The Market Already Wrote Half the AI-Visibility Dictionary. Someone Has to Finish It.

Open two AI-visibility dashboards and ask each one what "visibility" means, and you get two answers printed under the same word. One treats it as binary, a yes or no on whether your brand surfaced in a single AI answer, one reply at a time. Another reports it as a benchmark score from 0 to 100, tracking how often you appear against competitors across a whole basket of answers. Same word, printed on both boxes. One is a yes-or-no about one reply. The other is a ranking across hundreds of them. A brand that counts as "visible" in the first sense has told you nothing about where it lands in the second.

That isn't a two-product quirk, and the wider you pull the frame the worse it gets. "Share of voice" is where it turns ugly. One tool computes it as a plain ratio, your brand's mentions over everyone's. Another builds it from modeled impressions and weights those impressions by search demand, a step honest enough to call a modeling choice rather than a measured relationship. A third scopes the very same phrase per domain. One term, three denominators. Two dashboards both reporting "share of voice, 40%" are not reporting the same number, and nothing on either screen tells you so.

What the Market Already Agrees On

The lazy read is that this is chaos, snake oil with a dashboard, and the smart move is to wait a year for it to settle. That read is wrong, and being wrong about it is what's costing the field. Underneath the scrambled labels the market has already agreed on something, and agreed on it firmly. Nearly every vendor whose glossary bothers to spell it out draws the same first line. A mention is your brand's name sitting in the answer text. A citation is your domain actually linked as a source. Some tools state it as flatly as it can be stated, presence is not citation, a brand can be named and never linked. Others split brand visibility, where you're named, from source visibility, where your page fed the answer even though nobody typed your name. The distinction shows up in tool after tool, and nobody had to be talked into it.

So the trouble was never that nobody agrees, it's that the agreement only ever covered the two easy words and left the hard ones to whoever grabs them first. Somebody will. The likeliest author is whichever vendor raises the biggest round and gets to stamp its private glossary on the category, or a standards body that shows up late and hands the vocabulary to people who spent the last decade measuring banner ads. IAB shipped an AI content-disclosure framework in early 2026 and is busy standardizing agentic advertising, but it hasn't gone near the question of what separates a citation from a mention from whatever sits between them. That vocabulary has no owner.

That doesn't mean the ground is empty, and pretending it is would be its own dishonesty. The cleanest early work is academic. The 2023 paper that named generative engine optimization, by Aggarwal and colleagues and later accepted to KDD 2024, was the first to treat visibility inside an AI answer as something you could define and measure at all, a set of metrics for how a source shows up in a generated response. The idea it pried open, and the one no vendor has cleanly named since, is the gap between a source that merely appears in an answer and one that actually shaped it. Practitioners feel that gap constantly. Nobody has given it a word.

A Vocabulary No Vendor Owns

So here's the vocabulary this publication is proposing, deliberately small and owned by no one. Four states, one brand, one answer, and a brand sits in exactly one of them per answer. It's cited when the answer carries a clickable link back to its domain, the highest-confidence signal of the four because the link sits right there in the response and anyone can audit it. It's mentioned when its name is in the text with no link home, awareness it can't prove drove anything. It's assisted when its content clearly moved the answer, got paraphrased or used as background, and earned neither the link nor the name. And it's other, or uncredited, when the answer plainly knows something that traces back to no nameable source at all, the honest residual left when you subtract the first three from everything the model appears to know.

Notice what this is and isn't. It isn't another score to set against the ones vendors already sell, it's a key their scores can be read through. A share-of-voice number is a count of the mentioned bucket. A citation-frequency chart counts the cited one. An impression-weighted figure is a model laid over those same two states. Every metric in the field maps onto these four words, which is the whole point, because it lets you translate one vendor's dashboard into another's instead of being locked inside whichever glossary you happened to buy first. The vendors keep their math. They just get a shared set of nouns under it.

Assisted is the hard one, and it's worth saying so out loud. Most of the time you can't tell a source that genuinely shaped an answer from one the model happened to echo, not without the engines disclosing what they pulled, which they mostly don't. Name it anyway. A vocabulary that only covers the two things we can already measure cleanly just re-freezes the exact gap everyone keeps tripping over. Giving the unmeasured thing a name is what lets people start arguing about how to measure it, and that argument is overdue.

The market is already settling on a rough working vocabulary, share of voice, share of model, citation rate, answer presence, and the terms are useful. The trouble is that the ones catching on tend to carry a single vendor's fingerprints, coined inside one company's documentation and load-bearing for that company's product. A standard with one vendor's name stamped on it is marketing with better distribution, not a standard. The four words above are built to belong to nobody, which is the only way they ever come to belong to everybody. We're not adopting any house coinage here, ours or anyone else's.

The category spent two years proving its dashboards disagree. The more useful fact is that they already agree on half the vocabulary, and only the two hardest words are still unclaimed. Standards go to whoever writes them down first and gets cited, not to whoever signs up the most users. So here are four words that belong to no vendor, ours included. Use them, argue with them, map your dashboard onto them. Just don't wait for the company with the biggest round to name them for you.