Visibility in language models is the first question about reach in a long time that nobody in marketing has a practised answer for. The mechanics are new, the measurement is thin, and decisions are being made anyway. I think that makes it the right moment to build a position, before the agencies turn it into a product.
- Language models cite sources, they do not rank pages.
- Visibility comes from evidence, not from keywords.
- Five levers are workable today.
- Much of it cannot yet be measured cleanly.
Table of contents
Why visibility in language models works differently
In classic search a brand competes for a position. In a model it competes for a mention inside an answer that did not exist beforehand. That sounds like a small difference and it is a large one. A position can be counted, a mention is a formulation. Anyone reading visibility in language models as simply a new ranking list ends up measuring something that is not there.
The second difference lies in the source. Models do not reproduce what a brand says about itself but what has been written about it. The company website is one piece of evidence among many and rarely the strongest. Trade press, forums, retailer pages and comparison texts often weigh more than any carefully maintained brand page. That is uncomfortable, because it concerns exactly the part of communication nobody controls.
At least the technical side is documented by now. Google sets out in its guidance on AI features in search what content has to satisfy to be considered in generated answers at all. What is striking is how little of it is new. Accessibility, clean structure, clear statements. Much of what favours visibility in language models was already good work before.
The third difference concerns what happens next. In search, visibility is followed by a click, and a click can be counted. In a generated answer nothing often follows at all, because the question has already been answered. The brand was mentioned, the conversation ended, and no report shows it. Visibility in language models therefore produces an effect that simply has no place in most companies’ numbers.
Five levers I use for visibility in language models
The five levers below are a sequence, not a model. They come out of trying to make it traceable for a brand why it appears in answers or does not.
- Evidence. Is there a statement about the brand that sits outside its own channels? Without external evidence the brand remains an assertion.
- Distinctness. Can the brand be told apart from similar sounding names? Confusable names cost more in models than in search.
- Formulation. Does a sentence exist somewhere that explains the brand in one sentence? Models reproduce sentences, not claims.
- Currency. How old is the last reliable source? Outdated descriptions persist stubbornly because they were quoted often.
- Consistency. Does the brand say the same thing everywhere? Contradictions do not lead to a choice, they lead to omission.
The last point surprised me most. I had expected a model to settle on the most frequent variant when the sources disagree. In practice something else happens more often: the brand does not appear in the answer at all, because another one is described more clearly. Visibility in language models is therefore less a competition for affection than one for describability. How a brand can sound consistent across languages I have described in brand voice translation.
What I measured and what cannot be measured
Some time ago we started asking a fixed list of questions regularly and recording the answers. Not as a study but as an observation. About thirty questions a customer would actually ask, always phrased identically, in several languages. The effort is roughly an hour a month, and it is the only reading of visibility in language models I would defend. What comes out of it is not a ranking but a sense of whether the brand is mentioned, how it is described, and whether the description is correct.
The third part was the return. We discovered that an important product category of ours was consistently classified wrongly, on the basis of an old description that still sat on several retailer pages. That could be repaired, though it took a few months before it showed up in the answers. Without the observation we would not have known, because nothing in sales looked like it.
Since then I treat visibility in language models as a brand maintenance task rather than a search optimisation one. The difference is not academic. Search optimisation asks how a text gets found. Brand maintenance asks whether what circulates about the brand is still true. The second question can be answered even when the technology shifts again, and it leads to work that was due anyway.
What cannot be measured is the effect. I can say whether a brand appears. I cannot say how much revenue came out of it, and I treat every figure claiming otherwise as provisional. Anyone selling visibility in language models today with a hard contribution to the result is selling an estimate. That is no reason to leave it alone, but it is a reason to call it what it is. How I choose measures in general is in marketing metrics.
Where visibility in language models runs into limits
There are two cases where I advise against putting much energy in. The first is a brand whose category is barely asked about. Where nobody asks, the best description helps nothing. The second case is a brand that cannot yet state its own offer sharply. A model amplifies existing clarity, it does not create it. Visibility in language models is not a project in that situation but a consequence, and it only arrives once the positioning has been decided.
The second limit is a matter of time. The systems change faster than an organisation can agree on measures. Optimising a tactic today against a particular behaviour means working with a half-life of a few months. So I consider anything built only for this channel risky, and anything that makes the brand clearer safe. The second route is slower and survives every generation of models. Planning visibility in language models as a campaign, by contrast, books effort and builds nothing.
In the end visibility in language models is not a new discipline but a sharper test of an old one. Models reward brands about which something concrete can be said outside their own channels. That is the same requirement the press, the trade and word of mouth have always made, only without the courtesy. Working on it is not working on a channel but on whether the brand can be described at all. And that is the one form of visibility that cannot be bought.