correct attributes for a sample of your products, checked by a person, with a share held back for the final check.

  3. Days 8–13

    Compare

    Run your current setup and up to three alternatives, including at least one cheaper model, under identical conditions.

  4. Days 13–15

    Prove

    Confirm the results on the held-back products, check the AI judge against your reviewer's decisions, and write up.

What you send

  • A sample of 300–500 products across up to five categories, with images and the raw feed
  • Your current setup: prompt, model and settings, or access to your pipeline
  • Channel rules, and brand and terminology guidelines
  • About two hours of one catalogue reviewer's time

Your products and images stay private. They are never used in our public benchmarks.

What you get

  • A short decision memo: the recommendation, the evidence and its limits
  • A scorecard for each setup: field accuracy, invented values, critical failures, image and feed conflicts caught, channel compliance, estimated publish-without-edit rate, and cost and speed per 1,000 products
  • A catalogue of real failures, grouped by type and severity
  • A regression suite, so the next model or prompt change can be checked quickly
  • A 60-minute readout with your team
Free

Start with a spot check

Not ready for a full diagnostic? Send us 20 of your products and your current AI output. We score them with the CatalogBench checks and send back a one-page result. We run a limited number each month.

Ask for a spot check
No AI enrichment yet?

We can build it for you

If you do not have an AI-enriched feed yet, Incremento, our agency, can build one for your catalogue. Before it goes live, we test it with the same checks as the diagnostic.

Talk to us about a build

Not included

  • Conversion or revenue impact: offline quality does not prove conversion
  • Rebuilding your pipeline
  • Ongoing monitoring (available as a follow-on)

See the method first

CatalogBench runs the same checks on public test products, for every major model. It shows the kind of failure the diagnostic finds in your own catalogue.

CatalogBench results →

Talk to us about your catalogue.

Tell us what you enrich with AI, and the decision you are facing.

Talk to us