--- title: "How Do You Measure Whether An AI Knows Your Site?" description: "Search engines give you numbers. Language models barely do. We built a tool that asks the question instead of asserting the answer." author: "Matthias Meyer" published: 2026-07-06 updated: 2026-07-06 language: en tags: ["maschinenraum", "geo", "ki-sichtbarkeit", "messung", "eigenprojekte"] canonical: "https://studiomeyer.io/en/blog/geo-werkzeug-sichtbarkeit-messen" markdown_versions: ["https://studiomeyer.io/de/blog/geo-werkzeug-sichtbarkeit-messen.md", "https://studiomeyer.io/en/blog/geo-werkzeug-sichtbarkeit-messen.md", "https://studiomeyer.io/es/blog/geo-werkzeug-sichtbarkeit-messen.md"] publisher: "StudioMeyer, https://studiomeyer.io (llms.txt: https://studiomeyer.io/llms.txt)" --- # How Do You Measure Whether An AI Knows Your Site? With search engines the situation is comfortable. There are tools showing which terms a site appears for, at which position, how often it gets clicked. You can argue about accuracy, but there are numbers. With language models there is almost none of that. When somebody asks ChatGPT about a boating school on Mallorca, the site owner sees nothing of it. No hit in the log, no notification, nothing. The answer forms, the user reads it, and whether a particular site appeared in it is known to nobody but the user. That is a real problem, because part of search is shifting exactly there right now. ## The Only Honest Method Is To Ask You cannot measure what you cannot see. So you ask yourself. The tool puts to the models exactly the questions a customer would ask. Not "do you know this company", because a model will almost always answer that pleasantly. But the open question somebody would actually type: which provider for this thing in this area. Then the answer gets evaluated. Does the site appear. In which position. Who stands next to it. And with which description, because being named with a wrong description is worse than not being named. That is a sample, not a census, and it is reported as exactly that. Anyone claiming otherwise is selling a precision that does not exist. ## The Second Half: Is The Site Readable At All The second part of the tool does not look outward but at the site itself. Can a machine take in the content at all. Are opening hours, prices and offers marked up as structured data or do they sit there as running text. Is there a plain-text version for language models. Are the language versions cleanly separated. Does the site accidentally block the crawlers of the AI providers. This part is the more thankless and the more important one. A site can have excellent content: if a model cannot read it cleanly, it does not exist in that world. And it is the part you can actually influence. Whether a model names somebody is the model's decision. Whether the preconditions are right is yours. ## What We Explicitly Do Not Promise There is a market by now for promises in this direction. Guaranteed mention in AI answers, position one at ChatGPT, optimisation for a specific model. That is unserious, for one simple reason: nobody outside those companies knows how the selection comes about, and it changes with every model version. Guaranteeing a mention means guaranteeing something you do not control. What can honestly be said is: we make sure everything readable is easy to read, we measure regularly, and we show the curve. That is less than a promise and more than most deliver. ## Why This Has To Be A Tool Done by hand, a measurement like this has two drawbacks, both expensive. It costs time every time, so you do it too rarely. And it is not comparable, because the questions are phrased slightly differently each round. A comparison over time needs identical questions, otherwise you are measuring your own phrasing. So, a tool: same questions, fixed grid, history. It runs for our own site, for client sites, and can be subscribed to. ## The Case Nobody Expects The more awkward possibility is not "does not appear" but "appears, but wrong": the model names a business but files it under the wrong industry, quotes a service that no longer exists, or confuses it with a similarly named one. That is the good news in bad packaging. Being named wrongly is easier to repair than not appearing at all. In both cases the route is the same: make sure the unambiguous, structured answer is easier to find than the poor guess. ## Read more, Engine Room - [The Engine Room: What We Run Ourselves](https://studiomeyer.io/en/blog/maschinenraum-was-wir-selbst-betreiben.md) - [What We Actually Work With](https://studiomeyer.io/en/blog/arbeitssystem-editor-modelle-werkzeuge.md) - [An Agent Fleet, Sorted By Job](https://studiomeyer.io/en/blog/agenten-flotte-wer-macht-was.md)