Returns Policy for AI: Write One Agents and Buyers Can Use

A returns policy for AI has two jobs. It has to be readable by the shopping agents that now compare stores before a human ever visits, and it has to hold up against returns requests that AI tools helped write. If you run a Shopify or WooCommerce store, both jobs land on the same page: your return policy, in plain terms, backed by structured data.

This guide is for store operators. At the end, a short section covers the other thing people mean by this search: getting a refund on an AI subscription like ChatGPT or Claude.

TL;DR

• AI shopping agents weigh return terms when choosing between stores, so a vague policy can cost you recommendations. Put the window, cost, condition and method in plain sentences and in structured data.
• Mark up your policy with schema.org’s MerchantReturnPolicy so machines do not have to guess. Google documents the properties and requires it nested inside an Offer.
• Written claims are no longer proof of anything. Roughly half of consumers use generative AI to draft return requests, so verify with order data and photos, not prose quality.
• Say who owns the problem when an agent picks the wrong product from bad listing data. Fix the listing first, then the contract terms.
• Returns are a conversion tool. Treat the policy as part of the product page, not legal fine print.

Why AI changes what a return policy is for

For years a return policy was a footer link that shoppers read after they had doubts. Now a growing share of product discovery runs through assistants such as ChatGPT, Gemini, Claude and Perplexity. They gather facts about a product, price, shipping and returns, then recommend a few options. If your return terms are buried in a PDF, written in legalese or contradictory across pages, an agent has little to work with. A competitor with a clean, consistent policy gives it more to say.

That is why the policy now sits in two audiences at once:

1. Machines that parse, summarize and compare it.
2. Humans (and the AI tools they use) who invoke it after purchase.

A policy that serves only the second group is the old model. We cover how agents choose products in more detail in how AI shopping agents recommend your products.

What AI agents need to see in your policy

Agents extract facts. Give them facts in the same shape every time. At minimum, state each of these in one sentence near the top of the policy page:

Three habits matter more than the wording itself.

Be consistent everywhere. If the product page says 30 days, the policy page says 45 and the checkout says 14, an agent cannot tell which is true, and a buyer will quote whichever is best for them. Pick one set of terms and make every surface match.

Use real text, not images. A returns infographic is invisible to many crawlers. Keep terms in HTML text.

Keep product-specific exceptions on the product. If one item is final sale, say so on that product page, not only in the global policy.

Add structured data so nothing is left to inference

Plain sentences help. Structured data removes ambiguity. Google’s Merchant Center documentation says the schema.org MerchantReturnPolicy type must be nested within an Offer using hasMerchantReturnPolicy. The same page lists properties such as applicableCountry, which takes a two-letter country code like US, and returnPolicyCategory, which takes values from the MerchantReturnEnumeration, for example a finite return window.

In practice, a store with a 30-day window would declare:

• the countries the policy applies to,
• a finite return window with the number of days,
• the fee type for return shipping,
• the return method (by mail or in store), if relevant.

Check the exact property names against Google’s page before you ship, since they are the source of truth and can change. Most Shopify themes and WooCommerce plugins output product schema but leave the return policy block empty or generic, so test a live product URL with a structured data validator instead of assuming it is there.

One caution: markup must match the visible policy. If the schema says 30 days and the page says 14, you have created the exact contradiction you were trying to remove.

The broader data picture, including feeds and where agents pull facts from, is in our guide to simulating and fixing AI shopping data sources. If you also want the surrounding search work to hold together, this overview of why SEO for ecommerce drives sustainable growth is a useful companion.

Expect AI-written return claims

The other side of the policy is what happens after the sale. Riskified’s 2026 report, “Rewriting the rules on returns”, surveyed 2,091 consumers, and the company reports that about half now use generative AI to draft return and refund requests. It also notes that manipulated or AI-generated images of damaged items are showing up in claims.

The practical consequence is blunt: a polished, articulate, well-structured complaint tells you nothing about whether it is true. A messy one is not more honest either.

The wider abuse trend is real too. Signifyd reported abusive returns up 64% in May 2025 compared with January 2024, and says about 85% of consumers see some strategic or deceptive return behavior as acceptable. Those figures describe the market, not your store, so look at your own return reasons before changing anything.

Policy language that helps you verify

You cannot out-write a fraudulent claim, but your policy can set up checks that do not insult honest customers:

• Require the return to start in a digital channel. Fraud experts quoted in Digital Transactions argue merchants should push returns into digital channels, because they capture signals like email age, device history and IP consistency that help tell legitimate from malicious intent.
• State your evidence rules. For damage claims, say you may ask for photos of the item and the packaging, taken within a stated number of days of delivery.
• Tie eligibility to order data. Order owner, delivery date, window and final-sale status are things a system can check, whatever the message sounds like.
• Separate defect from change of mind. Different windows and shipping costs for each reduce the incentive to describe every return as a defect.

Keep the tone matter-of-fact. The goal is a fast path for the large majority of customers, and a verification step for the few who trip a signal.

Who owns the return when an agent picks the wrong product

A new gray area: a shopper tells an assistant what they need, the assistant picks your product, and it turns out to be wrong for the job. Analysts at Retail TouchPoints describe it as a liability storm: when an agent selects a product from flawed metadata or misreads it, it is unclear whether responsibility sits with the merchant, the AI provider or the consumer. Their advice is to clarify these roles in contract terms now.

We are not lawyers, and this is a question for your own counsel. What you can do without one is reduce how often it happens:

1. Make specs unambiguous. Dimensions, materials, compatibility and sizing, in text.
2. Say what the product is not for. Agents summarize; a clear limitation survives summarization better than a paragraph of caveats.
3. Offer the same return rights on agent-originated orders as on any other, and say so, so the buyer never has to argue about how they arrived.

Most mismatch returns are listing problems before they are policy problems. Our guide to product listing optimization covers the data side.

Know your legal floor by market

This article is written for US merchants, but agents answer shoppers everywhere, and rules differ by country, so state which country your policy covers.

• EU: consumers have a statutory right to withdraw from distance purchases within 14 days of receipt without giving a reason, with exceptions such as sealed goods opened for health or hygiene reasons.
• UK: GOV.UK says you must offer a full refund if an item is faulty, not as described or does not do what it is supposed to, and in some cases when the customer simply changes their mind.
• US: there is no single federal change-of-mind right for online purchases, and state rules vary, so check with a lawyer licensed in your state before you promise or exclude anything.

If you sell across borders, a short country section in the policy, plus the applicableCountry property in your markup, keeps both agents and customers from guessing.

Our perspective

We built ecentic around one observation: AI agents do not rank pages the way search engines do. They weigh whether a product is a safe recommendation, and a clear return policy is part of what makes it safe. When we simulate how ChatGPT, Gemini, Claude and Perplexity evaluate a product page, the win/loss diagnostics often point at plain, fixable gaps in the data an agent can read, not at anything exotic.

So our advice is unglamorous. Write the policy once, in plain sentences. Mirror it in structured data. Keep it identical everywhere. Then measure whether agents pick you more often, instead of assuming they do. A returns policy is not a cost center page. It is part of the listing, and it should be tested like one.

We would also resist the temptation to tighten policy terms in response to AI-written claims across the board. Verification belongs in the process, not in a harsher policy that punishes the honest majority and gives agents a worse thing to say about you.

— Xhurian

Test what agents see on your store

If you want to see how assistants read your product pages, return terms included, ecentic runs a simulation against the major AI shopping agents and shows where you win or lose against competitors, in plain English. You can start with a free scan, then move to a plan if the findings are worth acting on. Published plans are Compete at $14.90 per month, Scale at $44.90 per month and Dominate at $149 per month, plus a performance option of 3% per AI-attributed order. See the pricing page for details, or try the free WebMCP shopping audit to check how agent-ready your store is.

Once you have rewritten a policy, you will want to know whether it moved anything. Our guide on how to track AI shopping traffic shows how to separate agent-driven visits from the rest.

Turn the return experience into a reason to buy

It is easy to see returns as pure loss. Signifyd’s report says AI-driven traffic can be 1.8x to 2.3x riskier than organic search, yet also that 62% of consumers say they would spend more with a brand that offers a seamless return process. Both are true at once: the risk is real and so is the upside.

A short checklist to finish with:

1. One set of terms, identical on product pages, checkout, policy page and emails.
2. Window, cost, condition, method and exclusions in plain sentences.
3. MerchantReturnPolicy markup that matches the visible text.
4. A digital return path that collects verification signals.
5. Evidence rules for damage claims, applied to every request alike.
6. Country-specific notes if you ship abroad.
7. A review every quarter, using your actual return reasons.

If you meant refunds on an AI subscription

Some searchers want to get money back from an AI product they bought, not write a policy. The big providers are strict, so check the current page for your plan:

• ChatGPT: OpenAI says web-purchased individual subscriptions are generally non-refundable, though a discretionary refund may be considered for an unused subscription requested within 7 days of the charge. Cancelling stops future renewals but does not refund past charges. Subscriptions bought through an app store follow that store’s process, so see OpenAI’s cancellation guide.
• Claude: Anthropic’s refund page for paid plans says payments are non-refundable unless the terms or law say otherwise, and that customers in the EEA and UK may qualify for a pro-rated refund when cancelling within 14 days. It is also available in other languages, including Portuguese.

Before filing a dispute with your card issuer, note that Anthropic says disputes must be withdrawn before support can process a refund.

FAQ

Do AI shopping agents actually read return policies?

They read whatever text and structured data they can reach about a product and the store, and return terms are a standard part of that picture. We cannot say exactly how each assistant weighs them, which is why we test by simulation. Clear, consistent, machine-readable terms are the safe bet.

What should a return policy for an online store include?

The return window, who pays for shipping, the condition required, the refund method and timing, exclusions such as final-sale items, and how to start a return. Add country-specific notes if you sell abroad. Keep it in plain text and mirror it in MerchantReturnPolicy markup.

How do I handle return claims written or illustrated by AI?

Do not judge a claim by how well it reads. Verify with order data, delivery date and evidence rules, and steer returns into a digital channel where you can collect identity signals. Apply the same checks to everyone, so honest customers are not treated as suspects.

Can I get a refund on a ChatGPT or Claude subscription?

Usually not, with limited exceptions. OpenAI may consider a discretionary refund for an unused web subscription within 7 days, and Claude offers a pro-rated refund within 14 days for EEA and UK customers. Check your plan’s current help page first.

Sources

• Supported structured data attributes and values, Google Merchant Center Help
• The new returns reality, Riskified
• State of Fraud and Returns Report 2025, Signifyd
• Agents of Risk, Digital Transactions
• Agentic commerce and liability, Retail TouchPoints
• Distance selling: right of withdrawal, Service Public
• Accepting returns and giving refunds, GOV.UK
• How do I request a refund for my ChatGPT subscription, OpenAI Help Center
• Canceling your ChatGPT subscription, OpenAI Help Center
• Claude paid plan refunds, Anthropic Help Center (Chinese)
• Claude paid plan refunds, Anthropic Help Center (Portuguese)