Pick your best-selling product. Open its page on your own site, on momo, on Shopee and on the price-comparison site you check most, side by side, and look at four things only: the price, whether it’s in stock, the shipping cost, and how many days a buyer has to return it. If you can reconcile all sixteen at a glance, you can skip this piece.

An AI agent shopping on someone’s behalf does exactly this. The difference is that it won’t keep the tabs open and work through them patiently, and it won’t call you to ask which one counts.

Agents compare six facts, and your brand story isn’t one of them

An agent carries the user’s conditions: a budget, a size and colour, a date it has to arrive by. Its job is to find, among a pile of options, one it can be sure meets them. So what it compares are things it can check: price, specs, stock, shipping cost and delivery time, returns terms, and reviews.

The platforms name these fields in their own documentation. When OpenAI launched Instant Checkout in ChatGPT on 29 September 2025, it wrote that when several merchants sell the same product, ChatGPT’s ranking considers factors such as availability, price, quality and whether a merchant is the primary seller. Google’s agentic checkout, which began rolling out in the US in November 2025, has the shopper set the size, the colour and how much they’re willing to spend; when the price drops into that budget they get a notification, and once they confirm the purchase and shipping details, Google completes the checkout on the merchant’s site with Google Pay. Then on 24 March 2026, OpenAI said the first version of Instant Checkout hadn’t offered enough flexibility, let merchants use their own checkout instead, and put its effort into the comparison step: ChatGPT now lays products side by side with price, reviews and features.

Who runs the checkout keeps changing. What gets compared has stayed the same throughout. However well your brand story is written, by the time it reaches that comparison table it has shrunk to a single column: the brand name.

Machines pull those six facts from three places

The first is your product page itself. Crawlers read the visible text and the schema.org structured data embedded in the page: Product for the item, Offer for the terms of sale. Offer carries the price and how long that price is valid, the availability (in stock, out of stock, pre-order — there’s even a ready-made value for discontinued), and dedicated properties for shipping details and the returns policy.

The second is the product feed you submit. On Google’s side that means Merchant Center. On OpenAI’s side it runs through the Agentic Commerce Protocol (ACP), which OpenAI co-developed with Stripe. With the March changes, OpenAI extended ACP to product discovery, and merchants now use it to send product feeds and promotions into ChatGPT. OpenAI’s feed spec has nine required fields, price and availability among them; shipping, returns and reviews are optional. Google’s Universal Commerce Protocol (UCP), announced on 11 January 2026 and co-developed with Shopify, Etsy, Wayfair, Target and Walmart, lets people check out directly inside AI Mode and the Gemini app — and it, too, requires an active Merchant Center account with your products submitted.

The third is other people’s pages: marketplace listings, prices scraped by comparison sites, review sites and forum threads. Most of that isn’t in your hands, and some of it is run by resellers or by an agency that manages your marketplace store.

Each of these is updated by different people, on a different rhythm. Google said in May 2025 that more than 2 billion listings in its Shopping Graph are refreshed every hour. On your side, a price change might be waiting for next Monday’s meeting.

All of the agent checkouts described above are US-only for now. Google said in May 2026 that UCP-powered checkout will roll out to Canada and Australia in the coming months and to the UK after that; OpenAI’s standard feed format targets the US and is open only to approved partners. Taiwan isn’t on either list. The card-payment rails, meanwhile, are moving fast: Visa and Mastercard have both made it possible for AI agents to pay by card. The comparison step doesn’t wait for checkout, though. When someone asks ChatGPT, Gemini or Perplexity in Chinese which of two models is better value, the answer is built from data pulled out of these same places.

Faced with a contradiction, an agent’s easiest move is to switch sellers

Our view: in the agent era, brands mostly get dropped because their data contradicts itself, not because their price wasn’t low enough.

It comes down to the agent’s position. To choose for someone, it has to be able to say “this one meets your conditions.” If your site says free shipping and the marketplace says NT$80, it can’t say that. If one listing says in stock and another says discontinued, it can’t risk placing the order. At that point it has two options: spend effort working out which version is right, or move to a seller whose data is clean. The second is far cheaper, and there’s never a shortage of other sellers on the shelf. (The platforms haven’t published how their agents handle conflicting data; this paragraph is our inference.)

Google’s shopping system has already shown how little patience machines have for contradictions. Merchant Center’s help documentation says Googlebot routinely crawls your product pages and compares the price in your feed with the price on the page and in your structured data. Products with a mismatch may be disapproved, and the mismatch may lead to an account suspension. The system doesn’t stop to work out which price you actually meant.

One more line in OpenAI’s feed spec matters here: an empty shipping field means unknown, not free shipping. So you can offer free shipping on everything and still show up to the machine as a question mark. Put a question mark next to a competitor who spelled out free shipping, and the outcome isn’t hard to guess.

The four mismatches brands run into most

These are common among brands that sell on their own site and on marketplaces at the same time, and usually nobody meant for them to happen.

One product, five places, five prices

Your site shows the list price, the marketplace is running a campaign, members get a member price, the comparison site scraped last month’s sale price, and the structured data still carries the price from before the redesign. Each price makes sense in its own context. Side by side, they’re five different answers. When Google lists common causes of price mismatches, a sale price with the wrong effective dates or time zone is one of them: the campaign has ended, and somewhere the price is still stuck inside it.

Discontinued on one channel, in stock on another

The product is gone from your site but the marketplace listing is still live. Or it runs the other way: your site was never updated, and the marketplace sold out long ago. schema.org has a ready-made value for discontinued; the catch is that it has to be remembered and set, in every place the product appears.

Returns terms locked inside an image

Plenty of brands turn their returns policy into a beautifully designed long image at the bottom of the FAQ or product page. People can read it. Machines see a picture. And returns terms don’t sit still: Google’s own documentation says plainly that retailer return policies can get complicated and may change frequently.

Specs that don’t line up between your site and the marketplace

Your site says 1.5 litres, the marketplace says 1500 ml. Your site calls the colour “mist grey”, the marketplace calls it “grey”. Your site has the new model’s spec sheet, the marketplace still shows the old one. An agent about to place an order first has to establish that these pages describe the same model and the same option, and when they don’t line up, it can’t. That’s why OpenAI’s feed spec asks for one row per colour and per size, each carrying its own price, stock status and URL.

When these cases surface in an AI’s answers, either it’s reading an old version of you, or it’s piecing one together where you never spelled things out. We call this signal distortion.

Why “get someone to fill in the schema” won’t close the gap

The instinctive response at this point is to have someone complete the structured data on every page. That’s worth doing. It just doesn’t resolve the contradictions.

Structured data restates what’s on the page in a fixed format machines can read, which amounts to putting an official stamp on it. If the underlying data was inconsistent to begin with, finishing the schema only stamps one of the versions, and adds one more place that gets checked against your feed. The Merchant Center rule above compares exactly those three things: the feed, the page and the structured data.

Then there’s translation. The same fact has to be expressed in several formats, each with its own rules. Take returns. Google alone accepts four sources — the Content API, settings in Merchant Center or Search Console, product-level markup and organisation-level markup — in a fixed order of precedence. The example in its documentation: mark up a returns policy on your site and also set one in Search Console, and Google uses only the Search Console version. In practice, someone fixes the website and an old setting in the back office still wins. OpenAI does it differently again. Returns are split into three fields — whether returns are accepted, the window in days, and a policy URL — and supplying the URL alone doesn’t count as accepting returns.

The hardest part is people. Pricing belongs to merchandising, stock lives in the warehouse and the ERP, shipping rates are negotiated by logistics, returns terms need sign-off from customer service and legal, spec sheets come from the product team, and the marketplace store may be run by yet another agency. Every field has an owner. “All six fields match everywhere” has none. Which system is the source of truth for each fact? Who is allowed to change it? Who has to be told when it changes? Those are organisational decisions, and they sit outside the remit of whoever maintains the website.

A managed service fills the seat nobody owns: consistency

What a managed service can do, and what is hard to do inside your own organisation, is give this a dedicated owner who looks at it the way a machine reads your data. We measure it on two layers. The GEO audit shows how your site reads to machines, with structured data as one of its dimensions. AI citation backtesting puts the shopping questions your customers ask to each of the major AI engines and checks whether what they say about your price, specs and returns terms is right; that’s answer accuracy. Where things don’t line up, we fix your site and content. For the fields that sit on other channels, you’ll at least know which field and which version is wrong, and can take it straight to the person who owns it.

Bring a test question to a demo: something shoppers in your category genuinely ask, like which model is the best value at this price, and see on the spot whether the AI’s comparison includes you, and what it says in your column.

Further reading