An AI shopping agent comparing jackets leans on the attribute fields in your product feed. If "hood" isn't in them, the agent may not know your jacket has one, however clear the photo is.
In January Google added dozens of new Merchant Center attributes for its AI shopping surfaces, covering common product questions and related items such as accessories and substitutes. OpenAI's commerce documentation asks merchants for a regularly refreshed feed with identifiers, pricing, inventory, media and fulfilment details, and says its required fields "ensure correct display of price and availability."
If AI shopping agents send you any holiday buyers this year, they'll be reading your feed, so check what's in it before the holiday catalog is locked.
Cleanups don't last
When a catalog has holes, the usual answer is a cleanup project. Someone exports the catalog and fills blanks for a month, and the gaps come back as soon as new products go in, because nobody touched the form that makes them.
Look at how a product gets into your system. Usually it's a form in the ERP or the storefront admin, and several people fill it in. Say the buyer creates the record when the order goes out, and a week later someone in receiving adds what they can read off the box. Each knows different things about the product, and the form lets them skip whatever they don't know.
That's how "Stainless Steel", "stainless", "SS" and "18/10 steel" all end up in the material field for products made from the same metal. Sizes sit in the title for some products and in a variant field for others, and the compatibility note sits in the description as prose, where a feed can't filter on it.
Change the form
Make the form refuse to save a product without the fields it needs.
Required fields should depend on the category. A tent needs capacity, packed weight, season rating and pole material, and a kettle needs capacity and wattage. If the form asks every product for every attribute, people skip the lot.
Wherever the answer comes from a known set, swap free text for a picklist: material, colour family, size system, compatibility with a named product line.
Give each field a single owning system. Price lives in the ERP and marketing copy in the storefront, and neither gets edited in the other, because two teams with different deadlines editing one field in two places is how the copies drift.
Everyone who adds products goes through the same form. Close the bulk-import route, or make it run the same checks.
Then let an agent work through the backlog
Once new products come in clean, the existing catalog is a finite job, and an agent can draft any field the spec sheet or photos show.
Give it the manufacturer's spec sheet and the photos next to the current listing. It drafts the missing attributes and moves compatibility out of the description into a proper field. Where the sources disagree, like a description that says 2 litres next to a spec sheet that says 1.7, it sends the row to a person.
Someone who knows the products approves each batch before it publishes. A wrong safety attribute can mean an injury claim. Keep batches to around fifty rows so the reviewer reads every one.
Start where the revenue is
Pick the category that sold most last holiday season, and within it the products that earned most of the revenue. Get every required field filled and consistent for those, move to the next category, and leave the long tail for January.
Then check the feed in Merchant Center's Needs attention tab, which highlights the issues with the biggest impact and has a Priority fixes filter. Fix the ones that stop products from showing, starting with your best sellers.
This week
Watch three different people add a product to your system without telling them how. Note the fields they skip and the values they type that should come from a list.
Change the form for your biggest holiday category and have an agent draft what's missing for its top products, with a product expert reviewing each batch. Check how the first batch shows up in the feeds before you start on the next category.
Frequently asked questions
What product data do AI shopping agents use?
Structured fields such as identifiers, titles, price, availability, variants and attributes like size, colour, material and compatibility. OpenAI's commerce documentation asks merchants for a regularly refreshed CSV or JSON feed, and Google has added Merchant Center attributes covering common product questions and related items.
Why does a product upload form matter for AI?
Most catalog gaps start when a product is created. If the form lets people skip fields or type the same attribute in different ways, every feed inherits it. Fix the form first, so next month's new products don't undo the cleanup.
Should we clean up the whole catalog before the holidays?
Start with the products that earned most of last year's holiday revenue and get those complete. Work through the rest of the catalog in the new year.
Can an agent fill in missing product attributes?
It can draft them from the manufacturer's spec sheet and product photos. Have someone who knows the products approve each batch before it publishes, because a wrong attribute leads to returns, or worse if it's about safety.
Where should product data live: the ERP or the storefront?
Pick one system as the source of truth for each field and let the others read from it. Otherwise the two copies drift apart.