Agents can find your products. Can they trust your stock?

Every agentic-commerce checklist written in the last two years told D2C brands the same thing: get your catalog readable by AI shopping agents. That advice was right, and it is now incomplete. In March 2026 OpenAI pulled centralised checkout out of ChatGPT's product listings, and the reason it gave points straight at the part of the stack most brands have not touched - the accuracy of the stock, tax and price data their back office can produce on demand.

What actually changed in March 2026

OpenAI and Stripe published the Agentic Commerce Protocol (ACP) as an open standard under Apache 2.0 in late September 2025, and ChatGPT's Instant Checkout let shoppers buy without leaving the assistant. In early March 2026 that changed. An OpenAI spokesperson told Digital Commerce 360 that "Instant Checkout is moving to Apps, where purchases can happen more seamlessly."

Two things did not change. ACP continues, positioned as the infrastructure connecting shoppers to merchants across the shopping journey, and the assistant surfaces kept doing what they were already good at - discovery, comparison and recommendation. What moved was where the transaction is allowed to happen.

The read for a D2C brand: the checkout you own did not become less important. It became more important, because it is where assistant-driven demand now lands.

Why OpenAI's stated reason matters more than the feature

Feature retirements are normal; the reason given here is unusually specific. The company pointed to transaction complexity - inventory status, sales tax and pricing - as details requiring constant updating, and judged individual merchant apps better placed to handle them than one central checkout.

Strip the platform language out and that is a data-freshness problem. A centralised checkout can only be as correct as the merchant data feeding it. If your stock figure is a nightly export, your tax logic lives in a spreadsheet, and your listed price disagrees with the cart, no protocol design makes that purchase safe to complete elsewhere.

This is the same failure mode we see connecting an ERP to any external channel, and it is why we treat agent-readiness as an operations project. Being findable is a content problem. Being transactable is a systems problem.

The four data surfaces an agent judges you on

1. Live stock, not last night's stock

An assistant that recommends an out-of-stock item burns the shopper's trust and its own. Tolerance for staleness is lower than on a category page a human browses, because no shelf of alternatives sits alongside it. If your channels reconcile on a schedule rather than on an event, that gap is your exposure. Real-time multi-channel inventory sync is the unglamorous prerequisite, and it surfaces every place two systems quietly disagree about one SKU.

2. Tax that is correct for this buyer, not on average

Sales tax and GST are where "roughly right" stops being acceptable. The correct figure depends on the buyer's location, the product category and, in India, place-of-supply rules. A brand computing tax at the invoice stage in a finance tool, rather than where the price is quoted, cannot quote a trustworthy total to any external system. Moving that logic into automated GST and tax compliance inside the ERP makes the number reproducible.

3. A price that matches what checkout will charge

Price mismatch is the most common defect we find, and it is rarely a pricing error. It is a sequencing error: promotions, tier pricing and shipping thresholds are applied in the cart, while the price published outward is the base list price. To a human that reads as a discount at checkout. To an agent comparing options it reads as unreliable data.

4. One product identity across every channel

If the same product carries three titles, two attribute sets and a different identifier per channel, no external system can be confident it is one item. That is a product information management problem, and it compounds - every new channel adds a place descriptions can drift. Brands running a storefront and an ERP in parallel need a deliberate Shopify and Odoo integration so one record is authoritative and the rest are projections.

What this changes about where you spend

The 2025 planning assumption was that brands should prepare for purchases completing inside the assistant. For now that is not where the money lands. Discovery happens in the assistant; the transaction happens on your site. So the spend that pays off is the spend on the path the shopper actually walks.

Concretely, that reorders the list. Work that makes your product data accurate and queryable moves up, and so does work that makes your own product page and checkout convert a pre-qualified visitor who has already compared you against alternatives. Speculative work built specifically for in-assistant checkout moves down, because the surface it targeted has moved.

None of this replaces the discovery side. If an agent cannot read your catalog at all, none of the transactability work matters - the problem we covered in whether AI shopping agents can read your catalog. Treat the two as sequential: get readable, then get transactable.

How to tell whether you are actually ready

A short test, in the order we run it during an audit:

• Pick one SKU and follow its stock number. Where is it mastered, how often is it published outward, and what is the worst-case age a third party would see? If the answer is "overnight", that is a freshness gap.
• Quote the same product to two different delivery locations. If the tax component is identical, your tax logic is not location-aware.
• Compare the published price against the cart total for a product that participates in any promotion or tier. Any difference that is not shipping is a data-consistency defect.
• Diff the title and key attributes across your channels. Count the variants. That count is your PIM debt.
• Check whether you can report on assistant-referred sessions at all. If not, see the next section - this one blocks the measurement of everything above.

The measurement problem nobody mentions

Here is the gap we hit on our own site before client sites. Assistant traffic does not arrive labelled as a channel. It arrives as referrals from assistant domains, and unless you group those deliberately they scatter among ordinary referrals or fall into an unresolved bucket. We only got a usable read after grouping the assistant domains into one channel and watching it as its own line.

Two things became visible. The volume is small but real, and it moves independently of ordinary organic traffic - so judging it by your search numbers tells you nothing. More usefully, a share of what looks like direct traffic is nothing of the kind: assistant surfaces do not always pass a referrer, so some demand lands unattributed. If your direct channel has grown without explanation, investigate it before crediting brand strength.

The instruction is dull but load-bearing: define an assistant channel now, before the volume matters, so you have a baseline. Brands that start measuring only once traffic is significant cannot prove what caused the change.

FAQ

Did OpenAI abandon agentic commerce?

No. It moved checkout from a central place in ChatGPT into individual merchant apps, and the Agentic Commerce Protocol it developed with Stripe continues as the connecting infrastructure. The discovery and recommendation behaviour that sends shoppers to merchants was not withdrawn.

Do we still need to make our catalog readable to AI agents?

Yes, and it is still the first step. Readability decides whether you are considered at all; the data accuracy discussed here decides whether the shopper can complete the purchase once they arrive. Doing the second without the first wastes the work.

Is real-time inventory sync worth it for a small catalog?

It depends on how fast your stock moves, not how many SKUs you have. A hundred-SKU brand with high turnover across several channels has more exposure than a thousand-SKU brand with slow stock. The test is the worst-case age of the figure an outside system sees.

How do we separate assistant traffic from direct traffic?

Group the known assistant referral domains into one channel and track it separately. Accept that the grouping is incomplete - some assistant surfaces pass no referrer, and that traffic lands in direct. Watch both lines together rather than either alone.

The useful conclusion from March 2026 is not that agentic commerce slowed down. It is that the bottleneck turned out to be unfashionable: whether your systems can state, correctly and on demand, what you have, what it costs and what tax applies. Brands treating that as an ERP and data project will be transactable by whatever the assistant surfaces look like in a year. Brands waiting for a protocol will not.