What happens between "best dentist near me" and the answer you get

People tend to picture this as a lookup. You ask an AI a question, it "knows" the answer, it tells you. Simple as a phone book. It's not that. And the difference matters if you're trying to understand why some businesses show up in these answers and others don't.

Here's roughly what happens, stripped down. The model gets your question — "best dentist near me," say. Depending on the model, it either answers from what it absorbed during training, or runs a search first and reads what comes back, or some mix of both. Either way, it isn't holding one "correct" business in mind. It's holding a small cloud of candidates — names it's seen tied to that category, in that area, each with some signal of relevance attached. Then it does the part a search engine never had to do: turn that cloud into a sentence, in an order, and commit to it.

That ordering is the whole game. Most people skip right past it. Getting mentioned at all and getting mentioned first aren't adjacent outcomes on a scale — they're close to different outcomes entirely. A business named first in an AI answer and a business named fourth, buried in a "you could also consider" clause at the end, aren't both "visible" in any sense a business owner would recognize. When I built Foreground I first tried to fold that into one weighted 0-100 score. It didn't hold up: the number moved between runs more than it moved because of anything the owner did. So Foreground now just counts — named in how many of 12 answers, for one specific group of customers — and shows you who got named instead, in what order.

One more rule, added after I saw the first results. A question that names your business doesn't count. Ask any model "is Acme a good choice?" and it says "Acme" back. That proves it can repeat a word. It doesn't prove it knows you. Only the questions a customer would ask without your name in them are counted.

Run that same question through four different AIs — ChatGPT, Gemini, Claude and Grok, because they don't agree with each other nearly as often as you'd expect — and "AI visibility" stops looking like one number. It's a pattern. A business might get named first by one model and skipped entirely by another, and that gap alone tells you something. Usually it means the model that skips you doesn't have a clean, specific fact to attach to your name — nothing quotable, nothing distinct. So when it's forced to pick a short list, it reaches for whoever gave it something easier to repeat.

That's the mechanism worth sitting with: an AI answer isn't a ranking of your website. It's the model's best attempt at compressing "who should I tell this person about" into one sentence, under real pressure to be short and confident, using whatever specific, legible facts about you it managed to pick up along the way. Vague businesses lose that compression contest before it even starts. Not because they're worse. Because there's nothing about them sharp enough to survive being squeezed into a sentence.

It also explains something that confuses people at first: why the same business can get a completely different answer depending on how the question is asked. "Best dentist near me" and "dentist that does same-day emergency appointments" pull from different corners of that candidate cloud, because the model is matching the specific thing you asked, not running one fixed ranking it applies to every version of the question. There's no single "AI rank" the way there's a Google position — there's a different compression each time, built fresh around what was actually asked.

If you take one thing from this: stop asking "am I mentioned." Ask mentioned to whom, for which question — and who got named instead. Most people asking about their AI visibility right now aren't asking that question at all.

Next: SEO and AI visibility are not the same job — why the tactics that win a Google ranking don't automatically win a mention.

— Marc. Foreground runs this exact check for the customers you want — three questions per niche, four models, twelve answers — so you're not guessing at the pattern by hand. Try it free.