The First Thing I Got Wrong About Agent-Ready Product Data
I thought agent-ready product data was just a clean product feed. I was wrong. Here is the mistake I made, why it breaks agent checkout, and how I fixed it.
When I started, I thought agent-ready product data was easy. Just a clean product feed. Good titles, good descriptions, good images. Done.
I was wrong. And it took a broken checkout, in front of me, to teach me.
My mistake: I treated agents like humans or search engines
For years, product data has had two readers: humans and search engines. Humans want nice photos and friendly words. Search engines want keywords and tags. I have shipped plenty of both.
So I did the same thing for agents. I made the data read well. I wrote good descriptions. I added keywords. I felt quite pleased with it, honestly.
But an agent does not shop like a human, and it does not crawl like a search engine. An agent asks very exact questions, and it needs very exact answers. My pretty feed had none of them.
Where it broke
My test agent tried to buy a product. The order failed at the last step. The reason was simple, and a bit embarrassing: the agent thought the item was in stock at one price. The real store said something different.
To a human shopper, this is a small bump. You see "out of stock" at checkout, you sigh, you pick another item, you move on. To an agent flow, it is a wall. The agent had made a promise it could not keep. Trust broke in one second.
My data looked good. It just was not built to answer the questions an agent actually asks.
What an agent really needs
An agent asks things like:
- Is this exact item in stock right now?
- What is the exact price, and in which currency?
- Can it ship to this address, and by when?
- Which size or variant is actually available?
- Can it be bought right now, or not?
These are not marketing questions. They are plain yes-or-no facts. If the answer is missing or fuzzy, the agent either stops or picks wrong. Both are bad, and both are public.
The marketing words that help a human? To an agent, they are just noise. It wants clean facts it can trust.
The two fixes
I changed how I think about product data in two ways.
First: model the data around the agent's questions, not the human's page. I stopped copying the web page. Instead I wrote down the exact questions an agent asks, and made sure every one had a clear, structured answer — availability, price, currency, variant, shipping, and a simple "can this be bought now?" signal. Facts, not prose.
Second: keep it fresh and true to the real store. A feed that is even a little out of date will cause a failed purchase. So the data has to track the live store closely. I now treat "is this still true?" as the most important field of all. If the data and the store disagree, the agent fails — and it fails where everyone can see.
I also keep marketing copy and machine facts in separate lanes now. Humans read the first. Agents read the second. Mixing them was half of my original mistake.
Why this matters more than it looks
This sounds like a small data problem. It is not. Agent-ready data is the ground everything else stands on. If the facts are wrong, the best checkout protocol and the smartest payment system will still fall over. You cannot out-engineer a lie in your feed.
This is one of the first things we are building carefully into Sarvex: giving merchants product data an agent can actually trust, and keeping it honest against the live store. I learned why it matters the hard way, so the people using it do not have to.
The lesson in one line
Agent-ready product data is not a prettier feed. It is a set of exact, fresh, machine-readable facts, built around the questions an agent asks — not the page a human reads.
If you are preparing a catalog for agents, start there. For how these facts get used downstream, see agentic commerce examples. And save yourself one broken checkout.