ChatGPT product feed optimization: the fields that decide what gets shown

ChatGPT product feed optimization is the work of filling OpenAI's product feed with the facts shoppers ask about, in the fields built to hold them. Nine required fields get a product into ChatGPT. The optional ones, such as material, size, age group, dimensions, weight, reviews and returns, decide which shopper questions that product can answer.

Getting your products into ChatGPT is getting easier. Shopify and Etsy catalogs are integrated by default, and any other merchant can apply to share a feed through OpenAI's merchant program. Once most catalogs are in, just being listed won't set you apart. What will is what your product record says when a shopper asks for something specific.

This guide goes through OpenAI's product feed spec field by field: what each one is for, what to write in it, and the gaps we see most when retailers export their catalog as is. For how an assistant narrows a question down to a shortlist in the first place, see how AI shopping assistants pick products.

What does OpenAI's product feed require?

Nine fields, and every one needs a real value. OpenAI asks for “a real brand and seller name, not placeholders,” and a stable item ID per item or variant.

The nine required fields in OpenAI's product feed spec, and the most common way each one goes wrong.

Which optional fields answer shopper questions?

The optional fields are where ChatGPT matches the details a shopper asks for. When someone asks ChatGPT for a machine-washable merino sweater for a ten-year-old under $60, three of those four constraints have a field: price, material and age_group. The fourth, machine-washable, doesn't have a field, so it only gets matched if the description mentions it.

Optional fields in OpenAI's product feed spec, mapped to the shopper language each one answers.

Two rules in the spec trip up a lot of exports. First, “for optional fields, omit an unknown value.” A feed that writes “n/a”, “null” or “unknown” into every empty attribute is filling fields with noise. Second, “raw review entries and question-and-answer lists are not part of this discovery contract.” Anything shoppers usually learn from Q&A, such as care, fit notes or what a product works with, has to live in the description or on the product page.

For which attributes matter most by category and by surface, see our guide to product attributes.

What should go in the title and description?

Facts, in the order a shopper cares about them. OpenAI's guidance here is short: titles name the product “including the selected variant when relevant,” and descriptions are a “factual product description for this item.” OpenAI's help page on shopping adds that ChatGPT may generate simplified titles and descriptions to make results easier to read. But a rewrite can only work with the facts it's given. It can't add ones that were never there.

We saw this in a feed review this September with a large apparel and merchandise seller. One of their jackets came back with “black/black” removed from the title and the waterproof rating, 11,000mm, added. Their team asked why the rating went in and not the fit. The answer had to do with the whole record, not just the title. Google now looks at how complete the whole feed is, not only the title and description. So the title should say what the product is and the one fact that sets it apart, and the rest goes into the fields built for detail, without repeating itself. In that meeting we suggested keeping the description concrete about the product. Fit, use cases and technical detail went into the product details and highlights, each said once.

The same applies to ChatGPT. If your title leads with product type, brand and the one attribute that decides the purchase, any rewrite has something accurate to work from. A description that opens with the facts (material, fit, use, care) gives ChatGPT a way to match the requests that don't have a field of their own.

How should product_category be written?

Start from Google's product taxonomy and go one level deeper where Google stops. OpenAI's spec describes product_category as “your category path, from broad to specific, separated by >,” and its own example, “Apparel & Accessories > Shoes,” is a Google taxonomy path. Most retailers already map to Google's taxonomy for Merchant Center, so start there.

The catch is that Google's taxonomy stops early in places shoppers do not. It has no separate node for sandals, boots or slippers: all three sit at “Apparel & Accessories > Shoes.” It has no node for sweaters, polos or jeans either. A path that ends at “Shoes” tells ChatGPT nothing about whether the product is a sandal, so that difference has to be carried by your own deeper path (“Apparel & Accessories > Shoes > Sandals”), the title and the description. Our product taxonomy guide has a table of common products and where Google's taxonomy places them.

How do you keep a ChatGPT feed accurate?

Update price and availability when they change, not on a weekly export. The spec says to “update the feed when a sale starts or ends” and to update availability “when the item becomes available or sells out.” OpenAI's help page notes there may be a delay before updated pricing appears, and that when several merchants sell the same item, ChatGPT ranks them on factors including availability, price, quality and whether the merchant is the maker or primary seller.

That last one matters if you're a brand that also sells through retailers. Being the maker helps, but only if your own record is as complete and up to date as the retailer's.

A ChatGPT product feed checklist

1. 1.Confirm each item_id stays the same from one export to the next.
2. 2.Fill all nine required fields with real values, with no placeholders.
3. 3.Rewrite titles to lead with product type, brand and the one deciding attribute, then the variant.
4. 4.Open each description with facts: material, fit, use, care, and what the product works with.
5. 5.Fill every optional field you can state truthfully, and omit the rest rather than writing “n/a”.
6. 6.Write product_category as a path, starting from Google's taxonomy and going one level deeper where Google stops.
7. 7.Group variants exactly as they appear on your site, using group_id and variant_dict.
8. 8.Add review aggregates and the real returns policy.
9. 9.Sync price and availability whenever a sale starts or stock changes.
10. 10.Check that every fact in the feed also appears on the product page, which ChatGPT can read too.

What we don't know yet

OpenAI doesn't publish how ChatGPT ranks products. The spec says plainly that “eligibility does not guarantee display.” A complete record means a product can match more shopper questions. It doesn't guarantee a spot in any answer, and anyone promising placement is guessing.

It's also hard to measure. ChatGPT doesn't report impressions or ranking positions for products, so no retailer can run a controlled revenue test inside ChatGPT today. But the same product record also feeds surfaces where you can hold a control group, like Google Shopping and Meta catalog ads. That is where a change to the record can be tested before you trust it.

What this means for your team

Soon, almost every retailer will have its catalog in ChatGPT. Every Shopify catalog is already inside ChatGPT, and the waitlist for everyone else is moving. The difference will come down to how many of a shopper's needs each product record can answer, and in most retail teams, nobody owns the fields that decide that.

Lily Max writes and checks those fields across every surface that reads them, then tests each change against a control. Book a demo to see it in your catalog.