Reading time: 11 minutes
SEO for AI agents means preparing your site so an agent, like ChatGPT agent mode, can find, read and act on your pages. In practice, your site must allow access, show text and prices in the HTML, and name its buttons clearly. This guide shows what to adjust and how to measure it.
In short
- ChatGPT agent mode is gone. Now ChatGPT Work handles web tasks with a cloud browser that reads, clicks and fills in forms.
- Your own site can block the agent, because a CDN and a firewall restrict automated browsing. For that reason, OpenAI publishes how to allow its browser.
- In SEO for AI agents, price, stock, shipping and contact details sit in the page text, and buttons and fields have clear names.
- In a PYMNTS survey with Visa, only 15% of 584 US merchants had structured data that an AI agent can read.
What is ChatGPT agent mode today?
ChatGPT agent mode was the feature that let ChatGPT open sites, click and fill in fields to finish a task. It is gone, and ChatGPT Work took over its role. So for a site owner, the visitor is now the ChatGPT Work cloud browser.
According to the OpenAI Help Center, ChatGPT agent is no longer available, and the advice is to use ChatGPT Work. According to the ChatGPT release notes, Work arrived in July 2026.
The cloud browser is ChatGPT’s own browser, running on a remote computer. It is available on paid plans, except Free and Go. It reads pages, clicks buttons, fills in forms and works on public sites and sites with a login. OpenAI’s examples include asking companies for quotes and comparing product availability and local stock. How your SEO team can use these agents in its own work is covered in the guide on AI agents for SEO. Also, the agent’s full journey, from research to payment, has its own guide.
Which other AI agents visit websites?
The other big assistants have agents too, and each one reaches your site in a different way.
- ChatGPT Work (OpenAI): runs in a cloud browser and signs its requests with the Web Bot Auth protocol.
- Google-Agent (Google): according to the Google fetchers list, it is the user agent of agents hosted at Google that act on a user’s request.
- Claude in Chrome (Anthropic): this extension clicks, types and fills in forms in the person’s own Chrome.
- Browse with Copilot (Microsoft): it acts in the person’s Edge, with actions running in the browser itself, but for consumers only in the US for now.
The last two act inside the person’s browser, and their help pages name no user agent of their own.

How do AI agents browse and pick a store or supplier?
An AI agent opens pages, reads the text and follows buttons to the step you asked for, such as a quote or the cart. For example, when the task is to compare, it gathers price, stock and terms from each option.
According to OpenAI, however, when a connected app can do the task directly, ChatGPT may use the app instead of the browser. When a custom app is worth it is covered in the guide on apps in ChatGPT.
For purchases, ChatGPT considers structured metadata from merchants and third parties, such as price and description, according to OpenAI’s help. A merchant can also send a product feed, but only as an approved partner, per the OpenAI commerce documentation. Feeds and checkout are covered in the guide to GEO for ecommerce. What already works for stores is covered in the guide on how to sell on ChatGPT.
In the US, this preparation is still rare. According to PYMNTS Intelligence, with Visa, only 15% of 584 US merchants had structured data that an AI agent can read. Even so, 60% expect agents to drive at least 5% of digital sales within two years. The survey was fielded in March 2026. Visa sells digital payments, so it has a stake in the topic.
After all, SEO for AI agents does not replace the earlier step, which is being found. How ChatGPT chooses the sites it cites is covered in the guide on how to show up in ChatGPT. The basics of the topic are in the guide on what GEO is.
What blocks AI agents on a site: login, CAPTCHA and JavaScript?
An AI agent can stop at three points: your site’s security, a login, and steps it cannot finish alone. OpenAI warns that a site can block ChatGPT even when it works in the person’s own browser.
According to OpenAI, the site sets these restrictions, and each site decides whether to accept cloud browser traffic. To allow it, OpenAI publishes allowlisting instructions for Akamai, Cloudflare, HUMAN, Vercel and other CDNs. The browser signs its requests with Web Bot Auth, meaning HTTP signatures (RFC 9421) that your site can verify. So you can allow ChatGPT’s agent without opening the door to any bot that fakes its name.
At a login, ChatGPT’s agent pauses and asks the person to sign in through a secure form. When it cannot move forward, it asks the person to take over the browser. When a site is blocked, OpenAI suggests trying another site or opening the page in the person’s own browser. In our view, an anti-bot check in the middle of the funnel, like a CAPTCHA, can send the task to a competitor. OpenAI does not mention CAPTCHA on these pages.
Do robots.txt and JavaScript block the agent?
The robots.txt file matters less for the agent than for search. According to OpenAI, ChatGPT-User acts on a user’s request, so robots.txt rules may not apply. Likewise, both companies say Perplexity-User and Google’s user-triggered fetchers generally ignore robots.txt. GPTBot, in turn, collects content that may be used to train the models. Sites that block OAI-SearchBot stay out of ChatGPT’s search answers, though they can appear as links. The robot-by-robot choice is in the guide on robots.txt for AI.
JavaScript, on the other hand, matters more for the search bots that bring your site into answers. The cloud browser is a browser, so it tends to open the page as it looks on screen. By contrast, ChatGPT’s and Perplexity’s bots did not run JavaScript in the 2024 study by Vercel with MERJ. Vercel, by the way, sells site hosting. How to test what each bot receives is in the checklist on JavaScript SEO for AI crawlers.
Forms and buttons the agent can use
According to Anthropic, Claude in Chrome clicks and fills in forms the way a person would. But the agent must know what each field and button does.
Success criterion 4.1.2 of the W3C accessibility guidelines asks for exactly that. Every interface component, including those built by script, needs a name and role that programs can identify, as assistive technologies do. The criterion does not mention AI agents. Still, in SEO for AI agents, the logic holds for anything that reads the page through code.
- Labeled fields: link a
<label>to each field instead of using only placeholder text. - Buttons with text: write “Add to cart” on the button, and do not leave just an icon.
- Native elements: prefer
<button>,<a>and<select>over blocks that only respond to a click through script. - Errors in text: say which field failed and why.
Which structured data and pages help AI agents?
The pages that help are those that answer, in text, what the agent will compare: price, stock, variants, shipping, delivery time, returns and contact. Product structured data repeats these values in a machine format, and the feed sends the catalog straight to the platform. But all three must say the same thing.
The consistency rule comes from Google. According to its structured data guidelines, you should not mark up content that readers cannot see. If the JSON-LD describes something, the HTML must describe the same thing. In other words, the price in the product schema must be the price on the screen.
For SEO for AI agents, review these pages first:
- Product or service: price, variants, availability and delivery time written on the page.
- Shipping and returns: rules on their own page, linked in the footer and on the product page.
- Contact and quotes: a short form with labeled fields, plus a phone number or email in text.
Example: what an AI robot finds on the NotHuman blog
At NotHuman, the blog runs on WordPress. We made requests without JavaScript, as an AI robot does, and the text of each article already arrived in the server’s first response. As for structured data, Yoast generates BlogPosting and BreadcrumbList, plus FAQPage on articles with frequently asked questions. On access, robots.txt does not block GPTBot, Google-Extended or CCBot, by a GEO decision.
How do you measure visits and purchases by AI agents?
Measure in three places: server and CDN logs, GA4, and orders. Logs show who requested each page, and GA4 shows visits coming from assistants. None of the three separates every AI agent, because some act in the person’s own browser.
In your logs, look for the documented user agents: ChatGPT-User, Perplexity-User, Claude-User and Google-Agent. ChatGPT’s cloud browser, by contrast, identifies itself by its signature. So look for that count in a CDN that verifies Web Bot Auth. Anyone can fake a user agent, so also check the IP, since OpenAI, Perplexity and Google publish their agents’ ranges.
Visits from people who click a link in ChatGPT or Gemini, on the other hand, show up in GA4. The step-by-step, with the assistant channel and the regex, is in the guide on AI traffic in GA4.
For purchases, the question is where the order came from. According to the same PYMNTS survey, only 24% of US merchants can tell when a sale came from an AI platform. So match orders against the session source and against agent hits in your logs.
SEO for AI agents checklist
In short, SEO for AI agents fits in seven steps, from access to measurement. Start with the first two, because without access the agent never reaches your pages. Then review the content and forms on your best-selling pages and, finally, set up measurement.
- Test access: open key pages with each bot’s user agent and check that your CDN allows ChatGPT’s cloud browser.
- Review robots.txt: allow OAI-SearchBot, since it brings your site into ChatGPT’s search answers.
- Put the essentials in the HTML: text, price, stock and links must arrive in the server’s first response.
- Name buttons and fields: follow W3C criterion 4.1.2 on the quote, cart and contact forms.
- Align schema, feed and page: use the same price and availability in all three.
- Clear the path: keep logins and anti-bot checks only where security needs them.
- Measure monthly: track user agents, signatures and the AI channel in GA4.
Frequently asked questions
Yes. According to the OpenAI Help Center, ChatGPT agent mode is no longer available. Web tasks moved to ChatGPT Work. Its cloud browser, on paid plans except Free and Go, reads pages, clicks and fills in forms.
No. According to OpenAI, GPTBot collects content that may be used to train the models. ChatGPT’s agent, by contrast, uses the cloud browser, which signs its requests with Web Bot Auth. So its access depends on your site’s security rules, such as the CDN and firewall.
Not without approval. According to OpenAI, ChatGPT asks for confirmation before actions that are hard to reverse, such as making a payment. Also, the person sometimes has to finish the final step. So your checkout must work well for both.
Main sources
- OpenAI, Using cloud browser in ChatGPT (2026).
- OpenAI, ChatGPT Work’s Cloud browser allowlisting with Web Bot Auth (2026).
- OpenAI, Overview of OpenAI Crawlers (2026).
- Google Search Central, user-triggered fetchers and Google-Agent (2026).
- Anthropic, Claude in Chrome (2026).
- W3C, Understanding Success Criterion 4.1.2: Name, Role, Value (WCAG 2.2).
- Google Search Central, general structured data guidelines (2026).
- PYMNTS Intelligence, Global Digital Shopping Index: U.S. Playbook, with Visa Acceptance Solutions (2026).



