ChatGPT Operator Business Tutorial: Testing Arbitrage, Vendor Outreach, and Product Sourcing

An AI that gives you business ideas is useful. An AI that opens a browser, finds underpriced products, logs them in a spreadsheet, and contacts sellers is something else entirely.

That is what got me excited about ChatGPT Operator. Chris Koerner had been experimenting with the $200 per month ChatGPT Pro plan and shared an example of the agent working through Facebook Marketplace on his behalf. So we put it through three business tasks: finding resale opportunities, requesting catering quotes, and sourcing product samples.

Some parts worked. Some got stuck. One produced a real vendor reply before we finished.

What we did not establish was a business earning $500 a day. We tested the workflows that could support a business, not a verified revenue result. That distinction matters. But the underlying shift is still pretty wild: you can give software an outcome and have it start doing the browser work to get there.

Table of Contents

• What Makes Operator Different From a Normal Chatbot?
• Test 1: Find Underpriced Pizza Ovens on Facebook Marketplace
• Test 2: Automate Wedding Catering Inquiries
• Test 3: Source Products and Request Samples
• A Related Use Case: Sending Email From a Supplied List
• How I Would Approach Prompting a Browser Agent
• Is a $200 Monthly AI Agent Worth It?
• For Builders Turning Ideas Into Businesses

What Makes Operator Different From a Normal Chatbot?

Instead of simply explaining how to research a product, Operator can navigate websites, open listings, fill out forms, and enter information into a spreadsheet. In these experiments, Chris worked inside its browser, logged into the necessary accounts, and could take control whenever the agent needed help.

The appeal is not that every click happens faster than a person could do it. Often, it does not. The appeal is that the task can keep moving while you do something else.

Think about the work you would normally hand to a virtual assistant: compare prices, collect links, contact vendors, and organize responses. Those are the kinds of tasks we wanted to test.

The $200 monthly price was the price discussed during these experiments, not a guarantee of current pricing or access. Likewise, the results below describe what happened in these sessions, not what every account or website will support.

Test 1: Find Underpriced Pizza Ovens on Facebook Marketplace

Chris started with a category he understood: pizza ovens. He owns a Gozney, cooks pizza in his backyard at around 900 degrees, and knows what the equipment should cost.

That was a smart choice. If you already understand a product category, you can tell whether the AI has found a bargain or just confidently labeled something a bargain.

Give the Agent a Specific Search and Somewhere to Put the Results

The initial assignment was straightforward:

• Search Facebook Marketplace for Gozney pizza ovens priced well below comparable models.
• Begin within 500 miles of Dallas.
• Look for sellers willing to ship.
• Record promising opportunities in a supplied spreadsheet.

Operator was already logged into Chris’s Facebook account. It changed the search location to Dallas and began opening listings. Chris also supplied a blank spreadsheet and asked it to create the headers itself.

Within about two minutes, it reported a Gozney Roccbox listed for $250, including a peel and propane burner, in good condition. Chris estimated that the oven and accessories could be worth around $500 in that condition.

It eventually added two opportunities to the spreadsheet, including titles, locations, condition information, links, and brief reasons they looked promising. The explanations were fairly generic, but the useful part was concrete: there was now a small list of products to investigate.

Why Marketplace Arbitrage Exists

This is an old business model with a new sourcing tool.

A person clearing out a garage may choose Facebook Marketplace because it is easy and local. They want the item gone. They may not want to photograph it carefully, research national prices, package it, or ship it.

Another person can do that work and reach a larger audience on eBay. The opportunity sits in the difference between the seller’s convenient price and the price another market might support.

One of my early hustles used the same principle. I bought discounted DVD box sets in Canada and sold them on eBay to buyers in the United States. Product pricing and the currency difference created the opportunity. But I still had to go to the store, check eBay, and work out the economics myself.

Operator starts to take over the repetitive research portion of that process. The business model is not new. The amount of human attention required could change significantly.

Chris thought this approach could potentially support a six or seven figure business across suitable product categories. That was his assessment of the opportunity, not a result demonstrated by two pizza oven listings.

Messaging Worked Differently From Making an Offer

Next, Chris asked Operator to offer half the asking price on the shortlisted ovens and mention that he could pay cash.

It initially refused to make offers on his behalf, even though he had achieved something similar the previous day. It then proceeded toward seller messaging, but stopped short of clearly submitting the specific price offers requested.

That is an important limitation. Finding a listing, messaging a seller, and negotiating a purchase are separate capabilities. Success at one does not prove success at the others.

Chris also knew someone whose Operator session would not work with Facebook at all. Access and behavior were inconsistent. Breaking a task into smaller steps sometimes helped, but this was not a dependable, unattended buying machine.

The Proposed Resale Workflow

To turn the research into a business, Chris outlined a more complete process:

1. Research the average eBay price for a specific make and model.
2. Find matching Marketplace listings, particularly those that have been available for 30 days or longer.
3. Make lower offers, on the theory that an older listing may have a more flexible seller.
4. Once a seller agrees, arrange the resale and fulfillment.

His hypothetical example was an oven with a $300 market price: negotiate a $150 purchase, then offer it at a $250 Buy It Now price. That creates a $100 difference before fees, shipping, and other costs, not $100 of guaranteed profit.

He also proposed having the Marketplace seller ship directly to the eventual eBay buyer. That part was an idea, not a completed transaction. Before attempting a similar model, check the marketplace rules, including eBay’s dropshipping policy, rather than assuming a technically possible workflow is permitted.

The demonstrated win was deal sourcing and recordkeeping. Buying, reselling, and delivering at a profit remained unproven.

Test 2: Automate Wedding Catering Inquiries

The second experiment moved from earning money to saving time and potentially reducing costs.

The task: use The Knot to contact caterers in Dallas for a wedding with 50 guests, requesting a per-person price for steak and two sides.

This is exactly the kind of work people put off. You need to find vendors, open their pages, repeat the same details, fill out contact forms, and wait for replies. Then you need to compare what comes back.

Chris created an account for the experiment and gave Operator the basic requirements. The wedding itself was hypothetical, so this was a test of the workflow rather than a real purchasing project.

From a Short Instruction to Completed Contact Forms

Operator searched for caterers, moved its search to Dallas, opened vendor pages, and began filling out inquiry forms. It requested Chris’s contact details when it needed them.

It turned the rough instruction into a polite message describing the guest count and menu request, then asked for a per-head quote. It also selected the relevant guest-count range in the form.

The value here is not fancy writing. It is getting through the repetitive interface work.

The same general approach could be narrowed further: vegan caterers only, Dallas County only, or another specific requirement. Better constraints mean fewer irrelevant vendors and more useful responses.

The Agent Invented a Wedding Date

Here is where the experiment got interesting. Chris had not supplied a date, but the form needed one. Operator chose May 11.

When asked why, it explained that the date was a few months away and allowed planning time.

That is resourceful, but it is also a business risk. A vendor’s availability and quote may depend on the date. Filling a required field is not the same as filling it accurately.

An agent can complete a form while still getting an essential fact wrong. For a real inquiry, give it the actual event details and tell it to stop and ask if a required fact is missing. Do not use invented projects to generate unnecessary work for vendors.

What Happened When We Changed the Instructions?

We tried several adjustments while the task was running:

• Make the tone extremely friendly: The visible message did not change much.
• Stop asking before submitting: Later in the session, it submitted a request without another confirmation.
• Move faster: It appeared to speed up, although we did not measure a reliable improvement.
• Submit three requests per minute: It said it could not meet that pace.
• Use a rude tone or switch to Spanish: It declined those changes in that session.

The lesson is not that a particular phrase unlocks a secret performance mode. The lesson is that you need to check whether the requested change actually happened.

A numerical goal can make expectations clearer, but it cannot remove a tool’s limits. Likewise, a refusal to change languages in one running task does not establish that the product cannot handle other languages at all.

A Real Caterer Replied

Before we finished the broader experiment, CN Catering sent Chris a message. The vendor said the selected date was available and supplied sample menus.

That is a real outcome: the agent’s browser actions generated an external response.

It was not yet a finished set of comparable quotes. The next proposed step was to have Operator review the menus and compare pricing against the requirements. We did not complete that comparison.

We also discussed a related idea: a voice agent that calls vendors, explains that it is acting for a real customer, discloses that it is AI, and relays the answers back. That remained a concept. Operator did not make those calls in this test.

Test 3: Source Products and Request Samples

Chris has spent about 15 years sourcing products from China. One repetitive part of that work is contacting many potential suppliers with a similar message, then arranging samples.

So the third assignment was to visit AliExpress, identify a product featured as trending, and message 10 vendors asking whether they could send a free sample to Texas.

The dream is obvious: delegate the outreach, return later, and find supplier replies waiting. The actual test was messier.

AliExpress Failed, So the Agent Looked for an Alternative

Operator reported that AliExpress was inaccessible. Giving it the exact URL and trying a new tab did not resolve the problem, even though the site worked in Chris’s regular browser.

We did not establish why. It would be a mistake to turn that failure into a broad claim that Operator could not access Chinese websites.

Chris asked for a comparable site. Operator searched the web, opened an article listing alternatives, and navigated to Banggood.

This was one of the more striking moments for me. The agent was not just returning an answer. It was encountering an obstacle, finding another route, opening pages, and trying again.

That does not prove AGI. It does make the shift from software that advises to software that acts feel very tangible.

It Found Earbuds and Prepared a Sample Request

On Banggood, Operator selected BlitzWolf earbuds and found an “ask a question” form. It drafted a short request asking whether a free sample could be shipped to Texas.

The form had a checkbox indicating that the sender wanted a seller response. Chris explicitly told Operator to check it before sending.

Then the website required an account.

With authorization, Operator began registration using Chris’s email and a requested 20-character password. It encountered a registration-limit message and struggled with verification. Chris eventually took control and entered the code himself, after which the account was logged in.

The agent returned to the product inquiry form, filled it out, checked the response box, and submitted. It also captured a screenshot of the action.

We did not receive a confirmed free sample or demonstrate outreach to all 10 vendors. We demonstrated an alternative-site search, an account setup with human help, and a submitted product inquiry.

Featured Products Are Not the Same as Proven Demand

When asked why it picked the earbuds, Operator said they were prominently featured on the homepage and highly rated.

That explains its choice. It does not independently validate a product opportunity.

A stronger proposed workflow was to provide a defined list of trends, supported by a source such as Google Trends, and ask the agent to find products only within those categories.

That keeps the business judgment with you. The agent can execute research against a thesis rather than inventing the thesis from whatever a marketplace happens to promote.

A Related Use Case: Sending Email From a Supplied List

Chris also described an earlier cold email experiment. Operator refused a combined request to scrape email addresses from search results and conduct outreach.

When he changed the task to use a spreadsheet he supplied, containing email addresses, first names, and message text, it said it could send personalized emails through a Gmail account logged into its browser.

This was a reported prior interaction, not a completed email campaign in our session. Still, it illustrates a useful distinction: asking an agent to discover contacts and contact them can be treated differently from asking it to work from a supplied list.

It is another reason to define the task carefully rather than assuming every outreach workflow is interchangeable.

How I Would Approach Prompting a Browser Agent

The biggest practical lesson is to treat the agent like a capable but literal assistant. Be direct. Give it a defined job. Then inspect the result.

1. Start with a category you understand. Chris could evaluate the pizza oven listings because he knew the products and prices.
2. Specify the outcome and constraints. Include location, product model, guest count, menu, shipping requirements, or other details that determine usefulness.
3. Give it a destination for its work. A spreadsheet with links and relevant fields makes the research easier to review.
4. Separate research from external actions. Finding vendors, drafting messages, sending messages, and making offers are distinct steps.
5. Provide missing facts instead of inviting guesses. The invented wedding date is the clearest example of why this matters.
6. Take control when necessary. Human intervention helped complete account verification and unblock navigation.
7. Keep a record of what works. Repeated experiments reveal which prompts, websites, and task structures are dependable.

Chris described another useful approach from a Beehiiv task: when the agent paused, he took control and demonstrated the navigation sequence, then asked it to continue. That suggests demonstration can sometimes help when explaining a complicated interface in words is awkward. It is not proof that the agent permanently learns every action you show it.

We also discussed saving tasks, recording screenshots in a document, running multiple sessions, and leaving agents operating continuously. Those are interesting possibilities, but we did not establish reliable parallel operation, automatic record export, or uninterrupted 24/7 execution.

Is a $200 Monthly AI Agent Worth It?

Chris’s answer was basically: not if you have no idea what to do with it.

I agree. An unused subscription is expensive at almost any price. A tool that reliably removes valuable repetitive work is a different calculation.

What excites me is the combination: the internet already contains marketplaces, vendor directories, forms, inboxes, and payment systems. Now an agent can navigate parts of that infrastructure based on an outcome you describe.

It is easy to focus on the slow clicks, inaccessible pages, and awkward refusals. Those are real problems. But dismissing the whole category because it is imperfect misses the bigger change.

The opportunity is to learn which parts of a real workflow can be delegated reliably. Not to declare an entire business automated because the agent filled out a form.

The useful starting point is small: one product category, one research task, one spreadsheet, or one legitimate vendor inquiry. Learn where it succeeds, where it guesses, and where you need to intervene. That operational knowledge is the advantage you can build before the tools become easier for everyone else.

For Builders Turning Ideas Into Businesses

Startup Empire is my private membership for people developing startup ideas or trying to get more traction. It brings together practical training on topics such as email marketing and audience building, resources for execution, and other builders who could become collaborators or potential co-founders.