- AI agents search, compare and buy directly on websites on behalf of users. Agentic Engine Optimization makes sure AI agents do not fail on your website.
- Today almost all agents operate the page like a person in a browser. They mainly read the accessibility tree, so semantic HTML, labeled buttons and server-side content are what count.
- The stumbling blocks are well known and can usually be fixed without a relaunch. In one study an agent's success rate dropped from 78 to 42% as soon as it could only operate the page by keyboard.
A new type of visitor
Someone tells ChatGPT “Order me more printer paper”, and the agent does the rest. It finds the shop, puts the paper in the cart and goes through the checkout without a person touching the mouse.
Until now websites knew two kinds of visitors. People act, bots read. AI agents are a third kind: bots that act. For your website this raises two questions. Does the AI recommend you at all? That is GEO and the topic of the guide GEO vs. SEO. Can the agent operate your page once it is there? That is what this guide is about. The second question only pays off once the first is settled, because the agent goes where the AI sends it.
GEO, AEO, AIO sorted out briefly
| Term | Meaning |
|---|---|
| GEO, Generative Engine Optimization | Being mentioned and cited in AI answers |
| Agentic Engine Optimization | Building the website so that an agent can complete tasks on it. This is how we use the term here |
| AEO, Answer Engine Optimization | Google also uses this name for the field, in practice another word for GEO |
| AIO, LLMO | Further variants around being found |
Two ways an agent uses your page
- The agent operates the rendered page in the browser and clicks like a person.
- What gets optimized is operability, meaning the accessibility tree, semantic HTML and server-side content.
- Examples are AI browsers such as Perplexity Comet, extensions such as Claude in Chrome and agents in a cloud browser such as ChatGPT Agent.
- The agent calls a defined interface and needs no user interface.
- For commerce, protocols such as UCP from Google, ACP from OpenAI and AP2 handle search, cart and payment. Find out more in the guide Agentic commerce.
- Direct HTTP requests without a browser are reliable but still rarely used.
Which agents exist and how they differ is shown in the AI browser comparison.
How does an agent operate a page?
“Like a person” is only true of the result. The agent reads the page. It takes a screenshot and at the same time fetches the accessibility tree, the structure of roles, names and states that screen readers also use. From this it identifies the right element and triggers a click or an input.
It prefers the tree because it is unambiguous. Google writes that analyzing screenshots is slow and expensive. An agent reliably finds whatever is cleanly marked up in the structure. An element that only looks like a button remains a guessing game based on the screenshot.
The most common stumbling blocks
In April 2026 Google published recommendations for agent-friendly websites. Together with the patterns on which agents fail in practice, they result in this list. The guide Optimize your website for AI browsers shows step by step how to check and fix each point.
| Stumbling block | Consequence for the agent | Fix |
|---|---|---|
| Unstable layout | Elements jump while loading, the click misses | Consistent layouts, low CLS, fixed image dimensions |
| <div onclick> instead of <button> or <a> | The element barely shows up in the structure | Native elements, otherwise at least role and tabindex |
| Buttons and fields without labels | Icon buttons and fields without a label have no name | Descriptive labels, link fields via for/id |
| Transparent overlays | Covered elements are skipped | Avoid them or make them cleanly closable |
| Heavy client-side rendering | The page is empty until JavaScript runs | Deliver core content server-side |
| Hover- or mouse-only operation | Menus and drag and drop are unreachable | Make them operable by focus and click |
| Infinite scrolling without pages | The agent does not know when it is done | Offer pagination or a clear end |
| CAPTCHA on every action, hard bot blocks | Legitimate agents acting for users fail | Use them selectively at sensitive points |
Google names two further points. Clickable elements should be recognizable as clickable via cursor:pointer, and required elements must be visible and larger than 8 square pixels. This is what the difference looks like in the code.
One step further with WebMCP
Removing stumbling blocks means not standing in the agent's way. With WebMCP you actively help it. The page describes its important actions, such as search or cart, directly as a tool, and the agent does not have to guess what a button does. WebMCP is a draft and runs in Chrome from version 149 and in Edge as an origin trial. The guide WebMCP explains setup and measurement.
How big is the difference?
The A11y-CUA study, presented at CHI 2026, had a current computer-use agent complete 60 everyday tasks while changing the way it operated the page.
| Operation | Success rate |
|---|---|
| Standard, with mouse and image | 78.3% |
| Keyboard only, i.e. via the page structure | 41.7% |
| With screen magnifier | 28.3% |
The rate almost halved as soon as the agent had to rely on the structure. For websites this means that a clean structure decides a large part of the success.
Can you see agents in your data?
| Type of agent | Detectability |
|---|---|
| Extension in the user's own browser | Practically invisible. Same session, same IP, normal Chrome |
| Dedicated AI browser | A rough signal, the installation suggests agent use |
| Cloud agent | Most likely detectable via data center IPs and sometimes their own user agents |
Standard analytics and referrers hardly help. The referrer is missing, says “direct” or shows the AI platform. Log file analysis, behavior patterns, Web Bot Auth from the outside and WebMCP from the inside tell you more. Web Bot Auth is now a draft of an IETF working group with which bots identify themselves via signed HTTP requests.
Testing and getting started
- Have AI agents run through your most important journeys, meaning search, cart, booking and inquiry. Test several agents, because they behave differently.
- Build these tasks into a repeatable, automated test. That way you notice when a release breaks a journey for agents.
- Check the pages with Lighthouse in the “Agentic Browsing” category (from Chrome 150). It checks WebMCP, llms.txt, accessibility for agents and layout stability.
- In Chrome DevTools, look under Elements → Accessibility to see where names or roles are missing.
- Keep an eye on what your shop system and your payment provider offer for UCP and ACP.
- Start measuring. Log what can be logged and try out WebMCP.
Common mistakes
- Bot protection blocks all automation across the board and thereby also hits agents of real customers.
- Variant selection such as size or color is a custom widget without roles.
- The cookie banner sits as an overlay over the checkout.
- After a relaunch, labels that were there before are missing, and nobody checks.
Almost all of these fixes also help screen readers and search engines. You measure CLS anyway via the Core Web Vitals.
What has changed since?
- The Lighthouse category “Agentic Browsing” can be used since Chrome 150. In July 2026 it was only available in Chrome Canary. It is still considered experimental and, instead of a score from 0 to 100, gives the share of audits passed.
- OpenAI discontinued the AI browser ChatGPT Atlas on 9 August 2026. Its agent features now sit in the ChatGPT desktop app.
- According to Google, Gemini in Chrome with auto browse has also been coming to Android since the end of June 2026, so far only in the USA.
Frequently asked questions
What is Agentic Engine Optimization?
Agentic Engine Optimization prepares a website for AI agents completing tasks on it, such as searching, filling in a form or buying. The abbreviation AEO usually stands for Answer Engine Optimization, which is practically the same as GEO.
What is the difference between GEO and Agentic Engine Optimization?
GEO makes sure AI answers mention you. Agentic Engine Optimization makes sure an agent then reaches its goal on your page.
How do AI agents read a website?
Most take a screenshot and read the accessibility tree at the same time. From both they choose the right element and click or type.
Do I need a relaunch for AI agents?
Usually not. Labels, native buttons, stable layouts and server-side content can be retrofitted selectively.
How do I test whether agents can operate my page?
Have an AI agent run through your most important tasks, check the page with Lighthouse “Agentic Browsing” and look at the accessibility tree in DevTools.
Can I see agent traffic in Google Analytics?
Hardly. Extensions run in the user's browser and look like their normal visits. Log files and behavior patterns help more.
Sources
Retrieved on October 15, 2026. If something is out of date or missing, let us know.
- web.dev, Build agent-friendly websites: https://web.dev/articles/ai-agent-site-ux · As of 1 Apr 2026
- A11y-CUA Dataset, CHI 2026 (arXiv 2602.09310): https://arxiv.org/abs/2602.09310 · 10 Feb 2026
- Chrome for Developers, Lighthouse Agentic Browsing: https://developer.chrome.com/docs/lighthouse/agentic-browsing · As of May 5, 2026
- Chrome for Developers, WebMCP: https://developer.chrome.com/docs/ai/webmcp
- IETF, HTTP Message Signatures for Bots (Web Bot Auth): https://datatracker.ietf.org/doc/html/draft-ietf-webbotauth-httpsig-protocol · Draft, Sep 2026
- Google, Gemini in Chrome on Android: https://blog.google/products-and-platforms/products/chrome/bringing-chrome-ai-to-android/ · May 12, 2026
- OpenAI Help, Evolving Atlas into ChatGPT for browser-based agentic work: https://help.openai.com/en/articles/20001371-evolving-atlas-into-chatgpt-for-browser-based-agentic-work · Discontinued on 9 Aug 2026
- 9to5Mac, OpenAI is discontinuing ChatGPT Atlas, its standalone desktop browser: https://9to5mac.com/2026/07/09/openai-is-discontinuing-chatgpt-atlas-its-standalone-desktop-browser/ · Announcement, 9 Jul 2026