Original research · Pre-registered · Emerging surface

What Agentic Browsers Actually Fetch

ChatGPT Atlas, Perplexity Comet, and Gemini in Chrome are shipping now. What they actually request from a page — and whether it differs from a standard bot fetch — has barely been studied.

Ritik Namdev Ritik Namdev ·Published September 2026 ·v0 — instrumentation stage ·11 min read
The short version

Agentic browsers act on a specific page at a user's direction, rather than crawling autonomously at scale. Do they render JavaScript? Execute page interactions? Fetch differently than a standard retrieval bot? Largely unstudied, and worth measuring early, before conventions calcify.

A brand-new surface, already shipping

Agentic browsing moved from concept to shipping product within a short window. ChatGPT Atlas, Perplexity Comet, and Gemini's in-Chrome assistant are all live. Coverage of "AI SEO" has been slow to catch up to this specific surface. That's exactly why it's worth instrumenting a test site now, before the behavior gets assumed and repeated as folklore.

The current landscape

ProductOperatorInteraction model
ChatGPT AtlasOpenAIDedicated agentic browser
CometPerplexityDedicated agentic browser
Gemini in ChromeGoogleIn-browser assistant mode

What is known

Fact

Each of these products can navigate, summarize, and take actions on a live web page at a user's direction, distinct from a chatbot's independent web search tool.

Open question

Whether they render JavaScript, respect robots.txt for the specific fetch they perform, or leave a distinguishable trace in server logs, hasn't been rigorously documented in public.

What isn't known

Does an agentic browser's fetch behave like a standard retrieval bot? Like a full human-driven browser session? Or like something in between? Does it respect the site's robots.txt directives for the relevant user agent? Does it execute JavaScript the way a human's browser would, since it's genuinely rendering the page for a live user, not extracting content for an index?

Why this is a genuinely different question from crawling

A retrieval or training bot's incentive structure favors speed and scale. Fetch as many pages as efficiently as possible, generally without executing JavaScript, since running a full browser engine for each of billions of pages doesn't pencil out economically.

An agentic browser has the opposite incentive structure. It renders exactly one page, at exactly one moment, on behalf of one user actively waiting for a result. That's much closer to the resource profile of a real browser tab than to a batch crawler. This is precisely why this study doesn't assume an agentic browser will behave like a crawler, even when the same company runs both products.

Study design

An instrumented test site logs every request header, JavaScript execution signal, and interaction pattern when each agentic browser gets directed to visit it. It's a natural extension of the crawler JS-rendering experiment to this newer surface.

A worked instrumentation example

Here's the concrete version. A test page includes a button that only reveals a specific, traceable piece of text after being clicked. That content stays invisible to any bot that merely fetches the raw HTML, or even executes JavaScript on page load without further interaction.

Direct each agentic browser to "find the hidden detail on this page." Then observe whether it actually clicks the button, executes the resulting script, and reports the revealed content back to the user. That tests not just whether these products render JavaScript, but whether they interact with a page the way a human genuinely would. That's a distinct, higher bar than passive rendering alone, one no existing crawler-focused study addresses.

Pre-registered hypotheses

#HypothesisPrediction
AB1Agentic browsers render JavaScript on the pages they visit, unlike autonomous retrieval botsSupported
AB2Agentic browsers respect robots.txt directives less consistently than autonomous crawlers do, given their user-directed rather than bulk-fetch natureSupported

ChatGPT Atlas, Perplexity Comet, and Gemini in Chrome are already shipping. What they actually fetch from a page — and whether it behaves like a bot or like a human browser — has barely been studied.

Share on X

Agentic browsing sits alongside two other emerging agent-facing surfaces covered elsewhere on this site. One is MCP servers, where an agent calls a structured tool interface, rather than parsing a rendered page at all. The other is WebMCP, a proposed standard for exposing page actions to an agent in a structured way, rather than requiring it to infer them from visual layout.

An agentic browser today generally has to interact with a page the way a human does, by rendering and interpreting the visual DOM. That's the least structured, most inference-heavy of the three approaches. Whether agentic browsers increasingly adopt structured interfaces like WebMCP over time, instead of continuing to rely on visual page interpretation, is a natural follow-on question once this study's baseline measurement exists.

Why this matters now, not later

As agentic commerce and agentic browsing grow, understanding exactly what these agents can and can't see on a page will matter for the same reasons crawler behavior matters for classic AI citation. Except the conventions here haven't been established yet. That's precisely the window where first-party measurement has outsized value.

What to check on your own site today

You can start observing this behavior yourself, without waiting for the full study. Open your server logs and search for known agentic-browser identifiers, then check whether their requests differ from a normal human visit in headers, timing, or the pages they touch.

If you can, visit your own most important page using one of these tools directly, and watch what it actually does. Does it click through to secondary content? Does it stop at the first visible answer? That single observation tells you more about your own site's readiness than any general statistic can.

Limitations

  • These products are evolving rapidly — any specific behavior documented is a snapshot, likely to shift as the products mature.
  • A single test site's results may not generalize to every page type or interaction pattern these agents encounter in practice.
How to cite this
Namdev, R. (2026). What Agentic Browsers Actually Fetch (v1). Retrieved from https://ritiknamdev.com/blog/agentic-browsers-what-they-fetch

Published under CC BY 4.0 — reuse freely with attribution.

Related work on this site

Complements the JavaScript rendering experiment and the AI Bot Registry, which covers autonomous crawlers rather than user-driven agentic browsing.

FAQ

Frequently asked questions

What is an agentic browser, exactly?
A browser, or browser mode, with an AI agent embedded that can navigate, read, summarize, and take actions on web pages on a user's behalf. ChatGPT Atlas, Perplexity Comet, and Gemini in Chrome are current examples. They differ from a chatbot's web search tool because they interact with a live page, not a search index.
Isn't this the same as the existing AI crawler question?
Related, but distinct. Training and retrieval bots fetch pages autonomously at scale, to build an index or answer a query. Agentic browsers act on a page a specific user is currently pointed at, often executing JavaScript and interacting with page elements the way a human would. That's a genuinely different technical surface.
Why is this registered as a study now, if the products are so new?
Because being early is the entire value proposition here. Measuring behavior on a genuinely new surface, before conventions and folklore calcify around it, is exactly the kind of land-grab research opportunity this site is built to pursue.
Would an agentic browser visit be logged differently from a normal human visit in analytics?
Potentially, if the product identifies itself distinctly in its request headers or user-agent string. But whether it does so consistently, or presents as indistinguishable from a normal browser session, is itself one of the open questions this study is designed to document, not assume.
Could a site deliberately serve different content to an agentic browser than to a human?
Technically possible, if the agent is reliably identifiable, similar in spirit to dynamic rendering for classic bots. But whether that's advisable, and whether these products' own policies permit it, is a separate and more ethically loaded question. This measurement study doesn't take a position on it. It aims first to establish what actually happens under normal, unmodified serving.
Ritik Namdev
Written by

Ritik Namdev

Growth · SEO · GEO

Growth marketer documenting a brand-new site's climb into Google and the AI engines - in public, with real numbers. Every tactic here is tested on real sites before it's published.

The Lab · Weekly

One experiment. Every week.

The field notes in your inbox - one thing I tested, the raw numbers behind it, and what it means for getting cited by AI.

Free forever. Unsubscribe anytime.