What should an agentic commerce readiness audit cover?
Start with the commercial objective. A manufacturer seeking brand recommendations, a retailer seeking orders and a direct-to-consumer business seeking profitable customers need different evidence. Record the recommended product, seller and purchase destination separately. A recommendation that sends the shopper to a stockist can meet the manufacturer's objective without generating a direct-store order.
Define the markets, channels and buying tasks before commissioning a readiness score. Include important retailers of your products and credible competitors. Select products that expose different risks: bestsellers, high-contribution lines, complex variants, stock shortages and demanding compatibility or delivery requirements. A sample containing only straightforward, available products misses the cases most likely to change the investment decision.
Check channel eligibility before testing implementation. Shopify's ChatGPT route uses the merchant's checkout and requires selling to US customers. Google's Universal Commerce Protocol supports separate cart-transfer and native-checkout paths; its Merchant Center integration hub is initially rolling out in the US. Those conditions were checked on 23 October 2026. Record an unavailable capability outside the current pilot. Shopify's ChatGPT requirements, Google's UCP integration guide
Our agentic commerce overview explains the implications for brand choice. The audit turns that discussion into observable buying tasks and an investment decision.
What evidence belongs in the readiness checklist?
Ask the audit team to follow the same customer task through six areas. This is a proposed commissioning method, not an industry certification.
- Access and eligibility. Relevant pages can be reached and indexed; required channel settings and feeds are accepted. Name unresolved exclusions and their owners.
- Product understanding. The recommended variant meets the buyer's requirements. Specifications, identifiers and images agree across the site, feed and relevant retailer listings. Supporting claims and cited reviews are credible and relevant to that product.
- The current offer. Price, stock, delivery cost and timing, and returns terms agree for the selected variant and market. A controlled update reaches the customer-facing offer.
- Purchase and service. The link selects the correct item. The supported purchase route works, including payment failure and changed stock. Orders, tracking, cancellation, refunds and support reconcile.
- Customer authority. Where agents act, authentication, consent, spending limits and escalation work. Repeated requests do not create unintended orders. Qualified owners assess payment, access and customer-data controls.
- Commercial measurement. Identifiable channel activity can be connected to orders and contribution after relevant costs. State attribution gaps and exclude internal tests.
Marketing defines buying needs and assesses recommendations. Commerce, engineering, security and customer operations validate the systems they own. Inspect a price or stock change through the route and record when each system updates. An accepted feed can still leave a shopper seeing an old offer. Google lists conflicting website/feed data and incorrect variant attributes among causes of eligibility and display problems. Google's product-data specification
Require a documented reason for proposed technical work. Google Search requires no special AI file or schema for its AI features. OpenAI's product specification separates discovery from checkout, which requires its own enabled integration. Feed eligibility does not guarantee that a product appears. Google Search guidance, OpenAI's product specification
How do you test AI product recommendations?
Build cases from customer enquiries, buying objections and returns reasons. For each important requirement, use an open category question, a named-product question and a request the product cannot satisfy. These help distinguish failure to find the product from misunderstanding it, and test whether the assistant rejects an unsuitable match.
Record the prompt, assistant, market, date, account conditions, answer, cited source, seller and linked variant. Repeat important cases in fresh sessions. Keep conditions consistent on retest; a selected set of prompts cannot establish performance across all customers.
If the assistant understands the product when named but omits it from category answers, investigate discovery, competing offers and relevance. If it misstates a specification, inspect the cited page and variant. An error from an old retailer listing needs that listing corrected. Rewriting your own site may leave the source of the error intact.
Where the cited source is accurate, retain the facts and check for conflicting sources before attributing the error to interpretation. Repeat the named-product and unsuitable-use checks, preserving the results for the service's feedback or support route. Reporting an error does not guarantee a correction. Run the buying task through your existing website too. Failure in both routes points towards a wider product-information or commerce problem. This comparison narrows the investigation without proving the cause.
What does an actionable audit finding look like?
A shopper wants earbuds with a case that charges on a Qi pad. Apple's AirPods 5 specifications distinguish the USB-C case from the Wireless Charging Case, which supports Qi-certified chargers. The distinction is real; the audit outcomes below are illustrative. Apple's specifications
An answer claiming that the USB-C-only case supports Qi calls for a source and variant check. The source owner corrects an inaccurate listing. If the source is right, the team investigates interpretation and repeats the suitable-use and unsuitable-use questions.
A correct recommendation opening the wrong case calls for a different investigation. The commerce or feed owner traces the link and variant mapping, involves the platform if needed, and verifies the corrected destination through basket and checkout. If the correct offer reaches checkout but payment fails, commerce and payment owners investigate that step. Specify the expected charge, order count and order status for each authorised test, including any failed or pending records created by a declined payment. A successful purchase must reconcile once. Rejected authority must not produce an unauthorised purchase.
A purchase from another seller raises a commercial question. The owner assesses whether that outcome serves brand demand, retailer sell-through or direct sales. A visibility report could count a brand mention in every one of these cases while concealing the different decisions required.
How should the audit change investment decisions?
Mark each requirement as not tested, unavailable, failed, observed once or repeatedly passed, with its supporting record. An unresolved result stays unresolved. Prioritise material customer harm and broken purchases, followed by recurring faults affecting important products. An aggregate score cannot compensate for unauthorised purchases or orders the business cannot fulfil.
Before a transaction pilot, require evidence that consent, the final offer, payment, single-order creation and exception handling work within the defined scope. Name the owner and condition for resolving every open control. Use the platform's sandbox and required production validation. Appoint one executive to own the decision across teams, with a monitoring owner and retesting after material product, feed or integration changes.
The report should specify the next funded action, owner, cost and acceptance test. Set a pilot time and spending ceiling, success criteria and explicit reasons to stop or expand before launch. The decision may be to repair foundations, run a limited pilot, fund an integration or defer it because relevant demand and expected contribution do not justify the work. Correcting a shared catalogue defect may benefit existing retail channels too. A new checkout integration needs evidence of supported demand and enough expected contribution to cover implementation and operation. Record those assumptions before committing.
How do you measure agentic commerce performance?
Count correct, incorrect and unresolved recommendations within the defined test set. Separately reconcile identifiable orders with revenue, product and fulfilment costs, payment or channel fees, returns and additional service effort. Finance should agree the contribution definition. Exclude test orders and retain unknown attribution.
Google's UCP reporting distinguishes buy clicks, purchases and cancellations. Those records establish activity through the integration; they cannot alone establish additional sales caused by it. Where feasible, compare a staged rollout with similar unchanged products or markets, tracking other changes in price, promotion and demand. A non-random comparison still leaves uncertainty. UCP analytics
The return-on-AI Answer covers the wider investment assessment. The Centre's review of product information and agentic commerce examines why readiness findings and simulated recommendations cannot be treated as commercial returns.