What makes a B2B catalog ready for AI agents
By Jason MurphyPublished 3 min read
The short answer
An agent-ready B2B catalog states every fact once, agrees with itself on every surface, carries a unit and basis, names its source and confirmation date, and says who owns it. An agent cannot ask a clarifying question, so each of these has to be answered in the data before it reads.
In this article · 6
What an agent is trying to do
An agent acting for a buyer wants to find items that meet a requirement, confirm they do, compare suppliers and prepare an order. Each step needs a specific answer and none can be resolved by asking a person. Agent-ready means the answers are already in the data and the data can be trusted to agree with itself.
The criteria
- One identifier per item, mapped. Manufacturer part number, supplier SKU and GTIN where one exists, with the mapping between them stated.
- Agreement across surfaces. The website, the PDF, the feed and the marketplace listing state the same value for the same fact, or the difference is explained by a basis.
- Units and basis on every measurement. Value, unit, nominal or actual, tolerance where it applies.
- Revision and effective date on every specification and document, with superseded versions marked.
- Availability with a timestamp, and unknown as an allowed value.
- Source on every fact: a system, a document, a named person.
- An owner on every kind of fact, and an authorization statement for who may sell and install.
- Compatibility stated, not inferred: a relation between parts exists only where a source says so.
What the facts that change need
Some facts have a short clock. Stock, lead time, price, and promotions can be wrong within a day. These need a visible observation time and an explicit unknown when the reading is stale. A confident old value is worse than an honest gap. Static facts, such as a dimension on a given revision, need a revision and nothing more.
How to test it
- Pick a sample of items across product families and revisions.
- For each, collect the value of every criterion from every surface: site, data sheet, feed, distributor listing, marketplace.
- Mark every disagreement and every missing unit, basis, revision, date, source or owner.
- Ask the assistants questions about the same items and compare each claim with the record.
- Rank the failures by who can fix them and how many items they affect.
What it is not
Agent-ready is not a visual redesign, a chatbot, or a markup exercise. It is the condition of the underlying record. Structured data on a page helps machines read a value, and says nothing about whether the value is right. The work starts in the record and the chain that publishes it: the product information chain, part numbers and measurements and what AI agents need from a supplier.
Questions
Is agent-ready the same as having structured data?
No. Structured data lets machines read a value. Agent-ready also requires the value to be correct, consistent, dated and owned.
Why does an agent need an explicit unknown?
Without it, a missing value may be filled with a guess. An explicit unknown is treated as a gap to resolve with a person.
How big a sample is enough to test a catalog?
Enough to cover each product family and each current revision. The goal is to find the classes of failure, not to count them.
Who fixes the failures?
Whoever owns that kind of fact. The ranked list groups failures by owner so each party sees its share.
Labs · Field notes
Labs · Research
Labs · Field notes
- What AI agents need from a supplier
An agent comparing suppliers cannot call to ask. It needs the answer to already be data.
What AI says and why3 min read
- The product information chain: from PIM to page to feed
A spec is copied six times before anyone reads it. Each copy is a chance to change it.
Product data3 min read
- Part numbers and measurements that disagree across systems
The same item has four identifiers and three kinds of size. Systems agree until a person or a machine has to pick one.
Product data4 min read
- Structured data that shows where a fact came from
Markup tells software what a value is. A source and a date tell it whether to believe it. The vocabulary has room for both.
Product data3 min read
- Does ChatGPT read PDF spec sheets? The PDF problem
AI can read a public PDF. It is slower to trust an undated one two clicks from the page. Put the key facts in text and keep the PDF as proof.
Product data3 min read
- Can AI understand your product specifications?
Assistants do well with specs stated in text, and badly with values that depend on the configuration, lapsed approvals and series names only your reps know.
Product data3 min read
Where does your own information disagree?
Get Upstream
One email a week: what we asked the AI, what it got wrong, and how we checked.
Our research reaches 17,000+ followers across our channels. Upstream is the weekly version.
Weekly. Unsubscribe anytime. Privacy