I have been really interested in AI recently. So I thought now was a good time to see if it could build a business for me.
A business built by AI and using AI.
The Idea
So in this post, I will attempt to build a ‘business’, or rather a money-generating website, using AI. And then to complete the circle the website itself will use AI.
I need a business that can be contained in a website. And the business should use AI to create some value that we can monetize. One thing I have been doing recently is using ChatGPT to turn my images into colouring-in sheets that I can print off and give to my daughter. Here is an example:
So my simple idea is a website where you upload a photograph and it gets turned into a colouring-in picture for you. Then we monetize it in some way.
Getting AI To Build The Website
There are many AI’s that could be used to do the programming for the website. I have a computer science background and normally use Cursor with Claude for my AI aided development. But for the sake of this experiment I didn’t want to do any programming at all. Everything must be done by the AI and possible by someone with no coding background at all.
After a bit of research I decided to use Google AI Studio. It’s free to use and I’d heard good things about the results.
I started by typing a pretty simple prompt to see what it came up with.
It is pretty impressive, you can watch it programming in front of you and the end result ‘worked’ even if it wasn’t exactly what you wanted.
Once the website is created you can give it instructions on what to change or add. And eventually I ended up with something I quite liked:
You can see what’s going on. It’s a pretty simple website, you upload your photo and click generate and in exchange get a nice simple image that is reminiscent of the original picture. But why is the resultant image so different from original? You can tell it is related but it is almost like it was made from someone describing the original photo without ever seeing it itself.
Well it turns out that is exactly what it is doing. Google AI Studios decided to use it’s own AI, Gemini. There currently isn’t an option to use an image as the prompt for image generation. So what decides to do is feed in our original image into the chat machine and get a detailed description. Then use that description to create the final image. Here are the prompts it uses.
Analyze the provided image carefully. Generate an extremely detailed factual description of its main subject(s), essential elements, and overall composition. Focus on objective details that would be necessary to recreate it as a simple line drawing for a child’s coloring book. It needs to be detailed enough that the subjects are recognizable. Make sure to describe the positioning of the elements in the image exactly, for instance the exact angles they are facing and exactly where they are positioned relative to each other. You can use n object of measurement such as pixels to give positioning and sizing. Avoid mentioning colors, shading, complex textures, or artistic styles. Output only the description text, without any introductory phrases like “The image shows…” or “This is a picture of…”. Just the raw description. Include precise body positioning, specific features of clothing and accessories.
Then it feeds the result into an image generator:
Generate a black and white line art coloring book page specifically designed for young children (ages 3-7). The image should feature: ${description}. Ensure the drawing has very clear, bold outlines, simplified shapes, and absolutely no shading, gradients, or intricate textures. The style must be clean, uncluttered, and easy for a small child to color within the lines. The background should be simple or white.
I played with it a bit to get a more consistent result. This is the sort of results we ended up with:
This is ok and the result is fine for what I want. My daughter will happily colour any of those in and will understand the reference.
Looking into more options we could use ChatGPT from OpenAI to get something better. But If I use Gemini in Google AI Studio it is free! And we really need to keep costs down if we want to make this a viable business. Which brings me on to the hardest part of all this.
Working Out The Costs
Using Google AI Studios is free and so is using the Gemini AI inside the studio. But if we want to publish it to a website that others can access then we need to start paying, and obviously we do!
You can do a 1-click deploy to a real website using Google Cloud, which I did. It has pretty generous free tier limits and the website would have to get pretty popular before I needed to worry about them. But the AI on the other hand isn’t free. The input and text results actually have a generous free tier, and once we’re past it the AI queries are cheap enough not to be a worry. The text above that we parse in is about 170 tokens and the image about 258 tokens – 1 million tokens is about $0.30, so you get the idea.
But the image generation on the other hand is quite expensive and is not available for free.
3c an image doesn’t sound like much but it would very quickly add up if we don’t have some way to recouperate that from visitors to the website. For reference this website gets about 1000 visitors a day, if it cost me 3c+ per visitor it wouldn’t be worth it.
I looked into some free or cheaper AIs. I could host something like stable diffusion myself, or I could using a free service. Hugging Face has hosts who let you use a specific model. But unfortunately none were able to create something good enough for my needs. Here’s an example:
So instead let’s look at how we can make money given the relatively high costs.
Freemium Model
My ideal business model would be microtransactions. I remember reading a book by Bill Gates 20 years ago that predicted in the future most business online would be microtransactions. Pay 0.01p to send an email, that sort of thing. That hasn’t really happened and still the only way to spend money online is painfully.
Instead the way to get microtransactions is through advertising. Where we get paid some small amount per impression or click. Unfortunately the amount we’d earn from standard display advertising is far lower than what we need, more like 0.1c a visitor.
The model I decided to trial was to let the user generate the image for free upfront, costing me about 3c, and then putting a watermark on it and letting the user choose to pay £1 if they wanted to remove it. This was simple enough to add, I just asked in Google AI studio and they added the watermark and a Stripe integration which lets them pay for it. Here is the finished result:
Viable Business?
After doing some focus testing I decided that this business model wasn’t going to work, I need about 1 out of every 30 generations to convert to a sale and I don’t think we’d get that. Especially considering the inconsistent results and so each user would probably generate a few different versions before deciding on one to buy.
Given the current cost of AI I’d probably need to create my own specialised model to do it economically, and that’s a bit too much work for this test project.
For fun I made a simpler one that you can use for free. It gets the AI to create the colouring-in image using text rather than a real image. I quite like how goofy the results look! Because it doesn’t generate an image we can use the Gemini free tier and some free hosting from Vercel.
You can use my AI Coloring Page Creator here.
