Product Feeds for AI Shopping Agents

01What is agentic commerce, and what changes for you?

This is the sequel to when the buyer is a bot: that piece covered why agentic commerce matters and how the protocol stack works; this one is the operational how, how to structure your product data so an agent includes, compares and recommends you. The mindset shift is total. For twenty years, ecommerce optimisation meant persuading a human on a page. An agent is not persuaded by a page. It is fed a record, and it reasons over fields.

The 2026 settlement made this concrete. After OpenAI retired Instant Checkout in March 2026, the industry converged on a division of labour: AI handles discovery and comparison, merchants keep checkout in their own environment. So the job is no longer 'sell to the agent'. It is 'be in the set the agent surfaces, with data clean enough to be chosen', then convert the human when the agent hands them back.

02How big is agentic shopping, really?

The numbers describe a channel that is past the experimental stage. Around 50 million shopping queries a day run inside ChatGPT alone (about 2% of all its queries, per OpenAI's own research); Salesforce put AI agents behind roughly a fifth of global holiday orders; and Adobe measured a near-quadrupling of AI-referred traffic to US retail. The buyers are already asking agents what to buy. What decides the answer is not on your homepage, it is in your feed.

03How does a shopping agent actually find your product?

An agent assembles a shortlist the way a database query does: it reads structured records, filters on attributes (price, availability, category, eligibility), and ranks on signals it can compute (relevance, rating, price competitiveness). The primary source is the feed, a file you push to the platform, and the secondary source is corroboration from the open web, your PDP schema, reviews, and third-party listings. Both have to be clean and consistent, because an agent discards outliers the same way an AI answer does.

04What is a product feed, and what do you push?

Mechanically, you push a compressed feed to an endpoint the platform provides (OpenAI's, for example), with daily updates accepted. Each record tells the agent what you sell, whether it is in stock, the price with its currency, the images, and the eligibility flags for search and checkout. Get the format right and you are in the catalogue; get a field wrong and the record is filtered out before a shopper ever sees it.

# A single line of a product feed (JSONL), the unit an agent actually reads.
# Merchants push a gzip file (.jsonl.gz / .csv.gz / .xml.gz) to the platform endpoint,
# updated daily. This is the record, not your marketing page.
{
  "id": "SKU-10482",
  "title": "Aero Running Shoe, Men's, Volt",
  "description": "Lightweight road running shoe, 8mm drop, 220g. Breathable knit upper.",
  "link": "https://example.com/p/aero-running-shoe",
  "image_link": "https://example.com/img/aero-volt-1.jpg",
  "additional_image_link": ["https://example.com/img/aero-volt-2.jpg"],
  "price": "129.00 USD",              # price WITH ISO 4217 currency code
  "availability": "in_stock",         # in_stock | out_of_stock | preorder
  "inventory_quantity": 42,
  "gtin": "0195819023471",            # GTIN / MPN / brand = the identifiers agents match on
  "brand": "Aero",
  "condition": "new",
  "google_product_category": "Apparel & Accessories > Shoes",
  "shipping": "US:::5.99 USD",
  "return_policy": "30-day free returns",
  "enabled_for": ["search", "checkout"]   # eligibility flags the platform reads
}

05Which fields decide whether you get included?

The pattern mirrors the rest of GEO: inclusion is a gate you pass with clean, complete, machine-readable data, and only then does relevance and ranking decide the order. Most merchants lose products at the gate, an unpriced item, a stale stock status, a missing GTIN, long before any ranking logic runs.

06Does on-page Product schema still matter if the feed does the work?

Not every agent works from a pushed feed; some read the open web, and all of them cross-check. Product and Offer schema on the PDP, with the same GTIN, price, currency and availability as the feed, gives the agent a consistent second source. The schema playbook covers the full stack; for shopping, the load-bearing types are Product, Offer, AggregateRating and the shipping/return details, all matching the feed to the character.

<!-- Product + Offer schema on the PDP: the on-page truth an agent corroborates the feed against. -->
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Aero Running Shoe, Men's, Volt",
  "gtin": "0195819023471",
  "brand": {"@type": "Brand", "name": "Aero"},
  "image": ["https://example.com/img/aero-volt-1.jpg"],
  "offers": {
    "@type": "Offer",
    "price": "129.00",
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock",
    "itemCondition": "https://schema.org/NewCondition",
    "url": "https://example.com/p/aero-running-shoe",
    "shippingDetails": {"@type": "OfferShippingDetails", "shippingRate": {"@type":"MonetaryAmount","value":"5.99","currency":"USD"}},
    "hasMerchantReturnPolicy": {"@type":"MerchantReturnPolicy","returnPolicyCategory":"https://schema.org/MerchantReturnFiniteReturnWindow","merchantReturnDays":30}
  },
  "aggregateRating": {"@type": "AggregateRating", "ratingValue": "4.6", "reviewCount": "812"}
}
</script>

07What are the competing agentic-commerce protocols?

The takeaway is not to master all five. It is that they all consume structured product and offer data, so a clean, well-identified feed and matching schema are the portable asset that lets you onboard to whichever protocol your buyers' agents adopt. Chase the data quality, not the protocol of the month.

08Discovery or checkout, which is the job now?

OpenAI's retirement of Instant Checkout six months after launch reset expectations: the durable pattern is AI-led discovery, merchant-owned checkout. That is good news operationally, it means your conversion, payments and data stay in your environment, and the new work is narrow: get discovered and get chosen. The agent does the shortlisting; you win by being a clean, complete, competitively-priced record, and by keeping the checkout it hands off to fast and frictionless.

09What makes an agent recommend you, not just list you?

Once you are in the set, the agent ranks. It rewards what it can verify: a price it can compare cleanly, stock it can trust, an aggregate rating and reviews corroborated off-site, and complete attributes that match the shopper's constraints (size, spec, compatibility). It penalises ambiguity, an item with no reviews, a vague category, or a price that disagrees between feed, PDP and third-party listings. Recommendation is won on the same axis as citation: be the cleanest, most corroborated, least ambiguous option.

10Why do products drop out of AI shopping results after they were showing?

The most common silent failure is not a bad initial feed, it is drift. A product sells out but the feed still says in_stock (so the agent surfaces it, the shopper bounces, and the platform down-ranks you); a price changes on the site but not the feed (so feed and PDP disagree and the agent distrusts both); a daily push fails and the whole catalogue goes stale. Monitor push success, reconcile feed against live stock and price daily, and treat feed health like uptime.

11How do you measure whether agents recommend you?

Build a portfolio of the shopping questions your buyers ask an agent, 'best [product] for [use] under [price]', '[your product] vs [rival]', and run each several times across ChatGPT shopping, Google's agentic surfaces and the rest, because one run per prompt is noise. Log three things: are you surfaced, where do you rank, and are the price/stock/attributes the agent shows correct. A wrong price or stale stock is a feed-hygiene failure (section 10); being absent entirely is an inclusion failure (sections 4 and 5).

12The AI-shopping-feed checklist

13What a feed won't do

Two honest limits. First, an agent recommends on merits it can compute, so a clean feed surfaces an over-priced or poorly-reviewed product accurately as over-priced and poorly-reviewed; the feed removes the reasons you are wrongly excluded, it does not fix the product or the price. Second, because the 2026 settlement keeps checkout with you, the handed-off human still has to convert on your site, so a fast, low-friction pricing and checkout experience still matters. The feed wins the shortlist; you still close the sale.