The next visitor to your website might not click a single thing. It might be an AI agent doing the browsing.

ChatGPT Atlas, Perplexity Comet and Copilot Actions now browse, fill forms and complete tasks on real websites. What your site needs in place before 2027.

By Yannis Spanenburg · 2026-08-16 · 6 min read

Not long ago, "who's browsing your site" meant a person on a phone or a laptop, scrolling, comparing tabs, deciding what to click. In 2026 that assumption is starting to break. OpenAI's ChatGPT Atlas browser, Perplexity's Comet, and Microsoft's Copilot Actions in Edge can now open a real website, read it, fill in a form, and complete a task end to end, with no human touching the keyboard after the first instruction.

This is a different problem than AI search or agentic checkout, which deal with being found and being bought from. This one is about whether an agent working on someone's behalf can operate your website once it lands there.

The browsers already doing this

ChatGPT Atlas can work through multi-step flows across tabs: research a topic, compare options across sites, and act on the result. Perplexity's Comet does the same inside its own browser, treating each page as something to reason over. Microsoft has built agentic actions directly into Edge through Copilot, and Google has folded years of Project Mariner research into Gemini's browsing capabilities in Chrome. All of these now ship to real users.

Adoption is still early relative to total web traffic, but the direction is not in question. Every major browser maker is now building a version of "browse and act for me" on top of "browse and answer questions."

What actually changes for your website

An agent works out a page from its structure: what a button does, what a form field expects, and what happens after it submits. A site that is technically usable by a human but structurally ambiguous to code is exactly the kind of site an agent gets stuck on.

1. Swap div-soup with bolted-on click handlers for real semantic HTML and labelled form fields. An agent reading unlabelled inputs has to guess what "field_3" means, and it often guesses wrong.
2. Write button and link text that says what happens, such as "Book a 15-minute call", and retire vague labels like "click here" and "learn more" that an agent cannot act on with confidence.
3. Keep core content out of infinite scroll, hover-only menus and JavaScript that only renders after a specific human-like interaction. Anything behind a mouse gesture no agent would think to make stays out of reach.
4. Reserve CAPTCHAs and bot walls for the flows that need them, like checkout or account creation. Gating your pricing page or blog behind one turns away every agent, including the ones bringing you a genuine customer.
5. Publish structured data for the actions your site supports as well as for your content. Schema for services, availability and booking spells out what an agent can do on the site, the same way Product and Offer schema does for a store.

Deciding who gets in

Not every bot claiming to be an agent is one worth letting in, and blocking all of them isn't free either. A robots.txt and firewall policy written for 2019's scrapers will turn away 2026's legitimate agents acting for a real customer. Review what you're blocking on a schedule, with the same scrutiny you'd give a new ad channel.

Don't overbuild for a visitor that isn't fully here yet

Most of your traffic is still, and will be for a while, a person with a mouse. Spend the effort on the structural debt that also makes your site slower, less accessible, and harder to maintain, work you likely owed your human visitors already. Bolting a chatbot onto your homepage or rebuilding your stack around agentic browsing is overbuilding for a visitor that is still arriving.

Proof of Pixel is a marketing and consulting agency for businesses that live on enquiries and orders, building the website, the tracking and the follow-up as one system, with offices in Dubai, New York, London, Antwerp and Malaysia.

Proof of Pixel's position: clean, labelled markup on your highest-intent pages is now a requirement, because an agent can only do what your site's structure lets it do.

Where to start

Pick your three highest-intent pages, the ones where someone books, buys or contacts you, and check them with a text-only browser or a screen reader. If the primary action isn't obvious without images or a mouse, an agent will struggle with it too. Fix labelling and structure there first, then extend the audit to the rest of the site, using the same checks that were already good practice for human visitors.