A few weeks ago, we watched a demonstration of OpenAI's Operator. At first glance, it looked unimpressive: a small cursor moving across a grocery site, adding items to a cart and stopping to ask the user before paying. It was slow and somewhat clumsy, much like someone learning to use a tablet for the first time.
On reflection, however, the significance is clear. The clumsiness will be corrected. What remains is that a piece of software now reads a store's pages and decides what to buy.
OpenAI introduced Operator in January, for Pro subscribers in the US, and in late February it became available in Canada and a number of other countries. It is a research preview on a plan that costs US$200 a month, so the average customer is not using it yet.
What do AI shopping agents actually do today?
Operator is not the only one. Perplexity added a "Buy with Pro" option for its US subscribers in November, and Shopify stores appear in its product results. Google presented Project Mariner in December, a research prototype that works inside Chrome. Anthropic released "computer use" for Claude in October, based on the same idea and aimed at developers.
They all work in roughly the same way. The user states what they want in plain language. The agent searches, opens pages, reads them, compares and fills in forms. Some return control to the user at the payment step, which is a sound practice. Others get stuck on a pop-up and stop progressing.
In our testing, these tools are impressive and frustrating in roughly equal measure. They get lost in cookie banners. They miss size selectors hidden in unusual menus. They cannot find a price that only appears after an option is selected. These are precisely the same obstacles that cause real customers to abandon a purchase.
Why does product data matter more than your homepage?
An agent pays no attention to your hero image or your choice of font. It reads text, page structure and whatever data your site exposes. The stores that will perform well with these tools are therefore those with plain, clean product information.
This begins with a product title that states what the item is. "The Classic" tells a machine nothing. "Merino wool socks, men's, size 9 to 11" tells it everything it needs.
It also means size, colour, material and weight written out as text, not embedded in a photo. And it means structured data on the page, the schema.org Product markup Google already uses for rich results, so that price, availability and reviews can be read without guesswork. If you sell through Google Merchant Center, the same rigour applies to your product feed.
None of this is new. It is the same work that has supported search for years, and it forms a large part of what is now called AI search optimization. This is one reason we see no cause for alarm. If your product pages were already in good shape for Google, most of the work is done.
If they were not, there are now two reasons to improve them.
Why do clear prices, stock and store policies matter?
The other common obstacle for agents is imprecise information.
"Contact us for pricing." "Ships in 3 to 10 business days, usually." A return policy placed in a PDF that no one opens. An agent asked to find a gift under $50 that arrives before Friday cannot work with any of this, so it moves on to a store that states the information plainly. A human customer behaves the same way, only more slowly and with more frustration.
We recommend the following short list to any small store:
- Price visible on the product page, in the currency you charge, with taxes explained in an obvious location.
- Stock status that is accurate. If an item is out of stock, say so.
- Shipping costs and delivery times written out, by region if possible.
- Returns and exchanges in plain text, on a page with a normal URL.
- A checkout that does not require customers to create an account.
Each of these measures helps a human customer today. That is the test we recommend applying to anything presented to you as "AI-ready": if it only helps the machine, it can wait.
Should you block AI bots or let them in?
Some store owners are asking whether they should block all of this activity. The concern is understandable. A great deal of scraping is taking place, and some AI crawlers access small sites intensively enough to slow them down.
However, a crawler collecting training data and an agent making a purchase for a customer are fundamentally different. Blocking the second is comparable to refusing to hand over an order because the customer sent her brother to pick it up.
At present, there is no reliable way to distinguish them in every case, and the way agents should identify themselves is still being defined. Our recommendation for a small store is to avoid unnecessary barriers. Keep a CAPTCHA at checkout if fraud is a genuine problem for you, but do not place one on your product pages.
Should you rebuild your online store for this trend now?
Our experience in web work spans several moments that were presented as turning points. Some genuinely transformed the field, as mobile did. Others faded, or took ten years instead of two.
Agents shopping on behalf of the general public will likely take a few more years to become common. This requires trust, payment systems that work without human supervision, and tools that no longer stumble over pop-ups. The trend appears likely to materialize; its timing remains uncertain.
For this reason, we do not recommend investing in an "AI commerce platform" this spring. Any vendor offering one is ahead of your customers, and you would be financing their speculation.
What we do recommend are the low-cost measures. Clean up product titles. Write the specifications out in full. Add structured data if your platform does not already do so (many Shopify and WooCommerce themes handle part of it, so verify before paying anyone). Place your policies on proper pages. If your store is old enough that none of this is straightforward, that is a sign it needed a website redesign regardless of agents.
The stores that will work best for an automated agent are the same ones that were already easy for people to use: clear, transparent and quick to load. That was sound advice before Operator appeared, and we expect it to remain sound advice as the next developments arrive.