AI visibility is the practice of monitoring and improving how often your brand appears in AI-generated answers. It’s the AI-search equivalent of organic rank in Google Search: your brand is either visible in the AI-generated response or it isn’t. Appearing in the answer has three versions, though, named in passing, cited as the source, or recommended as the pick, and only the recommendation sends a buyer.
AI visibility is earned on the pages the assistant reads and measured by running the same buyer prompts repeatedly, because one answer is basically a coin flip. Monitoring and improving those pages is what an AI visibility optimisation platform is built for.
What AI visibility is
AI visibility, in practice, is a bundle of three different outcomes that most reporting treats as one thing. The term is new enough that vendors still define it their own way, so it’s worth being precise about which outcome you’re counting.
Frase’s guide treats it as how often your brand appears in AI-generated answers, from platforms like ChatGPT, Perplexity and Google AI Overviews. Visiblie’s definition page calls it “how often your brand is mentioned, cited, or recommended” and blends four metrics into one weighted visibility score. Both describe a real thing, and both count outcomes that aren’t worth the same to you.
The three outcomes behave very differently once you pull them apart. An answer can name you in passing, cite your page as the source for a claim, or put you forward as the pick. A mention means a buyer heard your name in a list of six, beside your competitors.
A citation means the assistant used your page to build an answer that may still recommend someone else. A recommendation means it told the buyer to choose you, and only that sends anyone to your site, or your client’s site.
You’ll also hear the term “LLM brand visibility”, which is the same thing named for the models, large language models, the engines behind the assistants. It’s the same work under a different name, so pick one term and hold it still in your reporting. A metric that changes definition mid-quarter is hard to defend in a budget review.
How AI visibility is decided
An AI assistant decides which brands to name by running searches before it answers, and the pages those searches return decide who makes it into the answer. Currently, when ChatGPT, Claude, Gemini and Perplexity run searches before answering, they read what comes back and build the answer from those pages, which means the brands that get named are the brands on the pages the assistant chose to read. As Frase’s guide puts it, there is no position 7 in an AI-generated response: you’re either in the answer or you’re not.
Two practical things follow from how an assistant decides what to name. The searches it runs are rarely your keywords, so “best CRM for a 40-person logistics company” doesn’t look like anything in your keyword tool. And the pages it keeps reading are often third-party pages you don’t own, so part of the work happens off your site.
We measured this directly in The Credibility Layer, a study of how ChatGPT recommends local businesses. It logged 200 reasoning sessions, 5,933 searches and 1,982 citations, and where a business ranked in Google barely touched whether it got recommended. Every recommendation came from a 30-search credibility audit, and Reddit, despite its reputation, was cited zero times.
AI visibility vs. SEO visibility
Does Google ranking predict AI visibility? Barely, so the two need separate numbers.
Classic SEO visibility is a rank index: Semrush defines its Visibility metric as “a metric showing how often your website is found on the Internet”, calculated from click-through rates across your tracked keywords’ positions in Google’s top 100. It’s a decent index of your Google standing, and it answers a different question.
In our 10x AI Visibility Refresh study, the correlation between Google rank and AI citation was basically zero (Spearman rho 0.000, p = 0.914). 41 cited pages had no Google impressions at all, so Google’s own measurement couldn’t see pages an assistant was reading. That’s one study, one 91-day window, and correlation rather than causation, so don’t over-read it, but a good Google standing clearly isn’t a ticket into the answer.
So measure AI visibility directly. If a tool hands you one blended number for both, ask which pages the assistant actually read, because that’s the part you can go and work. A rank index tells you where you stand in Google; it can’t tell you where you stand in the answer.
How to measure AI visibility
You measure AI visibility by asking the assistants the questions your buyers ask, several times each, and recording what each answer recommends. Running it by hand for a week teaches you the mechanics before you automate anything. The shape that holds up looks like this:
- Build a fixed prompt set from how buyers actually ask: comparisons, “best X for Y” questions, plain category questions. Twenty to fifty prompts is enough to start, and date any change to the set, because changing the questions mid-trend breaks the trend.
- Run each prompt several times per assistant, on ChatGPT, Claude, Gemini and Perplexity. Ask the same question twice and you can get two different answers, so let’s say 30 prompts at five runs each per assistant before you quote a rate.
- Record the brands each answer names, and mark whether each was put forward as the pick or just named in passing.
- Work out your share of recommendations against the named competitors, and hold that denominator fixed. If the competitor set changes every run, the number swings for no reason.
- Re-run on a schedule, weekly suits most categories, and read the trend rather than any single run.
A few things go wrong in practice. Answers vary run to run, which is why the repeats aren’t optional, and model updates move a whole category at once. Keep a few prompts untouched as a comparison: when everything moves together, the untouched ones tell you it was the platform, and the gap between the worked prompts and the untouched ones is the lift you can actually claim.
And know what the number can’t tell you. A recommendation rate is demand capture at the point of choice, and connecting it to pipeline or revenue happens on your side.
How to increase AI visibility
You increase AI visibility by improving the pages the assistant reads, and some fixes move the number much more than others. In the order that tends to work fastest:
- Put the answer in the page’s opening passage. AI engines score one chunk of a page at a time, per our study of the three numbers that decide citations, so a page whose answer is buried three headings down gives the assistant nothing to lift.
- Make the claim a model can point at. An assistant quotes a passage it can attribute, so give it a number with a source, a plain definition or a named result rather than a comfortable mission statement.
- Refresh the pages that already earn citations. In the 10x AI Visibility Refresh study we refreshed 98 posts, published nothing new, and citations went up 10x in 91 days. The pages already in the assistant’s reading set are the cheapest lift you have.
- Put the current year in the title tag, where the page is genuinely time-bound. Across our title-tag study, 1,000 searches over five AI engines, adding the current year lifted citations by as much as 192%. Stamping a year onto an evergreen page is a fake freshness signal, and it makes the page worse for a human.
- Show up in the third-party pages the assistant reads, the ones you don’t own: review pages, comparisons, expert roundups. Off-site presence is slow to build, so start it before the on-page work looks finished.
Measuring this in practice
Ten prompts, asked by hand, is the limit of what you’ll do manually, and past that the repeats make it a measurement job. Measuring AI visibility in practice is the job Omnipresence was built for: it tracks whether each assistant recommends you or a competitor for the questions your buyers ask, with every question asked many times, so the number doesn’t swing on one answer. It also keeps a few questions untouched, so when the rate moves you can show the lift was your work.
The limits, plainly: it covers four assistants, ChatGPT, Claude, Gemini and Perplexity, and that’s the list. You bring your own model key and pay the providers at list price. Not every answer puts a brand forward as the pick, so some runs don’t count toward that number, and the work itself, the rewrites and the outreach, stays with your team.
In practice, that’s meant 10,000+ AI citations for a client, from zero in three months. Your first quarter will look smaller, and the order of work is the same at ten prompts or a hundred.
None of this is quick, and it isn’t a one-off audit either, because the answers change as the pages change. Measure it, fix the pages it reads, measure it again next week.