LLM Visibility: Where Your Brand Shows Up Inside the Models
Ask a model what a business does, and it can answer with total confidence — the right city, the right service, sometimes even the founding year. Ask the same model, a minute later, who it would recommend for that exact service, and the business can be absent from the answer entirely. Two questions about the same business. Two unrelated outcomes.
Two different questions, two different measurements
Most checks for “does an AI know about my business” ask one kind of question: name the business, then ask the model to describe it. What is it, where is it based, what does it sell. A model that answers those correctly looks, on the surface, visible.
That’s recall, and it’s the easier of two things worth measuring. An accurate answer to a direct question says a model has stored something correct about a business somewhere in what it knows. It says nothing about whether that business ever comes up when nobody asked about it by name.
The harder question never says the business’s name
A second kind of question tests something else. Instead of naming the business, it asks the thing an actual customer would type: “best accountants in Leeds,” “who are the best providers for this?,” “I need one of these — who would you recommend, and why?” No name in the prompt anywhere.
One AI visibility audit engine runs both kinds of question deliberately, and keeps them apart on purpose — mixing “does the model know this business” with “does the model volunteer this business” produces a number that answers neither question cleanly. The identity questions get asked plainly, at a low setting that favors a consistent answer. The recommendation questions get asked with more randomness allowed in the model’s response, on purpose: how much a recommendation list shifts from one run to the next is part of what’s being measured, not noise to average away.
Every recommendation response also gets checked for who did get named, when the business itself didn’t. That produces a rough list of who a model actually reaches for in that category — which is often more informative than the absence itself.
Why a business can pass the first test and fail the second
This is the actual gap. A business can answer every identity question a model gets asked about it — correct location, correct services, no hallucinated details — and still never once get volunteered when someone asks who to hire. Recall and recommendation are not the same measurement, and a report that only checks one gives a business half a picture while sounding complete.
The reverse also holds, less comfortably: a model that gets basic facts about a business slightly wrong can still recommend it, because a recommendation question is really asking “who comes to mind for this category,” and coming to mind doesn’t require getting every detail right first.
Does it actually matter which one a business scores well on?
Recommendation is the one that produces customers, and there’s a real number behind that claim rather than just intuition. A March 2026 survey of US adults who’d used an AI assistant for product research, run by EMARKETER and Publicis Commerce, found 49% saying they were likely or very likely to try a different brand than usual if an AI assistant suggested one instead. Getting named unprompted, in other words, has a measurable chance of moving someone who wasn’t already a customer. Getting described accurately when someone already knows to ask doesn’t carry the same weight — that person had already decided to look.
How much does one answer actually prove?
Not much, and it’s worth being honest about the limits here rather than presenting a single number as settled. A model asked the same recommendation question twice, minutes apart, can return two different lists — that’s the reason the randomness on those questions is left high rather than tuned down for a “cleaner” answer. A single favorable mention doesn’t mean a business has secured a spot. A single miss doesn’t mean it’s been dropped from consideration.
There’s a second limit worth stating plainly: not getting recommended isn’t always a sign something is broken. A category can genuinely have two or three dominant, well-known names, and a smaller competitor can be described with complete accuracy and still lose out to size and reputation rather than to anything a technical fix would touch. Measuring the gap between recall and recommendation is useful precisely because it separates “the model has wrong or missing information” — fixable — from “the model has correct information and still isn’t reaching for this name” — a market position problem, not a data problem.
What a free scan checks here, and what it skips entirely
None of the model-calling work above runs on a free scan. Checking what a model already believes, and whether it volunteers a business unprompted, means paying for live calls to more than one model, more than once, and a free tool has no budget for that. A free scan checks the groundwork instead: whether the pages AI crawlers actually use are reachable, whether structured data resolves cleanly, and — covered in more depth in the companion piece on writing for a model rather than a search engine — whether the content itself is written in a way a model can lift a clean statement from, which a free page-level check can confirm on its own. That groundwork matters, but it’s a different question from the one this piece is about, and a report that implies otherwise is overstating what it checked. What a full audit covers end to end, model calls included, is laid out in the methodology and in a look at what a deep audit actually runs.
A business can spend months making sure a model states its details correctly and still have no idea whether any of it changes what the model says when nobody mentions the business at all. That’s the harder number, and it was never being measured by getting the easier one right.
Frequently asked questions
What does 'LLM visibility' actually measure?
Two different things, usually tracked separately: whether a model can answer accurately when it's asked about a business by name, and whether it brings that business up on its own when asked a category question that never names anyone. The first measures recall. The second measures recommendation, and they don't move together.
Is being recalled by a model the same as being recommended by it?
No. A model can describe a business correctly — the right city, the right service, the right founding details — and still never mention it when asked something like "who's the best option for this?" Recall means the model knows the facts when prompted. Recommendation means the model reaches for the name unprompted, and that's the harder bar.
Does asking a model about a business once tell you anything reliable?
Barely. Language models don't return the same answer twice, especially on open-ended questions like a recommendation list. One favorable answer doesn't mean a business is safe, and one bad one doesn't mean it's invisible. A reliable read needs the same questions asked several times, phrased a few different ways, across more than one model.
Can a free AI visibility scan check any of this?
No. Checking what a model already believes about a business, and whether it volunteers that business unprompted, means calling live models and paying for each call. A free scan checks the technical groundwork instead — crawler access, structured data, page content — and skips model testing entirely. That work sits behind a paid deep scan.
Why would a model recommend a competitor over a business it can describe accurately?
Because a recommendation question never names anyone. It asks the kind of thing a customer would actually type — "best accountants in Leeds" — and the model reaches for whatever comes to mind first, which is often whichever name shows up most across its training data and any live sources it can check. Being accurately described in a direct question doesn't put a name anywhere near the top of that list.