Somebody who has never heard of your store just typed a sentence into an assistant. Not a keyword. A sentence, the way you would ask a knowledgeable friend: "I need a waterproof dog bed for a 70-pound lab who chews everything, under two hundred dollars."
They got back four products. Yours was not one of them.
There was no search results page to scroll. No second page. No ads down the side to click instead. Four names, a short reason for each, and a link to go buy. The shortlist was assembled somewhere the shopper never saw, out of data they never looked at, and by the time they were reading it the competitive round was already over.
That is the part worth understanding, because it is genuinely new. And it is much cheaper to fix than the version of this story you have probably been sold.
Everyone Told You the Robot Would Buy It for Them
For about a year, the pitch has been that shopping is about to happen entirely inside the chat window. The assistant finds it, the assistant adds it to the cart, the assistant pays. Your store becomes a supplier at the back of somebody else's shopfront.
The platforms genuinely tried to build that. On 29 September 2025 OpenAI launched Instant Checkout in ChatGPT, letting US shoppers buy from US Etsy sellers without leaving the conversation, with over a million Shopify merchants to follow. It ran on the Agentic Commerce Protocol, an open standard OpenAI co-developed with Stripe. In January 2026 Google announced its own open standard for the same idea, the Universal Commerce Protocol, built with Shopify, Etsy, Wayfair, Target and Walmart, powering a checkout button on eligible product listings in AI Mode and the Gemini app.
Then it started coming apart.
By early March 2026, OpenAI had quietly deprioritized Instant Checkout in favor of merchant-specific apps, with a spokesperson saying the company was "evolving how we approach commerce in ChatGPT to better meet merchants and users where they are." The stated reasons were the ones any store owner would recognize: stock status, sales tax and pricing change constantly, and a checkout that lives somewhere else is always slightly behind.
A couple of weeks later the numbers arrived and explained it. Walmart had made roughly 200,000 products available through Instant Checkout from November. Daniel Danker, Walmart's EVP of AI acceleration, product and design, said that purchases completed inside ChatGPT converted at one-third the rate of purchases where the shopper clicked out to walmart.com to finish. He called the experience "unsatisfying." Walmart pulled it and replaced it with its own agent living inside ChatGPT instead.
Here is why that matters to a store doing a few hundred orders a month. It means you can stop worrying about the expensive half of this. You do not need to plumb your store into an agentic checkout protocol. You do not need a developer. The transaction is staying on your site, because that is where it converts, and the biggest retailer in the country just paid to prove it.
What did not stall is the discovery. That part worked, it is still running, and it is the half that decides whether you are in the conversation at all.
The Shortlist Is Built Before Anyone Reaches Your Site
The thing to understand about these results is the order they come in. The assistant surfaces a product first, and then shows you which retailers sell it. Product, then seller.
If you have run Google Shopping ads, you know that is backwards from what you are used to. A Shopping ad arrives as product and retailer together: your photo, your price, your store name, one click to your site. The assistant experience is closer to the organic side of Google Shopping, where a product gets named and then a list of sellers appears underneath it.
That difference changes what you are competing for. On a search results page you are competing for a click among ten or twenty visible options, and a shopper who does not like the first one just keeps scrolling. In an assistant answer, there is no scrolling. There are four or five things, chosen by a system, described in a sentence each.
Being fifteenth is not being fifteenth any more. It is being absent.
And the selection is not happening on your website. The assistant is not browsing your beautifully built homepage, admiring your photography and reading your About page. It is reading structured product data, cross-referencing what people have written about that product elsewhere, and assembling an answer. Your site is where the shopper lands afterwards, if you make the list.
The File That Decides This Is One You Already Have
Here is the genuinely good news, and it is the reason this is a Monday-morning job rather than a project.
The data source for Google's side of this is Merchant Center. That is the same product feed you already maintain if you run Shopping ads, or that your platform generates for you automatically if you are on Shopify or WooCommerce. Google's own announcement is explicit that the new conversational attributes "complement retailers' existing data feeds."
And OpenAI's product feed specification is, in its own documentation, a Google-compatible product data format. The required fields on every row are the ones you would expect and mostly already have: a unique ID, title, description, link, image link, availability, price, brand, plus who the seller is.
So the answer to "how do I get into AI shopping results" is not a new pipeline. It is the catalog you already own. Which is the same conclusion we reached about Google Shopping ads having no keywords - the feed is not paperwork you file to get approved, it is the thing being judged.
The problem is that most small-store feeds are thin, because until now nothing punished them for it. Open yours and look for these:
- Brand is missing or is your store name. If you sell other people's products, brand needs to be the manufacturer, or the assistant cannot match your listing to the product it is recommending. This one silently removes you from the running.
- One image. The feed accepts several. A single hero shot on white tells a matching system almost nothing about scale, use or fit.
- The title is your website's title. "The Dawson - Our Bestseller" is a name. A title in a feed needs to say what the thing is: brand, product, key attribute, size. Boring is correct here.
- The description is marketing copy. More on this next, because it is the biggest opportunity in the file.
- Availability and price are stale. If your feed updates once a night, you are wrong for most of the day every time something sells out or goes on sale. Fresh data is worth more here than it ever was for ads, because an assistant that recommends something out of stock stops recommending you.
- No GTIN or MPN on products that have one. These are how a system knows your listing and somebody else's listing are the same physical object.
None of that requires a developer. All of it is an afternoon in your product admin.
Write the Description for the Question, Not the Product
This is the part almost nobody has done, and it is the part with the most room in it.
When Google announced the conversational commerce work, it asked retailers to submit "dozens of new data attributes in Merchant Center designed for easy discovery in the conversational commerce era" - and named the examples: answers to common product questions, and compatible accessories or substitutes.
Read that again, because it is remarkable. The platform is asking you to file the answers to the questions your customers ask before buying, as structured data.
We wrote a couple of days ago that your support inbox is a sales channel - that the questions arriving before a purchase are a list of the reasons people have not bought yet. That was a conversion argument then. It is a distribution argument now. The pre-sale questions in your inbox are the raw material for the fields that decide whether an assistant considers you a match.
Go read your last fifty pre-sale messages and pull out what people actually ask:
- Fit and sizing. Not the measurement, the decision. "Runs half a size small" is worth more than a chart.
- Compatibility. What it works with, and what it does not. This is the single most valuable thing you can state plainly, because compatibility questions are exactly what people bring to an assistant.
- Materials and care. Waterproof or water resistant. Machine washable or not. Dishwasher safe or hand wash.
- What it is not for. Counter-intuitive, and it works. A system trying to match a request to a product is helped enormously by knowing where the product stops. Stating the limit also keeps you out of the recommendations you would have been returned from.
- What goes with it. Google asked for this one by name. If your product needs a bracket, a filter or a cable, say which.
Write those as plain declarative sentences in the description, and into the attribute fields where your platform gives you one. You are not writing to charm anybody. You are writing so that a machine reading a thousand products can tell that yours is the one that fits a 70-pound lab who chews everything.
The Agent Is Not Impressed by Your Marketing
There is a real shift buried in all of this, and it is worth naming plainly: increasingly, the thing evaluating your product first is not a person.
An assistant does not have an emotional response to your packaging. It has not seen your ads. It does not know your founder story and is not moved by the influencer you paid last quarter. It weighs the things that can be checked: whether your data is complete, whether your reviews are consistent, whether your price has behaved sensibly, whether other people on the internet have said your product does what you claim.
Two consequences follow, and both cut against habits that are common in small ecommerce.
The first is that reviews stop being a trust widget on your product page and start being an input somewhere upstream. We have written before about a stranger believing your customers rather than your store, and everything in that piece still holds. What is new is that reviews attached to the product now matter as much as reviews attached to your store, because the shortlist is built at the product level. Twelve reviews of your store are worth less here than twelve reviews of the item.
The second is pricing behavior. The old retail trick of raising a price in October so the November discount looks enormous works on a human who has not been watching. It does not work on something that can hold your price history. If you have read our argument that Black Friday is won in September, this is one more reason the September version of that work beats the November version: a promotion built on a real offer survives being checked, and a promotion built on an inflated reference price does not.
Reselling Somebody Else's Product Is the Hard Version
Everything above is easier if you make what you sell. If you resell, be honest with yourself about the position you are in.
When the assistant names a product first and then lists who sells it, a reseller is a row in a comparison. The product has already been chosen. What is left is which retailer, and the fields on offer are mostly price, availability and shipping.
That is not hopeless, but it is different work. Three things actually move it.
Show the true landed cost. These comparisons surface shipping alongside price, and being two dollars cheaper on the item while being two hundred dollars more expensive to deliver is a losing position that looks like a winning one in your own reporting. Free shipping thresholds and honest delivery estimates are competitive assets in this context, not margin leaks.
Be complete where others are lazy. Most resellers paste the manufacturer's copy and move on. If you are the only listing with real dimensions, honest stock status, compatibility notes and product-level reviews, you are the easiest one for a matching system to be confident about.
Own a category rather than a catalog. A store that sells four hundred unrelated things has no story a system can tell about it. A store that is unmistakably the place for one kind of product accumulates the off-site mentions, the reviews and the specialist coverage that make it the safe recommendation. This is slower than a feed fix, and it is the only durable version of the answer.
You Still Have to Close
Now the reassuring part, and it is the reason not to panic about any of this.
The traffic still arrives at your product page. That is the whole meaning of the Walmart result: the transaction went back to the retailer's site because that is where it converts. So the visitor who found you through an assistant lands in exactly the place your store has always had to perform.
They arrive differently, though, and it is worth adjusting for. They have already been told your product is a good match for a specific need. They are further along than a cold visitor from an ad, and they are carrying a claim your page now has to confirm. If the assistant said "waterproof, chew resistant, fits a large dog" and your product page opens with a lifestyle photo and a poetic paragraph, you have made them go looking for their own answer.
Which brings all of this back to the same page it always comes back to. Everything in your homepage routing and your product page selling still applies, and so does everything about not talking people out of it at checkout. The assistant is a new front door. The house is unchanged.
What to Do Monday
None of this is a project. It is a list.
- Open your product feed and look at it. Not the dashboard summary, the actual rows. In Merchant Center, the diagnostics tab tells you what is being rejected and what is missing. Most owners have never opened it.
- Fix brand, GTIN and MPN first. These are the matching keys. Everything else is refinement; these are the difference between being a candidate and being invisible.
- Rewrite ten product titles the boring way. Brand, product, defining attribute, size. Start with your ten bestsellers.
- Mine fifty pre-sale messages and turn the recurring questions into declarative sentences in your descriptions and attribute fields.
- Check your feed's update frequency. If it is nightly and you sell out of things, make it more often. Your platform probably has a setting.
- Get reviews onto products, not just the store. If your review app only collects store reviews, that is a configuration you can change today.
- Watch your referral traffic. In your analytics, look for sessions arriving from assistant domains. It will be a small number and it will be an undercount, because much of this is a black box and you will never see the answers you did not appear in. Watch the direction of travel rather than the absolute figure.
That last point deserves honesty rather than a workaround. You cannot see the searches you lost. You cannot A/B test your way into an assistant's shortlist. This is a channel where you do the legible, checkable work and accept that the feedback loop is slow and partial - which is genuinely uncomfortable if you are used to optimizing against a dashboard.
The consolation is that the work itself is not speculative. A complete, accurate, current product feed with honest descriptions was always the right thing to have. It made your Shopping ads cheaper, it made your product pages better, and it now decides whether a machine has ever heard of you. That is an unusually good deal for an afternoon's work.
If you would rather not spend that afternoon in a spreadsheet, this is the sort of thing we do for ecommerce clients every week. And if you would rather do it yourself, the list above is the whole job - we would just be doing the same seven things.
