What it is, how answer engines actually decide who to name, how to measure it so the number means something, and which changes move it. Written for founders and marketers, not engineers.
Your buyers used to search and get ten links. A growing share of them now ask an assistant and get one answer that names three or four companies. There is no page two. Either you are named or you do not exist in that conversation, and most brands have never once checked which it is.
The single most useful finding from running this measurement repeatedly: the brand winning AI answers usually does not have a better website than the brand losing them. The gap is almost always in third-party sources, not in the homepage.
Simplifying only slightly, three things happen. The assistant interprets the question and decides what a good answer looks like. It retrieves supporting material, some from its training, much of it fetched live from the web. Then it writes an answer that is consistent with the material it just read.
That third step is the one people miss. The model is not ranking pages against a query. It is summarising a body of evidence. So the question is not whether your page is optimised. It is whether the material the model reads about your category happens to mention you, in a form it can use.
In practice the material that gets cited is rarely a vendor homepage. It is listicles and roundups, review platforms, forum threads where real people compare options, trade press, documentation, and directories. These are the pages that read as neutral, that name several options in one place, and that are structured enough to quote.
Your own site still matters, but mostly in a narrow way: it has to confirm what the third-party sources say and be machine-readable enough that a crawler can extract a claim without guessing.
Measurement is where most of this falls apart, because a single question asked once produces a result that will not reproduce. These models are probabilistic. Ask twice, get two answers.
Write the questions a real buyer would type, not the ones you wish they would. Cover the buying stages: the category question, the comparison question, the objection question, the geography or segment question. Twenty to sixty questions is enough for most categories. Then freeze the set. If you change the questions you lose the ability to compare months.
Askulia scores every question 0 to 2. Zero if you are not named at all. One if you appear among other options. Two if you are named first or recommended outright. Sum it, express it as a percentage of the maximum, and you have a number that is comparable across months and across competitors.
ChatGPT, Perplexity and Gemini retrieve differently and will disagree about you. A single blended score hides the thing you most need to know, which is which engine you are losing and why.
Every zero is also a data point about somebody else. The competitor named most often in your zeros is the most useful single fact in the whole exercise, because their citations are a map of the sources you are missing from.
If you want the number without doing the work yourself, the free score gives you one figure and the competitor being named instead of you. If you want the full picture, the Visibility Report is the one-time version and the done-for-you track is the version where we also do the fixing.
AI visibility is how often, and how favourably, a brand is named in the answers that assistants like ChatGPT, Perplexity and Gemini give to buying questions. It is not a ranking position. There is no list of ten links to sit in. There is one answer that names a handful of companies, and you are either in it or you are not.
No. Classic SEO optimises a page to rank for a query. AI visibility optimises the evidence a model reads before it writes an answer, most of which sits on sites you do not own. The skills overlap but the surface being worked on is different. See the GEO vs SEO comparison for the full breakdown.
With a fixed question set, run across each engine on a schedule, scored on the same rubric every time. Askulia scores each question 0 to 2: 0 if you are not named, 1 if you are named among others, 2 if you are named first or recommended outright. The number is only useful if the questions and the rubric never move.
Because these models are probabilistic and are re-running retrieval each time. One run is an anecdote. This is why measurement uses a fixed set run repeatedly rather than a single check, and why anyone quoting you a single dramatic result is showing you noise.
No, and you should walk away from anyone who says otherwise. Nobody controls the output of a model they do not own. What can be influenced is the evidence available to it, and what can be promised is honest measurement of whether that influence worked.
Getting cited usually depends on third-party sources being updated or created, and those move on other people's timelines. Expect the first measurable movement over weeks, not days, and expect some questions to move long before others.
One score, and the single competitor being named instead of you. Free, no call needed.
Get your free score