Browser agents are often described as if browser use were a single capability. It is not. The model matters, but so do the browser engine, the control interface and the structure of the website.
We ran a small test to separate those parts. The task used Aiso's public website, required no login and created no external side effect. It was simple enough to verify but still required the agent to understand a page, navigate and operate an interactive control.
The test
The agent received five steps:
- Open the Aiso homepage.
- Understand the page and locate Pricing.
- Navigate to the pricing page.
- Identify the cheapest plan for brands.
- Switch billing from Monthly to Yearly and verify the new state and price.
The plan was to repeat those instructions in Chrome, Brave and Firefox. The comparison stopped before the second browser. The cloud environment listed Chrome as its only controllable target.
Browser availability was the first result
Chrome was available and completed the task. Brave and Firefox were not exposed. We have no result for how either browser would have performed.
| Browser | Availability | Outcome |
|---|---|---|
| Chrome | Available | Completed all 5 steps |
| Brave | Not exposed | Not tested |
| Firefox | Not exposed | Not tested |
Chrome did not win a browser comparison. It was the only browser allowed to compete. That is more useful than a made-up ranking because it shows where browser-agent capability can stop: at the environment boundary.
What Chrome understood
The agent opened Aiso's homepage, read the navigation and selected Pricing. It reached the live Aiso pricing page and recognized the brand and agency tabs, the monthly and yearly billing control, and four plan cards.
It identified Starter as the cheapest brand plan at $19 per month. It selected Yearly and checked two separate signals. The Yearly tab reported a selected state, and Starter changed to $16 per month.
That last check matters. A browser agent should not treat a click as proof that the requested change happened. It should inspect the result.
The measured timings
We recorded the elapsed time for each part of the run. These are observations from one run, not averages.
Browser connection and first homepage load
Navigation from the homepage to Pricing
Monthly to Yearly switch, including state verification
The 14.3 second first step included the initial browser connection and a cold homepage load. It is not a measure of Aiso page performance. Once connected, the agent reached Pricing in 3.2 seconds and completed the billing interaction and verification in 286 milliseconds.
This was more than text extraction
The agent used page structure to find named links, headings, tabs and selected states. It also inspected the rendered screen. The pricing headline, billing switch and plan cards had a clear visual hierarchy.
A cookie banner covered part of the first plan card in the first screenshot. It did not block the task because the page exposed the same information through well-labelled controls and readable content. The agent later rejected the banner and confirmed the full pricing layout visually.
This is the practical combination websites need. Semantic HTML gives agents reliable handles. Visual design tells them what the page emphasizes and whether the result looks coherent.
The constraint was not model reasoning
Chrome completed all five steps without a retry. The larger limitation was the execution environment. An agent can only use the browsers and control surfaces made available to it.
A claim that an agent can use the web leaves out several details: which browser families it can control, whether it can inspect structure and pixels, whether sessions persist, how state changes are checked, and what happens when a site presents bot detection or a CAPTCHA.
This run answered the navigation and interaction questions for Chrome. It did not answer them for Brave or Firefox.
What website teams should do
Websites now serve human visitors and agents that can see and click. The work needed for both audiences overlaps.
- Use descriptive headings and links.
- Give interactive controls accessible names.
- Expose selected, expanded and disabled states in the page.
- Make important choices visually distinct.
- Show a clear result after an action.
- Keep the page fast enough that agents do not spend their budget waiting.
A page should remain obvious when read as a set of structured controls and when seen as a screen. That is useful for accessibility, conversion and agent use.
Limits of this test
This was one run, on one website, from one cloud environment on September 4, 2026. It did not test authenticated flows, forms, downloads, CAPTCHAs, mobile layouts or recovery after a failure.
The browser availability result can change when the environment changes. The timing numbers will also vary with network and cold start conditions. They should not be used to compare browser products or general web performance.
What the result supports is narrower: a ChatGPT cloud agent completed a five-step navigation and interaction task in Chrome with no retry, while Brave and Firefox were unavailable in the same environment.
Aiso's AI search visibility platform helps teams inspect the other side of this shift: what people ask AI, which sources shape the answer, and where their brand appears.