Does AI Find My Studio?
It started with an Instagram carousel. It was advertising an online event about being visible on the web, about classic search engine optimisation and about what is now called GEO, which is all about turning up in the answers of AI systems. The first slide read, in German, “Die KI kauft ein und nicht mehr der Mensch”, roughly: AI does the shopping now, not people. It was said by Matthias Wesselmann, named in that release as CEO of the communications agency Ketchum in Germany, in a press release from March 2026. I wanted to know what lies beneath it. On October 27, 2026 that turned into a day of research with my digital team, and by the end there were 127 sources on the list.
What came out of it is set out in more detail in a journal note in German, “Die KI kauft nicht. Sie sortiert vor.” (AI doesn't buy. It pre-sorts.), and the title sums it up rather well. The surveys from 2026 mainly measure what people can imagine and where they let AI help them search and compare; none of them measures an AI buying without asking first. What AI already does happens one step earlier: whoever asks it gets a selection before clicking on anything themselves. Whether clients in Germany go about it this way when they are looking for someone for design or photography is not examined by any study we found. The research on visibility in AI answers, as far as we read it, works with English queries, with big brands or with consumer goods. So for a studio like mine there are no figures I could rely on, and if I want to know whether an AI finds it, I have to look for myself.
The question isn't entirely new, though. On July 15, 2026, I had already asked ChatGPT with web search and Perplexity about my studio, one round each. Both found my website, named it as a source and correctly captured all three areas of my work, art, design and photography, Perplexity even in great detail. Perplexity, however, also knew of a separate website for my art that no longer exists in that form; its address now only forwards to my main site. Nothing about it was on my site, the information came from old data.
And then there was a hidden test word. At the time it sat in the website's source code, in a comment and in a meta tag, places nobody ever gets to see in a browser. Neither of the two AIs mentioned it, even though Perplexity explicitly looked for it. That speaks more for the hiding place than against the AI. Apparently what gets read is what a person can read on the page as well.
For a first look that was pleasing, but it isn't an answer to my question. One round per system is not much, and the questions came from a check report in my analysis tool that mainly looks at whether name, address and phone number are given correctly; most of those questions already mention my name or my website. Anyone asking like that already knows who they are looking for. A different question is whether an AI names me when someone has never heard of me and only knows what they need: a visual identity for their company, art from the area, photographs of a new building or a tool that fits the way their business works. That is the question now.
For the experiment to yield more than two snapshots, it needs a set-up that can be checked. The questions are put the way clients might put them, in their own words as far as possible: which contemporary artists work in Celle, who builds small businesses their own tools with AI, how a company can present itself consistently on its website and on social media, or who photographs architecture around Celle and Hanover. Whether I really hit the words clients use, I can't say for sure; I lack data on how they search. There is also a question about me, so that it can be checked whether what is said about me is right. The questions go to several systems; planned are Google Search with its AI Overviews, ChatGPT with web search and Perplexity, starting with six questions.
The asking itself, of all things, is not automated. Otherwise I like building tools that take manual steps off my hands, and I would rather describe what I want than click my way through menus. Here I still ask myself, by hand, on the same interfaces a client sees. OpenAI's terms of use prohibit extracting answers by automated means, and Google restricts automated access too, and what a program retrieves through a provider's interface does not necessarily look like the answer on the screen. So every query starts in a new private browser window, logged out, so that neither an account nor a search history gets a say. Whether every system can be used without signing in is something a trial run will show first. The cookie banner gets a “Reject all” every time, the question is pasted from a list rather than typed, so that not a single comma goes astray, and the answer is recorded with date and time, once it is settled how long such answers are kept. Then the window closes and the next query starts from scratch.
My digital team prepares the list, extracts text and links from every saved answer and writes everything into a log that checks each line for missing details. For each file it calculates a checksum that will still prove, even after deletion, that the answer existed and how it was counted. Anything disputed I decide, and before any result appears here, a reviewer who wasn't involved in the counting looks at samples.
Each question goes to each system five times; two dates a few days apart are planned. A single answer from an AI says little. Researchers at Ruhr University Bochum and the Max Planck Institute for Software Systems compared which pages search systems link to, in summer 2025 and again in September, using English queries. In Google's AI Overviews only 18 percent of the linked pages were still the same the second time, compared with 45 percent in ordinary Google search. A research overview from July 2026 names seven to eight repetitions as a starting point, our research via Perplexity came up with three, and none of our sources sets a binding minimum. I decided on five, which sits almost exactly in between.
Five answers still aren't many, and it's worth working out honestly how few. If my studio appears in none of the five, that doesn't mean “never”, only: probably in fewer than about 43 percent of cases. Zero hits, and in reality it could still be a good two in five answers. If it appears in all five, the share is probably above about 57 percent, but not certainly a hundred. Two in five and three in five can hardly be told apart at all. The calculation behind this is called the Wilson interval, and it is designed not to suggest more certainty than there is when the numbers are small. That is why I will later write “two in five” and never “40 percent”, and between the two dates only an almost complete swing counts as more than chance.
What gets counted is whether an answer names my studio, whether it links to my website, whether it gives the site as the source for a specific statement that is actually there, and whether what it says about me is right. Right only means that it matches details I confirmed beforehand in a fact file; whatever is missing there stays open instead of being guessed. If no answer comes, or Google shows no AI Overview at all that day, this is noted separately and not counted as “not named”.
What the experiment cannot deliver is part of it too. There will be no third place in AI search to report afterwards, no market share either, and nothing about whether anyone gets in touch because of it. On this basis I can't promise anyone anything, myself included. If an answer names colleagues, I count them but don't name them here. And the results will be published even if I'm not in them.
Part 2 follows after the first measurement date. Until then, I know about as much about my studio and AI search as the answers from July can tell me.