Short answer: name the product, the sites, and the fields you want, then require a single comparison table with source URLs. Atlas will open the tabs, extract the numbers, and return a table you can spot-check in under a minute.

Why do naive comparison prompts fail?

Ask “find me the cheapest X” and you get a confident answer that may ignore shipping, tax, or stock status. The agent did what you said — you just did not specify what “cheapest” means.

Define the comparison dimensions up front. Price is rarely the whole story.

How do you structure a comparison prompt?

1. State the exact product and variant — model number, size, color.
2. List the sites to check by name.
3. List the fields you want: item price, shipping, estimated tax, stock, return window.
4. Demand a markdown table with a source URL per row.
5. Ask for the total landed cost as a final column.

A working prompt reads: “Open Amazon, Best Buy, Walmart, and Target for the Sony WH-1000XM6 in black. For each, return a table with item price, shipping cost, stock status, and the source URL. Add a final column with total landed cost to zip 94107.”

What does a useful output look like?

The value is in the columns, not the row count.

Always ask for the URL. When an agent gives you a table without sources, you are trusting extraction you cannot check — and product pages change layouts constantly.

Where does the agent still get it wrong?

Dynamic pricing is the big one. Pages that personalize price by location or fetch it via JavaScript after load can yield a number that differs from what you see in your own session. Stock counts also lag reality by minutes to hours.

Coupon fields and membership discounts are the other gap. Atlas rarely applies a loyalty rate automatically unless you tell it to. Add “include member pricing where shown” to the prompt if that matters.

When should you not automate this?

Skip automation for one-off purchases where opening three tabs takes 90 seconds. The prompt-writing overhead only pays off when you compare the same category repeatedly — laptops, flights, or recurring supplies.

Also avoid letting the agent complete checkout. Reading prices is low risk; entering payment details is where a wrong click costs real money. Let Atlas produce the table and you finish the purchase manually.

Quick Summary

• Specify product, sites, and exact fields — vague “cheapest” prompts fail.
• Require a markdown table with a source URL per row so you can verify.
• Include landed cost, not just item price, since shipping changes the ranking.
• Dynamic pricing and stale stock data are the main accuracy gaps.
• Let the agent compare; complete checkout yourself.

Short answer: install Atlas for macOS or Windows, sign in with a ChatGPT account, and start with read-only tasks like summarizing or comparing pages. Keep any task that touches payments on manual control.

What is OpenAI Atlas?

Atlas is a desktop browser built on Chromium, so it renders pages the way Chrome does and supports most extensions. What makes it different is the side assistant: it sees the page you are on, can open new tabs, and can act on your behalf.

It runs on GPT-5 class models and keeps per-site memory, so it remembers that you always check the return policy before buying, for example. It is available on macOS first, with Windows following.

How do you install and set up Atlas?

1. Download the installer from the OpenAI site for your OS.
2. Sign in with your ChatGPT account — Plus and Pro plans get higher agent limits.
3. Import bookmarks and passwords from Chrome during first-run.
4. Open the assistant panel and grant per-site permissions when prompted, not globally.
5. Turn on memory only for the sites where continuity actually helps.

Importing from Chrome takes a minute and saves you from rebuilding a browser profile.

What tasks is Atlas actually good at?

Read-heavy and comparison work is where it earns its place.

Ask it to “open these four tabs and tell me which has the shortest shipping time” and it will read each page, extract the facts, and compare. That is a genuine time saver.

Is OpenAI Atlas safe to use?

Treat it like a very capable intern with access to your logged-in accounts. The prompt-injection risk is real: a malicious page can contain text designed to instruct the agent to do something you did not ask for.

Three rules cut most of the risk. Keep agent permissions site-scoped. Never let it enter card details automatically. And review the action log after any task that changed something — a purchase, a form submission, a message sent.

Many enterprises block agentic browsers outright because they bypass traditional endpoint controls. If you handle regulated data, check policy before installing.

How does Atlas compare to a plain browser?

For pure browsing it is slower and heavier than Chrome. Its value shows only when you lean on the assistant. If you mostly read documentation, a normal browser plus a side-by-side chatbot gives you 90 percent of the benefit for none of the risk.

Atlas suits researchers, comparison shoppers, and anyone who runs repetitive multi-tab workflows. Skip it if you want a lean browser or you work under strict compliance rules.

Quick Summary

• OpenAI Atlas is a Chromium-based agentic browser with a built-in AI assistant.
• It is strongest at reading, summarizing, and comparing across tabs.
• Keep permissions site-scoped and never automate payment entry.
• Prompt injection is the main security risk — review the action log.
• For plain browsing, a standard browser is faster and lighter.