Every agency on LinkedIn is now selling AI visibility. Almost none of them will tell you what it is, because the vague version sells better. So here's the mechanism, in plain terms, and a test you can run on your own business before lunch.
The one-sentence version
AI visibility is whether a machine says your name.
When a buyer asks ChatGPT, Perplexity, Gemini, Copilot or Google's AI Overviews a question in your category — "who should I use for X" — the system returns two, three, maybe four companies with a sentence about each. That list is the entire consideration set. There is no page two. There is no scrolling past the fold to find you at position eleven.
Either you're in it or you don't exist for that buyer.
Why this isn't just SEO with a new name
Traditional search gives you a consolation prize for being close. Rank fourth and some people still click. Rank eighth and you get the occasional visitor. Position is a gradient.
Generated answers are a cliff. The model picks a handful of companies it considers credible and defensible, writes them into a paragraph, and discards everything else. Being the fifth-most-credible option in a three-name answer isn't a lower score — it's a zero.
Ranking is a gradient. Being cited is a cliff.
The second difference is who's reading. Traditional SEO optimises for a person scanning a list of links, deciding what to click. Answer engines optimise for a model deciding what to trust and repeat. Those are different jobs. A page can be perfect for the first and useless for the second.
What actually determines whether you get named
Strip away the mystique and three things decide it.
1. Entity clarity — can a machine tell exactly what you are?
Models don't reason about your company from your homepage headline. They reason about an entity: a stable, cross-referenced idea of who you are, what category you sit in, where you operate, who you serve and what you're known for.
Most B2B sites make this needlessly hard. The homepage says "we deliver transformative solutions that empower organisations to unlock their potential." A human skims past it. A model reads it and genuinely cannot determine what you sell, to whom, or where. So when a buyer asks about your category, you aren't a candidate — not because you scored low, but because nothing connected you to the category in the first place.
Entity clarity comes from boring, mechanical things: consistent naming everywhere, structured data that states your category and service area in machine-readable terms, an about page that reads like a fact sheet rather than a manifesto, and the same description of your business appearing consistently across your site, your LinkedIn, your directory listings and anywhere else you're mentioned.
2. Answer-shaped content — does anything on your site respond to a real question?
Buyers don't ask models for keywords. They ask things like "what should a 40-person recruiting firm expect to pay for RPO" or "is it worth switching from Netsuite at our size."
If your site contains a page that answers that question directly — with a real answer near the top, specific numbers, stated conditions and honest caveats — you become a quotable source. If your site contains a page called "Our Services" with three paragraphs of positioning, you don't.
The practical test: could someone extract a self-contained, accurate two-sentence answer from your page without needing the rest of it? If not, there's nothing for a model to lift.
3. Corroboration — does anyone else back up what you say?
A model weighs claims by how well they're supported elsewhere. Your own site asserting you're the leading provider in your region counts for very little. Independent mentions, credible directories, guest contributions, real customer reviews, press, comparison pages and industry listings count for a great deal.
This is the slowest of the three and the hardest to fake, which is exactly why it carries the most weight.
The ten-minute self-test
You don't need a tool for the first pass. You need a notepad and ten minutes.
- Write down five real buying questions. Not keywords — the actual sentences a prospect would type. Steal them from your sales calls. "Best X for Y", "how much does X cost for a company our size", "X vs Y for Z", "who does X in [your city]", "is X worth it for [buyer type]".
- Ask each one in three different engines. ChatGPT, Perplexity and Google's AI Overview will do. Use a logged-out or incognito session so your own history doesn't flatter the result.
- Record who gets named. Make a simple tally: which companies appear, how often, and in what order. Do not skip the tally — the pattern is the finding, not any single answer.
- Note what's said about you, if anything. Wrong service list? Wrong country? Confused with a similarly-named company? Described as something you stopped doing three years ago? Each of these is a specific, fixable entity problem.
- Ask the engine directly. "What do you know about [your company name]?" and "who are the main alternatives to [your company]?" The second question is the more revealing one — it tells you which competitive set you've been filed under, and whether it's the right one.
Three outcomes, and each points somewhere different:
- You're absent entirely. This is usually an entity problem, not a content problem. Nothing has connected your business to the category clearly enough for you to be considered.
- You appear, but described wrongly. Fixable, and often quickly. The machine has found you and misread you — that's a clarity and structured-data job.
- You appear correctly but rarely. A corroboration problem. You're understood, just not yet trusted enough to be a default recommendation. This is the slow, compounding work.
One warning about the test. Generated answers vary between sessions, users and regions. Run each question two or three times and treat the pattern as the signal, not any single response. If a competitor shows up in five of six runs and you show up in none, that's not noise.
What most companies get wrong next
They panic and publish. Twenty blog posts in a month, aimed at nothing in particular, on the theory that more content means more chances to be cited.
It doesn't work, for the same reason it stopped working in SEO years ago. Models are selecting for credibility and specificity, not volume. Three genuinely useful pages that answer real buying questions with real numbers will outperform thirty pages of restated positioning — and the thirty pages actively dilute your entity by making it harder to tell what you're actually about.
The right order is almost always: fix what you are, then fix what you say, then earn the corroboration. Skipping to the third step is why so much content marketing money disappears without a trace.
The honest timeline
Technical and entity fixes can show up within weeks — you're correcting things that are actively wrong. Content built around real buying questions typically takes one to three months to start being surfaced. Corroboration compounds over three to six months and doesn't stop.
Anyone promising to get you into AI answers in thirty days is either redefining the goal or guessing. The work is real, it just isn't fast, and the companies starting now will be the defaults their competitors have to displace later.