Guide · 19 July 2026

AI visibility

AI visibility and Glotier

A growing share of buying decisions now starts with a question to an assistant instead of a search. The assistant answers with a handful of product names. Everything not in that handful was not rejected, it was never considered. AI visibility is whether you are in it, and this is what the term means precisely, how to measure it without fooling yourself, and what the measurements actually look like when you run them.

A definition worth using

AI visibility is whether AI assistants name your product when someone asks them what to use in your category. That is the whole thing. It is deliberately narrower than the way the phrase usually gets thrown around, because the loose version, something like “your brand’s presence in AI”, cannot be measured and therefore cannot be improved.

The important structural difference from search is that there is no list. A results page has ten positions and a second page behind it, so being fifth still earns clicks and being thirtieth is at least a rung on a ladder. A synthesised answer names four or five products and simply does not mention the rest of the category. There is no position eleven. You are in the answer or you do not exist for that buyer, in that moment, on that question.

That binary quality is what makes the subject feel frightening, and it is also why the honest version is more encouraging than the marketing version. Because the shortlist is redrawn on every question, it is not a fortress that an incumbent holds. It is closer to a lottery whose odds you can genuinely change.

The four things people mean by it

Most writing on this topic collapses AI visibility into one number, usually a percentage with no denominator attached. It is really four separate questions, and confusing them is the source of nearly every bad decision made in this area.

Presence is the least interesting one

It is also the only one most people ever check. You ask an assistant a question, your name appears, and you conclude that things are fine. What you have actually established is that you were named once, on one phrasing, at one moment, by one engine. That is a data point with a sample size of one, and the section below on coverage explains why it is close to worthless on its own.

Coverage is the real score

Your buyers do not ask one question. They ask “best X”, then “X for small teams”, then “cheaper alternative to Y”, then “is Z any good”. Coverage is the fraction of those questions where you are named, and it is the number that corresponds to something real: the probability that a given buyer meets your name during their research.

Coverage is also the number that moves when you do good work, which makes it the only one worth putting on a dashboard. Presence flickers. Coverage trends.

Accuracy is the one that can hurt you

Being named is not automatically good. An assistant that describes your pricing model wrongly, attributes a feature you removed, or positions you for an audience you do not serve is doing you active damage at scale, and it is doing it in a conversation you will never see.

This failure mode is worse than absence because it is invisible in every metric you own. Absence shows up as silence, which at least matches the feeling of nothing happening. Being confidently misdescribed produces the same silence while also producing a buyer who has decided, incorrectly, that you are not for them.

Consistency is the one nobody checks

The assistants do not read the same pages, so they do not reach the same conclusions. A product can be reliably named by one and be entirely missing from another. Checking one engine and generalising is the most common measurement error in this field, and it is completely avoidable: ask more than one.

What this looks like when you actually measure it

Rather than assert any of the above, we measured it. We put 40 real software buying questions to three assistants, recorded every product named and every source behind every answer, and published the method, the limits and the raw shape of the result. Three findings bear directly on how you should think about your own visibility.

Nobody is winning, which means the position is available

The 40 answers named 266 distinct products. Of those, 223 were named in exactly one question, and the single most-named product in the entire study appeared in three questions out of forty. There was no category leader in any of the twenty-odd categories we sampled.

If you have been imagining AI visibility as a race with an unreachable incumbent at the front, the data does not support that picture. It looks far more like a wide, shallow field in which almost everyone is named occasionally and almost nobody is named consistently. Consistency is therefore unclaimed, and available to whoever treats it as a job rather than as luck.

A single mention is genuinely noise

That 223-out-of-266 figure is the strongest possible argument against the way most people check this. If five sixths of all named products appear exactly once, then your product appearing once tells you almost nothing about whether it will appear again on the next question. You have not measured your visibility. You have observed one outcome of a process with a lot of variance in it.

The sources decide, and they are not where people spend

Across those questions, Reddit appeared in the sources of 38 of the 40 answers. The five classic software review directories, the ones companies pay thousands a year to appear on, accounted for 0.7% of all citations between them, and four of the five were not cited once. Meanwhile, in 29 of the 40 answers, a named product’s own domain was also in the source list.

Read those together and you get the practical core of this whole subject. What gets you named is being present in the pages the assistant retrieves. Some of those pages you can write. Most of them you cannot, and you influence them by being worth mentioning rather than by buying a slot.

AI visibility, GEO, AEO, LLM SEO: which is which

Four terms are circulating for overlapping ideas and they are used interchangeably by people selling different things. The distinctions are real and small, and knowing them mostly protects you from buying the same thing twice.

AI visibility is the outcome: are you named. It is a measurement, and it does not imply any particular method of improving it. This is the term we use because it describes the thing a founder actually cares about.

Generative Engine Optimization (GEO) is the practice of improving that outcome. It is to AI visibility roughly what SEO is to search ranking: the work, not the score. If someone offers you GEO services, they are proposing to do things; if someone offers AI visibility tracking, they are proposing to measure things. You need both and they are not the same purchase.

Answer Engine Optimization (AEO) predates the current wave and originally described optimising for featured snippets and voice assistants, where a single extracted answer replaced a list. It has since been rebranded onto AI assistants by roughly everyone. In practice it now means the same as GEO, and the older meaning is worth knowing only so you can tell whether an agency’s case studies are actually about this problem or about 2019.

LLM SEO is the least useful of the four. It implies there is a separate optimisation surface for language models, distinct from the web, and there generally is not. Grounded answers read the ordinary web with ordinary crawlers. The phrase mostly signals that whoever used it is selling a repackaged SEO retainer.

One more distinction worth drawing, because it is where budgets get misallocated. Share of voice in the traditional sense counts how often you are mentioned across media relative to competitors. AI visibility is not that. You can have excellent share of voice in your industry press and be absent from assistant answers, because the publications that carry your mentions are not the pages being retrieved. The two correlate loosely and are measured completely differently.

A worked example, using our own bad numbers

It would be easy to illustrate this with a flattering hypothetical. Ours is more useful because it is real and it is not flattering.

Glotier sells AI visibility, and Glotier’s own coverage in its own category is poor. When we run our buyers’ questions, we are usually absent. Over 61 days we could identify exactly two occasions where an assistant named us unprompted and sent a visitor, which we wrote up in our case study. Two. Meanwhile the category questions come back naming the same handful of larger competitors.

Run through the four parts, that reads: presence, occasionally. Coverage, low and measured rather than guessed. Accuracy, fine when we do appear, which is the one part we are not currently losing. Consistency, poor, and the assistants disagree about us more than they agree.

There are two reasons to publish that rather than a tidier example. The first is that it is the honest state of a young product in a category with established names, which is the situation most readers are actually in, and pretending otherwise helps nobody. The second is that it demonstrates the point of measuring at all: we know these numbers precisely, we know which sources are deciding the answers we lose, and we can watch them move. The alternative is not a better position. It is the same position, unmeasured, with a nicer feeling attached to it.

Who should ignore this entirely

A category page that claims everyone needs the category is an advertisement. Some businesses should not spend an hour on this.

If you sell locally and in person, a barber, a plumber, a restaurant, then maps, reviews and local search decide your fate and assistant recommendations are a rounding error. If you sell through a channel where buyers arrive by procurement process or personal relationship rather than by research, the research step this affects does not exist in your funnel. And if you are pre-product with no page worth retrieving, there is nothing to be visible with yet: ship first, measure later.

Where it genuinely matters is a specific shape of business. You sell something a buyer researches before committing, the research happens online, the category has enough alternatives that a shortlist is needed, and you are not the name everyone already knows. Software, tools, apps, services with comparable competitors. If that is you, this is a channel that already has an opinion about you, and the default is not to look at it.

Why retrieval decides it

It helps to be concrete about the machinery, because the mystery around it is where most of the bad advice comes from.

When an assistant answers a buying question with search enabled, it does something close to: run a search, take the top handful of results, read them, and write a paragraph that synthesises what they say. Your product is named if it appears in those pages in a form clear enough to repeat. It is absent if it does not. There is no separate AI index, no submission endpoint, no ranking factor to game.

The alternative path, where a model answers from training memory with no retrieval, is effectively closed to any product that is not already famous. A small or new product is not in the memory, and nothing you do this quarter changes what a model learned last year. Which sounds bleak until you notice the implication: the retrieval path, the one that is open to you, is also the one that produces most answers to buying questions, because those are exactly the questions where an assistant reaches for fresh information.

So the game is not persuading a model to like you. It is making sure that when it goes looking, it finds you, in text, saying something clear.

What moves it

What does not move it, whatever you have been told

This category has attracted a lot of confident advice in a short time. Four things we would not spend money on.

Buying a directory listing in order to be found by AI. Directories may still sell for other reasons, and this is not an argument that they are worthless. It is an argument against one specific justification for the invoice. In our study those five directories together accounted for 0.7% of citations. If AI discovery is the line item that justifies the spend, the spend is not doing that job.

Hidden instructions aimed at models. Text meant to persuade a model rather than a reader, tucked into a page, is a bad bet in every direction. It does not survive the summarisation, it reads as manipulation if surfaced, and it is exactly the pattern retrieval systems get tuned to discount.

Schema markup as a standalone strategy. Structured data genuinely helps a machine parse what you are, and it is worth doing. It will not get you named on its own. It makes you legible, which is necessary and nowhere near sufficient. A perfectly marked-up page that nobody links to and no thread mentions is a perfectly legible absence.

Asking once and relaxing. Given that five sixths of named products appeared exactly once across our study, a single satisfying answer is the least reliable evidence available. It is also the most emotionally convincing, which is what makes it dangerous.

How to measure yours, by hand, this week

You do not need a tool for the first pass, and we would rather you did it manually once so that you know what any tool is claiming to automate.

Write down ten questions your buyers ask before they know you exist. Phrase them the way a person types, not the way a marketer writes. Critically, never include your own product name: a question containing your name guarantees an answer containing your name, and teaches you nothing at all.

Put all ten to each assistant with web search on. For every answer, record whether you were named, which competitors were named, whether anything said about you was wrong, and which sources the answer cited. Then count. Your coverage is the fraction where you appeared. Your accuracy problems are the wrong statements. Your consistency is how much the engines disagreed. Your roadmap is sitting in the source lists.

That is an afternoon, and at the end of it you will know more about your position than any article on this subject can tell you, including this one.

Where Glotier comes in

AI visibility and Glotier, since that is what this page is nominally about. We built Glotier because we did that afternoon by hand, repeatedly, in a spreadsheet, and it was both genuinely informative and genuinely tedious. Doing it once tells you where you stand. Doing it every week is what tells you whether anything you changed worked, and nobody sustains that manually.

So we automated the four parts. We put your buyers’ real questions to ChatGPT, Gemini and Perplexity as live web queries rather than from training memory, and record whether you were named, by which of them, who was named instead, and which pages each answer was built from. The last of those is the one people underestimate. Knowing you are absent is a fact you can do nothing with. Knowing which four domains decided the answer is an instruction.

What we will not tell you is why a model chose what it chose. Nobody outside those companies knows, and anyone selling you a list of “AI ranking factors” is selling a guess with a confident face on it. What can be measured honestly is the input and the output: the question asked, the pages retrieved, and the names that came back. That turns out to be plenty to act on, and it has the advantage of being true.

What nobody can tell you yet

Three open questions, stated plainly, because a page that defines a category should be honest about the edges of it.

How much of this converts to revenue. Being named in an answer is upstream of a visit, which is upstream of a sale, and the attribution between them is currently poor. Anyone quoting you a revenue-per-citation figure has modelled it, not measured it.

How stable any of it is. Retrieval changes continuously and models get replaced. A measurement is a snapshot, and the trend across many snapshots is far more trustworthy than any single one. This is an argument for tracking rather than for auditing.

Whether the assistants will keep sending traffic at all. An answer that fully satisfies the question does not need to be clicked. It is entirely possible to become more visible and less visited at the same time, which would make citation itself the thing to optimise for rather than the clicks behind it. Nobody knows how that settles, and treating it as settled is how you end up with a strategy built on an assumption.