Product Feed Requirements for AI Shopping Agents
On the protocol path, the AI shopping agent never opens your product page. It reads your feed, matches a handful of attributes, and buys - or it drops your product silently because one field is missing. Feed hygiene, not price or brand strength, is what decides whether the agent can see the product at all - a feed item that is incomplete or disapproved is not eligible to be shown.
“AI shopping tools don't guess. When a shopper asks ChatGPT, ‘What's the best lightweight waterproof jacket for hiking under £150?’, the AI looks for products whose data explicitly contains those attributes: weight category, waterproofing rating, intended use case, and price. If your product page doesn't include that structured information, your product won't appear in the answer.”
So this is the practical piece: the product feed requirements AI shopping agents actually check, why availability and price accuracy break sales first, and how to audit your own feed before an agent does it for you.
Stay in the loop
Get news and updates about GEO, AI search and new features. Unsubscribe anytime.
Is your brand a Ghost or a Guide on AI?
See if AI knows your brand. We ask Gemini and Claude live - in ~5 seconds, no signup.
This piece is built on the documented feed rules, not on a proprietary dataset. The requirements below come from Google's Merchant Center product-data specification, which most agents read through - a Semrush study found 83% of ChatGPT's shopping carousel comes straight from Google Shopping. Where a claim is about how an agent behaves rather than a documented field rule, I say so.
The protocol path buys from your product feed, not your page
There are two ways an AI agent checks out, and they read your store differently. On the protocol path the agent works from structured data - your feed, JSON-LD, a commerce endpoint - and assembles the order programmatically. On the browser-driving path it runs a real browser and clicks your actual cart like a human. I cover both in the agentic checkout piece; here I focus on the feed, because it is the surface the protocol path lives or dies on.
The distinction matters for one reason: on the protocol path there is no human to work around a gap. A missing GTIN or a stale in-stock flag does not earn a shrug and a manual check. The product just is not eligible, and you never see the lost sale.
“Product feed affects not only how your product appears in AI shopping experiences, but also how it is understood.”
What a product feed is to an AI shopping agent
To an agent, your feed is the catalog. Not the marketing page, not the hero image, not the review section - the machine-readable list of products and their attributes. Google Merchant Center is the most common format, and its reach goes well past Google now.
ChatGPT sources 83% of its carousel products from Google Shopping via shopping query fan-outs.
That Semrush analysis (reported by Search Engine Land) looked at 43,000 carousel products across 10 verticals, and in a separate 100-run test ChatGPT's top pick landed in Google Shopping's top 3 results 75% of the time. Put plainly: for most stores the Merchant Center feed is the gate to AI shopping visibility, not any single assistant's integration. Fix the feed and you fix eligibility across engines at once.
Product feed requirements AI agents actually check
Here is the attribute set that decides whether an agent can list and buy your product. The required rows are non-negotiable - miss one and the item drops. The identity and trust rows are what get you matched and ranked against competitors.
| Feed field | Why the agent needs it | Failure mode if wrong |
|---|---|---|
id |
Stable unique key per product and variant | Reused ids collapse variants; the agent buys the wrong one |
title |
Primary match signal for the query | Vague or keyword-stuffed titles miss the intent match |
description |
Secondary match and attribute extraction | Thin descriptions give the agent nothing to reason on |
link |
Canonical product URL to hand the buyer back to | Broken or redirecting links strand the handoff |
image_link |
The visual shown in the shopping carousel | No image, no carousel slot |
price |
Must match the live page exactly | Feed and page mismatch fails validation, drops the offer |
availability |
Current in_stock or out_of_stock |
Stale in_stock sells a product you cannot ship |
gtin / brand / mpn |
Identity match to a known product | No identifiers, no dedupe or comparison against rivals |
condition |
new, used, or refurbished |
Absent condition filters you out of condition-scoped queries |
| shipping / returns | Cost, time, and policy as structured fields | Locked in a PDF, the agent reads them as unknown |
The required base - id, title, description, link, image_link, price, availability - is table stakes. The block people skip is product identity: GTIN, brand, MPN. That is the one below.
GTINs and product identifiers AI agents use to match your catalog
An agent does not trust that your product is a real, comparable thing until it can tie it to a known identifier. GTIN (the barcode number), brand, and MPN are how it dedupes your listing against the same product sold elsewhere and decides where you rank in a comparison.
- Supply the GTIN wherever one exists. Google uses the GTIN to match your item to a known product, so for most manufactured goods a missing or invalid one is a common disapproval trigger - fix it first.
- Send brand on every item. Even for unbranded goods, Google's spec expects a brand value, and agents use it to cluster and compare products.
- Use MPN when there is no GTIN. Custom or handmade products without a barcode still need a stable manufacturer part number so the agent has an identity key to hold onto.
Availability and price accuracy: the two feed fields that kill agent sales
Every other attribute is about being found. These two are about not burning the buyer after you are found, and they fail quietly, which is what makes them dangerous.
Price mismatch is the fast one. If your feed says $49 and the live page says $59, the agent's validation step catches the gap and drops the offer rather than transact on a number it cannot trust. On the protocol path there is no human in the loop to notice and shrug.
Availability is the slow one. A feed that still reads in_stock on a sold-out product does not just lose one sale - it teaches the agent your catalog is unreliable, and unreliable feeds get sampled less. Keep availability and price synced to the live source, ideally near real time, not on a nightly batch that lags reality by hours.
The quiet killer is not a missing field - it is a present field that disagrees with the page. Feed price stale against a sale that already ended, in_stock against a variant that sold through. The agent does not email you about it. It just stops choosing you.
How to audit your product feed for AI agents - 7 moves
- Pull your Merchant Center feed and check the required fields first.
id,title,description,link,image_link,price,availabilityon every item. Any blank is a dropped product. - Add GTINs across the catalog. Map every product that has a barcode to its GTIN; it is the highest-leverage identity fix - see the Shopify product schema guide for the storefront side.
- Diff feed price against live price. Automate a spot-check; any row where the feed and page disagree is an offer waiting to be dropped.
- Sync availability to real stock, not a nightly batch. Near-real-time inventory beats a feed that lags your warehouse by hours.
- Move shipping and returns into structured fields. If the policy lives in a PDF or a help page, the agent reads it as unknown. Put cost, time, and the returns window into the feed.
- Confirm your structured data is server-rendered. Schema injected client-side with JavaScript is invisible to most agent crawlers - the same trap that breaks ChatGPT Shopping visibility. Check the raw HTML source, not the rendered DOM.
- Confirm the feed is the gate on Google's side too. No Merchant Center feed, no agentic surface - the point the Google AI Mode piece makes for the search-to-act flow.
The rendered cart still matters on the browser-driving path
One caution, because the tidy version of this story is wrong. "No agent ever renders your cart" is false today. The protocol path works from your feed, but browser-driving agents still load your actual checkout UI and click through it like a person. A JavaScript-gated add-to-cart button a human clicks fine can stop a computer-use agent cold.
So the feed is the priority - it decides whether you are eligible and recommended at all - but do not let the rendered cart rot on the assumption no machine touches it. Both paths lose you the same sale; they just lose it at different steps.
The feed is the asset that does not get scrapped in a vendor's product review. Checkout integrations come and go - OpenAI proved that this spring - but the store with clean, complete, accurate product data gets the agent's pick across every engine. Start there.
Want to see whether an AI agent can actually read and act on your product data? Run GEOlikeaPro's Visibility Vitals checker and start with the feed.
FAQ
Do AI shopping agents buy from my product feed or my product page?
On the protocol path, from your feed. The agent reads structured product data - your Merchant Center feed, JSON-LD, or a commerce endpoint - and assembles the order programmatically, so it never renders your page. Browser-driving agents are the exception: they load your actual cart UI and click through it like a human. Either way, a clean, complete feed is what makes you eligible, and a Semrush study found 83% of ChatGPT's shopping carousel comes straight from Google Shopping.
What product feed attributes do AI shopping agents require?
The non-negotiable base is id, title, description, link, image_link, price, and availability - miss any one and the item drops. On top of that, identity attributes (GTIN, brand, MPN) let the agent match and compare your product, condition scopes you into the right queries, and structured shipping and returns tell the agent what it otherwise cannot read. Google uses the GTIN to match your product to a known item, so a missing one is a common reason an item is disapproved or left out of comparison.
Why do AI agents drop products with wrong prices or stock?
Because they validate before they buy. If your feed price disagrees with the live page, the agent's checkout step catches the gap and drops the offer rather than transact on a number it cannot trust. A stale in_stock flag is worse - it sells a product you cannot ship and teaches the agent your feed is unreliable, so it samples you less. Keep price and availability synced to the live source, not a lagging nightly batch.
Do I need a GTIN for AI shopping agents?
For manufactured goods with a barcode, yes - a GTIN is how the agent ties your listing to a known product and dedupes it against the same item sold elsewhere. Google's spec expects it, and without it the agent cannot reliably match or compare your product, which is a common reason items get disapproved. For custom or handmade products with no barcode, send a stable MPN and a brand value instead so the agent still has an identity key.
Is a Google Merchant Center feed enough for all AI shopping agents?
It is the highest-leverage single fix, because most agents read through Google Shopping - Semrush found 83% of ChatGPT's carousel is sourced there. Get the Merchant Center feed clean and you fix eligibility across several engines at once. But keep your server-rendered structured data and your actual cart UI working too, since browser-driving agents load the page directly rather than the feed.