We have spent twenty years optimising for one thing: ranking high in Google search results. The right keywords, the right meta tags, the right link structure. Entire industries have been built on that knowledge.
But the playing field is changing. Not slowly, and not voluntarily.
Twenty years of ten blue links
Put the number in perspective: e-commerce accounts for roughly 20 percent of total retail sales today. That is a sobering figure for anyone who has spent their career building digital shopping experiences. All those investments, all those platform migrations, all those SEO campaigns — and still eight out of ten purchases happen offline.
But that is also the wrong way to read the number. E-commerce transformed all of retail, not just its own slice of the pie. Omnichannel strategy, inventory optimisation, return systems, customer data — all of that is e-commerce's legacy. Technology shifts do not bury channels. They add new ones. E-commerce did not kill physical retail. It forced physical retail to get better.
That is important context for what is happening now.
The bike tyre that never got bought
An analyst at Criteo described a test he ran recently. He lives in New York, cycles daily on asphalt that varies between bad and catastrophic, and needed new tyres. Instead of searching Google, he went to an AI agent and spent an hour specifying his needs: puncture-resistant, correct dimensions, durable but not too heavy.
The agent was excellent at that part. It asked the right questions, understood the problem, and presented options.
Then he wanted to buy.
Broken links. Products no longer sold. Items with no stock information. Prices that did not match. He closed the window and went to his local bike shop, possibly slightly more informed than usual, and bought from someone he trusts.
That is a near-perfect illustration of where AI agents stand today. Phase one: research and discovery. Exceptionally good at understanding a problem and mapping possibilities. But when it comes to actually recommending and completing a purchase, it falls short. Not because the agent is unintelligent, but because it lacks the right data.
An LLM is fundamentally a well-trained conversationalist. It knows a great deal about bike tyres in general. But it does not know which of the hundreds of items in your catalogue is in stock right now, which has a high return rate, or which was discontinued three months ago. That requires a different kind of data: transaction data, real-time feeds, normalised product information.
A sommelier can talk about terroir and vintages all day. But when you want to order, they need to look at the wine list. And that list needs to be accurate.
That gap is what three abbreviations are trying to close.
MCP, UCP and AEO
MCP, the Model Context Protocol, is the layer that lets AI agents communicate with external tools and systems. Imagine Claude speaks English, your ERP speaks another language, and Gmail speaks a third. MCP is the interpreter in the middle. Without it, the agent is isolated, however capable it may be. It is Anthropic's protocol and has become the de facto standard for how agents connect to tools.
UCP, the Universal Commerce Protocol, is the latest addition. Google launched it at the National Retail Federation conference, with Sundar Pichai appearing in person to present it. It is an open standard, not a Google product in the traditional sense — that distinction matters. The protocol was co-developed with Shopify, Etsy, Target, Walmart, and Wayfair. The documentation is already live at ucp.dev.
A useful analogy: imagine you own a restaurant and every food delivery app requires its own registration, its own menu, its own updates. UCP is the idea of registering once, in a format all apps can read. You expose what you sell and how you sell it. The protocol handles the rest. An AI agent working in Google's AI Mode or in Gemini can find your products, surface them in the right context, and — next step — help complete the purchase including payment and delivery details, without the customer filling in the same form for the fourteenth time.
AEO, Answer Engine Optimisation, is the consequence of all this for anyone selling online. SEO was about ranking high on a list. AEO is about being the answer.
Not link number three. The recommendation the agent gives when someone asks: which bike tyre should I buy?
That is a fundamental difference. You are no longer optimising for a crawler that indexes text. You are optimising for an agent that needs machine-readable, up-to-date, and reliable product data. The sign above your shop needs to be written in a language the GPS understands. Today, most signs are still written for humans.
Why it is not optional
This is a pattern we have seen before.
When mobile browsing took off, many said it could not be prioritised. The design process was complex, customers still bought mostly on desktop, and responsive design required investment that was hard to justify. Then Google announced that mobile optimisation affects ranking. Overnight, the discussion was over.
The same happened with HTTPS. With structured data. With Core Web Vitals.
That same pattern is now playing out around agentic commerce.
Google AI Mode and Gemini have hundreds of millions of active users. OpenAI launched Instant Checkout with Shopify, Walmart, Etsy, and DoorDash. Google's version carries the same roster of co-signatories. This is no longer an experiment.
Anyone who remembers Meta pulling back its checkout feature in 2022, or Google deprecating "Buy on Google" in 2023, has reason to raise an eyebrow. Those attempts were premature. The infrastructure was not there. But that is exactly what Bill Gates said: we overestimate change in the short term and underestimate it in the long term. Fully autonomous agents shopping without human approval are probably overstated in the near term. But the protocols that determine who is visible in the next version of search are being written right now.
That is a better position to be in than when the change is already done.
If your product catalogue is not clean, real-time updated, and structured to the new standards, the agent skips you. Not out of preference, but because it has nothing to work with. That is exactly what the bike tyre buyer encountered: not bad intent, just bad data.
What this means in practice
Starting with AEO does not require a complete overhaul.
Product data needs to be machine-readable and current. Not because it looks good, but because an agent cannot recommend a product that is out of stock, has a broken link, or lacks basic specifications. The structure needs to follow open standards. UCP builds on principles similar to schema.org and the product feeds you are probably already using for Google Shopping. This is not starting from scratch.
And the answer to "are we relevant in this context?" needs to be in the product content, not just the keywords. The agent reads context, not tags.
What separates the visible from the invisible in this transition is not budget. It is priority.
Your e-commerce operation is most likely optimised for a Google that existed two years ago. Three abbreviations are changing that. Most people do not know what they mean yet. That is a head start worth taking.