What Is AI Visibility? How to Measure and Improve Your Brand's Presence in AI Answers

AI visibility is how often, and how favorably, your brand gets named when someone asks an AI assistant, ChatGPT, Perplexity, Gemini, Claude, or Google’s AI Overviews, a question in your category. It is the AI-era equivalent of a search ranking, except there is no results page and no set of blue links to count. All you can measure is whether the assistant’s own generated answer mentions you, where in that answer it mentions you, and how often that happens across the questions your buyers actually ask.

Why it is not the same thing as a Google ranking

A Google ranking is a position on a page the search engine assembles from an index, and the visitor still has to click through to reach you. AI visibility is a mention, or an absence, inside a paragraph an assistant wrote itself, in its own words, for that one session, and often the visitor never clicks anywhere at all: the assistant already told them who to consider. That is a materially different event to earn and to measure than a ranking, and the two can move in opposite directions for the same page.

FoundCite’s own public teardown of the help desk software category shows this directly: across 24 real answers to 8 buyer questions, Zendesk was named most often, in 14 of 24 answers, but averaged position 4.6 within the answers that named it. Two free, open source tools named less often, Zammad and osTicket, averaged position 2.5 and 3.5 respectively, ahead of Zendesk, when they did get named. Being named most is not the same as being named first, and neither one is a Google ranking for either brand. A brand could be dominating page one of Google for “help desk software” and still be the fourth or fifth name an assistant reaches for when a buyer asks the same underlying question in a chat window instead of a search box. You can read every question and every answer in the raw teardown file.

Where it actually shows up

AI visibility is not one product’s behavior, it is a pattern that repeats with variations across several distinct surfaces:

• Chat assistants. ChatGPT, Claude, and Gemini, answering directly from a mix of training knowledge and, when the product supports it, live browsing. The same question can pull a different answer depending on which underlying model is currently serving it, since providers update these models on their own schedule, not yours.
• Search-grounded assistants. Perplexity, and browsing-enabled modes of the chat assistants above, which run a real search first and write the answer from pages they just read, with visible citations. Visibility here is closer to traditional SEO in mechanism, since a live search sits underneath it, but the output is still a written paragraph, not a list of links to click.
• AI Overviews inside Google itself. A generated summary sitting above the traditional results, pulling from the same index but presenting it as a written answer instead of a list of links. A page can rank well in the traditional results below the Overview and still not be one of the sources the Overview chose to summarize from.

FoundCite’s own measurement method covers three of these directly: ChatGPT, Grok, and Perplexity, run as fresh sessions with no memory or saved instructions. Claude and Gemini were not part of that specific run, but they answer buyer questions the same general way, a written response assembled per session, which is why the same question set and the same counting method apply to them without needing a different framework.

What actually generates a mention

Three broad forces sit behind whether an assistant names you: whether the model has absorbed a clear, specific association between your brand and the category (from training data, or from what it reads live), whether other pages it trusts already name you, and whether your own pages state facts about you plainly enough to be lifted into a sentence. None of these is something a traditional SEO dashboard tracks, because none of them is about your position in an index. They are about whether you show up inside somebody else’s sentence.

The help desk teardown makes the third-party point concretely: across the run, citations spread across more than sixty distinct domains, vendor docs and support pages, other vendors’ comparison posts, independent review sites like Capterra, and, for the compliance-focused question in the set, government sources like hhs.gov and marketplace.fedramp.gov. No single domain, including any one vendor’s own homepage, dominated the sourcing. A brand that only optimizes its own homepage is optimizing one input among dozens, and often not the one doing the most work.

Why one check is a snapshot, not a measurement

Ask an assistant a buyer question once and you learn whether you were named that one time, in that one session. It is a real data point, but it is a single one, and the whole point of a “fresh session, no memory” method, the same one behind the help desk teardown, is that every run starts cold. A model updated last week, a competitor’s new comparison page indexed yesterday, or a search-grounded assistant simply pulling a different set of live pages this time, can all change the answer to the exact same question between one afternoon and the next.

That variability is not a flaw in the method, it is the actual shape of the thing being measured. A search ranking is comparatively stable day to day. An AI answer is regenerated fresh, often per session, which means a single check tells you what happened once, and a repeated check, on a fixed schedule, is what tells you whether you are trending toward more mentions or fewer.

How to check your own AI visibility right now

You cannot infer this from your existing analytics or your search console. The only reliable check is to ask, directly, in a fresh session with saved memory and personalization settings turned off, using the questions your buyers would actually type. Not “tell me about [your brand],” which primes the assistant to talk about you. The real test is a neutral buyer question, the kind FoundCite’s teardown uses, like “what should a hospital or government team look for in help desk software for compliance reasons?” Ask it cold, in your own category, and read whether you appear at all, and where.

Doing that by hand for one question tells you something. Doing it across the full set of questions your buyers ask, on a fixed schedule, is what turns a one-time check into a measurement you can track. FoundCite runs a free version of exactly that analysis, using your domain and category, with no account required to see the first result.

From a mention to a number you can track over time

Once you have checked a handful of questions by hand, the natural next step is turning that into something you can watch move over weeks, not just a one-time snapshot. That is what a visibility or GEO score tries to do, compress the mention frequency, the position, and a few other signals into one trackable number. It is a useful shorthand, and also a number worth being skeptical of until you know exactly what went into it. See what a GEO score actually measures for what usually goes into that calculation, and why the same brand can get two different scores from two different tools.