Category Archives: ai

Investigating ChatGPT for advertising my software

Leave a reply

There is a common pattern with paid digital advertising channels. New platforms appear with opportunities for cheap ads. Over time, more and more advertisers start to use the platform. Supply and demand drives up the price per click. The platform owners also do everything they can to nudge click prices ever higher. Consequently ad prices rise until companies selling inexpensive products (like mine) can’t afford to bid high enough to get any clicks. But usually a new platform comes along and the dance starts again.

I’ve seen this play out over the 20 odd years that I have been using Google Adwords. In the early days I could get a decent number of clicks at an affordable price. But the price has risen now to the point where I get very few clicks for any price I am prepared to pay. I need a new advertising channel. I tried advertising on Reddit, but that was a resounding failure. I wondered if it might be worth advertising in the new hotness, ChatGPT. So I did some investigating. Here is what I have found out so far, from reading their documentation and other sources.

ChatGPT advertising is structured in a similar way to Google Adwords, with campaigns, ads, auction bids and conversion tracking. But, instead of matching keywords in search terms, you describe contexts in which your ads should appear. This should be less hassle than defining hundreds of search keywords. But it is hard to know how well targetted the ads will be or how well they will convert into sales. Some experimentation is required to answer that.

The ad format is fairly simple: name, headline, short description and small a image.

You can currently only advertise to customers on free plans based in USA, Canada, Australia and New Zealand.

In their own documentation, ChatGPT says: “Advertisers can set custom max bids for their CPC campaigns. We recommend a starting max bid of $3-5 USD per click.”. Yikes. I know they have massive costs to subsidize, but there is no way I can make a profit at $3 per click for my $99 data wrangling software Easy Data Transform. Given a typical 1% conversion rate I would be paying $300 per sale. However, bid recommendations are always very self serving, and can be taken with a large pinch of salt. It is likely that you can get clicks much cheaper. Especially given that there are currently relatively few advertisers compared to the number of users.

So far, so good.

The fly in the ointment is the minimum spend. $50k (down from $250k!). Ah. Maybe not. That minimum commitment may come down over time. But, by the time it is low enough for me to experiment, the bids will almost certainly be too expensive for it to be profitable to someone selling $99 software licenses. The search for affordable advertising channels continues.

** Update 29-May-2026 **

There are reports that the $50k minimum spend has now been dropped. However advertising is only open to US advertisers. More details.

Support is wild in an age of AI

4 Replies

Recently a customer emailed me:

Please consider re adding the option to convert from one time zone to another as was available in [Easy Data Transform] version 1.x. See the attached screen dump.

Easy Data Transform has never had a time zone conversion feature, so that is a bit strange. Although . But the screenshot really set off alarms bells as, despite saying “Easy Data Transform v1.11.2” in the title bar, that is not our software!

I thought that someone was trying to pass of their product as ours and did a search. But I couldn’t find any reference to another piece of software called “Easy Data Transform”. I emailed the customer to ask where they had got the software from. The reply came back:

The screenshot was from a ChatGPT output – maybe hallucinating.

Wow. Not only had ChatGPT hallucinated the feature, but also a fever dream screenshot of the user interface, with the non-existent feature. It looks like a real screenshot at first glance, but the icons are a giveaway if you look closer.

For reference, this is what the actual user interface looks like:

ChatGPT has got some of the input types, transforms and menus correct. But otherwise it looks quite different.

I had another brush with AI hallucinations when I asked MS Copilot how to perform ‘one hot encoding’ in Easy Data Transform. It came back with a very plausible and confident sounding answer, including this summary of the transforms required:

Just one problem – only 1 of these 4 transforms actually exists (Split Col). It hallucinated the other 3!

Customers are increasingly typing questions into AIs, rather than reading documentation or asking on a forum. That is good news if it means that the customer gets a quick and accurate answer without troubling busy developers. But it is very bad news if they are getting incorrect answers, especially when these answers look plausible and are confidently presented. It is galling enough that AIs are stealing all our web traffic, without them giving our customers bad support advice as well! Wild times are ahead.

Ps/ Time zone conversion and one hot encoding are now available in the latest version of Easy Data Transform.

A quick experiment with generative AI

WordPress offered to generate an image for my last blog post. Here is the prompt it suggested:

“Generate a high-resolution, highly detailed image capturing the essence of “20 Years of PerfectTablePlan Software.” The main subjects should be two screenshots side-by-side: one showcasing PerfectTablePlan version 1, reflecting a vintage desktop interface with a Windows aesthetic from 2005, and the other displaying version 7 with a sleek, modern design. The lighting should be bright and inviting, emphasizing the contrast between the older and newer software. The style should blend nostalgia with innovation, showcasing the journey of the product over two decades. Ensure the image has sharp focus and intricate details to attract the reader’s attention.”

And here 5 images it came up with, from that prompt:

They are simulateously very impressive and hilariously awful. Quite apart from the weird text (“sex 20”?), none of the screenshots look even slightly like PerfectTablePlan. I think I’ll pass!

The AI bullshit singularity

I’m sure we are all familiar with the idea of a technological singularity. Humans create an AI that is smart enough to create an even smarter successor. That successor then creates an even smarter successor. The process accelerates through a positive feedback loop, until we reach a technological singularity, where puny human intelligence is quickly left far behind.

Some people seem to think that Large Language Models could be the start of this process. We train the LLMs on vast corpuses of human knowledge. The LLMs then help humans create new knowledge, which is then used to train the next generation of LLMs. Singularity, here we come!

But I don’t think so. Human nature being what it is, LLMs are inevitably going to be used to churn out vast amount of low quality ‘content’ for SEO and other commercial purposes. LLM nature being what it is, a lot of this content is going to be hallucinated. In otherwords, bullshit. Given that LLMs can generate content vastly faster than humans can, we could quickly end up with an Internet that is mostly bullshit. Which will then be used to train the next generation of LLM. We will eventually reach a bullshit singularlity, where it is almost impossible to work out whether anything on the Internet is true. Enshittification at scale. Well done us.