# Your Store and AI Shopping Agents: The 5 Decisions You Now Own

Letting AI shopping agents into your store is no longer a hypothetical. Amazon blocks them, Shopify builds checkout tools for them, and retailers in between are deciding case by case. This playbook walks through the five decisions every merchant now owns: access, friction, answers, checkout, and attribution.

Canonical URL: https://nile.app/blog/prepare-for-ai-shopping-agents/
Author: Zhao Hanbo
Created: 2026-10-05T13:53:52.789Z
First published: 2026-11-10T15:34:05.200Z
Last modified: 2026-11-10T15:34:05.200Z
Language: en-US

## Key takeaways

- Access controls are moving from which agents to which actions: Cloudflare now lets sites allow personal agents, block them, or block them only on pages with ads.
- Retailers have split into named camps on agent access: Amazon blocks shopping agents including Meta's Muse, Shopify extended WebMCP so agents can complete checkout with buyer authorization, and others allow browsing but not purchasing.
- An agent access policy has three levels: block, browse-only, and transact. The decision is reversible, but defaulting into it through bot defenses configured years ago is still a decision.
- Revenue that depends on customer friction, such as unredeemed credits, forgotten subscriptions, and unclaimed price adjustments, is now an attack surface: agents with infinite patience collect what people abandon.
- Agents reach your checkout three ways: driving the site like a person, using platform checkout tools such as Shopify's, or sending a structured order over a commerce protocol. Browser-mode orders are hard to tell apart from human orders.

A year ago, "should AI shopping agents be allowed to shop here" was a panel question. This autumn it became a policy with named positions. Amazon blocks outside shopping agents, including Meta's Muse. Shopify went the other way and extended WebMCP so browser-class agents can complete checkout with the buyer's authorization.

In between, retailers are splitting case by case. [PYMNTS reported](https://www.pymnts.com/news/retail/2026/ai-agents-knock-retailers-decide-how-wide-to-open-the-door/), citing The Wall Street Journal, that:

* QVC and HSN let agents complete purchases and send their catalogs to OpenAI and Google
* Kohl's does not let agents buy
* Tapestry lets shoppers buy Coach and Kate Spade inside Google's AI Mode, while keeping agents off its own checkout

The agents knocking keep multiplying. Meta's Muse, [Instinct](https://techcrunch.com/2026/08/26/viral-ai-startup-instinct-has-raised-350-million-at-a-2-5-billion-valuation/), xAI's [Grok Bot](https://stripe.com/blog/helping-personal-agents-shop-more-intelligently-and-reliably-with-link), and OpenAI's [Dots](https://www.cnbc.com/2026/09/30/openai-follows-meta-into-the-red-hot-market-for-personal-agents.html) are all personal agents that act on one person's behalf.

We mapped the kinds of shopping agents and how each one buys in [a separate piece](/blog/ai-shopping-agents/).

Whatever you sell, you now own the same five decisions those retailers just made. This playbook walks through them in order: access, friction, answers, checkout, and attribution.

## Decision 1: your agent access policy

### Three positions you can take

**Block.** Bot defenses stop known agents at the door. This is Amazon's position, and it is not irrational: Amazon's ad business [generated $76 billion over twelve months](https://www.fool.com/investing/2026/10/03/metas-muse-ai-good-or-bad-for-e-commerce-shopify-a/), and an agent that shops by delegated intent does not look at ads. If a large share of your margin comes from monetizing attention inside your own property, agents are a threat to that specific revenue before they are a source of any other.

**Browse-only.** Agents may read, compare, and recommend, but checkout requires a human session. This preserves discovery value while you evaluate order quality, fraud exposure, and support load. Its cost is invisible: an agent sent to buy, not to browse, finishes the errand at a store that lets it finish.

**Transact.** Agents complete purchases through sanctioned paths. This is where Shopify has moved its merchants as a default direction, and where Meta's [announced Muse partners](https://www.cnbc.com/2026/10/03/meta-muse-shopping-ai-agent.html), including Walmart, Wayfair and Best Buy, have lined up, with Shopify and Stripe building the rails underneath.

![Retailers placed on a spectrum from blocking AI shopping agents to letting them buy, autumn 2026. Amazon blocks Meta's Muse, Kohl's does not let agents buy, Tapestry allows buying only inside Google's AI Mode, and QVC and HSN and Shopify stores let agents complete purchases](https://nile.app/blog-media/media/info_08_retailer_stances_a0855f5091/info_08_retailer_stances_a0855f5091.png)

### Why your current setting is probably an accident

Most stores already have an agent policy, and few chose it. The tools you run set defaults, and those defaults can point in opposite directions:

* **Shopify** keeps its [agentic storefronts channel](https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts) active by default for eligible stores, so your products can reach AI channels you never switched on. We walked through those switches in [a separate piece](/blog/already-in-chatgpt-what-you-control/).
* **Cloudflare** asks new sites during onboarding whether they run ads, then [recommends agent policies](https://blog.cloudflare.com/personal-agents-visibility-control/) based on the answer.
* **Older bot defenses**, configured years ago, may still be enforcing "block" on anything that does not look like a person.

Leaving those defaults in place is also riskier than it looks, because the asymmetry runs against waiting. A blocked agent does not file a complaint. It substitutes, silently, on every delegated errand, and the demand it carries goes to stores that said yes. You can reverse an access policy in a week. You cannot reverse the recommendation patterns an agent learned while you were closed.

How fast agent orders grow is still an open question. If it is gradual, there is time to make this decision deliberately. If it is not, a store that defaulted into blocking finds out from orders it never sees.

### What blocking looks like at Amazon

On September 20, 2026, Amazon [began cutting Muse off](https://ppc.land/amazon-blocks-metas-muse-ai-agent-from-shopping-on-its-site/), with a popup telling shoppers that continued access by an unauthorized AI agent violates its Conditions of Use.

The next day, Shopify CEO Tobias Lütke [announced a partnership](https://www.pymnts.com/commerce/ecommerce/2026/shopify-brings-shop-pay-checkout-solution-to-metas-muse-ai-agent/) to enable agentic checkout with Shop Pay for Muse on all Shopify stores.

In November, Amazon let select US customers [text Alexa for Shopping](https://www.retaildive.com/news/amazon-text-to-shop-ai-agent/874085/), its own agent, which can complete purchases and show products from across the web through Shop Direct. Blocking outside agents while building your own is a coherent plan for a marketplace that owns the customer relationship. Most independent merchants do not have that option.

### The legal ground is still moving

On August 4, 2026, the Ninth Circuit [vacated an injunction](https://www.jonesday.com/en/insights/2026/09/ninth-circuit-vacates-cfaa-injunction-against-perplexitys-comet-ai-agent) Amazon had won against Perplexity's Comet agent, holding that the user, not Perplexity, is the one who accesses Amazon's computers under the federal anti-hacking law.

The court limited its holding to the record in front of it, so it does not settle whether a store may turn agents away. It does make "the agent is trespassing" a weaker argument than it was a year ago.

### The next question is which actions, not which agents

The three positions above apply to an agent as a whole. The tools are starting to go finer. In November, Cloudflare [added per-agent controls for personal agents](https://blog.cloudflare.com/personal-agents-visibility-control/) with three settings: allow, block, and block on pages with ads. That last setting exists because an agent that reads a page without seeing its ads undercuts a publisher's business, while for a store the same agent may be bringing a buyer.

My read is that access policies will keep moving from agents to actions. Let agents discover products. Let them buy with the customer's authorization. Restrict the actions that bypass how you make money or create a security risk. Write your policy that way now, even if your tooling cannot enforce it yet, so that you are ready to switch it on when it can.

## Decision 2: audit your friction revenue

### What agents are being sent to collect

Some of the first errands people hand to agents are tedious money work. Instinct's founder, quoted by [TechCrunch](https://techcrunch.com/2026/08/26/viral-ai-startup-instinct-has-raised-350-million-at-a-2-5-billion-valuation/), said users had it cancel hundreds of dollars of subscriptions, alongside buying weekly groceries and concert tickets. Canceling subscriptions, chasing refunds and claiming price adjustments are all tasks where a business profits when the customer gives up.

Seen from the merchant side, some of that abandoned money is your margin:

* Subscriptions that renew because canceling takes effort
* Store credit and gift card balances that expire unused
* Price-protection and adjustment policies almost nobody invokes
* Return windows that lapse because the process is tedious
* Claims and warranties abandoned halfway

An agent with infinite patience collects all of it, at near-zero marginal cost, on the customer's instruction.

### How to run the audit

1. List every line of revenue or retained cash that exists because a process is tedious.
2. Assume each one trends toward zero as agent adoption grows.
3. Decide which of those you would rather fix on your own terms.

A subscription that is easy to cancel is also easy for an agent to verify, and "this brand's billing is clean" is exactly the kind of fact a recommendation system can hold about you. Brands that profit from friction will find agents working against them. Brands that do not can treat the same agents as a sales channel.

## Decision 3: make the store answerable

An agent executing "buy a tent under two kilograms that ships by Friday" does not read your product page the way a person does. It extracts fields, checks constraints, and moves on. We covered the product-page half of this in [One Product, Many Answers](/blog/pdp-ai-shopping/). The store-level half comes down to four things.

**Field-level product facts.** A constraint can only be satisfied if it exists as a field. Weight, dimensions, materials, compatibility, what the product is not for. Prose does not enter the comparison.

**Price and availability certainty.** An agent that cannot confirm stock and final price recommends a product it can confirm. An agent treats any uncertainty as a reason to pick something else.

**A stable path to purchase.** Guest checkout, predictable forms, no CAPTCHA wall at the finish line. Agents fail loudly at exactly the steps you have been meaning to simplify for humans.

**A deliberate promotion policy.** Agents hunt promo codes with perfect memory. An evergreen code that any visitor can find is, to an agent, simply your real price. Decide what price agents should find, because they will find the lowest one that works.

## Decision 4: pick your checkout paths

Juozas Kaziukenas's analysis of Muse's code, [covered by PPC Land](https://ppc.land/metas-muse-can-skip-ads-yet-still-shape-instagram-ads-mbi-argues/), found three checkout modes: a browser mode that types card details into whatever form the store presents, a Shopify mode that runs over APIs, and a Stripe Link mode. Across the wider field of agents, the paths a merchant actually sees group into three:

1. **Browser mode.** The agent drives your site like a person and enters payment details at your checkout. Muse, for example, can pay through Stripe's Link or Shop Pay after showing the shopper an [approval card](https://www.cnbc.com/2026/10/03/meta-muse-shopping-ai-agent.html) with the order details. It works on any site, and you can almost never tell an agent placed the order.
2. **Platform tools.** On Shopify, agents can increasingly complete checkout through sanctioned tools, including WebMCP's checkout functions. The order arrives through a structured path, so you can see that an agent placed it.
3. **Protocol.** Dedicated shopping agents can send a structured order over a commerce protocol such as the Agentic Commerce Protocol, and the order arrives labeled as one.

![Three ways an AI agent checks out at a store. Through the browser it enters payment details at checkout and looks like an ordinary card payment. Through platform checkout tools such as Shopify's, or over a commerce protocol, the agent order is clearly visible to the merchant](https://nile.app/blog-media/media/info_08_checkout_modes_v2_79ccffb98f/info_08_checkout_modes_v2_79ccffb98f.png)

### What to do about it

* **Support the sanctioned paths your platform offers.** Every order that arrives through a structured path is one you can see, verify, and if needed reverse.
* **Make the structured path the easy one.** Browser-mode orders will happen regardless of your preference. The goal is to shrink their share.
* **Revisit order confirmation and cancellation.** Agents can misread an instruction and order the wrong thing. A reversible confirmation is cheap insurance for you and the shopper.

### Payments are moving fastest

Stripe said on September 29, 2026 that [agentic purchases made with Link grew 38x over the previous month](https://stripe.com/blog/helping-personal-agents-shop-more-intelligently-and-reliably-with-link), and named Muse, Grok Bot and Instinct as agents using Link's wallet for agents.

The same week, OpenAI announced [Dots](https://www.cnbc.com/2026/09/30/openai-follows-meta-into-the-red-hot-market-for-personal-agents.html), always-on agents available from its $100-a-month Pro plan.

OpenAI did not launch Dots as a shopping agent, and it does not need to be one. Once a general agent can act on someone's behalf and pay, shopping becomes one more task it can take on, with no dedicated shopping feature needed. Your checkout will meet more agents that were never built for commerce.

## Decision 5: measure before you can attribute

Attribution is the decision you cannot fully make yet. A browser-mode agent fires your analytics tags like a person. Its visit registers on your site tag as the user's own activity, which, per [MBI's analysis](https://ppc.land/metas-muse-can-skip-ads-yet-still-shape-instagram-ads-mbi-argues/), is also how agent browsing can inform retargeting on Meta's ad surfaces even though Muse itself carries no ads.

Your dashboard will show a human who browsed efficiently, never scrolled idly, and converted. As agents take a growing share of traffic, several familiar metrics change meaning: session duration, pages per visit, add-to-cart rate.

### Build the baseline now

What to log, starting now:

* Log checkout path (browser, platform API, protocol) as a first-class order attribute
* Log payment method, since agent wallets such as Stripe's Link can be a useful agent signal
* Watch for the behavioral signature: zero dwell, direct paths, perfect form fills
* Keep AI-referral and agent-order numbers in a separate line, so when a platform ships proper agent identification, you have a before and after

### Agents can now be paid for recommending you

Some agent recommendations are about to become paid placements. Under Shopify's [promoted placements program](https://shopify.dev/docs/agents/catalog/promoted-placement), Shopify pays agent developers a base commission of 0.3% on attributed purchases of eligible products, and merchants can fund more through Shop Campaigns.

The merchant-funded part applies when the buyer completes the purchase in the Shop app or on shop.com. Access is invite-led during a developer preview, and Shopify says the rates might change before general availability.

For attribution, that adds one more split to track: orders that came from an organic agent answer, and orders that came from a promoted one.

It also raises a question your customers will start asking about the agents they use: who is this agent working for? If you fund promoted placements, decide what you are willing to pay for and keep those orders on their own line from the first one.

## Start here

Most of the risk in this playbook comes from defaults nobody chose. The three checks show you yours.

### Check today: see what your store already decided

* Open your Shopify agentic storefronts settings and note which AI channels and checkout options are already on
* Look up the agent policy in Cloudflare, or whatever bot protection you run, and note what it does to personal agents
* Search for your brand plus "promo code" and see which codes an agent would find

### Decide this quarter: make those choices on purpose

* Write down your access policy (block, browse-only, transact) instead of inheriting it from your bot defenses, and spell it out action by action
* List your friction revenue and decide which lines you will give up on your own terms
* Get product facts into fields, keep price and stock certain, and clear the path to checkout
* Prefer sanctioned checkout paths, and make browser-mode the minority by making the structured path easier
* Make orders reversible and confirmations explicit, for the agent's mistakes
* Start logging agent signals now, before you need the baseline
* Keep promoted-placement orders separate from organic agent orders from day one

The field-level work in Decision 3 is the part you do not have to do by hand: turning a catalog into facts every kind of agent can read is what [Nile](https://nile.app/) builds for brands. If you want to see what your catalog can answer today, [connecting it to Nile](https://nile.app/join) takes a few minutes and [costs nothing until it earns you a sale](https://nile.app/pricing).

Your store will answer these five questions either way, on purpose or by default.

## Frequently asked questions

### Should I block AI shopping agents from my store?

Blocking protects ad-dependent and marketplace revenue, which is why Amazon does it. For most independent merchants the calculus is different: an agent that cannot buy from you recommends a store it can buy from. A browse-only policy is a reasonable middle step while you verify order quality.

### How do AI shopping agents pay for purchases?

Mostly through agent wallets such as Stripe's Link or Shop Pay, after the person approves the order. When the agent pays by entering card details at your checkout, the transaction looks like a normal card payment to your payment stack.

### Can I tell agent orders apart from human orders?

Often not today. An agent driving a browser session fires your analytics tags like a person would. API and protocol checkouts are identifiable. Start by logging what you can: checkout path, session characteristics, and payment credential type, so you have a baseline when agent volume grows.

## Sources

1. [Meta's Muse can skip ads yet still shape Instagram ads, MBI argues](https://ppc.land/metas-muse-can-skip-ads-yet-still-shape-instagram-ads-mbi-argues/) — PPC Land
2. [Meta's Muse AI: Good or Bad for E-Commerce? Shopify and Amazon Are Taking Opposite Approaches](https://www.fool.com/investing/2026/10/03/metas-muse-ai-good-or-bad-for-e-commerce-shopify-a/) — The Motley Fool
3. [Shopify opens checkout to browser-based AI agents](https://techcrunch.com/2026/09/28/shopify-opens-checkout-to-browser-based-ai-agents/) — TechCrunch
4. [Meta's Muse can shop and check out for you. Here's how it works](https://www.cnbc.com/2026/10/03/meta-muse-shopping-ai-agent.html) — CNBC
5. [AI Agents Knock. Retailers Decide How Wide to Open the Door.](https://www.pymnts.com/news/retail/2026/ai-agents-knock-retailers-decide-how-wide-to-open-the-door/) — PYMNTS
6. [OpenAI follows Meta into the red-hot market for personal agents. But will users pay?](https://www.cnbc.com/2026/09/30/openai-follows-meta-into-the-red-hot-market-for-personal-agents.html) — CNBC
7. [Viral AI startup Instinct has raised $350M at a $2.5B valuation](https://techcrunch.com/2026/08/26/viral-ai-startup-instinct-has-raised-350-million-at-a-2-5-billion-valuation/) — TechCrunch
8. [Amazon blocks Meta's Muse AI agent from shopping on its site](https://ppc.land/amazon-blocks-metas-muse-ai-agent-from-shopping-on-its-site/) — PPC Land
9. [Shopify Brings Shop Pay Checkout Solution to Meta's Muse AI Agent](https://www.pymnts.com/commerce/ecommerce/2026/shopify-brings-shop-pay-checkout-solution-to-metas-muse-ai-agent/) — PYMNTS
10. [Amazon debuts text-to-shop AI agent feature](https://www.retaildive.com/news/amazon-text-to-shop-ai-agent/874085/) — Retail Dive
11. [Ninth Circuit Vacates CFAA Injunction Against Perplexity's Comet AI Agent](https://www.jonesday.com/en/insights/2026/09/ninth-circuit-vacates-cfaa-injunction-against-perplexitys-comet-ai-agent) — Jones Day
12. [See the agents, set the terms: new visibility and control for personal agents](https://blog.cloudflare.com/personal-agents-visibility-control/) — Cloudflare Blog
13. [Helping personal agents shop more intelligently and reliably with Link](https://stripe.com/blog/helping-personal-agents-shop-more-intelligently-and-reliably-with-link) — Stripe
14. [Earn with promoted placements](https://shopify.dev/docs/agents/catalog/promoted-placement) — Shopify developer documentation
15. [Agentic storefronts](https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts) — Shopify Help Center