AI shopping agents, search engines and recommendation systems read your product data and act on it — literally, with none of the judgement a human uses to read around a mistake.
CatalogueScore verifies whether that data is ready for them: correct, consistent and complete enough for an AI to trust.
It starts with a file you already produce — the Google Merchant Center product feed your team exports for Shopping. Nothing bespoke, nothing commercially sensitive.
The same product feed you give Google Shopping. One file. No system access, no integration project.
We model your catalogue as a structured set of facts and check each product — identifiers, brands, categories, prices, images, sizing — using open W3C validation standards. Every product, every rule, not a sample.
A readiness score, the issues that can be safely normalised, and a prioritised list of what needs fixing at source — named by product ID. Nothing is invented to fill a gap; missing data is flagged, never fabricated.
We don’t retain your catalogue or build a profile of it. You hold your own records and history; we provide the assessment.
A catalogue can look completely healthy on screen and still contain thousands of errors no one would catch by looking — a brand spelled three ways, a product in the wrong category, an invalid identifier. For years these were a quiet tax. AI changes that.
An AI doesn’t read a catalogue the forgiving way a person does. It takes every piece of data literally and passes it to the customer with complete confidence. The retailers who do well as AI commerce grows won’t be the ones with the flashiest assistant — they’ll be the ones whose underlying data is good enough for the AI to trust.
Make sure products carry the attributes an AI agent needs to match a shopper’s request.
Verify every product against explicit rules, with an itemised pass-or-flag result you can act on.
Re-check on a cadence, because catalogue quality decays the moment new products arrive.
We report what we measure. Missing or invalid data is flagged for you to fix at source — never invented.
Before running our engine against a real retailer’s feed, we answered a simpler question: does it reliably catch the errors it’s designed to catch? We built a controlled test feed of 220 laptop records, seeded 61 known errors across ten categories, and ran the engine blind against it.
Here are three of the contradictions it found — each a single product record disagreeing with itself. None would be caught by looking at the product page:
A controlled test on a synthetic dataset modelled on Google Merchant Center specifications. It demonstrates detection accuracy against known ground truth — a capability demonstration, not a study of any retailer’s data. The error rate reflects what we seeded, not a real-world finding.
Send us the product feed you already export, and we’ll assess it and report back — where your data is sound, where it isn’t, and exactly what to fix first. No obligation, and we don’t retain your data.
hello@cataloguescore.com