The previous version of this page did a sum for me. Take a small software product at $49 a month, find a thousand subscribers, and you arrive at almost $600,000 a year. The sum sat in the middle of a list of “AI business ideas”, and I had not tried a single one of them.
Since then I have actually built a small subscription product. It runs on a server I pay for, it has plans and a sign-up page, and on the day I’m writing this it has no paying customers. I would rather give you that number than the other one, and the rest of this page is built the same way.
Between June and August 2026 I launched five things with AI doing much of the work: this blog, a second blog about web hosting, a twelve-page PDF on Gumroad, a book on Amazon KDP, and that software product. Here is what each one needed, what it cost to start where I can document it, and where it stands on 21 September 2026, including the ones that went nowhere. Those are the useful part.
The five, side by side
| Project | Went live | What it cost to start | Where it stands, 21 September 2026 |
|---|---|---|---|
| This blog | 8 June 2026 (first posts) | A share of one Hostinger shared plan I use for three sites, plus the domain. I haven’t split it out. | Turned down by AdSense four times, the last on 10 September. Being cut back and rewritten. |
| A second blog, about web hosting | 13 July 2026 (WordPress version) | The same shared plan. | Five clicks from Google in three months. Back in AdSense review since 31 August. |
| 12-page PDF on Gumroad | 4 July 2026 (file date) | Nothing up front. Gumroad takes a cut of each sale instead. | On sale at $5. No sales figures published. |
| Book on Amazon KDP | July 2026 (cover file dated 6 July) | Nothing up front. | On sale. No sales figures published. |
| Small SaaS on a VPS | 20 August 2026 | €9.67 for the first month of the server, about €19 for the domain, and a data service at $29 a month before VAT. | Live. No paying customer. Parked since 19 September. |
Two blogs, about 190 posts, twenty clicks from Google
This blog published its first posts on 8 June 2026. By 24 August it had a hundred of them. Thirty-six went live on a single day, 1 July, almost all were built on the same skeleton, and many carried income figures I couldn’t back up. Google AdSense turned the site down four times for low-value content, the last time on 10 September. The reviewer was right every time.
The second blog replaced an older static HTML site of mine on 13 July, when I moved that domain to WordPress. It launched with about ninety posts: forty published on day one and fifty scheduled at one a day. It has had its own AdSense rejections, and was back in review when I last looked, on 19 September.
What did they take? Very little time per post, and that was the problem. What they couldn’t skip was the waiting, which turns out to be the real cost of a blog. In the three months to 19 September, Search Console showed fifteen clicks for this site (average position 67) and five for the hosting blog, from 6,710 impressions (average position 63). Twenty clicks, two sites, roughly 190 posts.
AI compressed the writing to almost nothing and did nothing about the months in which nobody reads you. The specific mistakes, counted from an export of every post, are in the AI content mistakes that got this blog rejected. Both blogs are now being cut back and rewritten.
Two products, one week in July
The PDF is a twelve-page guide to earning money with AI. A model wrote it and a script laid it out, on the evening of 4 July 2026, and I listed it on Gumroad at $5. Gumroad charges no monthly fee and takes 10% plus $0.50 per sale through your own link (Gumroad’s pricing page, checked 21 September 2026), so the start-up cost was the evening and nothing else.
The book followed within days, on Amazon KDP, which costs nothing up front. It is called Make Money With AI, and the title tells you the register it was written in.
I do not publish sales figures here, for my own products as much as for anyone else’s. What is worth saying is where they were sold from: the only place either was ever promoted was this blog, in a box under every post, the one Google sent fifteen clicks to in three months, and that box quoted the PDF at $12 while the checkout charged $5.
The mistake isn’t the tools, it’s the order. Both were made to be sold before I had done anything they could honestly describe, so they average every other guide on the subject and promise income I had never earned from the method inside them. I’ve gone through both files sentence by sentence, including the metadata of the PDF and what Amazon makes you declare about AI, in what nobody tells you before you sell a digital product made with AI.
The one that taught me the most: a subscription tool on a VPS
This is the only one of the five I would start again. It’s a tool for people who import used cars from Germany into Spain: it watches listings, works out the landed cost with Spanish taxes, compares it with what similar cars sell for in Spain, and sends an alert when the margin looks worth a phone call. React on the front end, a Node and Express API, SQLite, and Caddy in front with systemd keeping it up, on a Hostinger KVM 1 VPS running Ubuntu 24.04. I built it with Claude Code, and more than eighty automated tests were passing before it went live.
The first working piece, the cost calculator, ran on 10 August, and the site went live on 20 August. Those ten days are what AI genuinely changed. What it cost to start, from my own records:
- Server: €9.67 for the first month, renewing at €17.49 a month, because I chose monthly billing. For comparison, Hostinger lists KVM 1 at $6.49 a month on the 24-month term, renewing at $11.99 a month (Hostinger’s VPS pricing page, checked 20 September 2026). Two years up front is cheaper per month, but I had no idea whether the thing would last two months.
- Domain: about €19 a year at DonDominio. I meant to buy the plural form of the name, paid for the singular by mistake, and kept it.
- Data: a third-party service billed per result. I started on its free allowance of $5 of credit a month, then moved to its paid plan at $29 a month, which comes to $35.09 once Spanish VAT is added.
That last line is what I’d have wanted to hear first. The server is the cheap part: the data costs nearly twice as much, every month, whether anybody subscribes or not.
Then the things that went wrong, none of which were in any plan:
- A couple of days after launch, Safari on an iPhone refused to open the site. Caddy’s certificate used Let’s Encrypt’s newer ECDSA chain, ending at ISRG Root X2, which that phone rejected. Forcing an RSA key gave a chain ending at ISRG Root X1, which it accepted.
- On 4 September the registrar suspended the domain, because I had never clicked the registrant verification email it sent in August. The server kept running perfectly for about two weeks while nobody could reach it. I got the domain back on 19 September.
- The data source never reported a car as sold. The code handled “sold” events correctly, but none ever arrived, so long-gone cars stayed listed. A rule that retires anything not seen for fourteen days fixed it, and its first run removed 5,946 stale listings. No model flagged it, and I didn’t think to ask.
- In mid-September the data service hit the monthly spending cap I had set, and scanning stopped until the billing cycle reset.
Where it stands: live again, with no paying customer, and the payment account was still in test mode when I stopped working on it. On 19 September I parked it, along with most of what is on this page. How I keep it and my other sites running is in my working setup for a one-person web business.
What used to be on this page, and what I haven’t tried
The old version of this page, and several others that now redirect here, listed ideas: side hustles for students, weekend projects, low-cost businesses. I won’t rank ideas I haven’t done, so here is the honest split.
- Tried: websites for local businesses, which is my actual job and is covered in what AI changes about running a one-person web studio. Blogs built to earn from ads, a digital product, a book and a small subscription tool, all above.
- Not tried: chatbots for local businesses, faceless video channels, AI stock images and print-on-demand art, paid email subscriptions, prompt packs, freelance transcription or translation, and “AI audits” sold door to door. Some of them may work for somebody. I can’t tell you which, and the old page couldn’t either.
Whether any of this earns anything at all, and what it costs to find out, I tried to answer without numbers I can’t prove in an honest look at making money with ChatGPT.
What the five have in common
AI made the first version cheap, every single time. A hundred posts in eleven weeks, a PDF in an evening, a working subscription product in ten days. If your plan depends on the build being hard, it no longer is, for anyone.
It did nothing for distribution. Every one of the five stalled at the point where strangers have to find it, and no tool I use touches that point.
The costs that mattered were the recurring ones and the waiting. The start-up costs in the table are small. The monthly data bill, renewal prices and the months before anyone reads you are not, and none of them shows up on launch day.
The ones I made to sell did worst. The PDF, the book and a blog built for ad revenue were designed around the money before there was anything in them. The subscription tool was designed around a specific job whose result anyone can check. It hasn’t earned anything either, but it is the only one that taught me something I didn’t already know.
On 19 September I froze everything on this list and put my time into the web studio. A day later I unfroze the two blogs, on one condition: they get rebuilt from what I can actually show. This page is one of the results.
If you’re choosing a first small project: pick the one where you can already name the first ten people who would use it, and write down what it will cost every month after the introductory prices end, before you build anything.