This article makes no income claims. What anyone earns depends on their skills, their time, their market and some luck, and I have no figures of my own or anyone else’s that I would stand behind.
Can you make money using ChatGPT? Yes, in the sense that people are paid for work in which a chatbot does part of the job. That answer isn’t very useful, though, and it’s the one the previous version of this page gave before filling the rest with income bands it couldn’t source. The questions worth answering are which part of the job the chatbot does, whether that part is what the buyer was paying for, and what it costs to find out.
A disclosure of my own first: ChatGPT is not my main tool. I build websites with Claude most working days, for my own sites and in the one-person web studio I run in Valencia, Spain. Everything below applies to any chatbot; the brand in the question is the one people search for, not the one that matters. Which assistant I would start with for which kind of job, and which ones I have never tested, is in my notes on which chatbot suits which job.
What I’m paid for, split between me and the model
The paid work I know is building small websites for businesses. I don’t show or name client sites anywhere on this blog, so here is the job broken down by stage instead, with an honest split of who does what.
| Stage of the job | What the model does | What I do |
|---|---|---|
| Brief and scope | Nothing useful until the conversation has happened | Talk to the business, and decide what is in the job and what isn’t |
| Direction | Produces a competent, generic layout if asked | Choose the references and the level of finish, from research on real brands in the trade rather than from what the model thinks a typical site looks like |
| Site structure | Builds it quickly once it is decided | Decide how many pages the business needs and what each one is for |
| Per-page SEO: title, meta description, canonical, Open Graph, structured data, sitemap.xml, robots.txt | Most of the typing, and it is good at it | Check that every page’s tags are unique and match what the page actually says |
| Copy | First drafts | Rewrite until it reads like someone who works in the trade, not like a template |
| Facts: address, opening hours, phone numbers, prices | Nothing. Asked for them, it invents plausible ones | Get them from the business and check them |
| Photos | Can’t judge whether a photo is right for a page | Choose real photos from the business and check each one on the page it ends up on |
| Deployment, DNS, certificates, Search Console | Drafts config files and scripts | Run them, verify the result, and fix it when it breaks |
Look at where the model’s column is strong: markup, boilerplate, per-page tags, first drafts, configuration. Then look at where it is empty: what the business actually needs, the facts about it, its photos, and the responsibility when something breaks. The buyer isn’t paying for the first column. With a bit of patience they could get most of it by typing a prompt. They pay for the second column, and for knowing who to call when the site goes down.
Most of what sits in my column is written down as rules the model reads before it touches a project, each one added after something went wrong. I have printed the file in the studio post linked at the end of this page, so I won’t repeat it here.
Why the cheap part is the part you can’t sell
The idea from the old version of this cluster that holds up is that a chatbot multiplies a skill; it isn’t one. If what you sell is text the buyer could generate themselves, you are selling them access to a tool they already have, and the price of that tends towards what the tool costs. If what you sell is the result, checked and owned by someone who knows the trade, the tool simply makes you faster at something people already paid for.
That has an uncomfortable corollary for anyone starting from nothing. The chatbot doesn’t supply the skill, the judgement about what is wrong with its own output, or the customers. Those three are the business. The rest is typing, and typing got cheap.
The old posts in this cluster also sorted money-making into “active” and “passive” streams, with blogs and digital products filed under passive. My experience runs the other way. One of the most demanding things I run right now is this blog, which I am rewriting post by post after it was turned down for advertising four times, and the two small “passive” products I made, both in the list of small projects I actually launched, are the things I’m least comfortable pointing anyone to.
What it costs to find out, with dates
Testing an idea is cheap in cash and expensive in hours. The cash side, from pricing pages read on the dates given:
- The chatbot. Claude has a free plan; Pro is $20/month, or $17/month when billed annually at $200 up front (Claude pricing, checked 20 September 2026). OpenAI’s pricing page returned an HTTP 403 error to automated requests on 20 and 21 September 2026, so I’m not quoting a ChatGPT price from memory; check the current one yourself.
- A place to show the work. A static site can live on Cloudflare Pages’ free plan, which allows 500 builds a month; Cloudflare’s documentation says “Requests to static assets are free and unlimited” (Pages limits and Workers pricing, checked 20 September 2026). One of my own sites runs there. A WordPress site on shared hosting costs a few dollars a month at the entry price, as long as you pay for several years up front, and several times that once it renews; Hostinger’s pricing page, where my own blogs live, lists both figures (checked 20 September 2026).
- A way to sell a file. Gumroad has no monthly fee and takes 10% + $0.50 per direct sale, or 30% on sales through its Discover marketplace (Gumroad pricing, checked 21 September 2026). Amazon describes its Kindle Direct Publishing tools as free.
Add a domain and the cash cost of a first test is small enough that it isn’t the real decision. The real cost is hours, and the way to stop them leaking is to decide in advance how many you’ll spend and what you’ll count. If I were testing a service in the evenings alongside a job, I would give it eight to ten hours a week: two evenings of making the thing, one evening of putting it in front of people, and a couple of hours at the weekend looking at what happened.
And I would track three numbers each week: things finished, people asked, and money collected. The third can sit at zero for a while without meaning much. If the first two are also at zero, the third will stay there.
Three conditions, checked against my own projects
Across everything I have run, three things decided whether an AI-assisted project had any chance. I’ll score my own projects against them, because the failures teach more than the rule does.
1. The buyer couldn’t get it by typing the prompt themselves
A working multi-page website with correct facts, real photos and a sitemap that Search Console accepts passes this test: most business owners can’t produce that with a chatbot, however good the chatbot is. A generic article about making money with AI fails it completely. I know, because I published a hundred of them.
2. Someone accountable checks everything before it leaves
This is where this blog failed. The posts went out at the speed they could be generated, thirty-six of them on a single day, 1 July 2026, with nobody checking the claims inside them. The result was four AdSense rejections for low-value content, the last on 10 September 2026, and a rewrite that is still going on. The full count is in my tally of what went wrong here. In paid work, the checking is most of what the client is paying for, and where I draw the line between using AI and selling its raw output has a post of its own.
3. Buyers can find you
The chatbot contributes nothing here. My other blog, about web hosting, had 90 posts published when I audited it on 19 September 2026, and Search Console showed five clicks from Google search over the previous three months. Ninety posts, five clicks. Publishing is not the same as being found, and no amount of generated text changes how slowly trust gets built. Whatever you are testing needs a route to buyers you can name before you start: people you already know in a trade, a marketplace profile, a local business listing.
The job that doesn’t work
If I had to name one kind of paid work I’m confident doesn’t hold up, it is selling unreviewed AI output: articles, product descriptions, social captions or anything else that goes from the chat window to the client without someone who knows the subject checking it. It fails three ways at once. The buyer can make it themselves. It is exactly where confident invented details slip through, the statistic that sounds right and doesn’t exist. And the buyer usually discovers the problem after paying, which is the worst possible moment for your reputation.
It fails hardest where being wrong has consequences. Health, legal, financial or safety content written by a model and not reviewed by a qualified person isn’t a cheaper service. It is a liability you have sold to someone else. The old version of this page listed generic blog posts and generic social captions as the most saturated categories. I’d go further: for someone whose only advantage is the chatbot, they aren’t a market at all, because the buyer has the same chatbot.
A four-week test that doesn’t require quitting anything
- Week one: pick one service you could already deliver without AI, for one kind of buyer you understand. Make three real samples, and time two things for each: how long the chatbot part takes, and how long the checking takes.
- Week two: ask people. Put the offer in front of specific people, by name, with a note that shows you understand their situation. This is the step most people skip, and the old post about landing a first small payment had this part right: the difference is usually whether you asked anyone to pay.
- Weeks three and four: deliver and measure. If anyone says yes, deliver properly, then look at the hours. If the checking takes as long as the work used to, the chatbot saved you nothing on that job, and that is worth knowing early.
- At the end, decide on evidence. Either someone paid and you now know what the job really takes, or nobody did and you know which of the three conditions failed. Both beat another month of reading about it.
Keep your job while you do it. None of this needs a leap, and the people telling you it does are usually the ones selling the leap.
If the test works and you end up running something small of your own, the post on starting a one-person web studio covers scoping, quoting, the rules file I mentioned above and what I never let the model decide.