What 1,069 clicks on ChatGPT ads actually delivered
In September 2026 we ran ads inside ChatGPT. Six markets, two weeks, 183 euros. It produced 1,069 clicks and zero orders. This piece is the open ledger of that experiment: what we paid, what happened, and which mistake was entirely ours.
There is still very little independent data on advertising inside AI assistants. The platform is young, and most of what is written about it comes from parties with something to sell. So we are publishing our own numbers, including the part that makes us look bad. Anyone weighing the same decision is better served by an honest measurement than by a success story.
The numbers, unvarnished
Spread across the Netherlands, Germany, the United Kingdom, the United States, Canada, Australia and India:
- 1,069 clicks
- 183 euros spent
- 28 people started something
- 3 reached a finished design
- 0 orders
The cost per click varied wildly. In India a click cost 2 cents, in the United Kingdom 3.80 euros. Those same 24 British clicks cost 91 euros in a single day. India delivered 712 clicks, three quarters of all our traffic, and not one order.
That gap matters more than the total. Cheap traffic is not the same as cheap customers. Optimise for the click price alone and you buy exactly the audience that intends nothing.
The mistake we made ourselves
When we took the campaigns apart, every ad group turned out to have an empty field called context_hints. That is where you declare which kinds of conversations your ad may appear beside. All eighteen of our ad groups sat at zero.
So our ad could surface next to a conversation about tax returns, a recipe, or a block of code. People clicked out of curiosity and left again. In hindsight that explains the pattern better than any other hypothesis: plenty of clicks, almost nobody who actually began anything.
What makes systems like this awkward is that they fail quietly. The interface accepts an empty list without a warning. And when we did fill the field in, the API echoed the new value back while a check immediately afterwards still showed zero. That was lag rather than refusal, but you have to check twice before you draw the right conclusion.
The contrast that surprised us most
During exactly the same period, people also arrived because ChatGPT had pointed them to us on its own, with no ad involved. No cost, no campaign. That traffic produced 94 people who started a book, and 23 orders.
Same place, same kind of people, and the difference between zero and 23.
For us that explains the whole experiment. Someone who asks an assistant where to find a personal gift and gets an answer is searching. Someone busy with something else entirely, who sees a card beside it, is not. The ad interrupts; the recommendation answers.
That is not a new idea in marketing, but it shows up more sharply inside an assistant than in a search engine. On Google an ad sits among results for someone who was already looking. In a conversation it sits beside something else.
When this could still work
Our outcome is not a general verdict. There are situations where advertising inside an assistant probably lands better than it did for us.
For a low-friction product, where buying is a matter of seconds, interruption costs less. We ask for the opposite: you enter names, choose a photo, approve a storyline, and only then pay. That is an investment of attention, and you will not get it from someone who was in the middle of something else.
It also matters that we left the conversation-topic field empty. Anyone who fills it in properly from day one is running a different experiment than we did. Treat our 183 euros as a measurement of broad targeting, not of the platform as a whole.
What we do instead
The comparison that taught us the most was with Google Shopping. Our books appear there in the free product listings. That channel costs nothing and produces more orders per month than the entire advertising experiment combined.
If you are facing the same choice: start with the channels where people are already looking. A comparison of suppliers, a search for a personalised book with your child’s name, a recommendation from someone else. It is slower than buying clicks, but the audience is right.
So we are stopping paid advertising inside ChatGPT. Not because the platform does not work, but because for us it works precisely in the way that costs nothing. If you are curious how we handle AI in the making itself, we wrote earlier about what happens to your child’s photo once it enters our pipeline.
The lesson that stays: never read clicks as a result. Two days in we had hundreds of clicks and could have told ourselves it was working. Only when we counted through to the number of people who actually began something did it become clear that nothing was happening. That number, not the click, is what you steer on.