# AI agents > AI agents are AI systems that complete tasks independently on a person's behalf: they plan steps, read and operate websites, fill in forms and compare offers. Source: https://www.codaai.ai/en/knowledge/geo-glossary/ai-agents/ Published: 2026-10-05 May be quoted with attribution and a link to the source. --- AI agents are AI systems that complete tasks independently on a person's behalf. Google describes them as “autonomous systems that can perform tasks on behalf of people, such as booking a reservation or comparing product specifications” (Google Search Central, AI optimisation guide, July 2026). Unlike a chatbot, an agent does not just write an answer: it opens websites, reads them and operates them. ## Which AI agents use websites today? The major providers have released their own agents. The overview below only includes details from the vendors' own sources: | Agent | Provider | How it uses websites | |---|---|---| | ChatGPT agent | OpenAI, since July 2025 | Its own virtual computer with a visual and a text-based browser; asks for permission before consequential actions | | Claude in Chrome | Anthropic, pilot since August 2025 | Browser extension that sees pages, clicks buttons and fills in forms; on all paid plans since December 2025 | | Google Agent | Google, documented since April 2026 | Crawler used by AI agents running on Google's infrastructure | In addition, there are agents that run directly on a website and agents for the command line. Andre Bandarra (Chrome team, Google) names these three forms and stresses that the agentic web is not a separate web but an evolution of the existing one (Chrome for Developers, July 2026). ## How do AI agents work on a website? An agent does not see your page on a monitor but as a machine-readable representation. According to Google, there are three ways (web.dev, Kasper Kulikowski, April 2026): - **Screenshot:** a vision model identifies elements on the rendered page. This is slow and uses many tokens, so it serves as a fallback. - **HTML/DOM:** the agent reads nesting, attributes and text. This is how it knows that a “Buy” button belongs to a particular product. - **[Accessibility tree](https://www.codaai.ai/en/knowledge/geo-glossary/accessibility-tree/):** the browser-generated summary of the roles, names and states of interactive elements. Modern agents combine all three. The longer the task, the more context builds up — and, according to the Chrome team, the greater the risk that the model misreads a signal and takes a wrong turn (Chrome for Developers, July 2026). With [WebMCP](https://www.codaai.ai/en/knowledge/geo-glossary/webmcp/), a website can shorten this detour and offer actions directly as tools. ## How do AI agents differ from AI crawlers? [LLM crawlers](https://www.codaai.ai/en/knowledge/geo-glossary/llm-crawlers/) collect pages for training or a search index. They do not interact like users. Agents do exactly that and therefore need different conditions, Tory Gray explained in a Sitebulb webinar (June 2026). She also warns against confusing the two: an agent-ready website is not automatically visible in AI answers. ## Why do AI agents matter for AI visibility? Agents move part of the purchase decision into the machine. At SEO Week 2026, Crystal Carter described a new “validation layer”: the stage where agents assess whether a brand meets a user's preferences and requirements before surfacing it as a recommendation (iPullRank, June 2026). A site that offers missing details or an unusable interface at this stage drops out — even if it was found earlier in the [web search](https://www.codaai.ai/en/knowledge/geo-glossary/web-search/). In retail this is already concrete: agents compare prices, check availability and return conditions, and sometimes buy themselves ([agentic commerce](https://www.codaai.ai/en/knowledge/geo-glossary/agentic-commerce/)). ## What does this mean for your website? A website that works well for people also works for agents — that is the Chrome team's core message. In practice: - **Content in the HTML:** important text, prices and specifications are in the delivered HTML, not loaded later by JavaScript. - **Semantic controls:** `