AI Shopping Agents: How Brands Can Get Discovered, Convert and Measure Growth

A customer asks an AI assistant for a waterproof jacket that arrives by Friday, fits a certain budget and can be returned locally. The assistant may compare products before the shopper ever visits a retailer's site. The growth question is no longer only “How do we get the click?” It is also “Can an external system understand the right product, show an accurate offer and carry the customer safely to a purchase?”

This guide is for e-commerce, retail and growth teams deciding what to prepare now. It is an original editorial synthesis, not a claim that Akshay has implemented a particular agentic-commerce protocol for a client. Growthcraft Editorial prepared it with AI assistance; personal review by Akshay is not claimed. Examples and numbers are synthetic. Platform access, markets and specifications can change; check the linked documentation for your own setup.

Five decisions to make first

• Separate product discovery, checkout and post-purchase ownership. They can happen in different places.
• Make variant, price, stock, shipping and policy information consistent across the source of truth, feed and destination.
• Check eligibility and market coverage before assuming a feature is available to your store or customers.
• Measure retained orders and contribution alongside discovery and referral signals. A platform-attributed order is not automatically an incremental order.
• Keep B2B pricing, customer information and marketing permissions behind appropriate access controls.

Why this is topical in October 2026

On October 26, Salesforce announced that merchants using its B2C Commerce release can enable Universal Commerce Protocol (UCP) checkout on supported Google surfaces, including AI Mode and the Gemini app, while remaining the merchant of record. That is a vendor announcement about a specific integration, not evidence that every merchant, market or product can use it today. Source: Salesforce, October 26, 2026.

Earlier in the month, TikTok described a Shopping Assistant and Buy Direct experiences it is building with commerce partners. Salesforce's separate TikTok announcement says the phased rollout of its UCP-powered experiences is expected to begin in Q1 2027. Treat those statements as an announced roadmap, not a promise that the same checkout is live in your country. Source: TikTok, October 5, 2026; Source: Salesforce, October 5, 2026.

The discussion is broader than a single launch: The Verge reported on Shopify's opening of its catalog to more AI agents on October 13, and TechCrunch covered TikTok's shopping announcement on October 5. These are signals of industry attention, not proof of a particular search volume or sales uplift.

The journey has three separate layers

1. Discovery: A shopping surface receives a buyer's question and may find products through a catalog, feed, partner integration or the public web. A listing or citation is a visibility observation; it is not an order.
2. Decision and checkout: The buyer may be sent to the merchant's site, or an eligible integration may support checkout within another surface. Product eligibility, available payment methods, region and buyer approval depend on the specific integration.
3. Order and relationship: Someone must accept payment, fulfil the order, handle returns and answer customer questions. Map which organisation holds each responsibility and which information can lawfully move between systems.

Do not collapse these layers into “the AI agent sold it.” Shopify describes its ChatGPT channel as discovery followed by the merchant's own online-store checkout; its documentation separately describes direct checkout for eligible Google AI Mode and Gemini use cases. The exact experience and rollout can differ by platform, market and product. Shopify's ChatGPT channel documentation; Shopify's Google channel documentation, both checked October 27, 2026.

A practical operating map

Start with one product category and draw this sequence for each channel:

1. Buyer need: Which question or constraint is the buyer trying to resolve—fit, delivery date, compatibility, price or returns?
2. Product evidence: Which catalog, page or feed supplies the answer? Who owns it, how often does it update, and which market and variant does it describe?
3. Offer: Where are the actual price, currency, availability, shipping and return conditions determined?
4. Handoff: Does the shopper go to your storefront, or complete an eligible checkout in the partner experience? What confirmation does the buyer see?
5. Outcome: Which system records the order, cancellation, refund and customer-service handoff? What source information, if any, reaches your analytics and CRM?

For every arrow, name an owner and an exception path. If the agent shows an outdated price or a payment fails, the business needs a way to diagnose the mismatch and help the buyer. Do not assume all conversations create a referral that your analytics will recognise.

Build a product record an agent can use

Shopify says its catalog can provide titles, descriptions, options, images, prices and availability to eligible agentic channels, and that catalog data is updated as product information changes. The merchant still needs to make sure its own data and mappings are correct. Source: Shopify Catalog documentation.

Use a small, high-value sample before attempting a catalogue-wide project. For each product and variant, check:

• Identity: Stable product and variant identifiers; a size or colour should not inherit another variant's stock.
• Meaning: A plain title, category, dimensions, material, compatibility and relevant buyer constraints. Avoid descriptions that imply unsupported certifications or claims.
• Offer: Price and currency for the same market, current availability and any relevant sale conditions. Compare the feed with the actual storefront offer at the same observation time.
• Delivery and returns: Useful shipping, fulfilment and refund information for the markets served. Do not promise a delivery date that the merchant cannot meet.
• Ownership: A source system, responsible person, refresh process and escalation path when a field disagrees.

Google's product-data guidance requires merchants to supply accurate information and keep price and availability current for its supported products. Google also documents Product and Offer structured data for product pages. That is useful for its own systems; adding markup does not guarantee an AI recommendation or make a stale inventory feed correct. Google product data specification; Google structured-data guidance.

A small, reproducible offer check

The example below compares one synthetic feed offer with one synthetic storefront offer. Prices are integer minor units, such as euro cents. It assumes both records describe the same variant, market and observation window; checking those identities and timestamps upstream is essential. It does not call a platform API, check real inventory or determine whether an agent will recommend a product.

function checkOffer(row) {
  const issues = [];
  const money = value => Number.isSafeInteger(value) && value >= 0;
  const code = value => typeof value === 'string' && /^[A-Z]{3}$/.test(value);
  const states = ['in_stock', 'out_of_stock', 'preorder'];
  if (typeof row.sku !== 'string' || !row.sku.trim()) issues.push('missing_sku');
  if (!money(row.feedPriceMinor) || !money(row.storePriceMinor)) issues.push('invalid_price');
  else if (row.feedPriceMinor !== row.storePriceMinor) issues.push('price_mismatch');
  if (!code(row.feedCurrency) || !code(row.storeCurrency)) issues.push('invalid_currency');
  else if (row.feedCurrency !== row.storeCurrency) issues.push('currency_mismatch');
  if (!states.includes(row.feedAvailability) || !states.includes(row.storeAvailability)) issues.push('invalid_availability');
  else if (row.feedAvailability !== row.storeAvailability) issues.push('availability_mismatch');
  return issues;
}

checkOffer({sku:'jacket-m-blue', feedPriceMinor:12000, storePriceMinor:12000,
  feedCurrency:'EUR', storeCurrency:'EUR', feedAvailability:'in_stock',
  storeAvailability:'out_of_stock'});
// ['availability_mismatch']

Run this on anonymised snapshots only after normalising provider-specific availability values into the three example states. A real audit also needs tax and shipping treatment, sale dates, feed timestamps, duplicate variants, rate limits and a documented source of truth. A match in one snapshot is not proof of continuous freshness.

Checkout is a business and technical decision

UCP is an interoperability protocol: it describes how participating systems can exchange structured commerce information and negotiate capabilities. It does not itself guarantee a distribution partnership, access to every AI surface, a sale or a particular placement. Pin the version and capabilities supported by the actual partners; avoid treating an illustrative workflow as a universal integration. Source: UCP overview; UCP checkout specification.

Before enabling an integration, ask the platform and your commerce team: Who is the merchant of record? Where is the buyer's final approval? How are taxes, payment, inventory, order confirmation, cancellations and refunds handled? Which order identifiers and data fields are returned? What happens when a request times out or an item sells out? A retry after an uncertain order outcome requires reconciliation; it should not create a second order.

For Shopify's documented Google AI Mode and Gemini direct-checkout path, eligibility and customer-market restrictions apply; its current documentation lists the United States, Canada and Australia as customer markets for that setup. This is not a statement about every Google or Shopify integration. Verify terms, fees and availability with the providers before making a commitment. Source: Shopify channel requirements.

The B2B and privacy boundary matters

A public retail catalog and a negotiated B2B price list are not interchangeable. Shopify says its agentic storefronts support direct-to-consumer sales and warns that custom arrangements which hide B2B prices may not always be recognised as B2B-only by its detection. Audit what an external channel can actually see before exposing a shared catalogue. Source: Shopify's B2B considerations.

For B2B discovery, a useful first step may be public product specifications, compatibility information and a route to a qualified conversation—not public negotiated prices or autonomous purchasing. Keep contract pricing, customer records and order history behind authorised access. An order or enquiry does not automatically establish permission for marketing messages; retain and check the applicable purpose and permission in the system that owns it. Involve your privacy and legal owners for your markets and agreements.

Measure what happened—then ask what was incremental

Build a measurement table before launch. Keep these rows separate:

• Visibility: Product appearances or mentions from a defined set of prompts, markets, surfaces and observation dates. An appearance is not a click, a stable ranking or a sale.
• Observed referral: Sessions that analytics identifies as coming from a surface, subject to consent, tagging, redirects and missing referrers. A missing referral does not prove the assistant played no role.
• Orders: Confirmed and retained orders with the source and destination each system actually supplies. Reconcile duplicates, cancellations and refunds against the order system.
• Economics: Contribution after product costs, discounts, fulfilment, payment costs and the incremental cost of the activity. Keep returns, customer support and channel fees inside the agreed cost boundary.
• Incrementality: The difference from a credible comparison of what would otherwise have happened. Platform attribution assigns credit; it does not by itself establish that the order was caused by that channel.

Never put email addresses, phone numbers or conversation text into URLs or analytics tags. When a platform does not supply an event, keep it unknown instead of inventing a complete funnel. A direct-checkout order can occur without a conventional storefront session, so session conversion rate is not automatically comparable across the two journeys.

Worked example: a sale is not the same as incremental profit

Imagine a synthetic retailer observes 36 orders associated with an AI discovery route. Six are cancelled or fully refunded; the remaining 30 are retained. For simplicity, assume there are no unrecovered costs from the six excluded orders. Each retained item has a pre-discount, pre-tax list price of €120, a €6 discount, €60 product cost, €6 fulfilment cost and €4 payment cost. The associated incremental media or activation cost is €600.

Retained list value is 30 × €120 = €3,600. Subtract €180 discounts, €1,800 product cost, €180 fulfilment, €120 payment costs and €600 activity cost. The resulting example contribution is €720. This is an accounting scenario for the observed retained orders, not an estimate of causal profit.

Before the €600 activity cost, contribution is €1,320, or €44 per retained order under these uniform assumptions. If the €600 cost is genuinely incremental and each additional retained order has that same €44 contribution, at least 14 incremental retained orders are needed to cover it: €600 ÷ €44 = 13.64, rounded up. An attribution report showing 30 orders does not establish that 14 were incremental. Test different margins and include unrecovered return or support costs if they apply.

Use the existing incremental profit calculator to explore assumptions, and the CAC payback guide to think about refunds and time horizons. Neither tool proves that an agent caused a sale.

A 30-day starting plan

1. Days 1–5: choose the decision. Pick one market, product group and business question. Confirm whether each discovery and checkout route is actually available to your store. Record the current product and order data owners.
2. Days 6–12: inspect the evidence. Sample variants, compare catalog and storefront offers, and document mismatch reasons. Review private B2B products and permissions. Fix a known defect at its source before adding new automation.
3. Days 13–20: observe the buyer journey. Test representative questions and documented handoffs without placing real orders unless separately authorised. Record which product information appears, where the buyer goes next and which events are observable. Repeat observations; do not treat one generated answer as a ranking.
4. Days 21–30: review a bounded decision. Reconcile retained orders and costs where data is available. Decide whether to expand the sample, repair a data gap or design a controlled test. Record ownership, cost, risks and what evidence would change the decision.

If the channel, budget and volumes support a controlled experiment, work with an analyst to define the comparison, outcome, observation window, leakage and stopping rules before spending. A geographic holdout can be useful in some settings, but only when the design and data support it; two arbitrarily chosen locations or a before-and-after chart do not automatically establish causality.

Four mistakes to avoid

1. Treating every agent as the same channel. Discovery, checkout, buyer approval and data sharing differ across integrations.
2. Adding a discovery file while ignoring the product feed. Shopify explicitly distinguishes its store-information files from its product catalog; neither a file nor a schema guarantees inclusion.
3. Optimising for appearances alone. A prominent mention is not a retained customer, and an attributed purchase is not necessarily incremental.
4. Exposing information by accident. Verify product access and marketing permissions, especially for B2B price lists and customer data.

The next decision

Start with the customer question, the product evidence and the commercial boundary—not a promise to “optimise for every agent.” A clean catalog, reliable handoffs and an honest measurement plan are useful even if a specific platform rollout is delayed.

If you need help connecting the growth brief to the systems and evidence, explore forward-deployed marketing support or B2C growth consulting. Scope and delivery depend on your data, team and the integrations available to you.

Sources and scope: Platform documentation and announcements linked above were checked October 27, 2026. Product availability and protocol specifications can change. The worked figures, audit record and 30-day plan are illustrative; they are not client results, a legal assessment or a forecast.