From Field to Page: How Catalog Attributes Become Answers an AI Can Give
Follow seat depth from a supplier PDF to an attribute, a specs table, a feed field and an AI answer. Six steps, and where the chain breaks between teams.

Take one fact and follow it. A shopper asks an assistant for a sofa with a deep seat, "at least 23 inches". Your Harlow sofa has a 24-inch seat depth. Whether the assistant can say so depends on a chain of six steps, and the chain breaks quietly at any of them.
Step one: the fact exists somewhere
Seat depth is on the supplier's spec sheet, as a PDF, in a table that also covers the loveseat and the chair. It's correct. It's also, at this point, in a document that machines can index but struggle to parse into the right value for the right variant.
Breaks if: nobody ever extracts it. The PDF gets linked from the page as "Download specifications" and the fact stops here. This is the most common break, covered in where furniture product data hides.
Step two: the relevant family has a slot for it
Your sofa family has width, depth and height. It doesn't have seat depth. So even if someone reads the spec sheet, there's nowhere to put the number except the description paragraph, where it becomes prose rather than data.
Breaks if: the attribute isn't defined. Fix: create seat depth (and seat height, interior width, packed dimensions, fabric composition, cleaning code, frame, fill and lead time) as attributes and add them to the relevant family, with units. This is a one-time catalog design decision per category, and it's what makes everything downstream possible.
Step three: the attribute is filled, per variant
The 72-inch, 84-inch and 96-inch Harlow all have the same seat depth, as it happens, but the sectional's chaise doesn't. Seat depth has to be on each variant that differs, or inherited where it doesn't.
Breaks if: it's filled on the parent and wrong for a child, or filled on 40 products and empty on 1,960. Fix: fill it from the spec sheet (extraction from the PDF into fields, once), from the fabric supplier's data for fabric facts, or by measurement for the facts nobody wrote down. Across a range, this is bulk work and wants tooling: reading supplier files into attributes, flagging empties by category, sampling before committing. The point is it's done on the record, once, not on pages, repeatedly.
Step four: the page reads the attribute
The page template has a specifications table. The table is built from attributes, so when seat depth is filled, "Seat depth: 24 in" appears in the table on every sofa page. The FAQ template has a line, "How deep is the seat?", answered from the same field. The short description can mention it because it's generated from the fields, not typed.
Breaks if: the template has a hard-coded list of specs that doesn't include seat depth, or the table is an image, or the tab that holds it is fetched only on click. Fix: the template reads the attributes assigned to the relevant family, as HTML text delivered with the page.
Step five: the feed and structured data carry it
Product feeds and page structured data give machines a direct route to fields. If seat depth is an attribute, it can be mapped to a supported feed column and, where the implementation allows it, represented in the page's structured data. The visible page text should agree with the structured version.
Breaks if: the feed mapping was set up years ago with name, price, image and availability and nobody added the new attributes. Fix: map the attributes. This is usually a configuration task, not a build.
Step six: the assistant reads it and answers
The shopper asks for a seat depth of at least 23 inches. The assistant, working from your page text, your structured data or a feed, finds "Seat depth: 24 in" on the Harlow, attached to the right variant, and includes it in the answer with the reason.
Breaks if: any of steps one to five broke. Also if your site blocks AI crawlers, or the page needs script to render its content and the system reading it doesn't run scripts. And it's never guaranteed: assistants sample, differ and change, so a published fact can be found, but it isn't guaranteed a place. A fact you haven't published may still reach an assistant through a marketplace listing, a feed or a review, but you're no longer the source of it.
What the chain shows
Every step is a different team's responsibility in most retailers: buyers for step one, whoever owns the catalog for two and three, the web team for four, the feeds or marketplace team for five, and nobody for six. The fact has to survive five handoffs. Many don't, and when a fact doesn't arrive, each team can honestly say their part was fine.
The way to make the chain hold is to make steps two through five read from the same record. One attribute definition, filled once per product, with the page template, feed and structured data all pointing at it. Then adding a new fact for the whole range is: define the attribute, fill it, and it can flow to every mapped output. Changing a fact is: change it once, then check it published.
That is what "the catalog as the source" means in practice, and it's a different thing from a page CMS with a lot of fields. The existing guide to what product data fields furniture retailers need for ChatGPT Shopping and Google AI Mode lists what to carry; this is the path it travels.
How to find where yours breaks
Pick one fact you know is true about one product and that a shopper would ask about. Trace it: is it on the record as an attribute? On the page as text? In the feed? In the structured data? Then ask an assistant about it. The first "no" is where your chain breaks, and it's usually step two or three.
For who should own fixing it, see who owns being found by AI. For the whole argument, see being found by AI.




