How does agentic commerce work?
A shopper might ask an assistant to find a laptop bag that fits a particular computer, protects it from rain and arrives before a trip. The assistant needs to find available products, compare them against those requirements and identify a seller. Depending on the service, the shopper follows a link to the store or completes a supported checkout inside the assistant.
A recommendation does not mean that an agent has permission to buy independently. The checkout and any delegated spending depend on the service and the customer's authorisation. For a business, the practical question is whether the recommended offer accurately represents what it can supply.
OpenAI says ChatGPT shopping results can use product metadata and outside content, together with the shopper's query and context. Its product results are selected independently of advertising. Accurate information helps the assistant assess your offer; it cannot guarantee a recommendation. How ChatGPT selects shopping results
Can a small business sell through ChatGPT?
Eligible Shopify merchants already share product data through Shopify Catalog. Shopify's ChatGPT route sends buyers to the merchant's own checkout. The store must sell to US customers, although it can be based elsewhere. Check Sales channels → Agentic in the Shopify admin for your position. A UK-only seller should not assume it qualifies. These requirements were checked on 23 October 2026. Shopify's ChatGPT guidance
OpenAI also documents direct product feeds. Check the available access and your platform's existing integration before commissioning another feed. Native checkout inside an assistant is a separate capability with its own integration requirements. OpenAI's product-feed specification
Google's Universal Commerce Protocol supports transferring a basket to the merchant and native checkout on Google. Its Merchant Center integration hub is initially rolling out in the US. The existence of these routes does not establish that your store has access or needs a new integration. Google's integration guide
What should I do first to prepare for AI shopping?
For a first pass, choose three important products, one customer market and a shopping service relevant to your buyers. Include a straightforward product and one with variants or compatibility requirements. Save the baseline, inspect the buying route and identify the first fault to fix. This is a suggested starting scope; set a time and cost limit that fits your team before extending it.
Check channel eligibility and product-feed errors. Ask your website provider whether important product pages can be indexed and whether crawler rules block the services you want to reach. For Google Search's AI features, the normal SEO requirements apply; Google requires no special AI text file or schema. Google Search guidance
Use the first inspection to find the most consequential fault. A wrong compatibility claim or a link to the wrong size can justify immediate correction. An accurate recommendation for a competitor may simply mean its offer better meets that shopper's needs. Correct and retest a clear fault before expanding the product group.
How do I optimise product pages for AI shopping?
Start with the questions customers ask in enquiries, reviews and returns. Which model fits? What comes in the box? Will it arrive in time? Put the answers that could change a purchase on the relevant product page.
Reconcile the exact variant and identifier, photographs, dimensions, compatibility, price, stock, delivery terms and return conditions across the page, feed and important retailer listings. Give one person responsibility for checking that changes reach each place. Google identifies conflicting website/feed information and incorrect variant attributes among causes of product display and eligibility problems. Google's product-data guidance
For an illustrative laptop bag, “fits most laptops” leaves the buyer guessing. Publish verified compartment dimensions and explain how to compare them with the device. If rain protection matters, state the supported conditions and any relevant test evidence. A claim that the bag is waterproof needs substantiation.
When the choice spans several models, write a comparison guide and link to the exact products. Keep limitations visible and seek honest customer feedback. Adding more copy helps only when it supplies information the buyer needs. Our brand-building Answer explains why the product and service experience belong in the work too.
How can I test AI product recommendations?
Use a requirement drawn from customer enquiries and save three kinds of question:
- “Which [product type] can [required use] and is available in [market]?” Leave out your brand to test discovery.
- “Does [exact product and variant] support [required use]? Show the source.” This helps check understanding.
- Ask whether the product supports a use you know it cannot fulfil. A confident yes is an error even when it recommends your brand.
Record the question, date, assistant, market, recommended variant, cited source, seller and next action. Repeat important questions in fresh sessions under the same conditions. Providing your product link helps test interpretation but does not show whether the assistant would discover you unaided.
Apple's AirPods 5 illustrate the difference. The Wireless Charging Case supports Qi-certified chargers; the USB-C-only case does not offer that charging route. “Which AirPods 5 case can I charge on a Qi pad?” has a checkable answer. This is an illustrative exercise, not a test we have conducted. Apple's specifications
If the answer claims both cases support Qi, inspect the cited source and variant. Correct a wrong listing where it occurs, including a retailer listing if that supplied the error. If the source is right, check for conflicting listings and repeat the tests. Keep accurate specifications intact and preserve evidence for the service's feedback route; a report does not guarantee a correction.
If the recommendation is right but the link opens the USB-C-only case, ask your provider to trace the link and variant mapping, involving the shopping platform if needed. Retest the corrected route. If a competitor alone can meet the buyer's required delivery date, the offer needs attention. Better wording cannot supply an unavailable service.
How do I check the checkout works?
Follow the supported buying route on mobile. Check that the right variant, price and stock survive into the basket, delivery charges and timing are clear, and payment, confirmation and support work. Resolve a broken purchase before spending on more discovery content.
Ask your commerce provider to test changed stock, failed payments and duplicate requests in its sandbox or test mode, or with explicitly authorised test orders. Do not assume a feed appearing correctly means the purchase route works. A new native-checkout integration needs a commercial reason as well as technical access.
How do I measure whether the changes helped?
Repeat the saved questions after the update has reached the relevant page or feed. Keep a change log and count correct, incorrect and unclear answers across the same cases. Fewer errors establish improvement in those checks, with sales still to be measured.
Track the referred visits or channel orders you can identify, purchases, returns and the money left after relevant product, delivery, payment and service costs. Record the seller: a stockist's sale may help your brand without creating a direct-store order. Keep unknown attribution visible and exclude test activity.
Where catalogue size and order volume allow, improve one group of comparable products first and retain another temporarily for comparison. Track pricing, promotions and stock changes in both. If both groups improve equally, the page edits are a weaker explanation. A better result for the edited group justifies further investigation, although other differences may remain. A very small store can start with dated fault checks and leave sales attribution unresolved.
Expand only when the route works consistently and the likely commercial value justifies more work. If relevant demand is weak, retain the useful product-information repairs and defer further integration spending.
Our weekly brand-launch measurement Answer explains how to keep early results in perspective. The Centre's agentic commerce research review examines the evidence behind product-information effects and their commercial limits.