Returns and refunds dominate ecommerce support volume — and they dominate failure stories too. The question is rarely exotic: how long, who pays shipping, what condition, final sale or not. When those lines are missing or scattered, models still answer. Shoppers hear policy. Finance hears disputes.
Invention here is not random hallucination. It is pattern completion from “helpful” training data — thirty days because that sounds standard, restocking fee because other stores have one. Your bot is not malicious; your content left room for a plausible lie.
Questions that expose gaps
These show up in every inbox. If your canonical page does not answer them in plain language, assume the bot will try.
- Final sale / clearance — returnable or not?
- Opened cosmetics, underwear, food — hygiene exceptions.
- International orders — who pays return shipping and duties?
- Exchange versus refund — timing and price differences.
- Damaged on arrival — photo requirements and deadlines.
- Gift returns — store credit only or original payment method?
One page, numbered, no marketing fog
Consolidate returns, refunds, and exchanges into one maintained URL. Use numbered rules with explicit negatives: “Items marked final sale cannot be returned.” Repeat the return window in the first screen, not only in the footer. Marketing pages can link; they should not restate policy differently.
Index that page first. Exclude blog posts that discuss returns anecdotally. Wrong-page retrieval is how “30 days” becomes “14 days” when an old post wins search — see why your bot answers from the wrong page.
Refuse edge cases you have not written
When the question is “can I return worn shoes if…?” and the policy is silent, the correct product behaviour is escalation or a form — not empathy plus invention. Configure refusals for categories you will not automate: worn goods, partial sets, third-party marketplace orders.
Train support to treat bot transcripts as quotes customers believe. If the bot invented a window, honour-or-deny is a business call — but fix the content before the next shopper asks.
Citations shoppers will check
Returns shoppers are motivated clickers — they want proof before printing a label. Label citations honestly (“Returns & refunds policy”) and link the section that states the window. If you quote a number in chat, the same number must appear in the linked paragraph.
Marketplace and third-party seller questions need explicit routing: bot answers for your store policy, handoff or refusal for FBA or partner listings if your site does not govern them. Mixed inventory without clear labels is how “free returns” spreads to SKUs you never promised.
Ops loop tied to chargebacks
Review weekly: escalated returns chats, chargeback reasons citing “website said”, and playground tests on the top ten paraphrases. Any invented fee or window is a P0 content fix, not a model tweak. Update policy, recrawl same day, rerun tests.
- Maintain one canonical returns URL.
- Index it; exclude competing pages.
- Test ten paraphrases after every policy change.
- Refuse or hand off when the policy is silent.
- Log citation URL on every returns ticket for audit.
“Shoppers do not treat chat as advice. They treat it as your store speaking.”
Warranty versus return window
Customers conflate warranty claims with change-of-mind returns. If your site separates them, the bot must too, with distinct URLs or clearly headed sections. Otherwise “can I return this?” retrieves whichever page mentions “return” first, including manufacturer warranty copy that is not your policy.
Gift purchases and B2B orders
Gift receipts and corporate POs introduce rules consumer FAQs skip — store credit only, original purchaser must request, restocking on opened bulk. If you sell B2B or gifts, add a subsection to the canonical policy. Otherwise the bot applies consumer defaults to orders that never qualified for them.
Chargeback text that quotes chat should trigger the same review as chargeback text that quotes email — pull the citation URL, fix the page, recrawl, log the incident. Patterns repeat by SKU category more often than by model version.
Returns automation is worth doing, but only on policy you would sign in court. Until that page exists, the bot’s job is to point at it or step aside — not to improvise retail law.
