A customer used to find you, read your site, and form an impression from your own words. Now, increasingly, they ask an AI first. “Who does X in my city.” “Is this company any good.” “What should I use for Y.” The model answers in a few sentences, and those sentences are the first thing the buyer learns about you. You did not write them. You cannot see them. And they are not always right.
Is AI visibility about placement or description?
AI visibility is about description, not placement. The usual framing is “how do I rank in AI” or “how do I get recommended,” which treats an AI answer like a results page with slots to win. It is not one. When someone asks about your category, the model writes a short, confident description of two or three businesses and stops. When someone asks about you by name, it writes a description of you specifically. So the real question is not only whether you got named. It is what the answer says when you are named, and whether it is true. That description is measurable: in our first pilot, a single tenant over one point-in-time run, Ignite tracked 95 distinct claims a model could state about one brand across 1,658 answer samples gathered over 21 days. Those figures are one pilot, not a benchmark to expect.
How does an AI get your business wrong?
An AI describes your business wrong in three ways: it leaves you out, it states outdated facts about you, or it repeats a framing someone else shaped. None of them show up in your analytics, because the conversation happened somewhere you cannot watch.
It leaves you out. A buyer asks for the best option in your category and the model names others. You are not in the sentence, so you are not in the consideration set, and you never learn it happened. Poor retrieval is often the cause: one 2026 chunking study measured 24% precision-at-1 for well-chunked text versus 2 to 3% for poorly chunked text, so a page a model cannot cleanly parse is a page it leaves out[1].
It gets you wrong. The model names you but describes a service you dropped, a market you left, or a price from years ago. A confident wrong answer is worse than silence, and it is being told to people who will believe it.
It describes a version of you someone else shaped. Models read the whole web, not just your site. If the sources they trust describe your space in a competitor’s terms, that framing becomes the description a buyer hears, even when the question was about you.
Why can’t you just edit what an AI says about your business?
You cannot edit an AI’s answer because there is no ad slot inside it and no dashboard the model hands you. The answer is assembled fresh from whatever the model can read and remember, and the same question can return a different answer each time. Because of that nondeterminism, Ignite never trusts a single response: it samples each tracked prompt on a weekly cadence and withholds any number until it has at least 2 samples, the rule we call the ≥2-sample gate. You can only change the inputs the model reads, and then measure, across repeated samples, whether the description actually moved. That is slow, it is indirect, and it is the honest version of the work.
What actually changes how an AI describes your business?
Three unglamorous levers change how an AI describes your business: be readable, state what you do plainly, and earn corroboration beyond your own site. Being readable means the model can find and cleanly parse current facts about you at all. Stating what you do plainly, in the words a customer would use, gives the model a clean fact to quote instead of a guess. In one controlled study, adding statistics, citations, and quotations raised how much of a page’s own text generative engines quoted back by about 22% Position-Adjusted Word Count[2]. Earning corroboration from sources beyond your own site is the third lever, and this one we hold as an editorial stance rather than a cited metric: a model repeats what the wider web agrees on more readily than what one page asserts. None of that is a trick. It is making the true description of your business the easiest one for a model to tell.
How do you know an AI’s description of you actually improved?
You know it improved by asking the models the same questions your customers ask, reading what they say back about you across repeated samples, and watching the rate move. Not once, because the answer varies. Enough times to see a pattern, with the wrong and outdated descriptions counted as plainly as the good ones. Then you change something and ask again. In our first pilot, a single tenant measured at one point in time, one brand’s mention rate across tracked prompts moved from 3% to 17.8%, but that lift was mostly branded: in non-branded category questions the brand still appeared in 0 of 24 answers. That gap is the whole point of measuring instead of buying a slot you cannot audit, and it is one pilot result, not a number anyone should expect.
References
- [1] Shaukat et al., chunking study, 2026. Measures retrieval precision-at-1 of 24% for well-chunked content versus 2 to 3% for poorly chunked content. https://arxiv.org/abs/2603.06976
- [2] Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024. Reports Position-Adjusted Word Count gains of 30 to 40% on the GEO benchmark and about 22% Position-Adjusted Word Count from adding statistics, citations, and quotations. https://arxiv.org/abs/2311.09735