GEO-097·Issue 09·Technical Foundation·97 / 132
Agentic Commerce Readiness: From Being Recommended to Being Ordered
Five layers an agent reads before it orders — product ID, price, shipping, returns, payment — plus a seven-check audit and four steps to wire the chain.
Issue focusReady to Transact: Agentic Commerce and Product Feeds
At a September review, a US DTC espresso-machine brand finds itself in chat answers more often than in June while orders stay flat. An engineer walks the agent path: model picked, Chicago ZIP entered, cart assembled. The reply is empty — no shipping cost, no delivery date, returns buried in an FAQ as a “hassle-free promise.” Nothing throws an error. The assistant picks a competitor that returned a landed price.
An agent does not order because of a sentence. It orders when five layers return computable values — identity, price, delivery, returns, payment authorization — and one gap stops the chain without an error.
A recommendation is a sentence; an order is a call
A weak recommendation costs exposure you measure next month; a broken transaction costs an order and leaves no trace — no error, no ticket, no abandoned cart.
The sequence inside an agent is close to fixed: gather candidates, filter by constraints, compute landed cost and arrival, price the worst case, then authorize payment. Each step needs a value it can calculate. “Two business days” calculates; “ships quickly” does not. “30-day returns, buyer pays return postage” calculates; “returns accepted” does not. “€349 delivered to Berlin, VAT included” calculates; “plus shipping at checkout” does not.
So the work moves from copy to data: the agent may never open your page. The four entry points are mapped in the AI search traffic shift.
Five layers, and the chain stops at the first gap
| Layer | The agent’s question | What it needs | If it breaks |
|---|---|---|---|
| 1 Product | Which exact item is this? | Stable SKU or GTIN, variant relations, feed record | Never enters the candidate set |
| 2 Price | What does it cost right now, all in? | One write source for price and stock, currency and tax basis | Quote contradicts itself; model defers to “check the site” |
| 3 Fulfillment | Does it ship here, when, who clears customs? | Ship-to coverage, transit windows, cutoff, freight and duty responsibility | Landed cost impossible; item dropped |
| 4 After-sales | Can it come back, to whom, at whose cost? | Return window, address, refund timing, warranty coverage | Risk unpriceable; the safe answer is no |
| 5 Payment | How is it paid, who authorizes, what if it fails? | Method list, one-time authorization, order confirmation, human fallback | Last hop fails; four layers wasted |
Most teams stall at layer three. Price and stock have systems; shipping is “calculated at checkout,” cutoff time lives in a 3PL rep’s message, duty responsibility in a footnote. That knowledge sits in people’s heads, not an interface — the expensive part. Layer two has its own playbook in offers, prices, and stock.
Layer five is the most overestimated: accepting cards is not accepting agent payment. Protocols introduced in late 2025, such as Agentic Commerce Protocol, pass a scoped, single-use authorization rather than card data; as of 2026-09 this is still moving. The question is not how many methods you list but whether one authorization completes without anybody tapping confirm — and falls back when it fails.
The ten-minute check
Score 0 or 1 per line on a live SKU.
| Check | Passes when |
|---|---|
| Page, structured data, and feed name the same item | Model numbers and variants match |
| Four price exports agree, tax basis included | Pulled together: zero difference |
| A foreign postcode returns cost and an arrival window | A number or range, not “calculated at checkout” |
| Cutoff and handling time are written down | You can compute “order by 2 pm, ships today” |
| Return window, address, postage payer, and warranty are machine-readable | In a field or table, not an FAQ |
| Orders return a number, an ETA, and a service link | Without a person touching it |
| A dead end has a labelled fallback | A route, not a footer email |
0–3: fix layers one and two until the product is citable; 4–5: wire three and four; 6–7: run one SKU end to end. Sample answers rather than screenshots — see the AI visibility baseline.
Four steps to a working chain
- Pick one chain. One category, one market, twenty SKUs, all answerable by a single buyer question.
- Collapse the facts. Price, stock, freight, and transit from one quote endpoint; policies from one table.
- Write fields. Express shipping and returns with
OfferShippingDetailsandMerchantReturnPolicy, mirrored as a visible table; the five layers must survive JavaScript off. - Walk it. Play the agent from a US ZIP, a German PLZ, and a Japanese postal code, and log every value that would not compute.
Step four is the only falsifiable step; its gap list is the backlog.
Example: one grinder, two levels of orderable
Example — model, prices, transit times invented.
Before: “The G-2 brings café-quality grinding home. Shipping times are calculated at checkout; returns are covered by our satisfaction guarantee.”
An agent reads three unknowns — landed cost, arrival date, the cost of being wrong — and does not gamble.
After: “The G-2 is 110 V, 60 Hz, with 64 mm flat burrs. Ships from Newark; $349 plus $12 to 60614; order by 2 pm ET for same-day dispatch, two business days to Chicago, four to the West Coast. Returns accepted for 30 days to Newark, buyer pays return shipping, refund within five business days. Two-year warranty, US and Canada. Card and PayPal, with one-time authorization for agent checkout; orders return a number, an ETA, and a service link automatically.”
Every layer now holds a value computed in one pass, including who pays return shipping, and a person can check the same numbers.
Where this breaks
First, the product is not reliably identified: fix the feed; later layers have no reader.
Second, high-ticket, customized, or tactile goods: a checkable quote is the realistic goal.
Third, fulfillment you do not control: a field is a promise, and a missed one is a cancellation.
Fourth, marketplace-concentrated sales: the interface is the platform’s; your leverage is feed hygiene.
Fifth, readiness decays: price and transit change daily, so an unmonitored integration is not one.
This month, take one SKU you sell, walk an overseas address from “can I buy this” to “order placed,” and log every step that would not compute. The September line continues with the product feed as the new landing page; the wider stack is on the technical foundations hub.