Last week I built the first campaign for Sfinx in ChatGPT’s Ads Manager. We have not pressed start yet, so this is not a case study. No results to report, no cost per lead to celebrate, no graph pointing upwards. But it will come soon, because ChatGPT will soon have a given place in our media plan.
The moment, and its limits
Most advertising finds a person and shows them something later. A banner catches somebody who read about your subject three days ago and reaches them halfway through a news article. Search gets closer, because the person is looking right now, but it compresses their situation into the few words they typed.
In ChatGPT the ad appears underneath the answer, inside the conversation, while the problem is still open. Nobody has changed context. Nobody has moved on. That is what you are actually paying for.
It also explains the targeting. There is no job title targeting, and for a campaign aimed at Sweden you cannot use a customer or prospect list either. At ad group level you get a free text field for context hints, where you describe the conversations, needs and situations your offer may be relevant to.
Not a keyword list. Context.
One limitation is important, especially in B2B. Ads are currently shown on Free and Go, not on Plus, Pro, Business, Enterprise or Edu, and some of the senior people we would most like to reach are probably using those ad-free plans.
So this replaces nothing. LinkedIn helps us decide who to reach. Search catches somebody already looking for an answer or a supplier. ChatGPT reaches them while they are still working the problem out. Three different jobs, and the third is a moment we have not really been able to buy in quite this way before.
Which makes it a writing job
That is a real shift. Most channels reward budget, bidding and the discipline to keep the data tidy. This one rewards precision of formulation, which means the people who write for a living belong in the room while it is built. OpenAI’s own guidance points in the same direction: context hints should describe real needs and situations in clear, natural language, rather than broad audience labels or disconnected terms. The better you can describe when your offer becomes useful, the more useful the hint becomes.
One thing shapes all of it. Google gives you a Search Terms report and you refine from there. ChatGPT does not tell advertisers which conversations led to an impression or click. More of the qualitative research has to happen before launch.
So we looked where our buyers have already told us things. Searches that brought people to the site, questions from first meetings, why our own articles get read. Not a keyword list. A record of how people phrase the problem when nobody is selling to them.
That gave us four groups. What follows is not twelve phrases we expect ChatGPT to match word for word. They are ways of expressing the situations we want the hints, the ads and the pages to reflect.
The first is for people who have concluded that something is not working as it should:
- Our marketing is not producing enough business impact.
- How do we market a product that takes twenty minutes to explain?
- How do we build demand when hardly anyone is searching for what we sell?
The second is about growing across borders:
- How do we enter a new European market?
- Should our international marketing be centralized or run locally?
- How do we build awareness for a niche product in several markets?
The third is quieter. These are questions marketing people ask themselves rather than their agency:
- Does brand building actually work in B2B?
- How much should we spend on brand versus lead generation?
- How do we reach the people who are not ready to buy yet?
And a fourth, for those who have already got as far as looking for help:
- Do we need a strategic agency or a performance agency?
- How do we find one that understands technically complex products?
- When is it time to change?
Read them together and they are not four audiences. They are four moments in the same person’s year. Most of the time they are in the first three, working something out and telling nobody. The fourth is the short window when they actively look for help, and it is the easiest one to reach with traditional demand-capture media.
None of them were new to us. We have heard every one across a table for years. The work was choosing which to build around, and writing them the way a buyer says it rather than the way an agency summarizes it. That distinction is harder than it sounds.
We are starting in Sweden, which sounds odd given that the second group is about international growth. But the person asking can be sitting in Helsingborg while the problem is Germany, or twenty markets at once. The geography of the buyer and the geography of the problem are not the same thing.
One detail worth knowing. Context hints are optional. Even without them, matching still uses the conversation itself, your ad copy and your landing page. So the website is one of the signals that determines relevance, not just where the click lands. Our services page describes everything we do, which is right for a company overview and wrong for somebody who has just described one problem to an AI.
The decision I would not delegate
In our account, a setting called text adaptation was switched on by default. It lets AI produce adapted and translated versions of your headlines and descriptions, and those versions can be published without you seeing them first.
If you are responsible for how your company sounds, that is your decision and nobody else’s. We have turned it off for the first round, simply so we know what was actually shown. It can go back on later, once there is something to compare against.
What we expect to learn
We will not learn what people actually typed. We will see which of the four problem areas gets traction and what people do once they arrive.
If a group goes quiet, we look at the hints before we look at the budget. They might describe the problem in the wrong way, or describe a problem too few people are having. Relevance first. Page second. Money last.
Some of the hints we build from those twelve situations will change, but not because ChatGPT tells us to. It will be because the wrong companies are turning up on the page. We watch that in our analytics and in Leadinfo, which helps us identify some of the companies visiting even when nobody fills in a form. That signal comes from outside the channel, and for us it matters more than a cheap click.
The work starts with those situations and there is nowhere to hide if you get them wrong. No uploaded audience to fall back on. No job title proving you found the right person. Just the situation you described, and whether it is close enough to something somebody is going through.
So, why not try advertising in ChatGPT…
If you want to discuss effective marketing, just reach out to:
Joakim Ebstein
Head of Digital Marketing
+46 725 555 984
joakim.ebstein@sfinxagency.se
