How we check whether AI recommends your company

What we are measuring

Buyers increasingly ask an AI assistant before they search: "which accounting firm in Ljubljana handles foreign founders", "who builds industrial cold rooms in Slovenia". The engine answers with two or three names. Either you are one of them or the conversation never reaches you.

We call the share of those answers that include you AI Choice Visibility. It is measured on a fixed list of buyer questions, across several engines, on a known date. Nothing else on this site claims more than that number supports.

Step one: the unprompted question

We write the questions the way a buyer would, in the buyer's language, without your name in them. Typically 8–10 core questions cover one service in one market. The list is locked before the first run; a changed wording is a new question, not an edit.

Each question goes to each engine — ChatGPT with web search, Gemini with grounding, Claude with web search, Perplexity, and Google's AI Mode where we can capture it — two or three times. The set follows where buyers actually ask: by web visits in August 2026, ChatGPT, Gemini and Claude together carry about nine in ten AI-chat sessions (Similarweb). For every answer we store:

Step two: the brand question

Only after step one do we ask about you by name. If the engine describes you accurately, it knows you; the problem is not awareness but choice. If it describes you wrongly or not at all, the problem starts earlier — with the facts available about you on the public web.

The distance between step one and step two is what we call the unprompted choice gap. It is the single number a business owner needs from an AI visibility check, and it is the number we report before and after any work.

Why there is no shortcut

Google's own documentation says that AI Overviews and AI Mode draw on the same crawling, indexing and ranking systems as classic Search, and that no special markup is needed to appear in them. The controllable inputs are the ones that always mattered: can the engine read the page, are the facts consistent everywhere they appear, and do independent sources say the same thing about you.

Academic work points the same way. In the first controlled study of generative engine optimisation, adding citations and statistics to a page raised its visibility in generated answers by roughly a third, while keyword stuffing lowered it. Engines reward pages that look like evidence.

What you get from the check

A dated table of answers, the list of companies the engines choose instead of you, the sources behind those choices, and three to five causes we can actually act on. No composite score. If we cannot see a cause we can fix, we say so.