Summary of an article published on The Web Scraping Club on 2025-11-02 (free to read). Read the full article

Scraping has always relied on scripted browsers, Playwright or Selenium, imitating human behavior. Now the browser itself can read a page, interpret what it sees and act as an agent. ChatGPT Atlas, Browserbase Stagehand and Cursor’s browser utility all go in that direction, with different goals. For anyone building scrapers, the practical question is which of them changes the daily work today and which is still far from production.

We tried Atlas on Walmart and Saks Fifth Avenue, asking it to describe promoted TVs and homepage offers, including banners that exist only as images. I explain how Stagehand turns natural language into deterministic browser actions. Then we ran Cursor with Claude 4.5 on Harrods.com, asking it to paginate and look for internal APIs while the HTML and network calls sit in the LLM context. I close with what this means for prompt-centric scraping workflows.

What’s inside

  • My reservations about Atlas: sharing browsing activity with OpenAI, how that differs from sharing it with Google, and the risk of logging in to a bank account by prompt.
  • Niche use cases where an AI-reading browser could help, such as ad monitoring and share of virtual shelf, and the open question of running them on a daily schedule.
  • Why Stagehand’s mix of LLM reasoning and Playwright-level automation looks suited to production pipelines, and why I have not tested the new V3 yet.
  • The difference between asking an LLM to scrape from its training data and giving it the live network activity of the target site.

In the full article: Screenshots of the Atlas sessions on Walmart and Saks Fifth Avenue, the Cursor session on Harrods.com, and my assessment of how usable each tool is for production scraping.

Read the full article on The Web Scraping Club