AI shopping agents read your store in copies: make them agree

As of 21 October 2026. Picture a shopper who never visits your store. They ask an AI shopping agent for “a merino base layer under $90 that I can send back if it doesn't fit”. The assistant finds your product in a catalog feed that says returns are accepted within 30 days, compares it with two others, and recommends yours. Three weeks later the layer doesn't fit, and the shopper's agent opens a return, quoting 30 days. Your FAQ, rewritten in the spring, says 14. A banner in your theme says “free returns”. Your own chat assistant, asked the same question, gives the answer it read when it was set up. Which of those did your store promise?

That question is new, and most of the advice about AI shopping agents doesn't reach it. The usual advice is about being found: structured data, protocols, feeds. This piece is about being consistent, because your store no longer exists in one place. It exists in copies. The agents read the copies, and the copies disagree far more often than store owners think, because nobody ever reads all of them side by side. Nobody, that is, until a machine does.

• Agent orders are real and rising: on Forter's network they reached 5.3× their early-August level by early October, and 65% came from agents reading ordinary store pages, up from 1%.
• A single fact about your store, say the return window, now lives in at least six copies: the product page, the policy page, the Shopify Catalog that feeds ChatGPT, Gemini, Copilot and Meta, your Google Merchant data, the FAQ and banners in your theme, and your own chat assistant.
• A person reads one copy and never sees the others. An agent can arrive through any of them, and research finds agents lean heavily on the first thing they read. You don't choose which copy that is.
• The oldest machine reader of store data already treats disagreement as a defect: Google says a price that doesn't match the landing page “may be disapproved”, and that mismatches “may lead to an account suspension”.
• The fix is not a protocol. It is one source per fact, numbers instead of adjectives, and an afternoon spent checking that every copy says the same thing. The grid for doing it is below.

Forter, which screens orders for fraud across a large network of online merchants, published its data on orders placed by AI agents between 1 August and 7 October 2026. Two of its findings matter more than the headline growth.

The agents read the page. At the start of August, almost every agent order came through an integrated checkout protocol. By the last week measured, 65% came from agents that browse ordinary store pages, the same pages your customers see. One agent already reaches most stores that get any. Meta's Muse bought at 74% of the stores that received an agent order at all; OpenAI's ChatGPT Agent bought at 45%.

These are one network's numbers. Agent orders are still a small share of most stores' sales, and Forter publishes no comparison with human orders on value or returns, so we won't invent one. The average agent order was $223, and only 27% were $50 or less: these are considered purchases, not impulse buys.

Some of the largest sellers are still deciding whether to let agents in at all. Amazon began blocking Muse in September, telling its users that access by an “unauthorised AI agent” breached its conditions of use (Retail Gazette). Shopify went the other way. Its agentic storefronts make your products available in ChatGPT, Google AI Mode and Gemini, Microsoft Copilot and Meta, and in Shopify's own words the feature “is active by default for eligible stores”. For most Shopify merchants, then, the question is no longer whether agents will read your store. It is what they will read.

Count the places one fact about your store, your return window, is written down today.

• The product page and theme. What people read, and what most agents read too. It includes every banner, badge, icon row and accordion that your theme and your apps add to the page.
• The policy pages. The refund, shipping and terms pages in Shopify's policy settings.
• The Shopify Catalog. Through agentic storefronts, “your products are made available to AI channels through Shopify Catalog”.
• Your Google Merchant data. What Google's Shopping Graph knows about you. It now powers Universal Cart, one cart across retailers in Search and the Gemini app, drawing on “60+ billion” product listings. Merchants, Google says, “will remain the seller of record.”
• Feeds you send directly. In OpenAI's product feed, availability is required (a row without a recognised value is rejected), and the return window is a number in its own field, return_deadline_in_days.
• Your FAQ, help articles and emails, written at different times, often by different people.
• Your own AI assistant, which answers from whatever it last read.

Every one of those was accurate on the day it was written. They were written on different days, by different people and apps, in different formats: prose in one place, a number in another, a slogan in a third. That is all drift needs.

When a person shops, the copies rarely collide. They read the product page, perhaps the returns page, and buy. If your FAQ is a season out of date, they never open it, and the disagreement sits there harmlessly for months. Inconsistency has always existed in stores. It was invisible because no reader ever saw more than one copy.

An agent changes that in two ways. First, it can arrive through any copy: a catalog in ChatGPT, a listing in Google's cart, or your product page loaded directly. Second, the evidence so far says it does not weigh the copies carefully once it has one. In Microsoft Research's Magentic Marketplace, a simulated market where customer agents deal with business agents, every model tested showed a strong bias towards the first proposal it received, worth a 10–30× advantage to speed over quality. A separate study of AI shopping agents in a simulated store, “What Is Your AI Agent Buying?”, found strong position biases that vary across providers and model versions, and found that a model update alone can shift market shares significantly.

Put those together. The copy an agent acts on is decided by things you do not control: which channel it came through, which page it loaded first, which model version it runs this week. The usual instinct is to optimise for a channel. That is the wrong game, because the channel, the ordering and the model all move under you. You can't choose which copy of your store an agent reads. You can make sure it doesn't matter.

You can't choose which copy of your store an agent reads. You can make sure it doesn't matter.

None of this is a forecast about future agents. The longest-running machine reader of store data, Google's Merchant Center, has enforced agreement between copies for years. When the price in your product data doesn't match your landing page, Google's help page says “the product with the mismatch may be disapproved”, pulled from ads and free listings until the two agree, and that mismatched prices “may lead to an account suspension”. Googlebot re-crawls the page over the following hours or days and reapproves the product once the copies match. Google's own suggestion is to update the feed on a schedule so the copy follows the source.

Google's rule is narrow: price and availability, page against feed. But it shows how a machine reader reasons. To a person, two different numbers are a puzzle to ask about. To a machine, they are a defect in the data, and the safe response to a defect is to trust the source less. We expect AI agents to move the same way, for a plain commercial reason: every purchase an agent makes that comes back as a return or a dispute is a cost to the people who run it. That is our reading, not a measured result. But it is the direction every machine reader before them has taken.

Drift isn't carelessness. It is what happens to any store with more than one writer, and every Shopify store has several: you, your team, your theme, your apps, your feeds. These are the common ways it happens.

• Timing. You change a price, a policy or a stock level in one place, and every copy picks it up on its own schedule. Between the change and the last copy catching up, your store is saying two things at once.
• Prose against fields. Your policy page is prose; feeds want numbers. Change the prose to “14 days” and a feed field set to 30 when it was first set up will go on saying 30 until a person edits it.
• Text baked into the theme. “Free 30-day returns” in a header banner, an icon row or a product template is a copy that no policy change ever touches. It is also the copy everyone sees.
• Apps that write their own claims. Shipping-estimate widgets, upsell blocks, review badges and countdown timers each publish statements about your store. They are copies too, and each has its own settings page.
• Assistants that read once. An AI assistant answers from what it was given. If it read your store at set-up, it will keep quoting that version, fluently and confidently, after the original has changed. We found this one in our own product while writing this piece; more on that below.

Consistency decides which version of your store an agent acts on. What it acts on within that version is becoming clearer too, and it points the same way: towards facts that can be checked.

Research published in Harvard Business Review in May ran thousands of simulated shopping rounds across four leading models and four product categories. Star ratings were the only cue that consistently increased the chance of a product being chosen; price reliably decreased it. Classic persuasion tactics (scarcity, countdown timers, strike-through prices, vouchers, bundles) “do not reliably influence AI agents and can even reduce selection”, depending on the model. OpenAI's own help page for shopping in ChatGPT says, plainly, that “ChatGPT may consider available options, price, reviews, and ease of use”.

One finding cuts the other way, and it matters for this piece. SusBench, a benchmark of 313 tasks on 55 real consumer websites, found that computer-use agents, like people, are particularly susceptible to preselection, trick wording and hidden information. The last of those is drift by another name: a fact that exists in one copy (the small print) and is missing from the copy the reader actually used. It may get you the sale. It does not get you out of the return. It moves the argument to later, when the agent comes back.

Google's line about Universal Cart is worth reading twice: the merchant remains “the seller of record”. The decision may move into someone else's interface. The delivery, the return, the refund and the complaint still come back to you.

With a person, they come back as a memory: “I thought it said 30 days.” With an agent, they come back as a citation. As we wrote in A prompt is not a contract, an agent keeps what it is told and treats it as a record. When it comes back for a refund, it may quote the 30 days from the feed it read three weeks ago against the 14 in your FAQ, with both your copies in hand. Which copy binds your store is a legal question this article won't answer, and nothing here is legal advice. The commercial answer is plainer: you are now arguing with a customer who has your own words in writing, and the cheapest way out is usually to honour the more generous copy.

A person forgets what your FAQ said. An agent keeps a copy.

You can do this without any tool, in about an afternoon. Open your Shopify admin in one tab and your storefront in another.

1 · Pick the five facts that end up in disputes

• Price, including every variant.
• Stock status, including variants that are sold out.
• Delivery time to your main regions.
• Return window and conditions, including sale or final-sale items.
• Who pays for return shipping.

2 · Name one source for each

For each fact, decide where the truth lives: the Shopify admin for price and stock, the refund policy in Shopify's policy settings for returns, your shipping settings for delivery. Write that down. Every other copy should either be generated from the source or be removed.

3 · Read every copy, and fill the grid

Go through each copy in turn and mark it against the source. Count vague claims as disagreements: “fast shipping” cannot be compared with anything, so for a machine it is as unhelpful as a wrong number.

4 · Fix towards the source, never towards the copy

• Numbers, not adjectives. “Delivered in 3–5 business days in the UK” instead of “fast shipping”; “returns within 30 days of delivery, unworn” instead of “easy returns”.
• Delete what you can't keep in sync. A theme banner that repeats your return window is a copy you will forget. Make it link to the policy, or say something that can't go stale.
• Update the fields, not only the prose. In any feed you control, set availability, accepts_returns and return_deadline_in_days to match the source (OpenAI's spec lists them all).
• Mark sold-out variants as sold out with inventory tracking on, rather than hiding them in a dropdown that still accepts the order.
• Check Merchant Center's diagnostics for price and availability mismatches. They are the closest thing to a free machine reader's verdict on your store.

5 · Read your store the way an agent does

• Ask an assistant to shop for you in plain shopper language (“a merino base layer under $90 that ships to the UK in under a week”). Note whether you appear and what it says about you.
• Ask it about your policies. “What is this store's return window, and who pays return shipping?” If it hesitates or answers with a number you no longer use, it has found a stale copy. Ask it where it read that.
• Ask it to compare you with a competitor. The reasons it gives are the facts it could extract. The facts it says it couldn't find are your to-do list.

6 · Put a date on the next audit

Copies drift again. Re-run the grid whenever you change a policy, a price structure or a theme, and once a quarter regardless. It takes far less time the second time, because the columns are already there.

We ran the same audit on Lumnya while writing this, because an assistant is one of the copies, and found exactly the kind of drift this piece describes.

When you install Lumnya, it reads your store's published refund, shipping, privacy and terms pages and your full catalogue, so the assistant answers from your own words rather than from invention. Product changes reach it again every time Shopify sends an update. Policies, today, are read at install. If you later change your refund policy in Shopify, the assistant keeps quoting the version it read until it is refreshed. That is the “assistant that read once” in Figure 3, and we are changing it so that a policy change reaches the assistant the way a product change does.

Everything after the click (order tracking, cancellations, returns) is what Lumnya's storefront assistant is built to handle, and anything that touches a refund waits for your approval, at every setting. How approvals work is in our FAQ. We are in private beta, and we would rather you fixed your copies first.

For twenty years a store was one page with many visitors. Now it is many copies with one owner, and only the owner can make them agree.

Agents read your store's text, and some of that text is written by strangers: see prompt injection in ecommerce. And when your own assistant states a stock level or a deadline, it should match a record.

• [1] Forter, AI agent order data, published 1 October 2026, updated 8 October 2026; Forter's merchant network, 1 August – 7 October 2026 (accessed 19 Oct 2026).
• [2] Shopify Help Center, Agentic storefronts (accessed 19 Oct 2026).
• [3] Google, Universal Cart, 15 October 2026 (accessed 19 Oct 2026).
• [4] Google Merchant Center Help, mismatched price between product data and landing page (accessed 19 Oct 2026).
• [5] OpenAI, Product feed specification (accessed 19 Oct 2026); OpenAI Help Center, Shopping with ChatGPT search (read in a browser, 18 Oct 2026).
• [6] G. Bansal et al., “Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets”, Microsoft Research, arXiv:2510.25779, 27 October 2025 (accessed 19 Oct 2026).
• [7] A. Allouah, O. Besbes, J. D. Figueroa, Y. Kanoria, A. Kumar, “What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce”, arXiv:2508.02630, v3 17 December 2025 (accessed 19 Oct 2026).
• [8] J. Sabbah, O. A. Acar, “Research: Traditional Marketing Doesn't Work on AI Shopping Agents”, Harvard Business Review, 12 May 2026; findings as stated in the HBR listing (accessed 19 Oct 2026).
• [9] L. Guo et al., “SusBench: An Online Benchmark for Evaluating Dark Pattern Susceptibility of Computer-Use Agents”, IUI 2026, arXiv:2510.11035 (accessed 19 Oct 2026).
• [10] K. Howells, “Amazon blocks Meta's new AI shopping agent from its website”, Retail Gazette, 22 September 2026 (accessed 19 Oct 2026).
• [11] Lumnya product behaviour described here was verified in Lumnya's source code on 19 October 2026: the install-time store scan (policies and full catalogue), product re-sync on Shopify's product-update webhook, policies imported at install only, and the owner-approval rule for refunds.

Lumnya is a team of ecommerce AI agents for Shopify stores, now in private beta. Join the waitlist.