Agentic Commerce Fails on Thin Product Data

    Agentic Commerce Fails on Thin Product Data

    Merchi Team •

    Someone asks a shopping agent for a jumper they can wear to the office, in a cloth heavy enough for November, and a tap that will fit the pipe in the wall. The agent is ready to help. Your catalogue offers it a title and a photograph. The retailer who published the cloth weight and the bore size gets the recommendation.

    That is agentic commerce without the briefing. An assistant, on your site or inside ChatGPT, Gemini, or a similar tool, chooses a product or places the order. It can only choose what it can read.

    The agent is matching fields, not admiring the brand

    A person will skim a paragraph and still take a chance. An agent matches the request to fields: material, weight, size, compatibility, care, price, and whether it is in stock. An empty field is nothing to match. A sentence about premium quality does not contain a cloth weight.

    Google Shopping already runs a version of this test. A missing required field gets the product suppressed. An agent that cannot answer the question moves on to a product that can. Filling those fields is product data enrichment, and it is the work that has to happen before either channel asks.

    A guess is worse than a blank

    When the specification is missing, the product is skipped, or something fills the gap. The fill is a guess. A guessed bore size, a guessed tog rating, or a guessed age band means the shopper receives the wrong product, and the return arrives with your name on it.

    Publish the values you can stand behind. In merchi.ai a confidence score sits on every field, so a number taken from the supplier PDF is not treated like a number inferred from a blurry photo. If the source did not state it, the field stays empty. Document upload is how that PDF gets into the record.

    The record needs a shape, and a source

    An agent cannot depend on every description being a fresh piece of prose. It needs the same fact in the same place on every product. Collar, rise, and cloth weight for a shirt. Load rating and bar fit for a plate. Wear layer and room suitability for a floor. Those are schema blocks: the fields you decided matter, filled the same way across the range.

    Later, someone will ask where a value came from. A buyer, a marketplace, or the agent. AI provenance records that the content was generated, which model did it, and whether a person has reviewed it. A shopping agent should prefer a reviewed specification. You can only offer that preference if the trail exists. The standard behind the trail is the AI Provenance Protocol.

    Writing the catalogue and shopping from it are different jobs

    merchi.ai already runs an agent that processes products. It takes images and supplier files and writes the attributes, the copy, and the categories. That agent is not the one that shops. The writing pipeline is covered in agentic AI in retail. Agentic commerce is what happens afterwards, when a shopper’s assistant reads the result.

    The shopping agent itself belongs with the commerce platform, or with a specialist. It still fails when the catalogue is thin. Preparing the catalogue is the part a retailer can do now, and it pays off in ordinary search and in feeds before any agent is switched on.

    A UK flooring retailer reached that point with a backlog of 1,000 products that could not be published, because the descriptions and attributes were not there. Once the records were complete, those products could be found, and online revenue grew 976%. The detail is in the case study. The result did not depend on a shopping agent. It depended on a catalogue an agent would be able to read.

    What to finish before you brief an agent

    Agree the fields for each product type. Fill them from the image, the supplier file, and the data you already trust. Keep a confidence score, so a blank stays a blank. Keep the provenance, so a reviewed value can be told from a fresh generation. Then point an assistant at the catalogue.

    If the range is still a folder of PDFs and a spreadsheet of titles, an agent project will demonstrate the gap and call it a failure of AI. The gap is the product data.

    Book a 20-minute walkthrough and we will take one of your products and show the record an agent would actually see. Or start a 30-day free trial and run it on your own catalogue.

    Frequently Asked Questions

    What is agentic commerce?

    A shopping agent that chooses a product, or buys it, for a person. It might sit on the retailer’s own site, or inside an assistant such as ChatGPT or Gemini. It matches the request to product data. It does not invent a specification the catalogue never stored.

    Why does agentic commerce need structured product data?

    The match is made on fields. Cloth weight, bore size, compatibility, and care are the match. A paragraph of brand language does not contain them, so an agent cannot use it to decide.

    What happens when the attributes are missing?

    The product is passed over, or a gap is filled with a guess. A wrong guess becomes a return. A blank field, marked as unknown, is the safer record. Confidence scoring is how merchi.ai keeps those two apart.

    How is agentic commerce different from the agent that writes the catalogue?

    One agent writes the product record. Another reads it in order to shop. Agentic AI in retail explains the writing pipeline. This article is about the reading.

    Will a shopping agent complete missing specifications?

    It may produce an answer. That answer is a guess unless it came from the supplier file or a person. Specifications should be extracted, scored, and reviewed before an agent is allowed to treat them as facts.

    How should a retailer prepare a catalogue for agentic commerce?

    Decide the fields, fill them, score them, and keep a provenance trail. Product data enrichment is that preparation. An agent project that follows it has something to read.