Should You Let AI Agents Shop on Your Site?
Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
For most stores, the workable answer is yes to shopping, no to checkout: permit shop so agents can compare your products and build carts for their users, and disallow checkout so no agent completes a purchase unattended. That is the published store sample's exact shape — openai/agent may browse, read, search, and shop on a paid tier at 60 requests per minute, while checkout, form submission, and API calls are disallowed in every rule. Your policy is where the decision becomes public.
What does "letting an agent shop" actually grant?
Less than it sounds like. In the AGENT-POLICY vocabulary, shop is the browsing-with-intent verb: viewing products, comparing options, assembling a cart. It is distinct from checkout, the verb that spends money. The distinction is the entire decision: an agent that can shop is a tireless comparison shopper working for its user inside your catalogue; an agent that can check out is a purchasing officer. Most stores want the first and not the second, and the two-list structure of permitted and disallowed actions exists to let them say so in one rule.
The case for yes
Agent-assisted shopping routes demand. A user who asks their agent to find running shoes inside a budget gets a shortlist built from the stores whose policies and pages let the agent work — and a store the agent cannot read is a store the agent cannot recommend. Permitting shop also moves agent traffic from the shadows into the declared open: instead of guessing which automated visitors are customers' agents, you publish terms and watch named agents operate inside them, at your stated pace, under your data-use terms. The store sample prices this realistically — the shopping rule carries paid-tier terms, because catalogue access at working speed is worth something, a calculation the charging guide develops.
The case for fencing checkout, always separately
Checkout concentrates every risk a store cares about: payment, fulfilment obligations, discount and inventory edge cases, and a customer relationship that begins with a transaction the customer may not have reviewed. An agent that has compared faithfully can still misread a size, a delivery window, or a final price. Fencing checkout costs the honest agent nothing it cannot recover — it presents the cart and hands the purchase to its user, which is how the approve pattern in the examples guide works — while removing the entire risk class from your policy. Treat any request to permit checkout as a separate, later decision, made agent by named agent, never as the default consequence of permitting shop.
Deciding for your store, concretely
Ask three questions. Is your catalogue your shop window or your trade secret? Window: permit broadly. Secret: the answer tilts toward conditional and tight data-use terms. Can your operations absorb agent-shaped demand? A cart built by an agent is demand like any other; your pace cap is the throttle. Who do you want at the till? If the answer is "the customer", fence checkout and be done. Write the answers as one rule per agent class, mirror them in the catch-all, and publish — the decision only exists, for agents, once it is in the file at /.well-known/agent-policy.json.
One audience for this decision is easy to overlook: your own future customers, arriving through agents they have not adopted yet. Shopping behaviour is migrating into assistants quickly enough that a store's policy posture is becoming part of its merchandising — as real as its delivery terms or its returns page in deciding where agent-mediated demand lands. The store that fences thoughtfully today is not merely avoiding risk; it is writing the terms on which a growing channel will know it, while the terms are still cheap to set. The store that refuses by default will eventually write a policy anyway, under competitive pressure, in a hurry. The decision this guide asks for is rarely yes or no in the end. It is now, on purpose, or later, by accident.
Worked example: a garden centre says yes to shopping, in writing
Fictional example. Fernbrook Garden Centre, a fictional garden retailer with 3,200 products across plants, tools, and outdoor living, permits shop for 2 named agents and its catch-all (the latter at conditional), with checkout and submit-form disallowed in all 3 rules, a 40-per-minute cap for named agents, and data-use terms limiting catalogue use to customer assistance. In its first spring season under the policy, the centre tracked 312 agent-built carts handed back to customers at the checkout step; 96 completed as orders. The number its owner quotes, though, is the zero beside it: agent-completed purchases requiring unwinding — impossible by construction, because the verb that creates them was never granted.
The let-them-shop checklist
- Separate the verbs. Decide
shopandcheckoutindependently — yes to the first implies nothing about the second. - Fence the money step. Disallow
checkout,submit-form, andapi-callunless a named agent has earned otherwise. - Price the catalogue. Set commercial terms that reflect what working access to your products is worth.
- Cap the pace. Give shopping rules a rate limit your operations can absorb on the busiest day.
- Publish the answer. A decision agents cannot fetch is a decision they cannot follow — put it in the file and validate it.
Keep reading
Let Agents Browse but Not Check Out: the Split Rule · How to Charge AI Agents for Access to Your Site · What Is an Agent Access Policy (v0.1)? · Spec samples · Conformance · Pricing · All guides
Frequently asked questions
If I permit shop, can the agent complete a purchase?
Only if checkout is also permitted — they are separate verbs, separately listed. The standard store posture permits shop and disallows checkout in the same rule: the agent compares and builds the cart, then hands the purchase back to its user. Nothing about granting shop leaks into granting checkout.
Will letting agents shop increase my sales?
Nobody can promise that, and this guide does not. What permitting shop does is make your store legible to the agents your customers are already sending: your products enter their comparisons on your stated terms. Stores that refuse all agent access are not in those comparisons at all — which may be the right choice for some catalogues, but it is a choice with a cost side.
What terms should a shopping rule carry?
The store sample pairs its shopping rule with paid-tier terms at 60 requests per minute. Free terms suit stores treating agent traffic purely as demand; paid-tier suits catalogues with standalone value; contact suits owners who want a conversation first. The terms field and the pace should tell one story, whatever it is.
Can I let one agent shop and not another?
Yes — that is the ordinary use of per-agent rules. Name the agents you welcome with their own yes rules, and let the catch-all and the default carry your posture for everyone else. Each rule is independent: verdict, actions, terms, and pace are all set per agent.
What if an agent checks out despite my disallow?
Then it has acted outside your published terms, and the policy's role becomes the record: your terms were public, dated, and fetchable. Enforcement — the mechanical prevention — lives in your store's own checkout defences, as the ignores-your-policy guide explains. The declaration and the defences are complements, not substitutes.
Generate a policy — free · Validate a policy · What an agent access policy is