Guides · November 7, 2026 · 9 min read

Product Feeds for AI Agents: What to Fix First

person using MacBook Air

The common advice on product feeds for AI agents is to build a new “AI feed,” rewrite every description in a chatty voice and wait for ChatGPT to sell for you. That gets the order backwards. AI shopping agents mostly read feeds you already have: your Merchant Center feed, plus OpenAI’s format if you’re approved for ChatGPT. Products usually fail on dull problems: a price that differs between feed and page, a GTIN that lost its leading zeros, stock data that’s a day old. Fix the facts first, then product identity, then freshness, then context. Checkout flags come last. This guide follows that order, based on Google’s, OpenAI’s and Microsoft’s official specs as of November 2026.

Where AI Shopping Agents Actually Get Product Data

  • Google (AI Mode, Gemini). Google says shopping in AI Mode is built on its Shopping Graph. At I/O 2025 it said the graph had more than 50 billion listings, with over 2 billion refreshed every hour. You get into it through Merchant Center. Google’s UCP help page says Merchant Center “will continue to be the central hub to prepare your product data.”
  • OpenAI (ChatGPT). The products feed spec is OpenAI’s own format, pushed over SFTP. OpenAI’s docs say feed onboarding is “currently available to approved partners,” and you apply at chatgpt.com/merchants. That page says Shopify and Etsy catalogs are already integrated and that shopping is live for US users.
  • Microsoft (Copilot). Microsoft Advertising says the first step is “a UCP-ready feed” in Microsoft Merchant Center. You need a Merchant Center account to onboard to Copilot Checkout.

Your product pages still count. Crawlers and browsing agents check the feed against them, and Google can update feed data from them. But the feed is the primary source, and the page confirms it.

Why Agents Punish What Shoppers Forgive

A person who sees “in stock” in an ad and “ships in 3 weeks” on the page grumbles and carries on. An agent can’t do that. To answer “waterproof trail shoe, size 10, under $120,” it filters on structured attributes. If your feed never says waterproof, the filter has nothing to match. If the feed and the page show different prices, the agent has to pick one, and both Google and OpenAI tell merchants to keep them in sync.

That’s what product data quality means in AI shopping. Every fact should agree across feed, page and markup, and every field should be filled in. Our guide to structured data for AI agents covers the markup. This one stays on the feed.

Fix 1: Make Price and Availability Tell One Truth

Start with price and availability. They break most often, and checkout depends on them most. Google’s product data specification says the feed price must match the landing page, your structured data and checkout. OpenAI’s spec adds rules of its own:

  • Price is a string like 79.99 USD: a decimal amount with no thousands separators, then an uppercase ISO 4217 currency code.
  • sale_price must be lower than price and in the same currency, or OpenAI ignores it.
  • availability must be in_stock, out_of_stock, pre_order, backorder or unknown. An empty or unrecognised value rejects the whole row, and the spec warns that unknown doesn’t mean in stock.
  • Date fields don’t schedule changes. When a sale ends, you send the new price yourself.

Google’s automatic item updates can fix price and availability using your pages’ structured data, with one condition many storefronts fail. The markup must be in the HTML your server returns. Google says it “can’t be generated with JavaScript after the page has loaded.” If your theme adds the Product JSON-LD client-side, this fallback never runs. SEO Rocket’s site audit runs a Screaming Frog-backed crawler with JavaScript rendering, which helps you spot product pages whose markup only appears after scripts run.

Fix 2: Get Product Identity Right

Identifiers tell an agent that your shoe is the same product a review site or another retailer describes. If they’re broken, the agent has only your title to match on.

  • GTINs. OpenAI requires 8, 12, 13 or 14 digits, a valid check digit and preserved leading zeros. Spreadsheets strip leading zeros, so check any export that passes through one. Google’s advice: “When in doubt don’t provide this attribute.”
  • MPNs. OpenAI says not to invent an MPN to cover a missing GTIN.
  • identifier_exists. Set it to no only for products that truly have no identifiers. Don’t use it to silence errors.
  • Variants. Give every size and colour its own row with a unique ID, linked by a group ID. In OpenAI’s spec, the group ID must differ from every item_id, and offer_id must never contain the price.

Here’s an illustrative OpenAI feed row with placeholder values. It’s one line of JSONL, wrapped here so it’s easier to read:

{"item_id": "TS-01-BLK-10", "group_id": "TS-01",
 "listing_has_variations": true,
 "variant_dict": {"color": "Black", "size": "10"},
 "title": "Example Waterproof Trail Shoe, Black, Size 10",
 "description": "Waterproof membrane, rock plate.",
 "url": "https://example.com/trail-shoe?variant=blk-10",
 "brand": "Example", "seller_name": "Example Store",
 "image_url": "https://example.com/img/ts-01-blk.jpg",
 "availability": "in_stock", "price": "89.00 USD",
 "gtin": "00012345678905"}

That’s all nine required fields, plus the variant fields and a GTIN. Drop the two leading zeros and the GTIN no longer matches.

Fix 3: Freshness and Delivery Mechanics

OpenAI feeds are full snapshots. Each upload is “a complete catalog export treated as the source of truth.” Refresh it at least daily, using the same file name each time. A product you drop from the snapshot can stay live for up to 14 days, because OpenAI keeps the last processed record that long. To pull a recalled item sooner, set is_eligible_search to false.

Google’s Content API for Shopping has sunset. According to Google’s Merchant API changelog, the Content API reached sunset on August 18, 2026, and progressive degradation started on September 1. If a custom script still sends price and stock through it, those updates may be failing without anyone noticing. Platform partners migrate for merchants on hosted platforms. Custom integrations have to move to the Merchant API, which Google also recommends for real-time updates if you use UCP checkout.

Fix 4: Titles and Descriptions Built for Attribute Matching

Copy comes fourth. Google and OpenAI share the limits: titles up to 150 characters, plain-text descriptions up to 5,000. Treat both fields as places to state buying attributes, not as ad copy.

  • Order the title the way people search: brand, product type, the key attribute (waterproof, wide fit, 64 GB), then the variant.
  • Put specs in their own fields as well as in the prose. Google has product_highlight and product_detail. OpenAI has material, size, dimensions and return fields. A spec in its own field is easier to filter on than one buried in a paragraph.
  • OpenAI’s best-practices guide asks for “concise, factual copy.” Write facts, not adjectives.

Find the attribute words in query data instead of guessing them. SEO Rocket’s keyword research runs on real Ahrefs data. The long-tail modifiers it surfaces (“for flat feet,” “with USB-C,” “machine washable”) are the same kind of attribute language shoppers use in AI shopping prompts. Our product description guide covers the page side.

Fix 5: Add Google’s Conversational Attributes

This is the layer that’s actually new. In January 2026 Google announced “dozens of new data attributes in Merchant Center designed for easy discovery in the conversational commerce era,” and the Merchant API added them in May. Google’s help page lists six:

  • question_and_answer: question-and-answer pairs about the product.
  • document_link: manuals and spec sheets.
  • related_product: accessories and substitutes, given as relationship type, identifier type and identifier.
  • item_group_title and variant_option: the parent product name and each variant’s options.
  • popularity_rank: how popular the product is relative to your inventory.

A question_and_answer question or answer can be up to 1,000 characters. You can add up to 30 pairs, with a cap of 10,000 characters per product. Google says to leave out prices and dates and not to repeat what’s already in the title or description. It recommends a supplemental data source in TSV format, because commas in answers complicate CSV parsing. Adding these attributes doesn’t affect approval of existing products, so test them on your best sellers. Start with the questions your support team answers every week.

Fix 6 (Last, Not First): Checkout Eligibility Flags

  • Google UCP. The “Buy” button in AI Mode and Gemini appears only on listings flagged with native_commerce checkout eligibility. It’s open to participating merchants, for products eligible in the US, Canada and Australia. Subscriptions, pre-orders, personalised, age-restricted, rental and virtual items aren’t eligible. Google says the native_commerce group “is fully overwritten with each upload,” so send every flag each time, through a supplemental source. You also need return cost, return window, a policy link and at least one support contact. See our UCP explainer.
  • OpenAI. is_eligible_checkout is off by default. Turning it on requires a checkout integration and seller_privacy_policy and seller_tos URLs. OpenAI’s merchant page now says it is “moving away from a standalone Instant Checkout experience,” and by default shoppers buy on your site. Our ChatGPT checkout guide tracks the current setup.
  • Microsoft. Copilot Checkout is open only to English-language merchants selling to US buyers in USD.

Outside those programs, checkout flags do nothing yet. Discovery, which runs on the feed, still decides who gets seen.

A One-Week Triage Plan

If you’re auditing product feeds for AI agents from scratch, this order fits most stores into a week:

  1. Day 1: Export Merchant Center diagnostics. List price mismatches, availability mismatches, invalid GTINs and missing identifiers.
  2. Day 2: Crawl your product URLs and compare the rendered price and stock with the feed. Check whether the Product JSON-LD is in the raw HTML.
  3. Day 3: Fix identifiers. Restore leading zeros, remove guessed GTINs and repair variant groups.
  4. Day 4: Confirm nothing still uses the Content API, and match your refresh cadence to how often prices change.
  5. Day 5: Rework titles and attributes for your top 50 to 100 products by revenue.
  6. Days 6 to 7: Add conversational attributes to those products through a supplemental feed.

Then measure. SEO Rocket’s AI visibility tracking shows whether your brand appears in AI answers for your target queries. It sits next to your audit, rankings and keyword data in the same chat.

Do Now, Watch, Ignore

  • Do now: get price and stock to agree across feed, page and markup. Repair identifiers. Move off the Content API. Render Product markup on the server. Add conversational attributes to your best sellers where your account offers them.
  • Watch: the UCP checkout rollout in Merchant Center, whether OpenAI opens merchant onboarding beyond approved partners, and Copilot Checkout outside the US.
  • Ignore: paid “AI feed formats” that no platform documents, and keyword-stuffed titles written for chatbots.

FAQ

Do I need a separate feed for ChatGPT?

Not if you’re on Shopify or Etsy, whose catalogs OpenAI says are already integrated. Otherwise yes: OpenAI’s format is separate and currently limited to approved partners. Generate it from the same product database as your Merchant Center feed so both have the same facts.

Can schema markup replace a product feed?

No. Google can use markup to update price and availability, but only if it’s server-rendered and matches the page. The checkout programs from Google, OpenAI and Microsoft all run on feeds.

How often should product feeds for AI agents update?

OpenAI asks for a full snapshot at least daily. On Google, if prices change during the day, push those changes through the Merchant API instead of waiting for a nightly file.

Do conversational attributes guarantee AI Mode visibility?

No. They’re optional context, and nothing in Google’s documentation promises placement.

Clean product feeds for AI agents win for the same reason clean sites win in search: every fact agrees, identifiers match and nothing is stale. It’s a playbook proven across 1,000,000+ ranking pages, applied to your feed. Start with the crawl-versus-feed check in SEO Rocket, then read our guides to agentic commerce and Google AI Mode shopping.

Keep reading

Questions? Chat with us