# What Is an Agentic Browser and What Does It Do in Your Store? > What an agentic browser is, from Gemini in Chrome to ChatGPT Work, how it moves through a store, how it identifies itself and what to check so it can buy. URL: https://clione.ai/glossary/agentic-browser/ Lang: en An agentic browser is not a browser with a chat in the sidebar that summarizes the page for you. It is one where an [AI agent](/glossary/ai-agent/) takes the controls: **it opens your store, picks a size, applies a discount code and gets as far as the cart** because someone asked it to. For a merchant, that changes who walks the [product page](/glossary/product-page/). It is no longer just a person reading or a [crawler](/glossary/web-crawler/) downloading the HTML: it is a program that uses the page the way a customer would, and it gets stuck where a hurried, impatient shopper would get stuck. ## How it works The agent looks at the page, decides the next step and carries it out in the browser: a click, a field it fills in, a new tab. Vendors describe their products with tasks like these. What changes from one to the next is where the browser runs and how much it is allowed to do. **Google, inside Chrome.** On January 28, 2026, Google introduced auto browse, a [Gemini](/glossary/gemini/) feature in Chrome that handles multistep tasks on the user's behalf. Its examples include comparing hotels and flights, filling in forms and, starting from a photo, **finding similar items, adding them to the cart without going over budget and applying discount codes**. With permission, it uses Google Password Manager to sign in. It is in preview for Google AI Pro and AI Ultra subscribers in the United States, and Google announced that Chrome will support the [UCP (Universal Commerce Protocol)](/glossary/ucp/).[^google-auto-browse] On August 18, 2026 it reached Android, again for those subscribers in the United States.[^google-android-auto-browse] **OpenAI, first in its own browser and then in ChatGPT.** OpenAI launched ChatGPT Atlas on October 21, 2025, a browser with ChatGPT built in and an agent mode in preview that could, for example, look up a grocery store, add a recipe's ingredients to the cart and place the order.[^openai-atlas] On July 9, 2026 it announced the browser's retirement, **with the shutdown set for August 9, 2026**: the agent features moved to ChatGPT and Codex, and for longer browser work it pointed people to the desktop app.[^openai-release-notes] Since August 25, 2026, the ChatGPT Work browser can complete tasks on sites that require sign-in, and it asks for confirmation before actions such as a booking or a payment.[^openai-release-notes] **Where it runs matters to your server.** An agent can work in the user's own browser or in a cloud browser run by the provider. The ChatGPT Work cloud browser **signs every request** with the HTTP Message Signatures standard (RFC 9421), with a `Signature-Agent` header set to `"https://chatgpt.com"`, and OpenAI publishes the keys to verify it. Cloudflare recognizes it as a signed agent in its bot directory.[^openai-cloud-browser] That is the [Web Bot Auth](/glossary/web-bot-auth/) method, and it is what lets you tell a legitimate agent from one that only copies its name. **Pages are starting to talk to agents.** Today an agent acts on the interface, pressing what it sees. In February 2026 Chrome proposed WebMCP, a standard way for a website to expose structured tools: a declarative API over HTML forms and an imperative one with JavaScript. Its commerce example is an agent that finds the products, configures the options and walks through the checkout with precision. It is in preview for enrolled participants.[^chrome-webmcp] Since August 31, 2026, ChatGPT Work and Codex use those tools in the desktop app's built-in browser, on compatible sites.[^openai-release-notes] ## In a store A request like "buy me trail running shoes in size 10 for under $130" turns into steps on your site. It is a **made-up example**, to see what the agent needs at each one: | Agent step | What it has to find in your store | | --------------------------------------------------- | ----------------------------------------------------------------------------------------------------------- | | Reaches the [product page](/glossary/product-page/) | Name, [price and availability](/glossary/price-and-availability/) written on the page, not only in an image | | Picks the size | A variant selector with sizes as text and the sold-out ones marked | | Checks the total | Shipping costs and delivery time visible before payment | | Applies a discount | A code field that works without hidden steps | | Reaches payment | A checkout that does not block it for looking automated | At the last step the vendors stop: auto browse is designed to ask for confirmation before completing a purchase,[^google-auto-browse] and ChatGPT Work asks before consequential actions such as a payment.[^openai-release-notes] The agent prepares the purchase; **the person confirms it**. There is a second reason for that pause: OpenAI acknowledges that an agent can fall for malicious instructions hidden in a page or an email, meant to make it do something the user did not ask for, and that its defenses will not stop every attack.[^openai-atlas] Google says it trains its models to detect known threats such as this prompt injection.[^google-android-auto-browse] When the agent can buy without leaving the conversation, the term is [agentic checkout](/glossary/agentic-checkout/). Believing that an agentic browser is just another crawler and is controlled the same way. **A crawler walks the web on its own; an agentic browser goes where a person sends it.** The rules in [`robots.txt`](/glossary/robots-txt/) were designed for the first: OpenAI, for example, warns that `robots.txt` rules may not apply to visits a user initiates from ChatGPT.[^openai-bots] And blocking an agent of this kind is not stopping a training bot: it is turning away a customer who has sent someone to buy. ## How to check it 1. **Make the request yourself.** Ask an agent to find a product in your store, pick a variant and get to the cart, and stop it before payment. Write down where it hesitates or gets it wrong. 2. **Review your bot protection.** If you use a CDN or a firewall, see how it treats signed agents: the ChatGPT Work browser can be verified by its signature, and Cloudflare, Akamai and HUMAN recognize it.[^openai-cloud-browser] 3. **Look for the signature in your logs.** Requests from the ChatGPT Work cloud browser carry the `Signature`, `Signature-Input` and `Signature-Agent` headers.[^openai-cloud-browser] 4. **Read your product page without images.** If the price, the sizes or the shipping costs only make sense by looking at a photo, an agent has a harder time than a person. 5. **Follow WebMCP if you develop your store.** It is a proposal in preview: worth trying, not worth depending on yet.[^chrome-webmcp] ## Related concept: AI agent The AI agent is the one that decides; the agentic browser is one of the hands it acts with. The same agent can work with a browser, call the tools a site exposes or talk to a store through a commerce protocol, without opening a single page. For merchants, **the agentic browser is the path that requires no integration**: it uses the store as it is, with its strengths and its snags. Protocols like UCP and the [product feed](/glossary/product-feed/) are the structured path, and [agentic commerce](/glossary/agentic-commerce/) brings them all together. To go further, the [agentic commerce reading path](/learn/agentic-commerce/) orders the rest of the terms, and the [ChatGPT product feed](/glossary/chatgpt-product-feed/) explains OpenAI's structured path.