AI visibility is how often, in which position and with what tone a brand appears in the answers generated by assistants such as ChatGPT, Gemini, Perplexity and Google AI Overviews. It is not a search ranking: a brand is either named inside the answer text or it is absent from the decision entirely. Rixe measures all three dimensions daily and reports them as one visibility score per category.
The reason this behaves nothing like a search ranking is that there is no second page. A results page gives ten links and a user who scrolls; an AI answer gives one paragraph and three or four names. There is no position eleven to climb from. A brand is either inside that paragraph or it played no part in the decision at all.
The three dimensions move independently, which is why measuring only one is misleading. A brand can appear in 60% of answers and still lose, because it is named fifth and described as complicated to set up. Another appears in 25% and wins the ones it appears in, because it is named first and described as the fastest option. Presence gets you considered, rank decides whether you are remembered, and tone decides what the reader concludes.
What moves these numbers is rarely your own homepage. The model assembles its answer from the sources it trusts about you — review profiles, comparison articles, forum threads, documentation. Changing what those sources say is the work; measuring is only how you find out which of them matters.
Why it is a rate, not a rank
There is no second page in an answer. A results page offers ten links and a reader who scrolls; a generated answer offers three or four brands in a paragraph, and everything below that line does not exist. So the useful question is not where you rank but how often you are in the paragraph at all.
That single fact changes what a measurement has to look like. A rank is one number on one page at one moment; a rate is a share of many answers and therefore needs a sample, an interval and a date. A brand named in eleven answers out of forty is at twenty-eight per cent, plus or minus a margin you should be told about.
What actually moves it
Three things do most of the work, and only one of them is your website. The sources a model retrieves from — review platforms, comparison articles, forum threads — decide most of what it can say about you. The structure of your own pages decides whether a useful sentence can be lifted from them. And the consistency of your basic facts across the web decides whether the model is confident it is talking about one company.
None of the three moves overnight, and a programme that expects it to will misread its own data. A page rewritten today is re-read when the crawler next comes; a new listing on a source enters the retrieval pool later still. Four weeks is the earliest honest window to judge a change, which is why measurement and publishing belong on the same calendar.
Putting a number on it
A defensible reading needs four decisions written down: which questions count, which engines and markets are in scope, how many times each is run, and what counts as being named. Change any of them mid-quarter and the trend line becomes an artefact of your own settings rather than a fact about the category.
In practice the rhythm that survives contact with a real team is weekly, not daily. Engines are probabilistic and a two-point move on a Tuesday is usually noise; a four-week window with the same questions and the same engines is where a change becomes something you can attribute to work somebody did.
+2
