What is AI visibility? A plain guide for founders

AI visibility is whether and how your product gets named, cited and described when someone asks an AI assistant for software to solve their problem. When a founder types "best tools for onboarding emails" into ChatGPT, Gemini, Perplexity or a Google AI Overview, the assistant returns a short list of named tools with a sentence of context for each. AI visibility is your presence in that answer: whether you appear at all, whether you are described accurately, and whether the assistant points to you as a credible option. It is distinct from a ranking. There is no page two to scroll to and no list of ten blue links. There is one synthesised answer, and you are either in it or you are not.

The jargon, in one place

Four terms get used for overlapping things, often interchangeably and often by people selling something. They are not synonyms, and the differences decide what work you actually do.

If the distinction between AEO and GEO feels thin, that is because it is. The split worth keeping runs elsewhere: AI visibility is the thing you measure, and the optimisation work is what you do about it. Measure first, because without a measurement you cannot tell whether your problem is that you are absent, that you are described wrongly, or that you are present on one engine and missing from another.

Why it matters, measured rather than asserted

More people are starting their search for software inside an AI assistant rather than a search box. That much is widely claimed. What is rarely shown is what actually happens inside those answers, so here is what we found in our own, which we can show because measuring them is what our product does.

The first finding is the cliff. We asked the leading answer engines the same thirty buyer questions about each of 152 companies. The 18 established companies were named in 36.7% of their answers. The 118 young ones were named in 6.4% of theirs, and 48 of them were never named once. Before any question of ranking or description, most young products are simply absent from the answer.

The second finding is about what happens when you are named. Answer engines cite heavily: 95 to 100% of the answers that engines produced carried sources, typically seven to seventeen of them. But naming and linking are separate events. Of 1,068 answers that named the company they were asked about, 653 also linked to that company's own website and 415 did not. Two answers in five name you and send the reader everywhere but your site.

And either way, the answer is built from other people's pages. Across all 130,694 sources the engines cited in this run, the asked-about company's own website was 0.76% of them. This is why publishing another page on your own site moves so little on its own: the engines form their opinion of you elsewhere, and use your own site to confirm it.

How AI assistants decide what to name

It helps to reason from the mechanism rather than treat the assistant as a black box. Two things are happening, and they have different time horizons.

What the model already knows

A language model is trained on a large slice of the public web frozen at a point in time. If your product was described clearly and repeatedly across that web before the training cut-off, the model carries a representation of it: roughly what it does, what category it sits in, who competes with it. This is why incumbents appear effortlessly. It is also why a new product can be invisible no matter how good it is. The model cannot name what it never read. This part of visibility is slow to influence, because it only updates when the model retrains, and it rewards a long, consistent paper trail rather than a single recent push.

What the assistant retrieves live

Most consumer assistants now also search the live web at the moment of the question, read a few pages, and ground their answer in what they just retrieved. This is the faster lever. Here the assistant behaves like a very literal researcher: it issues a query close to the user's phrasing, opens the pages that look most relevant and credible, and lifts named products and descriptions from them. If a current, well-structured page names your product as a strong option for that exact use case, the assistant can cite you even though the underlying model had never heard of you. Live retrieval is how a new product breaks in before any retraining could have noticed it.

So AI visibility has two surfaces. The trained surface rewards a durable, coherent presence across the web. The retrieved surface rewards clear, current, quotable pages that match how people actually ask. You want both, and they compound: the more often you are correctly described in places the assistant retrieves, the more coherent your representation becomes the next time a model trains.

The engines do not read the same internet

The advice you will read treats "AI search" as one destination. Our measurements say it is not. Every engine in this run was asked the same questions about the same companies in the same window, and they still built their answers from different internets: community and user-generated sites are 2.6% of ChatGPT's sources, 4.5% of Gemini's, 6.3% of Perplexity's and 14.9% of Google AI Overviews'.

Look at what leads each list. ChatGPT's most cited source is techradar.com, and its top sources carry press-release wires. Gemini's is medium.com. Google AI Overviews' is youtube.com. Perplexity's is reddit.com. Every engine leads with a different source: one mostly reads what the press wrote about you, one reads long-form blogs, one watches video, and one reads the forums.

The overlap tells the same story. Comparing the ten most cited sources for each engine on the same questions, ChatGPT shares three of them with Google AI Overviews, three with Perplexity and two with Gemini, while Google AI Overviews and Perplexity share six with each other. Two of the four are close relatives, and ChatGPT is still the outlier.

How AI visibility differs from search rankings

AI visibility and search engine optimisation overlap but are not the same thing, and conflating them leads founders to do the wrong work. A search ranking is a position in an ordered list of links: your job is to earn the click, and the user then forms their own impression of you on your own page. An AI answer is a synthesis: the assistant has already formed and stated an impression of you on your behalf, in a sentence, before the user ever reaches your site. The unit of the contest changes from a position to a description.

Two practical consequences follow. First, accuracy becomes as important as presence. Being named but described as the wrong kind of tool, or paired with the wrong use case, can cost you the click you would have won. Second, the levers shift. Clear, factual, well-structured content about what you do and who you serve matters more than narrow keyword tactics, because the assistant is reading for meaning and lifting claims, not matching strings. Good fundamentals still help both surfaces. The framing is what changes.

How to tell if you have AI visibility

You cannot improve what you have not measured, and AI visibility is easy to measure crudely and worth measuring carefully. The crude version takes ten minutes. Open two or three assistants, then ask them the questions a real customer would ask: name the problem you solve, ask for the best tools for it, ask for alternatives to your nearest competitor, and ask the assistant to describe your product directly. Read the answers as evidence, not as verdicts.

• Are you named at all when someone asks for tools in your category, or only when you name yourself first?
• When you are named, is the description accurate, or does the assistant put you in the wrong category or attach the wrong use case?
• When you are named, does the answer link to you, or does it link to a review site, a forum thread or a competitor's comparison page?
• Which competitors appear consistently, and what is being said about them that is not being said about you?
• Does the picture change from one engine to the next? Given how differently they source, it usually does.

Run the same prompts across more than one assistant, because they retrieve and weight differently, and run them more than once, because answers vary between sessions. A single absence is noise. A pattern of absence, or a pattern of being described as the wrong thing, is signal. That pattern is the gap you are going to close.

How to start improving your AI visibility

The work is less exotic than the term suggests. It is mostly about making true, specific, machine-readable evidence of your product exist in the places assistants read, and keeping that evidence consistent. What the measurements above change is the order you do it in.

Start with your own pages, but treat them as the confirmation rather than the campaign. State plainly what your product is, the category it belongs to, the specific problems it solves and who it is for, in language a customer would actually use rather than internal branding. Assistants lift clear claims and struggle with vague ones. That work is necessary and it is quick, and on its own it will not move much, because your own domain is a small fraction of what gets cited.

Then spend the bulk of your effort off your own site, because that is where the citations are. The descriptions of you on third-party pages, comparison posts, directories, community threads and video are exactly what live retrieval reads. Where a competitor is named and you are not, the usual reason is that someone wrote a clear page comparing options and your product was missing from it or thinly described in it. And weight that effort by engine: press and publication coverage does more for ChatGPT, long-form writing does more for Gemini, and community and video presence do more for Perplexity and Google AI Overviews.

Treat it as ongoing rather than a one-off project. AI answers shift as the web shifts and as models retrain, so visibility is a position you hold by maintaining the evidence, not a box you tick once. Our own numbers move between measurements, which is the clearest argument there is for watching rather than auditing once.

How we measured this

Every figure on this page comes from one measured run, on a fixed, documented cohort. We asked the leading answer engines thirty real buyer questions about each of 152 companies: 118 young companies drawn from Y Combinator's public directory by a seeded random draw, 18 of YC's own designated top companies as the established control, and 16 AI-search products. That produced 11,383 answers carrying 130,694 sources across 38,630 distinct domains, measured on 22 and 23 August 2026, counting each source once per answer. Nothing here is a survey, an estimate or a figure taken from someone else's article.

The limits matter as much as the numbers, so here they are. The questions are software buyer questions, so these are claims about software discovery, not the whole internet. Company mentions are detected generously, so a company with a common-word name can be under-counted. The run is a two-day window, and answers vary between runs, which is why we publish the interval rather than a lone decimal. And one correction, stated rather than buried: an earlier version of this page guessed that a young product very likely does worse than a famous one at being linked once named. Measured on this cohort, that guess was backwards. Once named, the young companies were linked to their own site more often than the established ones, 69.8% against 51.4%, because an engine tends to name an unknown company precisely when it has just read its site. Obscurity does its damage at the naming step, not the linking step.

On confidence: the per-engine differences are large relative to the sample and consistent across the cohort's legs, so they are not noise. The linking rate is 61.1% with a 95% interval of 58.2% to 64.0% across 1,068 naming events. These numbers describe where engines source, not what causes them to source you, and they will move as the engines move, which is the argument for measuring on a cadence rather than auditing once.

Where to go next

The honest first step is to find out where you actually stand, per engine, on the questions your own customers ask. That pattern tells you whether your problem is presence, accuracy, or a gap on one engine that the others hide.