What is agentic optimization?

Agentic optimization prepares a website for AI agents: software that takes a task from a user and carries it out on the web by itself, like finding a product, comparing specs, filling in a form, booking a table or getting as far as checkout. The work comes down to three questions: whether the site lets the agent in, whether the agent understands what is on the page, and whether it manages to click the right button and fill in the fields.

How agentic optimization works

An AI agent opens the site in a browser. At some companies that browser runs in the cloud, like ChatGPT Work’s cloud browser, and at others it runs inside the user’s own browser, like Anthropic’s Claude in Chrome or Microsoft’s Browse with Copilot in Edge. From there the agent reads the page and operates it like a person: clicking, typing, moving between pages.

To understand the page, the agent leans on three layers: screenshots of the page as displayed, the DOM, meaning the page’s HTML structure, and the accessibility tree, the version of the page the browser builds for screen readers, with the role and name of every element. Google writes that browser agents may use all three, and Anthropic and Microsoft document screenshots. OpenAI writes that ChatGPT Atlas uses ARIA tags, the same labels and roles screen readers rely on, to interpret page structure.

What agentic optimization covers today

There is no agreed list. These are the approaches discussed today for agents that take actions, and where each one stands:

• Semantic HTML and accessibility. Real elements like <button> and <a>, labels tied to their fields, and ARIA attributes that describe role, name and state. Google and OpenAI both recommend this in their documentation, and Google adds a stable layout with no transparent layers over elements.
• Not blocking agents at the security layer. Agents that run in the cloud look like bots to the site’s firewall (WAF), so they can get blocked. Some of them sign their requests with web-bot-auth, so a site can verify them and let them in without opening the door to scrapers that impersonate them.
• WebMCP. The site defines tools for the agent, for example a flight search with parameters, instead of the agent filling in the form field by field. WebMCP is still experimental, and WebKit, Apple’s browser engine, opposes it.
• Commerce protocols. OpenAI and Stripe’s Agentic Commerce Protocol, and Google’s Universal Commerce Protocol, which define how an agent completes a purchase. ACP is in beta, and UCP is in early access for select merchants.

How agentic optimization relates to SEO and GEO

GEO deals with whether AI engines find the site, understand it and cite it in an answer. Agentic optimization starts after that, when an agent is already on the page and trying to complete a task there. llms.txt and Markdown versions of pages belong on the GEO side: they serve content to bots that read the page, and OpenAI, Anthropic, Microsoft and Perplexity document no agent or crawler that reads them. A site can be cited in ChatGPT and still fail an agent trying to book on it.

An AI agent arrives at a site with a user who has already decided to act: to book, to buy, to leave their details. If a WAF rule blocks the agent at the door, or the form’s button is not recognized as a button, the task stops on that site at the last step, the one where the conversion was supposed to happen.