Product feeds for AI shopping assistants

Feed structure for AI surfaces

An AI-ready feed has one row per variant. Each row has a stable ID, and a shared group ID ties the sizes and colors of one product together. Google and OpenAI both read the feed as their main record of what you sell, and the scale shows why. Google says its Shopping Graph holds more than 50 billion product listings, and 2 billion of them are refreshed every hour. A listing that's sparse or out of date loses to the many listings that aren't.

The two specs overlap more than they differ. That means you can keep one master catalog and send a version shaped for each channel. The table sets the core structural rules side by side.

Before you build anything, check that your IDs never change when a product is edited or restocked. When an ID changes, both platforms treat the item as brand new and its history is lost.

Attributes AI assistants parse

Google's AI shopping needs the same core attributes as a standard Shopping feed. Each row needs an ID, with a title and description to describe the item. Supply a link and image link for it. Price and availability are required, as are brand and a product identifier. Apparel also needs color and size, with gender and age group specified. The difference with AI surfaces lies in how hard the optional fields work, because an assistant breaks one conversational question into a set of attribute filters.

Required fields both platforms check

Both platforms require an ID and title on every row. A description must accompany the landing page link and main image. Price and availability are also mandatory; a row missing any of these fields is rejected. Google asks for brand only when a product has a clear brand or manufacturer, but OpenAI wants it on every row. A product that passes Google's checks can still fail the OpenAI profile, as CommerceClarity notes in its field-by-field mapping.

Attributes that answer conversational queries

CommerceClarity illustrates how assistants use optional fields with the query "waterproof trail shoe, black, size 10, under 100 euro." To answer it, the feed needs color and size. Material must also be specified, along with price and availability. Treat material and pattern as required for any category where shoppers ask about them. Apply the same rule to weight and product dimensions.

Identifiers and when to leave them out

The Global Trade Item Number (GTIN) is the barcode number that lets Google match your item to the same product sold elsewhere. Google's own help page says "products with missing or incorrect GTIN attributes may have limited visibility." Follow its guidance by sending brand plus the Manufacturer Part Number (MPN) for store-brand and custom goods without a barcode. A made-up GTIN does more damage than a missing one.

Titles for AI retrieval

To optimize titles for AI retrieval, put the product type and its deciding attributes first, and write the title for the attribute order your category uses. Google allows 150 characters but displays only about 70. In its help article on long text, Google writes that it "[strongly encourage[s] you to submit titles with 70 or less characters](https://support.google.com/merchants/answer/14989281?hl=en) whenever possible." An assistant reads the whole string, though. So the first 70 characters have to work for a human, and the rest can carry attributes for the machine.

Google trims "Nexus 7 tablet in black with 32 GB of storage" to "Nexus 7 tablet, 32 GB, black" in its own example. Each comma-separated chunk is an attribute a model can lift out. Write each title in this order:

1. Start with the brand if shoppers search by it. If they don't, open with the product type.
2. Add the attribute that decides the purchase in your category. For electronics, use specs; for clothing, use fabric and fit.
3. Add the variant's color and size if they match the row's own attribute fields. Do the same for capacity.
4. Remove promotional text such as "free shipping" or "best deal" because it triggers disapprovals. Remove ALL CAPS for the same reason.
5. Read the first 70 characters alone and check that they still name the product.

If you supply ChatGPT through its Google-compatible path, keep the full title in the plain `title` field because OpenAI maps that field exclusively. Copy any title text held in Google's `structured_title` into the plain field.

Custom labels that help ranking

The custom labels that improve rankings sort products by business value, such as margin band or season. Sales performance and price band are other useful criteria. They work through bidding. Google's spec says custom labels 0 to 4 let you build campaign filters; it supports them in Performance Max, Shopping, or Demand Gen campaigns. It also states that "the information you include in this attribute won't be shown to customers." Labels let you bid harder on the products worth winning, which lifts their paid placement. Their ranking benefit is confined to paid results.

A common scheme gives each of the 5 slots one job:

• `custom_label_0`: margin band (high, medium, low)
• `custom_label_1`: season or promotion window
• `custom_label_2`: sales performance (best seller, long tail, no sales)
• `custom_label_3`: price band
• `custom_label_4`: release year or collection

Plan the values before you fill them in. Google caps each label at 1,000 unique values across the account and won't raise the limit. Values past the cap are ignored for reporting and bidding. That's why a SKU or product ID never belongs in a label. Labels group products, while the `id` field identifies them.

Google product feed setup

You set up a Google product feed in Merchant Center. The steps below take a new store to approved listings. Time varies with catalog size, and most of the effort goes into step 4.

1. Create a Merchant Center account. Then verify and claim your website domain.
2. Set shipping rates and your return policy at account level. Google won't approve products without them.
3. Choose a data source. Store-platform integrations suit catalogs that already live in Shopify or a similar platform. A scheduled file or Google Sheet suits small catalogs. Use the API for catalogs where prices change during the day.
4. Map your catalog fields to Google's attributes. Fill in every attribute your category uses, and pay most attention to identifiers and variant fields.
5. Set the fetch schedule to match how often prices and stock change. Daily is the floor for most stores.
6. Turn on automatic item updates, so Google can correct price and availability from your page's structured data when the feed lags.
7. Open the diagnostics tab after the first fetch, and fix item-level issues before you launch campaigns.
8. Enable free listings so the products can appear on organic Shopping surfaces such as AI Mode.

You'll know the setup worked when diagnostics show the large majority of items as approved and the Products view lists no account-level issues.

ChatGPT shopping feed requirements

To structure a feed for ChatGPT shopping, follow OpenAI's product feed spec. It uses the same core fields as Google, plus two flags. `enable_search` lets ChatGPT show the product, and `enable_checkout` lets shoppers buy it inside the chat. Checkout only works when search is also on. Turning on checkout also makes two more fields required, `seller_privacy_policy` and `seller_tos`.

ChatGPT requires an application through its merchant portal and business verification by OpenAI, a different process from Google's. Next you push a sample of about 100 products for schema checks. Alhena's 2026 setup guide says the full process takes one to two weeks from application to live products. Accepted formats include compressed JSON Lines and CSV, all in UTF-8.

Your fields also decide which store ChatGPT sends the buyer to. OpenAI's release notes say that when a shopper picks an item, ChatGPT ranks merchants "by factors like inventory, price, quality, primary seller status, and Instant Checkout availability." Accurate stock counts and current prices therefore carry real weight. Check three things before launch:

• Every row has a brand and a `product_category` path.
• Return policy and return window are set on each product, because OpenAI doesn't read Google's account-level settings.
• Review count and rating fields are filled in wherever you have real review data.

Fixing common feed rejections

Mismatches between the feed and the landing page cause most feed rejections, while missing identifiers or rule-breaking text account for the rest. Google's crawler compares your feed price with the price on the page and in its structured data. As Google's help page on mismatched prices puts it, "the product with the mismatch may be disapproved." Once the prices match again, the product is reapproved after the next crawl, and requesting a website check speeds that up.

Fix the cause in your source data. A manual patch is overwritten at the next fetch, and the rejection comes back.

Measuring feed performance

You measure feed performance on two levels. Use the channels' own reports for clicks and impressions, and track visibility in AI answers, where AI product recommendations matter as much as clicks. Merchant Center's performance reports cover Google surfaces. Google Analytics 4 (GA4) shows referral sessions from AI assistants landing on product pages.

The traffic is small but growing fast. Adobe Analytics found that traffic to US retail sites from generative AI tools rose 693.4% year over year in the 2025 holiday season. Digital Commerce 360 also cited Adobe's finding that revenue per visit from AI referrals increased 254% year over year. Measure those referrals separately from organic search, or the gains disappear into your totals.

Clicks only show the visits you won. To see the answers that left you out, you need prompt tracking. That means running the same set of shopping questions every week and recording which brands each assistant names. Snoika's tracking runs weekly prompt sets in ChatGPT and Google AI Overviews to record mentions and share of voice. It also records sentiment and citations. The before-and-after numbers make a feed fix measurable. Snoika's founder Anton Vedeshin reported to Startup Reporter 1StopVAT's move from 2.7% to 28.3% ChatGPT visibility in 16 weeks.

Who should manage the feed

Choose who runs product feed management based on catalog size and how often your data changes. Building the feed and checking whether AI assistants actually show your products are separate jobs, and each option below covers a different part of the work.

For a store under a few thousand SKUs, use a feed platform with one owner on staff, and add an agency only when ad spend justifies the fee. Large catalogs earn back a specialist's salary. Whichever option builds the feed, you still need weekly generative engine optimization data. That data shows whether feed optimization improves visibility on the surfaces where AI shoppers now start.

What to do next

Check your feed against the OpenAI spec this week, even if Google already approves every item. The gaps that block ChatGPT are fields Google allowed you to skip, such as brand and category. Per-product return policies are another gap. They take a few hours to fill and decide whether the product feed reaches a second assistant at all. Once the feed passes, measure whether ChatGPT and AI Overviews actually name your products, because approval alone doesn't tell you that. We track brand mentions and share of voice across AI assistants every week, with a human review layer on top of the automated runs. To see where you stand today, get a free AI visibility report.