Your next user is an agent
In 2026, “user” stopped meaning “human”. Shopping assistants and browser agents now use your website on someone’s behalf. And what helps them is what you should have built anyway.
On this page
A parent asks an assistant to check the school’s events page against the family calendar. The page is lovely: a poster-style layout, dates set inside the artwork, a button that is only an icon. The assistant fails, and the model is not to blame: the page was built for eyes only.
That is the new situation. Agents already use the same websites and apps people do, on their behalf, and they have become a user type of their own. They read your page instead of looking at it. And what they need to read is, almost line for line, what accessibility has been asking for since 2008.
This piece covers what changed this year, where agents stumble, and what is now standard in my projects.
What changed in 2026
Three things happened within twelve months. Shopping assistants gained a way to buy inside the conversation: merchants publish a product feed and a checkout interface, and the purchase completes without the person ever opening the shop. Browsers gained a proposed standard that lets a site declare its actions to an agent instead of making it guess from pixels, now in an early trial. And the first controlled study put a number on design: the same shop, rebuilt for agents, raised strict task success from 49 to 89 percent and cut the steps per task by a third.
The numbers are early and the setup was friendly. But the direction is clear, and shoppers are following it. In a survey of 2,273 people this summer, a majority already preferred an assistant for comparing products and checking stock. For a damaged parcel, two thirds still want a human. The research moves to agents. The trust is still earned by the shop.
What an agent sees
Instead of your layout, an agent reads the accessibility tree: the names, roles and states of your elements, plus whatever text is actually in the page. Screenshots are the expensive fallback, slow and error-prone. To an agent, an icon-only button without a name is simply a button called nothing. A price rendered as an image is no price. A green dot that means “in stock” means nothing at all.
The fix is the fix a screen reader needed all along: a name on every control, semantic HTML, headings that say what the page is for, and the important facts as text. Accessibility became a legal requirement across Europe last year. This year it became a revenue question.
Trail Runner 42
A name, a price, a button the agent can call
Trail Runner 42
An image and a button called nothing
Facts, not decoration
Agents are asked to compare and to decide. They need the evidence a careful human would want: the price as a number, the stock as a word, the policy in plain text, and a date that says how fresh the information is. Structured data in the page lets an agent read facts directly. Timestamps let it judge whether to trust them.
One rule above all: show agents what you show people. The moment an assistant quotes a price the customer cannot find on the page, trust is gone, for the assistant and for the shop.
Trail Runner 42
Lightweight, waterproof, 280 g
Price as text, stock as words, a timestamp
Trail Runner 42
Lightweight, waterproof, 280 g
Price as pixels, stock as a colour
Decide whether you want them
Not every product should welcome agents. If your business lives on browsing and ads, if regulation demands a human in the loop, or if your prices are competitive secrets, limiting agent access can be the right call. But opting out has a price too: the competitor who lets the assistant finish the task will get the task.
What I would not do is build two websites. A separate feed for machines that drifts from the page people see is a trust problem waiting to happen. Build one honest site, make it readable, and expose actions where a defined task benefits from them.
Readable page: five steps, done
Eyes-only page: seven steps, then gave up
The agent-ready checklist
The same list, in my experience, lifts the accessibility score too:
- Every control has a name a machine can read.
- Prices, stock and dates are text, not pixels.
- Headings state what the page is for and what you can do here.
- Forms work with keyboard and labels, and errors stay visible.
- Volatile facts carry a timestamp.
- Agents and people see the same data.
- Once a quarter, run your three key tasks with a browser agent and log where it fails.
Design for the reader who cannot see, and you have designed for the agent. No coincidence: the web was built to work this way.
Sources
- Nielsen Norman Group: AI Agents as Users (2026)
- Baymard Institute: AI Engagement: Quantitative UX Insights (October 2026)
- arXiv: Designing Agent-Ready Websites for AI Web Agents (ICEME 2026, preprint)
- Chrome for Developers: WebMCP origin trial in Chrome 149 (June 2026)
- OpenAI: Agentic Commerce: product feeds, checkout and delegated payment