AI browsers and SEO: Comet, Atlas and agent modes (2026)

AI browsers change SEO by letting an assistant read, compare and act on websites for a person. Being found still matters. Your site also needs to explain its products accurately and let that assistant finish a useful task without getting stuck or inventing missing details.

There’s an important update. OpenAI has retired Atlas. Its ChatGPT agent notice directs former users to Work and its cloud browser.

I’m Brenden House, founder of Searchmaxxed. We do SEO and AI search optimisation, including the pages, technical fixes and measurement behind getting found and turning that attention into enquiries.

In summary

• Keep investing in search. AI browsers add page reading and task completion to the discovery work SEO already supports.
• Check current vendor documentation. Atlas has retired; Comet and ChatGPT’s cloud browser work differently.
• Fix the buying pages. Specifications, integration details and usable forms matter more than an article announcing that your company is ready for AI.
• Separate citations from visits. Neither a crawler request nor an automated form submission proves that a qualified prospect has arrived.
• Measure the business outcome. Record which enquiries reach sales, alongside the sources and browser tasks you can actually observe.

What do AI browsers and agent modes actually do?

An AI browser combines ordinary browsing with an assistant that can interpret pages and, with permission, perform actions. An agent mode delegates a task.

The distinction matters. A browser can summarise an open tab without finding your company through search, and a remote agent can visit without using the customer’s browser.

This isn’t a browser ranking. It’s a distinction between the environments your prospective customers could use.

What changed for ChatGPT Atlas SEO?

Atlas SEO is now a historical label. OpenAI’s original Atlas announcement remains online, but its retirement notice takes precedence over the launch instructions underneath it.

The old ChatGPT agent help page also says that product is no longer available. It points to Work and cloud browser.

Avoid building around old menus. Your useful work survives these changes: clear product pages, search access, working forms and accurate information that a buyer can verify against the original source.

How does Perplexity Comet SEO differ?

Comet can use the page a person already has open, then research across other tabs or act on the site. Search discovery is only part of that interaction.

Perplexity’s access controls let users block particular websites from assistant tasks or context. Access isn't automatic.

ChatGPT’s cloud browser has separate sessions. OpenAI says it does not inherit the user’s device tabs, history, cookies or existing sign-ins, so an authenticated customer journey needs its own test.

For Searchmaxxed’s content work, that distinction changes what belongs on the public page: enough accurate detail to assess fit before a prospect enters a portal.

Does AI browser SEO replace traditional SEO?

AI browser SEO adds a task-completion requirement to familiar search work. It doesn't remove the need for accessible pages, useful content or evidence behind your claims.

Google’s current guidance is explicit. Its generative search features use core search ranking and quality systems, and pages must meet its indexing, snippet and AI-feature eligibility requirements.

Keep the foundations. At Searchmaxxed, that includes crawl and indexing fixes, canonical tags, internal links, title tags, site speed and structured data that matches the visible page.

Think about a B2B software purchase. A growth leader asks an assistant to compare platforms that support their identity provider, then find the right person to discuss deployment.

Each stage needs different evidence:

Keep these records separate. A mention inside an answer isn't a completed demonstration request, and successful crawling doesn't show that an assistant understood which product the buyer needed.

How do AI browsers read your pages?

Browser agents can interpret screenshots, page HTML and the accessibility tree. Different products combine those inputs differently, so a text-only crawl isn't a complete usability test.

The DOM is the browser’s representation of the page’s structure. The accessibility tree describes controls through their names, roles and states.

Google’s agent-friendly website guide explains these methods and recommends semantic HTML, labelled inputs and stable layouts.

Give the page a real test. Can an assistant distinguish a manufacturing product’s operating limits from its shipping dimensions, or does the same unlabelled number appear in both places?

A SaaS integration page needs equal care. State whether the connection is native, partner-built or available through an API, and explain any prerequisites alongside the relevant claim rather than elsewhere.

That is ordinary on-page work with a demanding reader. Searchmaxxed builds product, industry, comparison and integration pages around those distinctions, rather than repeating a broad feature claim across the website.

For the implementation detail, our agent-ready website guide covers the technical checks. Here, the question is whether the assistant understands and completes your prospect’s task.

Does schema or llms.txt make a site agent-ready?

Neither replaces a usable page. Google says it ignores llms.txt for search and requires no special schema markup for its generative search features.

Keep useful structured data accurate. It can describe the product or organisation, but a label in JSON doesn't repair a broken form or explain an unsupported integration to a buyer.

Google’s agent-friendly website guide also discusses WebMCP. It describes a proposed standard for exposing website tasks to compatible agents and warns that it remains in active development.

Treat that as an engineering option. Start with a specific supported workflow and a measurable problem, rather than adding a protocol because an AI-browser article happened to mention it.

Which crawlers should a business allow?

Decide separately about search discovery, model training and user-requested browsing. OpenAI documents different controls for each; allowing one isn't the same as allowing them all.

Source: OpenAI crawler documentation.

Check the firewall too. A permitted robots.txt rule doesn't prove that a cloud browser can pass your security controls, load the product page and open its enquiry form.

OpenAI’s cloud-browser documentation says websites can block these visits even when the same page works in a person’s browser. That explains some failed tests.

Searchmaxxed’s technical SEO work covers access faults. Any access change should preserve authentication, rate limits and protection for private data, with the actual blocked request identified before rules are changed.

What should enterprise teams test first?

Test a meaningful buying task on the pages closest to revenue. Start with public information, then stop before sending an enquiry or making a commitment.

Use a controlled test environment for submissions. For production browsing, a suitable instruction is: “Find the information and prepare the next step; do not submit anything”.

These are suggested tests, not reported client outcomes:

Write down the starting conditions. Record the browser, account state, market, prompt, date, pages visited, extracted answer and the exact step where the task stopped or went wrong.

Then change something specific. If the assistant confuses product variants, fix the comparison; if it can't identify a field, fix the label; if access fails, investigate access.

Keep the baseline and repeat. The same task after a release is more useful evidence than a different prompt that happens to produce a flattering answer about your company.

For Australian and US pages, check regional details. A correct product answer attached to the wrong deployment region, certification scope or enquiry team still sends a prospect down the wrong path.

Searchmaxxed connects page and technical changes to measurement. That makes this testing useful within search work, rather than a separate browser experiment nobody owns after the demonstration ends.

How do you measure AI browser traffic without fooling yourself?

Measure referrals, observed agent requests, task success and qualified enquiries separately. Your analytics channel name doesn't prove that a session came from a human using an AI browser.

We have a useful example. This approved finding comes from an enterprise B2B client’s GA4 review covering August to October 2026:

A client's GA4 showed AI-assistant visits jump from 354 to 6,050 in a month. Almost none of it was real. About 5,700 came in three one-hour bursts: same 1920x1080 desktop Chrome, all new users, no referrer, about 20 seconds each, landing on two pages. Real AI traffic was 385 sessions against 354 the month before, roughly flat.

That wasn't a Comet adoption study. It was a measurement problem, and treating the headline increase as new demand would have sent the marketing team towards the wrong conclusion entirely.

Use a report that keeps the evidence distinct:

Review suspicious spikes. Review timing, repeated paths, device patterns and missing referral details together; a missing referrer doesn't prove automation, and neither does a browser name.

Keep unknown traffic unknown. Don't rename all Direct sessions “AI”, or count an OAI-SearchBot request as a visitor who arrived ready to discuss your product with the sales team.

Searchmaxxed’s managed plans include checked Search Console and GA4 measurement. For the page-level conversion work after a visit, see our LLM referral landing-page guide.

Is Searchmaxxed right for your AI browser SEO work?

Searchmaxxed fits companies that need the website changes implemented alongside SEO and AI search optimisation. Our work covers the pages, technical foundation and measurement described here.

The scope matters. Expanding SaaS integration content differs from fixing manufacturing documents across markets or simplifying public-sector eligibility pages, even when browser usability matters to all of them.

You own the website, repository, content and data. Managed work follows a plan with weekly releases, and you have direct access to the people doing it rather than an account-manager relay.

We aren't the fit for everything.

Our AI search optimisation service covers the search work. Bring the product pages and buying tasks that matter, so the proposed scope starts with your business rather than a browser brand.

Frequently asked questions