Episode 14: Why AI Search Keeps Ignoring Your Products

A conveyor motor fails at a mid-size manufacturer. The replacement has to be IP65 rated, five to ten kilowatts, three phase, EU voltage compatible and fit for outdoor washdown. The procurement manager has no time to work through catalogues or email five suppliers.

So they ask an AI agent, in plain English, and get back two specific models with a note about checking lead times.

If your product was not one of the two, you did not lose on price, or service, or relationship. You were never in the conversation, and nobody in that business will ever know you existed.

That result was not magic. It came from structured, interpretable data. AI does not guess. It reasons, and it can only reason with what you give it.

The buyer stopped being a person

We have lived through two versions of this already. Early ecommerce search was passive: you had to know what you wanted, type the keyword, filter and scroll. If you did not ask, you did not get. Then came the recommendations era, where systems started nudging with people also bought and you might also like, driven by behavioural data but still resting on surface level tags and hierarchy.

Agents are different in kind. You express a goal, a need and a constraint in natural language, and the system interprets it. Not IP65 motor as a keyword, but a quiet, weatherproof motor for a food-safe environment, available next day.

They do not browse category pages. They reason a shortlist and explain it back to you.

So the journey no longer starts with a click. It starts with an expression of intent. The agent translates that into a structured query, works out what the buyer is actually trying to solve, and matches it against a product data graph of specs, tags, compliance and compatibility. That is the moment your product becomes visible or disappears. Then it recommends you or skips you, and the buyer never knows you were in the mix.

You are not optimising for search results any more. You are optimising for intent answers.

Structured is not the same as smart

There are three levels of product data, and most businesses are stuck on the first.

Descriptive data is the average record in a legacy ERP. A title and a description, and that is it. It was written for a human, so there are no attributes, no standard units, no values. The system cannot tell the IP rating, whether the product is waterproof, whether it is single or three phase, or what the frame size is. Vague to a person. Completely invisible to a machine. This is what most product data still looks like, especially in B2B.

Structured data is what most people picture when they hear the term. Defined attributes, values in the right format, units and standards in place. It powers filters, comparison tables, ecommerce feeds, PIM systems and quoting tools. It is a genuine step forward.

But structured only means readable. The machine now knows what the product is. It still cannot tell whether the product solves the buyer's problem, because there is no application logic, no context and no connection to intent.

Intent-ready data is the difference between here is a motor, and here is a motor that runs outdoors, resists humidity, works with your controller, ships next day in France, and will not void its warranty if you pressure wash it.

The questions AI is answering are not only what is this product. They are: is it suitable, is it available, and is it supported. That pulls in application context, environmental conditions, compatibility, operational detail, logistics, warranty, regulatory fit and real world proof.

Structured data gets you into the conversation. Intent-ready data gets you shortlisted.

This is already live

None of this is a prediction. Between October and January last year, Amazon was the top domain referred by ChatGPT, for the straightforward reason that its product data is structured in exactly the way these models can consume.

Google's search generative experience is rolling out across a substantial share of searches. Search it for a good bike for a five mile commute up hills and you are not entering a keyword, you are expressing intent that implies terrain, use case, budget sensitivity and comfort. What comes back is a summary of the factors that matter, a shortlist of products, and snippets pulled from specs and reviews. It runs on the Shopping Graph, a data set of some 35 billion product listings. That is the queue you are standing in.

ChatGPT has quietly become a discovery engine as much as a writing tool, returning product cards with real items, real pricing and real sources. OpenAI's own position is that any merchant can appear, provided the data is discoverable, structured and reachable by their crawler. If yours is unstructured, incomplete, hard to parse or sitting behind a portal, which describes a great deal of B2B product data, you are simply not in the index.

Perplexity has gone furthest, moving from answering questions to completing purchases. Ask it for the best price on a specific watch and it surfaces structured pricing and spec data, returns a shortlist, and offers checkout inside the agent. Retailers and distributors can submit products through its merchant programme and influence what buyers see.

This is what we mean when we say your product data is now your storefront. There is no page layout in that model. No imagery, no brand styling, no calls to action. Just machine readable data, and if yours does not tell the story, you do not appear.

How to get your data ready

The same process applies whether you hold a thousand SKUs or a million, and the order matters. Skip the foundation and everything above it collapses.

Build the foundational schema. Define the right attributes, units and values for each product type. This is the blueprint that tells a machine what your products are, and without a shared structure no amount of AI will rescue you.

Audit what you actually hold. Against that schema, work out what is missing, what is inconsistent, and what is unusable by a machine. You cannot fill a gap you have not measured.

Enrich with facts, not guesses. Use supplier input, internal sources, subject matter experts and AI tooling to close the gaps. This is the heavy lifting, and it is genuinely intensive in time, cost and effort. Pretending otherwise is how these programmes stall.

Align the data to buyer intent. Do not stop at spec attributes. An expression of intent is an outcome, a set of constraints and a context, so tag products with use cases, applications and buying context to match. That is the material AI actually uses to recommend or reject you.

The takeaway

The uncomfortable part of this is the tension between where the technology has got to and where most catalogues are. Agents are already researching, comparing and, increasingly, buying. Most businesses are still trying to get basic attributes structured, let alone the contextual data these systems reason over.

The gap is not going to close on its own, and it is not a marketing problem. No amount of content, campaign spend or site redesign compensates for a product record that a machine cannot read.

It is also not about future proofing, which implies time you do not have. Your buyer is already an agent in a growing share of cases, and that share is only moving one way.

So the question worth taking to your leadership team is a short one. Is our product data AI ready? If nobody can answer it with evidence, that is the answer.

Listen to the full episode

Episode 14 of Product Data Weekly is available now. For more episodes and the weekly newsletter on operational issues inside product data and ecommerce teams, visit productdataweekly.com.