On 22 June, Google released Gemini 2.5 Pro with Deep Think, by several accounts the most capable public model any lab has shipped, with a two-million-token context window. The benchmark isn't the story for a small online retailer. The story is what these models are increasingly used for: doing the shopping. As buyers start asking an AI to “find me a pair of gold hoop earrings under £200 that ship to the UK,” the question stops being “where do I rank on Google” and becomes “can the agent actually read and understand my products.” For most small shops the answer right now is “not really,” and that's the gap worth closing this quarter.
Every time a lab ships a more capable model, the coverage fixates on the benchmark and misses the consequence. This week it was Google's turn: Gemini 2.5 Pro with Deep Think landed on 22 June, with a two-million-token context window, double what anyone else offers, and enough reasoning muscle that it's being called the most capable public model going (the Google DeepMind blog has the headline numbers). Impressive. But if you run a small online shop, the spec sheet isn't your business. What these models get used for is your business, and increasingly that's acting on the shopper's behalf: comparing, shortlisting, recommending, and in some cases buying.
That shift changes the question you should be asking about your own store. For fifteen years the question was “how do I rank higher when someone searches.” The emerging question is different and quieter: “when an AI agent is doing the looking, can it understand what I sell well enough to put me on the shortlist?” Those are not the same skill, and the businesses that notice early get an unfair head start.
What “the agent does the shopping” actually looks like
Picture a customer who wants a gift. A year ago they'd open three tabs, search, scroll, and squint at photos. Increasingly they type a sentence to an assistant instead, “something handmade, silver, under £150, ships in time for Saturday,” and let it come back with a handful of options. The agent doesn't browse the way a person does. It reads structured information: the title, the description, the price, the materials, the stock status, the delivery promise. If that information is clear, machine-readable and honest, your product is a candidate. If it's a beautiful photo with a three-word title and the real detail living only in your head or a PDF, you're invisible to the thing now making the shortlist.
I worked with a jewellery e-commerce business on exactly this. Gorgeous photography, real craftsmanship, and product pages that said almost nothing a machine could use, “Aurora Hoops, £165” and a stunning image. To a human browsing, lovely. To an agent asked for “gold-plated hoops under £200 for sensitive ears,” silent, because the metal, the plating, the hypoallergenic detail and the dimensions weren't written down anywhere a model could read. The photos carried the information; the data didn't. We didn't change a single product. We changed what the page said about each product, in plain structured terms. That's the whole move.
Why this favours the small shop, if you move
It's tempting to assume the big retailers win this by default. They have the budgets, the SEO teams, the catalogues of millions. But agentic recommendation rewards clarity, not size. A focused shop with two hundred products, each described precisely and honestly, is easier for a model to understand and recommend than a sprawling marketplace of vague, duplicated listings. The advantage isn't your scale; it's that you actually know your products and can describe them properly. The constraint has always been that doing so by hand is mind-numbing. That constraint is exactly what's now cheap to remove.
This is where AI helps you defend against AI. Describing two hundred products in rich, consistent, structured language used to be a fortnight of soul-destroying copywriting. Now it's a job you can largely hand to a model: feed it your photos, your existing notes and your specs, and have it draft consistent structured descriptions for every line, which a human then checks for accuracy and voice. The bottleneck was never knowing what to write. It was the tedium of writing it for the hundredth SKU. Remove the tedium and the moat is suddenly buildable in an afternoon or two rather than a quarter.
The three things I'd fix first
- Write the details a person can see but a machine can't. Materials, dimensions, fit, what's in the box, who it's for. Every fact currently living only in your photographs or your head needs to exist as text on the page.
- Add structured product data. The behind-the-scenes markup that tells any system, in a standard format, “this is a product, here's its price, here's its availability.” Google's product structured data guide and the underlying Schema.org Product spec are the canonical references. Most shop platforms support it; many small shops have it switched off or half-done. This is what lets an agent trust and use your information.
- Make stock and delivery truthful and current. An agent that recommends something out of stock, or wrong on shipping, won't recommend you twice. Accuracy is now a ranking signal, not just good manners.
What I'd actually do this week
Pick your ten best-selling products and read their pages as if you were a machine that can't see the photos. Strip the images away in your head. Is the metal there? The size? The materials? Who it's for? If the answer is mostly “it's in the picture,” you've found your project, and it's a contained, cheap one, not a rebuild.
The frontier-model news will keep coming every few weeks, and most of it won't touch your shop. This thread does, because it changes who does the looking. When the shopper hands the search to an agent, the win goes to whoever wrote their catalogue down clearly. That's not a budget you can't match. It's an afternoon most of your competitors haven't spent yet.
Frequently asked questions
What is “agentic shopping” in plain English?
It's when a buyer asks an AI assistant to do the looking for them. Instead of typing into Google and scrolling through ten blue links, they tell an agent what they want in a sentence, and the agent returns a shortlist. Increasingly the agent will also handle checkout. For retailers, the practical effect is that an AI is now the first thing reading your product page.
What is product structured data, and why does an agent need it?
Structured data is information on your product page written in a standardised, machine-readable format (typically JSON-LD using the Schema.org Product type). It tells any system unambiguously “this is a product called X, it costs £Y, it's in stock, here are its dimensions.” AI agents rely on it to compare like with like across hundreds of shops in milliseconds. No structured data means the agent has to guess, and guessing usually means skipping you.
How do I tell whether my own catalogue is agent-readable?
The fastest test: open one of your product pages and turn off images in your browser. If the page is still informative — title, materials, dimensions, audience, key specs — you're close. If it's a blank space with a price, the photos are doing all the work and an agent can't see them. A second test: Google's Rich Results Test will tell you whether your structured data is present and valid.
Will AI agents really replace Google Search for shopping?
Probably not all at once. The more useful framing is that they add a layer on top: a chunk of buyers who used to start at Google now start at ChatGPT, Gemini or Claude. You don't need to stop ranking for the old game. You need your catalogue to be legible to both audiences (humans browsing search results and agents synthesising recommendations), and the same structured fixes serve both.
How long does it take to fix this for a 200-product shop?
With AI help, days, not months. You feed your existing photos, notes and specs to a model and get a consistent draft description per product, which a human reviewer then corrects for accuracy and voice. The unglamorous bit is the human pass, and it's fast once the drafts exist. Without AI, the same job used to be a fortnight of soul-destroying copywriting, which is why most shops never finished it.
Do I need to add structured data manually, or does my e-commerce platform handle it?
Most modern platforms (Shopify, WooCommerce, BigCommerce and others) generate product structured data automatically when the right fields are filled in. The common failure is that fields like material, dimensions, or availability are left blank, so the generated markup is technically valid but practically useless. Filling the gaps in your existing platform usually beats bolting on a new tool.