Agentic commerce for merchants: Get listed and prepare for AI checkout

What is agentic commerce?

Agentic commerce lets software act on a shopper's request by finding products, comparing offers, preparing a cart or completing an authorized purchase on a supported channel. The customer supplies the goal and constraints. How far the software can act depends on the channel, merchant and payment setup.

Where to start

Say a shopper asks for a replacement water filter that fits a specific refrigerator, arrives before the weekend and stays within budget after delivery. The assistant can identify the compatible part and compare sellers. An offer link leaves the buyer to select and pay on the store. Cart transfer can carry the chosen item and quantity into the merchant's checkout. Native checkout can let the buyer approve and pay within the shopping surface if the offer and merchant are eligible.

How does agentic commerce differ from product recommendations?

An ecommerce chatbot can answer questions or guide shoppers through a fixed workflow. An agentic shopping assistant can act on a new instruction, such as comparing offers or passing a selection to a supported cart. A chatbot can include those capabilities. On-site personalization usually chooses what to display from one retailer’s assortment, based on browsing history or cart contents. An assisted checkout might open a store's normal payment form and leave the next action to the buyer. A delegated purchase requires a documented authorization and a transaction path that supports it. Identify the action before deciding which engineering or merchandising team should own it.

Which retailers are using agentic commerce?

Google announced that UCP-powered direct checkout began rolling out to Canadian shoppers in AI Mode and Gemini on November 4, 2026. It named lululemon, Shoppers Drug Mart, Simons and select Shopify merchants as examples of eligible retailers. Google said simplified Merchant Center onboarding would follow in Canada. A shopper rollout with named merchants and the account tools for a wider group can arrive on different dates. Google Canada rollout announcement.

Why start with discovery?

Start with product discovery when your storefront already carries the right variant into checkout. Adobe Analytics counted visits from generative AI tools to US retail sites during the 2025 holiday season. Track those visits separately from completed orders when deciding where to invest.

Adobe projects 130% growth in AI referrals to US retail sites for the 2026 holiday season compared with 2025. Adobe Analytics 2026 forecast.

On Google, Pew Research Center found traditional result clicks on 8% of search visits with an AI summary and 15% on visits without one in its March 2025 browsing panel of 900 US adults. The panel covered general Google searches. Use it as context for measuring appearance and visits separately, with your store's own product and order data for the sales question. Pew browsing study.

A merchant with limited development capacity can start with offers shoppers already buy. Confirm the right destination, make the selected variant's facts easy to compare and use the current storefront to finish payment. A checkout build earns a place when an eligible surface can carry those offers farther and the team can support the orders it receives.

How do you get products listed in ChatGPT Shopping?

OpenAI says Shopify and Etsy catalogs are already integrated with ChatGPT. Merchants using other setups can apply to share a direct product feed. ChatGPT can also discover product pages through web search. Its public merchant shopping experience sends buyers to a merchant-owned website or app to finish the purchase. OpenAI merchant guidance.

Use the route for the storefront you sell through. Start with the public listing if you sell on Etsy. Once you confirm access, test a live offer in the US shopping experience. Check availability separately for any other market you serve.

Shoppers can describe preferences or start with a reference photo. Shopify’s ChatGPT demo pairs a request for sustainable skin care with product cards showing names and prices. The merchant’s destination is where you check the actual offer.

ChatGPT ShoppingShopify example

The shopping request

Suggested products

Select a card and check the seller, variant, currency and availability on the destination page. A visual match does not establish that the linked item is the option the shopper intended to buy.

Where to check Shopify Agentic Storefronts

To verify your Shopify agentic commerce setup, open Sales channels → Agentic in Shopify Admin. Under Sources → Shopify Catalog, check the product count. Open the menu on that row and choose Review product data mapping. If you see Ineligible, open the warning and record the reason. Then check ChatGPT’s catalog access, your US selling market and acceptance of the channel terms. In Settings → Policies, confirm the terms of service, privacy policy and return and refund policy required for the channel. Verify the mapping for the specific offer before treating it as ready.

Shopify Catalog · product fieldsShopify example

The channel switch and catalog count locate the connection. Open the menu beside Shopify Catalog and choose Review product data mapping to inspect a particular product. Compare its title, image, options and URL with the storefront, keeping both the product ID and variant ID.

An eligible catalog gives the channel access to item data. Selection still depends on the shopper's request and the shopping experience. If the mapped offer is eligible but you cannot find it, use the missing-product checks to isolate the issue.

What should Etsy sellers check?

OpenAI identifies Etsy as a connected catalog. Start with the live Etsy listing and the option the buyer asked for. Select it signed out and check stock, delivery country and price. If several options share one listing, save the option name and the page state with the URL. That gives marketplace support a specific record to inspect if the result points to the wrong offer. OpenAI merchant guidance.

If you sell through your own storefront as well, record which seller and destination appeared. The same brand or item can surface through different merchants. Follow the public URL you control and verify that the relevant search crawler can reach it. If your store also targets Google, use the separate Merchant Center route.

How to sell on ChatGPT with merchant checkout

A product card can send a buyer to a merchant checkout even when the website opens inside an app browser. The store still needs a destination that loads the correct variant and offers the right payment and delivery choices. Follow the result on a phone as well as desktop. A popup, forced country switch or link that opens the parent product can change the buyer's next step without changing what the card says.

OpenAI and Stripe introduced Instant Checkout in ChatGPT on September 29, 2025. OpenAI’s current merchant guidance prioritizes product discovery with merchant-owned checkout. Stripe’s launch announcement. Its technical checkout specification describes a flow for approved partners. A merchant exploring a separate integration should confirm availability for its own account with OpenAI. Current merchant checkout guidance and OpenAI’s checkout specification.

OpenAI puts direct feed applicants on a waitlist. While you wait, open a few live variant URLs signed out and verify their prices and availability. Fix any destination that drops the buyer’s selected item before checkout.

What goes in a ChatGPT product feed?

A ChatGPT product feed gives OpenAI stable item IDs, accurate offer details and public URLs for each product and image. The stable file specification lists nine required discovery fields. A variant row should describe the option the buyer can select and the offer the destination will show. OpenAI stable product schema.

Build a crosswalk from your source system to the export before changing the file. A parent product record can provide the brand and common description. The sellable variant should provide the selection, its URL and its offer. Keep a known example next to each mapping so a future change to a theme, pricing app or feed connector has something concrete to test.

Give each sellable option its own record

A parent record groups variants. The stable OpenAI file uses a distinct item_id per selection, a shared group_id, listing_has_variations set to true, and a variant_dict with the selected options. Each row carries its own price, availability, URL and images. Use the same option names across the group. One item can be a blue mug and another a white mug, with the grouping visible when buyers compare the two.

Before choosing the export ID, look for product families where the printed manufacturer identifier is shared by several merchant listings. A case quantity or bundle can use a different sellable item in your system. The feed must keep that orderable unit distinct. If your connector keys by parent SKU, test that it exports separate rows for the variants with different prices, inventory or destination URLs. Keep the identifiers stable when a product goes on sale.

Use the actual assigned product identifiers where supported. Keep leading zeros and leave an unassigned GTIN empty. Write out the capacity, material or compatible model on the product page as well as in any structured attributes. An item named for a color such as “Storm” still needs the product type and actual color when those facts affect the buying decision. Variant and identifier rules.

Operate the feed as a catalog export

OpenAI’s documented SFTP route prefers Parquet and also accepts compressed JSONL, CSV and TSV files with the .gz extension. Use UTF-8 for text exports. OpenAI recommends a complete snapshot at least daily, using stable filenames or a stable shard set. Its product URL rules allow absolute HTTP or HTTPS and prefer HTTPS. Validate a sample first, then compare the exported records with the merchant's current offers before relying on scheduled delivery. OpenAI file upload overview.

Keep the publication timestamp with the export. If you can see processing or acknowledgement status, record that too. An automation can finish sending a file while a missing field prevents a row from being used. A price change that has reached the page can still be absent from the last export. Log the first place the new value appears and the first place the old value remains.

Give discontinued items an explicit removal workflow. The documented snapshot behavior can retain a missing record for up to 14 days. For search ineligibility on the next processed snapshot, OpenAI documents is_eligible_search=false. Compare the processed status with the desired listing state after the update. This distinction matters when a merchant retires an item, reverses an incorrect option or needs to stop surfacing an old product URL. OpenAI retention and removal rules.

How do you verify a ChatGPT product feed update?

Compare the exact variant in the SFTP snapshot with the same item or variant returned by OpenAI’s Products API, where your feed has API access. Read the feed’s current products at GET /product_feeds/{id}/products. Submit product data updates through SFTP. Check the submitted eligibility flags and any processing or item status your connection exposes. A public product card needs its own test. OpenAI Products API and API overview.

When a price changes, save the new file’s timestamp and the exact variant ID, destination, currency and price you intended to publish. During a sale, compare the file’s regular price with the API’s list_price, and the file’s sale_price with the API’s active price, when those fields are present. The API uses minor currency units, so an amount of 7999 in USD represents $79.99. Check that the feed’s target country and the variant’s availability also match your intended offer. API price fields and SFTP price fields. If the updated record is missing, ask the feed owner to check file delivery and processing. When the record is current and the card differs, move to the seller and destination checks.

Which product records expose feed mapping errors?

Include a multi-variant product, a temporary promotion, a low-stock item, a preorder and a product restricted to one market. A test on a simple in-stock item can pass while the export sends a subscription price as the regular one-time price. Take the item URL from the exported row, open it in a clean session for the target market and compare the selected item, currency and quantity.

For a product available as a one-time purchase, send its public regular price in price. Reserve the stable file’s sale_price for a current lower price a target shopper can get without a membership, code or subscription. Keep conditional prices on the store and verify that your channel supports the specific condition before sending it. Update the export when a public sale starts or ends, then open the destination in a clean session to check the displayed offer. OpenAI’s price and promotion fields.

Set an owner for each kind of discrepancy. Catalog operations can correct a missing option name, engineering can fix a variant link that drops its selection, and the pricing owner can decide how subscription and promotion terms should be presented. A single assignment to “fix the feed” obscures where the first incorrect value entered.

How do you get products into Google AI Mode and Gemini?

Start with accurate, eligible items in Google Merchant Center. Google uses Merchant Center product data for shopping discovery, and its Universal Commerce Protocol integration can add cart transfer or native checkout on eligible offers in AI Mode and Gemini. The channel, merchant account and product must support the capability you intend to use. Google's UCP guide and Merchant Center onboarding.

Open the item in Merchant Center for the market you serve. Compare its processed price, currency and availability with the exact landing page and the selected item in checkout. Check free listings status, shipping services and returns for the same country. A product can be valid in one market and have an issue in another, so start with the country from the observed shopping result.

Check the product and account before choosing an integration

In Merchant Center, check the item’s visibility and status for the country and marketing method. Open Products → Needs attention for a limited or disapproved offer and choose View products on the issue card to find the affected items. If the item is hidden, check the visibility setting. An approved item is eligible for its marketing method. Sample AI Mode and Gemini separately to see what those surfaces display. Merchant Center product visibility and status.

Merchant Center · Needs attentionGoogle Help

Availability of Google’s integration hub varies by merchant and market. Google’s US rollout guidance describes a direct route in which the merchant prepares the implementation, registers interest, waits for selection and uses Merchant Center’s sandbox to validate capabilities. Ask an existing platform or provider which route it manages before you plan a separate build. A channel toggle can depend on the account, merchant and market. Google's UCP integration hub.

Shopify stores can send product data through the Google & YouTube sales channel or directly to Merchant Center through a feed provider or manual upload. Confirm that the exact item and its shipping data reach the intended Merchant Center account. For Google direct checkout on Shopify, also check the item’s Shopify Catalog eligibility and the checkout options available to your store in the Agentic sales channel. Google’s product feed and Shopify Catalog eligibility are separate checks. Shopify’s Google setup reference.

Which products can use Google native checkout?

Google’s native checkout excludes subscriptions, personalized items, preorders, used or final-sale goods, bundled services and products needing non-standard delivery fees, among other restrictions. Screen the proposed assortment against the current account guide before opting it into checkout. A restriction on native checkout does not by itself settle discovery or cart-transfer eligibility. Google’s checkout product restrictions.

Configure return policies and at least one customer-support contact method in Merchant Center. Supply applicable consumer notices with the product data. The notice requirements depend on the item and market, so include that review in account preparation before testing checkout. Account settings and consumer notices.

If you integrate directly, the UCP profile declares the capabilities your store supports. Declaring dev.ucp.shopping.cart makes free-listings-approved items eligible for cart transfer by default. Use native_commerce(cart_eligibility)=FALSE to exclude an item. Native checkout requires account onboarding and native_commerce(checkout_eligibility)=TRUE for the item. If you use both capabilities, send both values in native_commerce(checkout_eligibility:cart_eligibility) on every upload. The upload replaces the whole flag group. Test the buyer’s route after processing. Merchant Center account preparation and UCP profile setup.

Make Google item IDs resolve to the intended checkout item

When the Merchant Center id differs from the ID expected by your checkout API, set merchant_item_id to the checkout ID. It takes precedence over id in Google’s checkout requests. Test with a product whose variants differ on both price and availability. The intended variant should arrive in the cart with the correct option values, and changing it should cause the appropriate price and stock check. For Google UCP, the documented variant_option fields let merchants describe the option dimensions used in the selector. Google's checkout item mapping.

For the replacement filter, the pack count can be a variant_option when it distinguishes selectable offers. A question_and_answer field can address whether the part fits a specific refrigerator model, using the compatibility already stated on the page. Add a document_link if the installation sheet resolves a real question. These Google conversational attributes are optional. Use Merchant Center’s status screen to resolve approval issues. Merchant Center conversational attributes.

Keep discovery work moving while an account is waiting for checkout. Repair disapproved items, confirm variant URLs in a signed-out browser and publish shipping and return terms for the actual destination market. Read Google's AI features guidance for search inclusion and the Merchant Center specification for shopping data. They cover different surfaces. A web result citation, a free listing and a native checkout button each have their own acceptance evidence. Google Search guidance and Google's product data specification.

What can a shopping agent verify about one product?

Variant-level product data connects the selection a shopper asks for with the offer they can open and buy. Publish the deciding facts in visible text, then give the exact selection a stable destination and matching price, currency and availability in the markup and feed. A buyer should be able to recognize the same offer after following the link.

Death Wish Coffee's public Dark Roast Coffee page shows how to test this. Its base page markup listed eight offers by grind and weight. The five-pound whole-bean URL displayed an image of the whole-bean bag and a price of $84.99. On the same visit, the product controls showed Ground and 2 lbs selected. The base markup listed that ground option at $39.98, and the conflicting selection was visible before checkout. Public product page and Five-pound variant URL.

Dark Roast Coffee · Death Wish Coffee

What shoppers seePublic product page

What agents can read in the product markup

Requested by URL

Whole Bean · 5 lbs50255W · USD 84.99 in markup

Selected on page

Ground · 2 lbs50252GK · USD 39.98 in markup

When the price and checked options point to different offers, inspect which variant ID the Add to cart form sends. Match that ID and the cart price against the option the shopper intended to buy.

Make the buyer's deciding facts easy to check

Build the description around the decision. For coffee, grind, bag weight and one-time purchase or subscription terms affect the offer. Replacement parts may need a model number, dimensions and an explicit compatibility range. In apparel, buyers may need garment measurements, material and the return rules for the sale color. Use facts a buyer could ask customer service to verify.

Put the short decision facts near the item selection. A photo can show a bag or a connector, and the accompanying text can say which weight or socket the image represents. If a table is needed, label its units and state which selections its rows apply to. Name the product and its options so buyers can tell similar variants apart. The product-page writing guide goes deeper on the copy and markup.

Review the text for substitutions that make offers harder to distinguish. “Large” may describe dimensions, capacity, or a package of multiple units. “Compatible” needs the model or standard it refers to. “Ships fast” could mean orders leave the warehouse quickly or arrive by a certain date. Give the buyer the specific constraint and the condition that changes it.

How can price expose the wrong product option?

The coffee page’s recorded prices distinguish bag weight, while both grinds share the same amount at each weight. Use price to flag a possible pack error, then check the selected grind separately.

The amount can help you spot the wrong pack weight. To confirm grind, compare the destination’s selected option and variant ID with the offer record.

Make structured data identify the offer on the page

Product schema describes the offer a search crawler can read on a public product page. A merchant feed sends catalog records through a channel’s approved connection. Web crawlers can discover an offer from the public page. Merchants using a feed also need to maintain the channel’s fields. Keep both aligned with the visible selection, since they can update on different schedules. ChatGPT product discovery routes and Google’s product data specification.

Use product structured data to communicate the relationship among related variants and offers. Google's ProductGroup guidance distinguishes variants selected on one page from variants with their own pages. A single-page group can have a parent canonical URL and distinct variant links that open preselected options. Separate variant pages need enough local markup for Google to understand the product without relying on an entity defined elsewhere. Google's product variant documentation.

Inspect the markup for the offer corresponding to the URL. The first Offer may describe the base variant. A minimum price across all offers can belong to a different selection than the one a buyer requested. Compare price and stock at the level of seller, market and selection, then test the displayed selection itself. A structured-data validator can show that the fields parse. Opening the variant checks a different part of the journey.

A shopping agent can compare shipping and returns when the merchant publishes the terms for the exact variant and market. Keep the machine-readable signals aligned with what a shopper can see. State where you deliver, how fees depend on destination and which categories have a different return rule. When the final shipping quote needs a postal code, say how the buyer can get it. If a return window varies by item or market, give the policy a clear route from the affected offer. Treat a shipping rate that depends on the address or the full cart as provisional until checkout returns the actual quote.

Once you can reproduce the price or variant error, check the rest of the catalog for missing category facts that could cause similar mistakes.

Why is an AI shopping result wrong or missing?

When an AI shopping result shows the wrong price or variant, record the seller, selected option, currency and destination URL from the result. Compare that same offer across the rendered page, structured data, processed feed and cart. The first recorded disagreement tells you where to investigate. A product that never appears needs an eligibility and access check before you diagnose its wording.

Save the prompt, channel, country, date and product card or answer in a ticket. Copy the final destination after any redirect and note the price a signed-out customer sees. Start with one symptom the next person can reproduce. “Whole bean 5 lbs opens with Ground and 2 lbs selected” gives a developer a narrower job than “our product data is wrong.”

Why is my product missing after the feed is accepted?

Check four stages when an accepted feed produces no visible ChatGPT Shopping result. Find the exact variant in the processed catalog, check its eligibility for the market and discovery surface, open the destination as a signed-out buyer, then test selection with fixed shopping prompts. File acceptance covers the submission. Each later stage needs separate evidence.

1. Processed variant. Use the stable item ID and variant ID or SKU. If the item is missing, compare the latest file with any processing details you can see. A feed may be accepted while the specific item is absent or needs correction. For Shopify or Etsy, start with the corresponding connected catalog or listing record.
2. Market and surface eligibility. For a direct ChatGPT discovery feed, inspect the submitted is_eligible_search value and any available item status. Confirm the country and channel route. Check discovery separately from checkout or ads. For Google, use the item’s Merchant Center status for free listings in the target market. OpenAI product flags and Merchant Center status.
3. Exact public offer. Follow the destination in a clean session and record the final URL, selected option, price and currency. Check relevant crawler access when the web discovery route matters. Send the first visible discrepancy to the storefront owner with the URL and observed values. OpenAI crawler documentation.
4. Selection test. Run a branded lookup for the item and a non-branded prompt with a real buying constraint. Keep the answer, cited links and product cards distinct. If the right seller appears for the lookup, use the constraint prompt to check whether the offer communicates the details needed for comparison. For an absent offer, save the tests and report the outcome as a sample.

When escalating, include the market, channel, stable item and variant IDs, latest feed timestamp, known eligibility, destination and the exact prompt. If any account-level fields are unavailable, label them unknown. That lets a channel owner investigate without repeating the public storefront work.

Find the first place the offer changes

Read the exact variant URL first. On the coffee page, the matching price and gallery image would pass a quick spot check. The checked options reveal a different selection at the same URL. For your own store, inspect which variant ID the purchase form will send before you authorize any transaction. Check the destination in a private browser and with the market selection your buyer uses.

Then open the processed record available to your team. In Google Merchant Center, start with the item and its product data source. A Google mismatch diagnostic concerns Google's copy of the offer. For ChatGPT, check the catalog connection or direct feed you use and whatever processing evidence it exposes. An error in one channel can reveal an issue worth testing elsewhere, but the channel's warning establishes only what it observed. Google's offer data rules.

If the storefront and cart show the current price but the processed feed still shows last week’s sale, compare the promotion end time with the export timestamp. Check the source price if the export still has the sale. If the export carries the current price, check processing errors and which feed or integration supplies the channel’s price.

If two merchants sell the same product, open the result’s merchant options before comparing prices. Match the seller, exact variant, currency, sale terms and final destination. A price from a reseller or another pack size belongs to a different offer. If the result names your store and the values still disagree, use the captured card and processed item to isolate which system has the older value.

How to get recommended by ChatGPT Shopping

To help ChatGPT Shopping match your product to a buyer’s request, publish the exact variant, price and availability. State the buyer’s deciding constraint in plain language, such as which refrigerator models a filter fits. For a direct feed, check that the intended variant has search eligibility enabled. OpenAI’s product fields and eligibility rules.

Try a branded product query with an exact variant, then repeat with a non-branded requirement your product can meet. Use a clean conversation for each run. If you get only the branded result, note which comparison facts and offers appear in the non-branded answer. A product can meet the technical channel rules while another seller offers faster delivery or shows a clearer fit for the buyer's request. Change the product facts you can substantiate and test the same prompt again.

If neither query finds the item, check channel status before tuning descriptions. On public pages, confirm a successful response, visible product facts, index controls and crawler access. In a 2026 merchant-side readiness preprint, Adrian de Valois-Franklin and colleagues found that 105 of 555 homepages in its sample were unreadable to both its declared bot and declared headless browser. The authors drew their sample from clients named by two payment processors and describe the results as a pilot. For a specific store, evidence comes from the actual page and access route. Merchant-side benchmark preprint.

Close the ticket against the layer you repaired. For a storefront fix, open the URL in a clean session and verify the selected option and the ID the form uses. Check the processed item to confirm a feed fix. Date the separate shopping-result retest so the ticket shows what the channel displayed that day. Separate these outcomes so a missing recommendation cannot conceal a corrected feed or a persisting checkout bug.

How do you audit agentic commerce readiness?

Audit agentic commerce readiness by testing access, product accuracy, purchase terms, the checkout path and order handling for a defined set of offers. Record what passes, what fails and what remains unknown because the data sits inside a merchant account. Start with a market you serve and products an actual buyer could order.

Choose a best seller, a multi-variant item and an item whose stock or offer changes frequently. Include a product with a shipping or return exception that applies to your business. This selection exposes distinct failure modes. A low-stock variant can reveal an update lag that a best seller with ample inventory will miss.

What should an ecommerce AI-readiness audit check?

Record five checks for each offer: access, offer accuracy, purchase terms, the cart or checkout handoff, and order handling. Assign the public destination to the storefront owner and the processed offer to catalog operations. Fulfillment verifies delivery and returns. Engineering tests the handoff; payments and operations reconcile the order and refund evidence.

A signed-out shopper can inspect the page and its markup, try the public cart and read the delivery and return policies. A public check can also sample an AI answer for a specific question. Use the merchant account to inspect processed items and run controlled order, payment and refund tests. Record the evidence and the person who can act on it in the same sheet.

Use one row per item and market. Include the parent product, variant ID, direct URL, observed price, currency, stock state, shipping destination, return exception and last observed date. Add a pass, fail or unknown for each check. Mark missing shipping terms as unknown. Assign an owner and a retest date.

Work through the dependencies in order. Repair the variant URL before troubleshooting a payment test. Wait to send a larger assortment until the processed items carry accurate prices and stock. If the catalog and handoff pass and native checkout is unavailable, keep discovery and merchant checkout active as the pilot path. Revisit native checkout once the account supports it and the team can test it.

Decide how far the integration needs to go

For a retailer with complicated configuration, account pricing or recurring shipments, the merchant site may give buyers the clearest explanation and support. A catalog of straightforward repeat purchases may be easier to pilot in a supported native checkout. Set the decision around the buyer's needs and the failures you can handle. Platform work earns its budget when a verified requirement exceeds what the current storefront can do. See the ecommerce website development guide for scoping a build decision.

How do ACP, UCP and AP2 fit together?

OpenAI’s product feed and checkout specifications use the Agentic Commerce Protocol (ACP). The Universal Commerce Protocol (UCP) defines capabilities such as cart and checkout, while the Agent Payments Protocol (AP2) handles authorization for payments an agent initiates. Choose the specification that matches the action and channel you can support.

Start with the business operation you need to support. Use stable catalog data and a documented delivery route for a product feed. The cart handoff must map the item to a working merchant checkout, while payments initiated by an agent require a record of the shopper’s authorization and a way to reconcile the order. These jobs can involve overlapping specifications.

Confirm which capability your merchant account can use and check the channel’s participation requirements. Then open the protocol for that operation. Google announced AP2 on September 16, 2025. AP2’s closed Checkout Mandate records authorization for a finalized checkout. A closed Payment Mandate binds a specific amount to that checkout and payment instrument. Open mandates capture earlier constraints, such as the purchase goal, budget or allowed instruments. Ask which record your checkout and payments teams can retrieve if a shopper disputes the cart contents or the charge. Google’s announcement. ACP specification, UCP specification and AP2 documentation.

How to scope a Universal Commerce Protocol handoff

For a UCP cart or checkout integration, list the item identifier, quantity, price and currency at the boundary between the shopping surface and your store. Document which service owns the shipping quote and tax calculation, and which identifier appears in the receipt and the order dashboard. When a shopper changes an address or quantity, the order flow should recalculate before final approval and record the new state.

Apply the same discipline to status and support. A tracking link in the shopping surface is useful only if it follows the order your fulfillment team sees. Define who communicates a cancellation and which system issues the refund. An implementation scope that ends at the payment screen leaves those operating questions for the first customer who has an exception.

When should a store enable AI checkout?

Enable native AI checkout for a defined set of eligible offers once test orders reconcile through shopper approval, payment, fulfillment and refunds. When product discovery works and merchant checkout already handles the necessary choices, use that path while measuring demand. Cart transfer is worth testing if shoppers benefit from keeping their selected items as they move to your store.

Shopper approval and payment location answer separate questions. Native checkout can still require a human to confirm a purchase. An in-app browser can display a merchant checkout where the customer pays on the seller's site. A delegated purchase depends on the authorization a person has granted and the payment path that can enforce it. Label the path by what the shopper can do and where the merchant receives the record.

How do you test an AI checkout before rollout?

Test with the constraints that change the order

Run a normal test order, then exercise the failures that a valid request can encounter. While the cart is open, try a changed shipping address, a sale of the last unit, an expired promotion and an updated quantity. If an expected term changes, the buyer needs to see the updated offer or a clear error before the transaction proceeds.

Ask the technical owner how the checkout responds when a payment or order call times out. Retrying an unknown result can create duplicate payments or orders if the integration treats the retry as new. In the channel's supported test environment, submit the same authorized attempt again and reconcile its reference with the merchant's order and payment records. Have finance and fulfillment review the same result.

Include support in acceptance. Give someone the buyer-facing receipt and ask them to locate the order, state who ships it and identify the refund route. If support requires a special internal ID the buyer never sees, build the crosswalk before opening the channel to more products. A buyer should be able to solve a post-purchase problem even when the original shopping conversation is closed.

How do you measure ecommerce AI visibility?

Measure ecommerce AI visibility with Merchant Center’s organic shopping data, Search Console’s AI supporting-link impressions and a dated sample of real answers. Track ChatGPT referral sessions separately in GA4, then reconcile paid orders in your commerce and payment systems. Each source measures a different event.

Start with the question you need to answer. Merchandising may need to know whether products surface for shoppers comparing fit or availability. Search teams can check which product pages received links. A checkout owner needs to know what buyers did after landing or approving a purchase.

Where can you see Google AI shopping visibility?

Open Analytics → Products → AI performance in Merchant Center to track cumulative product appearances in Google’s organic AI shopping results. The report covers organic, shopping-intent queries in AI Mode and AI Overviews for English-language queries in eligible accounts in Australia, Canada, India, New Zealand and the US. Filter by product category and country. Review shopping-stage share of voice, top terms and popular attributes to see which product facts deserve a check. The same product can count more than once. To identify the exact offers shoppers see, sample the answers directly. Merchant Center AI performance insights.

Merchant Center · AI performanceGoogle example

In Google Search Console, select the Generative AI report for Search results under Performance. It measures impressions for links to your site in Google AI Mode and AI Overviews, with breakdowns by page, country, device and date. Properties with insufficient impressions may lack the report. Use it to check which product or guide URLs appeared as supporting links. The report groups impressions without exposing the exact prompts that produced them. Search Console Generative AI performance report.

Search Console · Generative AIGoogle example

Use the two reports for different repairs. If Merchant Center shows interest in filter fit and your item data omits compatible model numbers, improve the offer data and the page. For a guide that earns supporting-link impressions while the intended product pages rarely do, inspect the guide’s product links and the destination facts. Compare the same market and date range when you retest.

How do you check whether AI answers recommend your store?

Keep a branded lookup, a non-branded discovery question and a question with a constraint a buyer would use. For the filter example, that could be a specific refrigerator model, a multipack and delivery to a named location by a certain day. Run the same wording on the same surfaces after a relevant catalog or page update. Record the market, date and signed-in state so the team can repeat the check.

Save the answer and mark the type of appearance. Log a text mention when the answer names the store, a citation when it links to a page as evidence, and a product card when it shows a shopping result. Record the seller, variant, currency, price and final URL when an offer appears. Count correct appearances out of the sampled runs and keep missing and incorrect results separate. A run can show an accurate product card and cite a different site, so both fields belong in the log.

How to track ChatGPT traffic in GA4

In GA4, open Reports → Acquisition → Traffic acquisition and use Session source / medium. Filter for chatgpt.com and inspect the landing page in a compatible report or exploration. If a ChatGPT Search link contains utm_source=chatgpt.com, follow it to see which parameters survive redirects. Then confirm how the session is recorded in your property. Keep other observed source values separate until you can verify how they arrived. GA4 Traffic acquisition.

GA4 · Traffic acquisitionGoogle Help

A shopper can see a recommendation and return directly later. On a supported native surface, the purchase can produce an order without a tagged storefront session. Use the recorded order source when available and keep unassigned orders separate. Join the order ID to payment and refund records before reporting paid or retained sales. Low tagged traffic describes what analytics attributed during the period. Check the live prompt sample for visibility.

How do you measure an AI shopping pilot?

Run an AI shopping pilot on a fixed assortment and market, then decide whether the next investment belongs in the catalog, storefront handoff or checkout. Count offers eligible for the intended action, sample accurate appearances, and reconcile orders and refunds for the route you tested. Set the pass conditions before expanding the assortment.

Use the readiness checklist to choose the offers and the visibility evidence to keep appearance, sessions and orders separate. Count the pilot variants with confirmed eligibility for the named market, surface and action. Log accepted files and unknown eligibility as separate statuses so the catalog owner can see the next item to resolve.

Which result should drive the next pilot decision?

If accepted items still lack confirmed eligibility, assign the next change to the channel or catalog owner. Repair the storefront handoff when shoppers reach a mismatched variant. When buyers reach checkout and orders reconcile, examine the purchase and support results before making the integration larger. Compare each change with the same set of offers and the same operating conditions where possible.

Compare orders at the same fulfillment stage and account for returns and cancellations. If you estimate contribution, subtract the costs you can trace to the channel, including payment fees, discounts, shipping support and added service time. A larger number of product clicks can justify better catalog coverage while the economics still favor the existing checkout. Use the evidence to assign the next task to merchandising, engineering or operations.

When does AI shopping traffic justify checkout integration?

Test cart transfer or native checkout when shoppers repeatedly hit a handoff problem you can reproduce and the channel supports a route that could fix it. Use the same offers and market in both groups. Where the channel records comparable checkout starts, compare retained orders per eligible start after the same return window. Join channel order IDs to payments and refunds. A native order can bypass the storefront session, so use tagged visits to measure referrals separately. If you cannot match the starting event or have too few comparable orders, hold the expansion decision.

Suppose two matched groups each start 50 eligible checkouts. The store route retains 35 orders and the native route retains 40 after the same return window. That is 70% versus 80% in this sample. If the native route also produced a duplicate order, stop and fix the retry before expanding. A wrong variant or a refund the team cannot reconcile also blocks expansion. Once those checks pass, compare contribution and support work before investing in the new route.

Keep one named owner for the next check and its date. For a feed release, keep an expected accepted row and a way to detect a rejection. Storefront work can be tested at the variant URL. Use a reconciled order to verify a checkout release.

Agentic commerce FAQs