GEO-103·Issue 09·Content & Semantics·103 / 132
Shipping and Returns Policies: AI Reads Them Too
Delivery windows, shipping cost, and return terms are fields an agent screens on. Seven fields, a real rewrite, and a five-step policy playbook.
Issue focusReady to Transact: Agentic Commerce and Product Feeds
Consider a composite US cookware brand. Its CX lead pastes an assistant’s answer into the Tuesday growth thread. A shopper in Ohio asked whether a 12-inch skillet would arrive before Thanksgiving, and whether she could send it back if it did not sit flat on her induction hob. The answer: “The merchant does not state delivery times or a return window.” Two marketplace listings follow. The brand does not.
Delivery windows, shipping cost, and return terms are not after-sales copy; they are the fields an agent screens on, and a policy it cannot parse is one it records as missing.
A policy is a filter, not a disclaimer
An agent carries the buyer’s risk and prices landed cost and regret cost; policy feeds both. The buyer asks whether the pan is worth it. The agent asks whether it arrives, when, and who eats the return.
It opens the same page at three moments: screening (do you ship here), confirmation (total cost, arrival date), after-sales (who pays the label, to which address). When those answers conflict, the agent takes the most conservative one — usually “not stated.”
The policy page is one of the few pages an agent opens on purpose: not for reassurance, but to fill a cell in a comparison table. Whether it gets read is a layer underneath, mapped in agentic commerce readiness.
One paragraph, five conditions, one cell
The failure is usually a sentence, not a missing page: “We accept returns within 30 days of delivery. Items must be unused. Return shipping is the customer’s responsibility. Some exclusions apply. Final interpretation rests with the merchant.”
Extraction keeps “30 days” and drops the rest, which have no cell to land in. The final-interpretation clause voids the cell on purpose.
Sentences that survive read like if/then: which market, which category, what condition, who pays, counted from when. The closer a clause sits to a field, the likelier it is lifted verbatim.
| Field | What to state | Where it hides |
|---|---|---|
| Reach | Countries, excluded postcodes, islands | “Some regions excluded” in an FAQ |
| Transit time | Handling and transit apart; business or calendar days | One “3–7 days” mixing both |
| Shipping cost | Threshold, method, who pays duty and VAT | Checkout page only |
| Return window | Start point, calendar or business days | A terms PDF |
| Return shipping | Who pays, local address, prepaid label | Four words: “buyer pays” |
| Condition | Opened vs. used, hygiene, custom, final sale | Rendered into a product image |
| Refund | Original method or credit, when it starts | “Processed as soon as possible” |
Replace the three paragraphs about valuing the customer with this table. A policy inside an image, a modal, or one language only is, to extraction, absent.
Example: the same return policy, written twice
Example — brand, market, and figures are invented; only the sentence shape matters.
Before: “We offer worry-free returns. If you are not satisfied, request a return within 30 days of delivery. Items must be new and unused. Return shipping is paid by the buyer. Some special items are not eligible.”
After, for the US storefront:
| Field | US storefront |
|---|---|
| Window | 30 calendar days from the day after delivery, per the carrier record |
| Condition | Unused, or opened but unused: returnable. Filters, food-contact parts, engraving: not returnable |
| Return shipping | Defects: merchant pays, prepaid label. Change of mind: buyer pays, carrier rate |
| Refund | Original payment method, after warehouse inspection |
| Exceptions | Oversized and battery items need a return arrangement confirmed before dispatch |
Three edits: the final-interpretation clause went, “some special items” became a list, the clock got a start point. The shape held; readable conditions went from one to four.
Five steps from policy page to checkable fields
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Harvest | Pull policy fields from page, FAQ, checkout, listing, and feed into one sheet | Content | All five reconcile to one table |
| 2 Split | Label each sentence field, condition, or attitude; cut the attitudes | Content | Every condition fits a cell |
| 3 Name exceptions | Oversized, battery, custom, drop-ship: trigger plus fallback | Content, fulfillment | Support answers all three unaided |
| 4 Split by market | Master table plus one row per market for the moving fields | Content, local ops | Every live market has a row |
| 5 Verify | Ask four questions: ship, when, return, who pays | Content | Answers carry conditions, not “not stated” |
Step 5 is the only falsifiable one; one screenshot is not a result, so sample across days, accounts, and regions. A clean policy still has to match the live arrival estimate at checkout: see what AI checkout needs.
Step 3 is the one teams skip: an agent that cannot classify an oversized item tends to treat the whole policy as inapplicable — harder to undo than one wrong sentence.
One master table, one row per market
Policy is the least translatable content on the site. The window’s start, the statutory withdrawal right, the return-shipping habit, and whether a local address exists all move by market. Translate a US 30-day clause into German and both versions are wrong: the EU withdrawal right has its own clock and exceptions, and the US text neither states nor excludes it correctly.
The shape that holds: a global table for what is common, a per-market table for what changes. Do not fork a page per market: forks go unmaintained, versions drift, and the model cites whichever one it retrieved. Marketplaces carry their own policy fields, so settle who fulfills before publishing either surface; the trade-offs are in marketplace vs. DTC, and the general pattern in translation vs. localization.
Where this breaks
First, the policy does not exist. If a market has no return address, write “not supported” instead of inventing a row.
Second, the category carries too many exceptions — oversized furniture, perishables, batteries, made-to-order. Split subpages rather than smoothing them into one sentence.
Third, a policy is a promise. A generous window raises expectations and returns may rise; that is a fulfillment cost, not a GEO failure. Run the margin first.
Fourth, the page changes and the process does not: uninformed support and a warehouse that does not execute turn the sentence into a one-star review.
This month, paste one market’s return paragraph into a table, one sentence per row, and ask which cell each fills. Delete what fills none; for empty cells ask the warehouse and support rather than writing something plausible. Then rerun the four questions at two assistants and check whether “return policy unclear” is gone. More on sentences that survive extraction sits on the content and semantics hub.