Shopify Product Data for AI Shopping Agents: What to Fix

Quick answer: Most AI shopping channels connected to Shopify get your products from Shopify Catalog, which draws on your titles, descriptions, images, categories, variants and options, prices and inventory (Shopify Help Center, as at 31 October 2026). A product becomes hard to match when it fails an entry rule (Shopify lists uncategorised products as excluded), when it cannot be told apart from similar products (vague titles, no barcode), or when the facts a shopper filters on, such as colour and size, are missing or buried in free text. Fix them in this order: category, title, identifiers, options, description. No one can promise a product will be named in a given AI answer, because Shopify says each channel controls the final ranking and wording.

For ChatGPT, OpenAI’s help page says that for merchants on Shopify, product data is already integrated through Shopify Catalog, and that no additional work is required from individual merchants. What is left, and what you control, is the quality of the data in your admin.

This guide walks through the admin fields in the order I would audit them: what the documentation states, what it does not, and which gap makes a product unmatchable. Claims are limited to what Shopify, Google, OpenAI and the ACCC publish, read as at 31 October 2026. Where they are silent, I say so. Shopify’s AI-channel features change often, so confirm a field’s name and location in your own admin before you act on it.

How an agent gets your product data

Shopify describes Shopify Catalog as a structured source of product information gathered from eligible products sold by stores on Shopify. The Help Center lists what it draws on: titles and descriptions, images, categories, variants and options, prices, and inventory and availability.

The developer documentation shows the other end of the pipe. Agents can search a Global Catalog across Shopify merchants or a Storefront Catalog scoped to one merchant. The Global Catalog tools are search_catalog, lookup_catalog and get_product, and results include a title, description, price range, media, options and variants, with price, availability and seller details on each variant. Two details turn into admin tasks:

• Items that are not sale-ready drop out by default. The available filter defaults to true, which the documentation describes as “only sale-ready items”. Stock that your admin shows as unavailable is likely to be what this removes.
• Agents filter, not just read. The documented search_catalog filters include categories, attributes, rating, price_tier (low, medium or high) and ships_to (a country, with optional region and postcode). The supported attribute names are Color, Size and Target gender.

That second point is the heart of this post. A shopper asking for a black 4-cup pour-over dripper in a mid-range price tier that ships to Australia gives an agent a category, two attributes, a price band and a destination to work with. Those are the sort of inputs the documented filters accept. The pages do not say how an agent fills a filter from free text. A filter can only match a fact that exists as data, so a product that states its colour and size only in a paragraph of marketing copy is the one most at risk.

Shopify also produces some fields itself. Fields tagged Inferred include description, options and metadata.attributes (attributes like material, style and occasion). The developer documentation describes them as discovery and merchandising signals, not merchant-authored source text, which might not always be present or might vary in accuracy depending on available product data. The pages do not say how they are generated. If your data is thin, the inferred layer has less to work from, and what an agent sees may not be a copy of what you typed.

Three limits keep this honest:

• Inclusion is not placement. Shopify says being in Catalog does not guarantee a product will appear in a specific AI answer, and that each channel controls the final ranking and wording.
• Channels differ. Shopify’s agentic storefronts page lists ChatGPT, Google AI Mode and Gemini, Microsoft Copilot, Meta surfaces and personal AI agents connected to Shop (as at 31 October 2026). Google AI Mode and Gemini products come through the Google & YouTube sales channel or Google Merchant Center. The page names no countries, so check Sales channels > Agentic in your admin. Our posts on Shopify’s agentic storefronts and AI shopping in Australia cover the channel side.
• Weighting is unpublished. Shopify’s optimisation page says optimising can improve a product’s chances but that other factors specific to the AI agent or platform can influence results. It does not say how any field is weighted. Anyone offering a ranked list of “the 7 fields that decide visibility” is guessing.

Gate zero: the rules that remove a product entirely

Before fields, there are entry rules. A product that fails one is not a weak result. It is not in the pool. As at 31 October 2026, Shopify’s requirements page for Catalog lists these (it says the excluded categories can change at Shopify’s discretion, so check it):

• The store is on the Starter plan or higher and is not password protected.
• Products have a title and at least one image.
• Products have a price greater than 0.
• Products are published to the online store, Hydrogen or Headless channel.
• Products are not Unlisted and not hidden from search engines.
• Products are not in excluded categories. The list includes Uncategorized products, alcohol, tobacco and vaping, prescription and medical products, science and laboratory products, adult content, event tickets, live animals, and “Services and product add-ons”.
• The account is in good standing: terms followed, genuine sales history, prompt fulfilment, low chargebacks, no open policy issues. Eligibility is reviewed over time.

Three consequences:

1. Uncategorised is a real gap. With no category in Shopify’s Standard Product Taxonomy, the requirements page lists the product as excluded.
2. Add-ons can be excluded. The excluded “Services and product add-ons” category covers services and add-ons, and the page names extended warranties, installation and maintenance plans among them. If you sell something like that as a separate product, it will not be in Catalog. The page does not name gift wrap or shipping protection, so check how yours is classified.
3. Classification can differ from your admin. Shopify says it classifies products from their product information, which may differ from the category you set, and that images or descriptions with sensitive content can exclude a product even in an allowed category. Read the whole excluded list against your range.

The fields, in audit order

1. Product category

Shopify says all products should be assigned to a specific category in the Standard Product Taxonomy, for example Apparel & Accessories > Clothing > Clothing Tops > Shirts. Products with no category save as uncategorised. Shopify can suggest a category with AI, based on the product name, description and images.

Check four things:

• Is it set at all? Given the exclusion above, this is the first filter to run across your catalogue.
• Is it the most specific category that is true? Accept the AI suggestion as a start, but read it first. It is generated from your name, description and images, so a vague title produces a vague category.
• Have you checked your channels before changing a published product’s category? Shopify warns that new categories might not be compatible with sales channels. Change in batches.
• Do you know it is your only lever? Shopify’s mapping page says you cannot map the product category for Catalog. The admin category is the one that counts.

2. Title

Shopify, Google and OpenAI’s feed specification all require a title. They differ on quality.

• Google’s product data specification says the title should accurately describe the product and match the landing page, with no promotional text, all capitals or gimmicky characters, and a limit of 150 characters. For variants it says to include distinguishing features such as colour or size.
• OpenAI’s help page says ChatGPT may generate simplified titles and descriptions from what it receives, because merchants often use varying titles for the same product. Your title is raw material, not final copy.
• Shopify’s mapping page lets you source the Catalog title from a product field or a metafield, falling back to the default if yours is empty. It does not change your online store.

My working rule, not Shopify’s: a title names what the thing is, in the words a shopper would type, plus the one or two details that tell it apart from your other products. “Stoneware Pour-Over Coffee Dripper, 4-Cup” does that. “BEST SELLER Summer Sale Pour Over Set Free Shipping” does not, and it breaks Google’s rule on promotional text.

Common title problems: offer words, code-only titles (“ST-204” is not something a shopper types), and identical titles on different products.

3. Barcode (GTIN)

Shopify’s barcode page says you can add barcodes to a product or variant. Types are Custom (any value up to 255 characters), GTIN (8, 12, 13 or 14 digits), UPC, EAN, ISBN, ASIN and NS-PID. A barcode registered with GS1 is set to GTIN automatically, and one product or variant can carry up to 20 barcodes. The page says some sales channels might require a GTIN before a product can be published.

Shopify’s Catalog optimisation page lists the barcode (ISBN, UPC, GTIN and so on) among the product details that AI platforms or shopping sites consider. It does not say Catalog requires one. The firmer language comes from Google, which is one of the channels. Google’s GTIN help says products without any unique product identifiers are difficult to classify and may not be eligible for all Shopping programmes or features, and that without a GTIN accurate matching can’t be assured. (Prism3 does not run Shopping ads. I cite the spec as a clear statement of a good product record.)

What to do:

• Enter the real GTIN on every variant of a product that has one. Google says each variant, such as a different colour or size, needs its own.
• Never invent one. Google says to provide a GTIN only if you are sure it is correct, and that a store-brand product generally won’t have one (it points to brand and manufacturer part number instead). OpenAI’s feed specification says to omit the GTIN if unassigned and not to invent a part number to replace a missing GTIN.
• Do not paste your SKU into a field that claims to be a GTIN. A made-up identifier is worse than a blank, because it can match the wrong product.

The audit is quick. Export your products and filter the “Barcodes” column for blanks. For each blank, ask whether the product has a real GTIN you simply have not entered. For resold branded stock the answer is usually yes. For your own unbranded products, a blank is usually correct.

4. Variants and options

Variants are where problems hide, because the page looks fine to a person.

Shopify explains that each combination of option values can be a variant, and its optimisation page says variants and option names are considered. The Global Catalog documentation shows variant-level price and availability, and says get_product returns option values with available and exists signals. An agent can ask whether the 4-cup in black exists and is in stock. That only works if the options are built to answer it:

• One fact per option. Colour in one option, size in another. A single option named “Style” with values like “Black / 4-cup” hides two facts in one string. Google’s specification keeps colour and size as separate attributes, and says to condense a multi-dimension size into one value.
• Real colour names. Google’s specification says not to use values like “multicolor” or “various”, and to list the primary colour first.
• Options linked to the taxonomy where Shopify offers it. Shopify’s category help says category metafield entries can be connected to variant options. Colour, size and target gender are the three attribute names the Global Catalog documents as filters, so get those right first. The documentation does not say how a freehand option name is read, so use the linked option where it exists.
• Availability by variant. If only the 2-cup is in stock, the 4-cup should show as unavailable.

The opposite mistake is splitting one product into several. If “Pour-Over Dripper (Black)” and “Pour-Over Dripper (Sand)” are separate products, an agent may see two listings. Shopify’s option is custom variant grouping in the Catalog mapping settings. By default Catalog uses your Combined Listings configuration (Shopify’s variants page describes combined listings as a Shopify Plus feature, so confirm what applies to your plan). Custom grouping replaces that setup for Catalog, and you can group by product title with a delimiter, by a product metafield, or by a product tag with a prefix. Shopify flags products that cannot be grouped, for example when two variants share the same option combination.

My rule: if the items are one product in different colours or sizes, use variants; if you must keep them separate, use custom grouping. Our product page SEO checklist covers the product page side.

5. Description

Shopify’s optimisation page asks for accurate, detailed and descriptive product information and says clear descriptions help AI platforms match products to searches. Google’s specification adds limits worth following even if you never use Merchant Center: only information about the product (no store links, sales information or competitor details), plain text, up to 5,000 characters.

The common failure is a description that sells a feeling and states no facts. “Elevate your morning ritual” gives an agent nothing to filter or compare, and leaves the inferred layer little to work from.

Here is a test I use (mine, not Shopify’s). Take your best seller and answer these from its page text alone: What is it made of? What are its dimensions, capacity or weight? What is in the box? Who or what is it for? How is it cared for or installed? What variants exist and how do they differ? If you cannot answer most of them, neither can a shopper’s agent. Write the missing answers as plain sentences or a short specification list, not only in an image or a PDF.

Do not stuff keywords, and do not paste supplier text. The Global Catalog documentation says results are clustered by Universal Product ID and include offers from multiple merchants, so your listing sits alongside other sellers of the same item. It does not say whose description is shown, so I cannot claim copied text costs you a match. It does add nothing that tells your listing apart. Write your own, even if it is shorter.

6. Category metafields and other metafields

The evidence here is mixed. Documented: each category has attributes called category metafields. Shopify’s example, the Shirts category, has nine, including size, neckline, fabric and colour. You can connect them to variant options, and editing an entry (renaming a colour, say) updates every variant that uses it. Based on AI predictions, Shopify adds some category metafields automatically and suggests others. Custom attributes are not currently supported. Separately, the mapping page lets you source the Catalog title and description from a metafield.

Not documented: Shopify’s Catalog optimisation page does not mention category or metafields, and no page I read says any agent reads category metafields or how they rank. Vendor blogs say they do, but I cannot back that with a primary source.

My position: filling them costs little, and they hold the facts shoppers filter on. But do them after category, title, barcode, options and description, and check any AI-predicted default: a wrong “fabric” value is worse than a blank.

7. Price, compare-at price and inventory

• Price must be greater than 0. Shopify’s requirements page excludes free products, and Google’s specification says not to submit a price of 0, with narrow exceptions.
• Prices should agree everywhere. Google says the price must match the landing page, the structured data and checkout. Our post on schema markup for AI citations covers the structured-data side.
• Compare-at price is conditional. The mapping page lets you choose whether the list price is the regular or the compare-at price, and says compare-at is used only when higher than the current price. The Global Catalog documentation adds that list_price appears only if the merchant shares compare-at prices with Catalog and the compare-at is higher. That is a pricing-honesty point as much as a data one. The ACCC says any information a business gives about its products, including prices, must be accurate, truthful and based on reasonable grounds (ACCC, “False or misleading claims”, as at 31 October 2026). This is not legal advice: the ACCC does not provide it either, so take your own advice before running a compare-at price you have not genuinely charged.

OpenAI’s help page says that when merchants update pricing or shipping terms there may be some delay before ChatGPT reflects it, so for a short sale assume the agent’s price can lag your page.

Because the default filter is sale-ready items only, a product wrongly showing as unavailable in your system is likely to drop out. One showing stock you do not have sends shoppers to a sold-out page, and the ACCC lists false claims about whether goods are in stock among the things businesses shouldn’t make. Reconcile stock before any big promotion.

8. Type, vendor, collections and tags

Shopify’s optimisation page lists Type, Vendor, Collections and Tags among the details AI platforms or shopping sites consider, without saying how. Keep vendor as your brand, spelt one way: Google’s specification wants a clean brand (max 70 characters) and says not to submit values such as “N/A”, “Generic” or “No brand”, and OpenAI’s feed specification requires a brand field. Shopify says each product has one category and one type, so use type for your own organisation. For collections, see the Shopify collection page SEO guide.

Outside the product: shipping and returns facts

Shopify’s agentic storefronts have a knowledge base for non-product information such as shipping, returns and company policies. Shopify describes the published facts as the source of truth for AI agents on those topics, and says they are publicly available through the Storefront MCP. The page also says the Knowledge Base app is being replaced and is deprecated on 15 January 2027 (as at 31 October 2026), so check the current setup before investing time.

A shopper’s agent needs to answer “when will it arrive?” and “can I return it?” as well as “what is it?”. Shopify says to keep the storefront and knowledge base consistent. Our guide to shipping and returns messaging for Australian stores covers the wording.

The five kinds of gap that make a product unmatchable

After the field-by-field walk, the gaps fall into five groups. This is my framework, not a Shopify classification. Work down the table in order.

A worked example (illustrative, fictional product)

This product is invented to show the process. It is not a client store, and the figures mean nothing outside this example.

A shopper asks an assistant: “Black 4-cup pour-over dripper, mid-range price, ships to Australia.” Broken into what a system needs, that is a category (coffee dripper), a colour (black), a size (4-cup), a price band and a destination. The Global Catalog documents filters for categories, Color and Size attributes, price tier and ships_to, so these are the sort of things an agent can use. Here is the same product before and after an audit.

The “before” version fails at the first gate (no category), then at filtering, identification and answering.

With 120 products, the work is not 120 equal jobs. Sort by revenue, do the top 20 properly, and apply the cheap structural fixes (category, barcodes, option structure) in bulk. Shopify’s product CSV has columns for “Product category”, “Barcodes” and option names, so most can be prepared in a spreadsheet. Back up the original export first and import a small batch to test.

A first-pass audit you can run this week

Do it on your top 20 products by revenue first.

1. Check the gates. Status Active, published to the online store, price above 0, at least one image, category set, not hidden from search engines.
2. Export the products CSV. Filter for blank category, blank barcode, and option names that are not clear attributes.
3. Read ten titles aloud. Would a shopper type this? Does it differ from your other titles?
4. Open each option list and run the description test on your top five products. One fact per option, real colour names, and write the missing facts.
5. Compare price, compare-at price and stock on the page, in the cart and in your structured data for three products.
6. Check your knowledge base facts for shipping and returns against your policy pages.
7. Log what you changed and when. If AI referrals move later, you need the dates.

What not to do

• Do not chase a field you cannot verify. If a blog says “metafield X boosts ChatGPT visibility”, ask for the primary source. Shopify’s optimisation page does not say how fields are weighted.
• Do not confuse product data with schema markup. Admin data feeds Catalog; structured data sits on the page. Fix the admin data first, then check they match.
• Do not treat llms.txt as a substitute. That is a separate question, covered in llms.txt on Shopify.

How to tell whether it is working

You cannot see the agent’s side directly, so measure what you can. Shopify says agentic storefronts give you insights into product search and sales performance, and that orders show channel or referrer attribution. Count AI-referred sessions and orders in GA4 using the method in how to track ChatGPT traffic in GA4 on Shopify, keep them separate from your normal channels, and judge changes over a long enough window. For how ChatGPT chooses products, see how ChatGPT shopping picks products.

Key takeaways

• For ChatGPT, OpenAI says Shopify merchants need no extra channel work, because product data comes through Shopify Catalog. The part you control is the product data in your admin.
• Catalog draws on titles, descriptions, images, categories, variants and options, prices and inventory. Shopify does not publish how any field is weighted, and inclusion does not guarantee placement.
• Entry rules come first: uncategorised, Unlisted, hidden-from-search and zero-price products are not eligible.
• Set a specific category on every product. It is your only category lever.
• Enter real GTINs on branded products, variant by variant. Never invent one.
• Keep colour and size as separate, clean options.
• Write descriptions that answer factual questions: material, size, contents, fit, care.
• Category metafields are cheap, but no primary source says agents read them, so do them after the basics.
• Keep price, stock and policies consistent everywhere.

FAQs

What product data do AI shopping agents use from Shopify?

Shopify says Catalog draws on titles, descriptions, images, categories, variants and options, pricing, and inventory and availability. Its optimisation page also lists Type, Vendor, Collections and Tags, barcodes such as GTIN, UPC or ISBN, and variant option names among the details AI platforms consider. Shopify does not publish how each is weighted.

Do I need a GTIN or barcode on every product?

No. Add a barcode where the product has a real, assigned one, on each variant. Google says a store-brand product generally won’t have a GTIN, and to provide one only if you are sure it is correct. Shopify lists the barcode among the details AI platforms consider but does not state that Catalog requires one.

Does the product category matter for AI shopping?

Yes. Shopify’s requirements page lists uncategorised products as excluded from Catalog, and the Global Catalog documentation includes a category filter. You cannot map the category for Catalog, so set it in your admin and review the AI suggestion first.

Do I have to fill in metafields?

Not as a first step. Category metafields hold facts such as colour, size and fabric, but Shopify’s Catalog optimisation page does not mention them and I found no primary source saying agents read them. Do category, title, barcode, options and description first, then fill the metafields that apply.

Will fixing my product data get my products shown in ChatGPT?

It improves the chance a product can be matched, and nobody can promise more. Shopify says inclusion does not guarantee a product will appear in a specific AI answer and that each channel controls final ranking and wording. OpenAI’s help page says ChatGPT considers the shopper’s query and context.

Is this the same as Google Merchant Center?

Not exactly. Shopify says products reach Google AI Mode and Gemini through the Google & YouTube sales channel or Google Merchant Center, and describes Shopify Catalog as the source for other AI channels. Google’s product data specification is still a useful model of a clean record. Prism3 does not run Shopping ads.

Want a second pair of eyes on your catalogue data?

Product page and structured-data work sits within our SEO and GEO services; we include GEO alongside SEO rather than selling it on its own. Nobody can guarantee that a product will be named by an AI assistant, so we do not. If you want to start with a conversation, book a free 30-minute diagnostic or tell us about your store: we look at your store, and at what the AI assistants say about your category, before we recommend anything. No pitch.