What is agentic commerce?
Agentic commerce describes AI systems carrying out parts of a shopping task on someone’s behalf: understanding a need, investigating options, comparing offers and, where the system supports it and the person authorizes it, progressing a purchase. These stages involve different technical capabilities and different commercial questions.
My working distinction is simple: discovery gets a product considered; recommendation gives a reason to choose it; a transaction requires an offer that can actually be fulfilled. An impressive answer does not establish that the last step works.
Six moments worth investigating
Trace one realistic customer need through the whole journey. At each step, record the information available, the decision made and the next action. This turns an abstract discussion about AI into an observable commercial process.
| Moment | The question to test |
|---|---|
| Intent | What does the buyer need, and which constraints are essential? |
| Discovery | Which products enter consideration without being named? |
| Comparison | Are differences accurate, relevant and supported? |
| Offer | Is the right variant available from an eligible seller? |
| Authorization | Does the customer understand the price, terms and action? |
| Fulfilment | Can the order, delivery and returns be handled correctly? |
One journey. Different responsibilities.
For a brand, the immediate question is whether a product is understood accurately across the places where it is represented. For a retailer, the question also includes selection, local stock, seller identity, delivery and the path to checkout. The two perspectives overlap, but the work belongs to different owners.
A useful programme connects commercial, ecommerce, product-information and engineering teams around a small number of buying scenarios. A generic visibility score cannot tell each of those teams what to fix.
A focused first 90 days
Treat this as an illustrative planning sequence, not a promised delivery schedule. The scope depends on catalog size, access to data and the systems involved.
- ↗Days 1–30: choose one market and category; define customer scenarios; document the baseline and the product-data owners.
- ↗Days 31–60: investigate recurring failures; correct a limited set of product and offer inconsistencies; keep a change log.
- ↗Days 61–90: repeat the measurement under comparable conditions; assess operational improvements separately from recommendation changes.
What should you do first?
Choose a buying decision that matters commercially and is narrow enough to investigate. Write down what you would do differently if the evidence changed. That gives the measurement a purpose and prevents a pilot from becoming a collection of interesting screenshots.
The distinctions that matter.
Is every AI recommendation agentic commerce?
A recommendation can be part of an agentic buying journey. It does not establish that the system can also complete a purchase. Specify which part of the process you are investigating.
Is structured data enough?
No. Correct structured data represents product information. A useful offer also needs accurate facts, current terms and working handoffs.
How do we know whether it works?
Define the outcome you want to improve first. Measure recommendations, information quality and operational functionality separately, and document the limitations.
Original documentation and published evidence. The working approach and priorities on this page are editorial recommendations, not platform guarantees. Reviewed 7 November 2026.
OpenAI · Commerce documentation↗OpenAI · Checkout specification↗Universal Commerce Protocol · Specification↗