The Product Feed Attributes ChatGPT Shopping and Google AI Mode Actually Read

If your product feed already passes Google Shopping and Meta review, it is tempting to assume it is also ready for ChatGPT Shopping and Google's AI Mode. On the Google side, it probably is not. OpenAI's Agentic Commerce product feed specification requires 9 fields to make a product eligible for search and checkout inside ChatGPT, according to OpenAI's own developer documentation. Google's Merchant Center Help, updated as part of the May 2026 Universal Commerce Protocol push, names 6 specific attributes it calls "conversational attributes" for AI Mode in Search. The two lists do not share a single field name. This guide walks through what each surface actually reads, what you likely already have, and what you still need to add, plus a scoring checklist to run against your own feed.

Two different "AI shopping" problems, not one

"Getting ready for AI shopping" gets treated as one project, but it is really two, built by two companies with two separate specs and two separate goals.

OpenAI's Agentic Commerce Protocol (ACP) lets a merchant sell directly inside a ChatGPT conversation through Instant Checkout. The product feed's job here is narrow: give ChatGPT enough structured data to display a product correctly and let a real transaction happen, with the merchant, not OpenAI, acting as merchant of record and handling fulfillment, tax and payment through their own payment service provider.

Google's conversational attributes, by contrast, sit inside the Merchant Center feed you already run for Shopping ads and free listings. There is no separate checkout protocol to implement, no new feed to build from scratch. Google is asking you to enrich the feed you already have so its own AI Mode and Gemini-powered shopping surfaces can answer more specific questions about your products.

Confusing the two leads to two failure modes: teams that build an entire second feed for ChatGPT they did not need to, and teams that assume ticking a few OpenAI-style boxes also covers Google's AI Mode, when it does not.

What OpenAI's product feed spec actually requires

OpenAI's specification (last verified against the live docs in September 2026) defines a base set of required fields, a smaller set needed specifically for checkout eligibility, and a longer list of recommended fields that improve ranking and trust.

Three more fields gate checkout specifically, on top of the base nine: is_eligible_checkout (boolean, must be true), seller_privacy_policy (a public privacy policy URL) and seller_tos (a public terms-of-service URL). A product can be search-eligible without these, but not checkout-eligible.

OpenAI accepts JSONL, CSV or TSV, including gzip-compressed variants. A single JSONL row looks like this:

{"item_id":"SKU-10432","title":"Merino Wool Crew Socks - Charcoal, M","description":"Mid-weight merino wool blend crew socks with reinforced heel and toe.","url":"https://example.com/products/merino-crew-socks-charcoal-m","brand":"Example Co","seller_name":"Example Co Online Store","image_url":"https://example.com/img/merino-crew-charcoal-m.jpg","availability":"in_stock","price":"18.00 USD","is_eligible_checkout":true}

There is no fixed update interval in the spec, but OpenAI is explicit that price and availability need to be current whenever a sale starts or ends, which in practice means syncing on the same cadence you already use for Google Shopping, not a slower one.

Why most Google- or Meta-ready feeds are already close to OpenAI's bar

Look at that field list again: title, description, price, availability, image_url, brand. Every one of these has a near-identical counterpart in the standard Google Shopping and Meta catalog specs you are almost certainly already populating. The renaming is mostly cosmetic: Google's image_link becomes OpenAI's image_url, Google's id becomes item_id.

The genuine gaps are the fields nobody's standard advertising feed carries, because no ad platform has ever asked for them: seller_name as a distinct identity field, seller_privacy_policy, seller_tos, and the checkout flag is_eligible_checkout. These need a mapping rule that pulls from your legal or company-info pages rather than your product catalog, which is exactly the kind of static, non-per-SKU field that a feed rules engine (Feedance includes one under feed enrichment) handles better than hand-editing a spreadsheet column.

In short: if your feed already clears Google Shopping and Meta review, building an OpenAI-compliant export is closer to a mapping exercise than a data-collection project.

What Google actually added for AI Mode

Google's own help documentation names exactly six attributes as "conversational attributes": question_and_answer, document_link, related_product, item_group_title, variant_option and popularity_rank. All six are optional, and Google states directly that adding them "won't impact the approval status of your existing products."

• question_and_answer — up to 30 question-and-answer pairs per product, each up to 1,000 characters, 10,000 characters total across all pairs. Written for exactly the kind of follow-up question a shopper asks a chat interface instead of scanning a spec sheet.
• document_link — up to five HTTPS links to PDFs only (manuals, spec sheets, care guides), 50MB max each, URL itself capped at 2,000 characters, so AI Mode can answer from your own documentation instead of guessing.
• related_product — up to 30 entries per product, each one typed by relationship (part of a set, required part, often bought with, substitute, different brand, accessory), so a conversational answer can suggest the right companion item.
• item_group_title — a shared title for a variant group (1–150 characters, truncated beyond that) that stays generic where each variant's own title gets specific about size or color.
• variant_option — name/value pairs (each up to 250 characters) for variant dimensions outside Google's standard six (color, size, material, pattern, age group, gender), submitted alongside a shared item_group_id.
• popularity_rank — a number from 0.0 to 100.0, one decimal place, reflecting a product's actual relative sales performance in your catalog, not a marketing claim.

A question_and_answer row in a tab-separated feed looks like Google's own example:

'Does it have a headphone jack?':'This version doesn't have a headphone jack.'

None of this is submitted through a new endpoint. It rides in the same Merchant Center feed, the same way product_highlight or gtin already do.

The two attributes everyone mislabels as "AI Mode attributes"

Several widely shared guides, including a popular community breakdown on Google's own support forum, describe "8 attributes for AI Mode" by adding product_highlight and product_detail to Google's list of six. That is worth correcting: Google's official conversational-attributes page names six, not eight, and product_highlight is a separate, older attribute (see Google's own spec) that has existed since well before AI Mode and was built for regular Shopping listings, not for conversational surfaces specifically.

That does not make product_highlight useless for AI Mode. Short, factual benefit statements (Google's own examples: "Supports both 2.4 Ghz and 5 Ghz Wi-Fi networks") are exactly the kind of structured, skimmable data an AI system prefers over a paragraph of marketing copy, and Google's May 2026 announcement pushes merchants toward more conversational product descriptions in general. But treating it as one of the "official six" overstates how new or AI Mode-specific the attribute is, and can lead a team to skip the six that actually are.

Feed readiness checklist, scored against both specs

Score one point for each item you can already confirm true, without guessing. This is a snapshot, not a certification: passing a field's presence check does not guarantee the content quality bar either platform enforces separately.

0–4: your feed is built for legacy Shopping ads only. 5–9: you are checkout-ready for OpenAI but Google's conversational layer is essentially empty. 10–13: you have done the enrichment work most catalogs skip.

Where this breaks down

This guide covers feed data, not the rest of the stack. OpenAI's Instant Checkout also requires a payment service provider compatible with its Delegated Payment Spec and your own handling of order validation, fulfillment and tax, none of which a feed tool touches; that is a payments and engineering project, separate from feed enrichment.

It also does not cover every AI shopping surface. Perplexity runs its own Merchant Program, with separate onboarding and its own product data submission outside both specs discussed here, and, as of this writing, Meta has not published a conversational-attributes-style spec comparable to Google's; Meta's AI-driven shopping features still run on the standard Commerce Catalog fields covered in our guide to Meta catalog disapprovals. Treat both as open questions to revisit, not gaps in this checklist.

Other feed platforms have published their own take on Google's rollout; Channable's breakdown is worth reading alongside this one, since Google's own documentation on rollout timing and measurement is still thin and a second interpretation is useful while the spec settles. None of this replaces core feed hygiene, either: a product missing GTIN or miscategorized will not surface in AI Mode or ChatGPT regardless of how well you fill in conversational attributes, because it is not eligible to appear at all.

FAQ

Do I need to build a separate feed for ChatGPT Shopping?

No. You submit OpenAI's required fields as their own feed file, but the underlying data is a subset of what you already maintain for Google Shopping. Most teams generate it as an additional export from the same product data, not a second catalog.

Will adding Google's conversational attributes affect my existing Shopping ads?

Google states explicitly that adding them does not change the approval status of existing products. They are additive fields, not a replacement for your current required attributes.

What format does OpenAI's feed spec accept?

JSONL, CSV or TSV, including gzip-compressed versions of each. JSONL is the format shown in most of OpenAI's own examples.

Is popularity_rank just my bestseller list?

It is meant to reflect actual relative sales performance on a 0–100 scale, not a manually curated "featured" list. Feeding it a marketing-driven ranking rather than real sales data defeats its purpose.

Do I still need item_group_id if I add Google's item_group_title?

item_group_id is what actually links variants together; item_group_title is a label layered on top of that grouping. You need both if you want the conversational layer to work correctly.

Does Meta have an equivalent spec I should be building toward?

Not as a published, AI Mode-style attribute set as of this writing. Meta's catalog requirements are still the standard Commerce Catalog fields; focus there on passing standard review rather than chasing a spec that does not yet exist.

How often do these fields need to update?

Match your price and promotion change frequency. OpenAI is explicit about updating price and availability when a sale starts or ends; Google's conversational attributes change less often but should be reviewed whenever you add new products or update specs.

Can I check whether ChatGPT or AI Mode is actually reading these fields correctly?

There is no public, real-time preview tool from either company at the time of writing. The practical approach is to validate the feed against the published spec field by field, then periodically search for your own products by name and specific attributes inside ChatGPT and in Google's AI Mode to spot-check what surfaces.

Run a free product feed audit to see which of these fields your current feed already covers and where the gaps are before you start mapping new attributes by hand.