Checkin's marketing department is two people. In ninety days they more than tripled their AI visibility score, and ten percent of new customers now say they found them through AI. Here is what they did, in what order, and the part their CMO says they should have started a year earlier.
Start with what sent them looking. Checkin handles registration and event finances for conferences and festivals in Norway and Denmark, and more than four thousand organisers use it. Specific enough, you would think. The models thought otherwise.
ChatGPT was quoting them, and what it told people about the company was flatly wrong. So they wrote a file for the AI models to read. It listed the founding year, the products, the prices and the customers. Then it listed the part almost nobody thinks to include: what the company is not. Erik Hedding-Reijrink, who did the work, told Caroline Bræin Klette on The GEO Shortlist that they had to state that boundary clearly, because otherwise the models kept inventing one. The file sits on their website, and no human will ever read it.
A description you did not write
Search engines never asked you to describe yourself. They asked you to be findable. Then they handed the visitor a link and stepped aside, and the visitor decided what to think.
AI models decide for the visitor. They produce a description of your company whether or not you gave them one. They build it from whatever they can find: your own pages, an old directory entry, a competitor's comparison page, a forum thread where somebody confused you with a different company. The description exists either way. The only question is how much of it came from you.
Camilla Reinhardsen, CMO at Checkin, calls this marketing to machines alongside marketing to humans. Two audiences. One of them reads structure rather than sentences, and it never tells you when it has misunderstood.
It also answers differently from one time to the next. Erik noticed this early. Ask a model the same question twice and the second answer can look nothing like the first. That single fact rules out checking by hand. You are looking at a range of possible answers, so you have to ask many times before a pattern appears. A screenshot of one good answer proves that a good answer is possible. It proves nothing else.
And a model that has never heard of you still answers the question. It answers with somebody else's name.
The half you can finish
What followed at Checkin was ninety days of technical work, and the order mattered more than the individual tasks.
They started by checking whether AI companies were allowed to read the site at all. It takes about two minutes, and everything after that step is wasted if the answer is no. Then they added schema markup, which is code placed on a page to tell a machine what kind of page it is and who wrote it. They did this across all four language versions of the site. They wrote an llms.txt file, a plain text summary written for AI models, and that is where the section on what the company is not ended up. They cut the loading time on a slow mobile connection from 18.2 seconds down to 1.2.
Over those ninety days, their AI visibility score more than tripled.
Real work, and it moved a real number. It also deserves an accurate name. Everything on that list makes a company easy to read. Machines can now reach the site and sort it correctly. The score rose because the models stopped guessing.
It is worth asking what a score like that measures. When a platform gives you a checklist and then scores you on that checklist, a rising number confirms that you completed the tasks. Useful. Not the same as a buyer choosing you.
Erik used a good image for the order of the work. You cannot decorate a Christmas tree before you have put it in its stand. He is right about the order. Getting the stand right is still not the same as having a tree.
The half you have to earn
Caroline asked both guests what they would do differently if they started again. Erik said he would build the technical foundation sooner. Camilla answered something else entirely.
She said they should have worked much harder on getting customers to recommend them. Getting mentioned on other people's websites. Making content together with customers and letting those customers tell the story in their own words. Setting up their profile properly on G2, the software review site, instead of leaving it half finished.
Read that list again. None of it is work they can do alone.
This is the second half of the work, and it decides whether a model recommends you or merely describes you correctly. When a buyer asks for three suppliers for a job, the model does not open your llms.txt file to pick the names. It draws on what independent sources say about companies in your category.
Ahrefs, which analyses links and search data, put 15,000 long tail questions to ChatGPT, Gemini, Copilot and Perplexity, then compared what those models cited against what Google and Bing ranked for the same questions. The overlap averaged around one citation in ten. The models are pulling from a different pool of sources than the one most marketing teams have spent ten years optimising.
You can fix your own code in an afternoon. You cannot write your own reviews. That difference is the whole reason the second half is slower and worth more.
Checkin seem to know it. Their next steps are a Wikipedia entry and more reviews on the review sites, which is the right direction and the part that cannot be rushed. And that ten percent of new customers who say they came from AI counts the people who arrived. It says nothing about the buyers who asked for three options, received three names, and never saw Checkin among them.
What ninety days buys
The useful thing about this story is that it comes from a company that did the work properly and is honest about where they are. Erik's advice to anyone who feels behind was that there is still cake left at the party. Fair enough.
But their own answers mark the line better than any framework would. The first half of AI visibility is a project. It has a checklist and an end date. You can finish it, and finishing feels like arriving.
The second half never ends. It grows or it stalls, depending on how many other people are willing to say your name in public. Ninety days buys you the right to be described correctly. What you do next decides whether anyone recommends you.
