What is AI visibility?
AI visibility is how often, how early and how kindly AI assistants bring up your brand when someone asks a question in your category. It is measured across many answers, because no two answers are quite the same.
Three things worth measuring
Presence. Out of the answers you sampled, how many named you? This is the core number and the hardest to fake.
Prominence. When you are named, are you the first recommendation or the fifth "also consider"? In a list-style answer, readers act on the top of the list.
Framing. Is the sentence around your name a recommendation, a neutral listing or a warning? A model repeating an old complaint about your pricing is visibility you don't want.
AI brand visibility vs LLM visibility
You'll see "AI search visibility", "AI brand visibility" and "LLM visibility" used for the same idea. The useful distinction is between two kinds of model:
- Models answering from memory (a plain API call to GPT, Claude or Gemini). What they say reflects their training data, which has a cut-off date. Changing it is slow.
- Models that search first (Perplexity, ChatGPT with search, Google's AI features). They fetch pages at question time, so what's on the web today can change the answer tomorrow.
A brand that shows up in search-first answers but not memory answers has a web presence the models haven't absorbed yet. The reverse usually means the brand was well known in the past but its current pages are hard to retrieve.
Why one check is never the answer
Language models pick each word with some randomness, so the same question can produce a different shortlist on the next run. The slider on the right makes that concrete. With small samples the swing is large enough to invent a trend that isn't there.
Your brand hasn't changed between Tuesday's run and Wednesday's; only the sample has. A fair parallel is the Moon: it never actually changes shape, only how much of its sunlit side faces us, which ahaboo walks through in a narrated explainer.
A measuring routine that holds up
- Write 10–20 unbranded prompts in your buyers' words and keep them fixed.
- Run them across several models, a few prompts at a time, over a week.
- Add up the totals: named in X of Y answers, per model.
- Compare month to month with the same prompt set. Change the prompts and you have started a new series.
The mention check saves each run in your browser and exports CSV, so you can build that series in a spreadsheet. The share of voice calculator turns the totals into shares with honest error ranges.
Try it: same brand, same prompts, six runs
Set how often a model "really" names a brand, and how many answers each check collects. Each row below is one simulated check.
Six checks of a brand that is really named 40% of the time came back anywhere from 13% to 63%.
From run 1 alone (3 of 8), the honest statement is "somewhere between 14% and 69%" (a 95% Wilson interval). Push the answers slider up and watch that range tighten.