How many people actually research this way?

Nobody outside the model providers can tell you precisely, and anyone quoting a confident percentage is guessing. What is observable is that the assistants have hundreds of millions of users, that search engines have built answer layers of their own, and that the behaviour is spreading from software buyers into ordinary local services.

Is this not just a smaller version of SEO?

The inputs overlap heavily, but the shape of the result does not. A search page is a shortlist you can be fourth on. An answer names two or three businesses in a sentence. There is no fourth place that still gets a look, which changes what you optimise for and how you measure it.

Why not wait until the picture is clearer?

Waiting is reasonable if nothing is compounding, and here something is. The sources assistants lean on are thin in most sectors right now, so a few clear pages and some genuine corroboration go a long way. Once competitors are described consistently everywhere, displacing them is slower and dearer.

Can anyone guarantee I will be named?

No, and you should be wary of anyone who says otherwise. These models are probabilistic, they are retrained and updated without notice, and the same question can produce different names on different days. You can improve your odds materially; you cannot buy a place in the answer.

Why checking your AI visibility matters now

Checking your AI visibility matters now because a growing share of "who should I use for this" research happens inside an assistant that names two or three businesses rather than listing ten links. A business that is never named has not been ranked low; it is absent, and absence leaves no trace in your analytics. The habit is forming while the answers are still cheap to influence.

The research step moved

For roughly twenty years the shape of a buying decision was stable. Somebody realised they needed a service, typed a rough description into a search box, and was handed ten links and a map. They opened three or four, compared them, and chose. Almost every piece of marketing advice written in that period quietly assumes that list exists.

A growing share of that first step now happens somewhere else. The same person opens ChatGPT, Gemini, Copilot or Perplexity and asks the question the way they would say it out loud: "who should I use for a loft conversion in Mansfield?" or "which accountants are good for a small limited company?"

What comes back is not a list. It is a short answer naming two or three businesses, each with a line of reasoning. The ten links are gone. The map is gone. The comparison the person used to do themselves has already been done for them, by a model, before they saw anything at all.

Nobody outside the model providers knows exactly what share of research starts this way, and any precise figure you are quoted is almost certainly invented. The direction is not in much doubt, though: the assistants have very large user bases, the search engines have built answer layers of their own, and the behaviour has spread from people buying software to people buying plumbing.

Being absent is not the same as ranking badly

This is the part that catches people out, and it is worth being precise about. A poor search ranking is a position on a spectrum. Eighth place still appears. Somebody scrolling, or searching a slightly different phrase, still finds you. You are losing, but you are in the game, and the gap between you and the winner is legible: three positions.

An answer has no spectrum. Two or three businesses are named and everyone else is not mentioned. There is no near-miss, no second page, no "also considered". If you are not in the sentence, the outcome for you is identical to not existing. You were not beaten by a competitor on quality or price. You were never entered into the comparison.

The asymmetry runs the other way too. The business that is named receives something far stronger than a link: a recommendation, in the assistant's own voice, which most people read as advice rather than advertising. One sentence of that is worth a great deal of impression share.

Why this is urgent now rather than in two years

Two things make the present moment different, and both of them decay.

The first is habit. People are deciding right now which questions they take to an assistant and which they still take to a search box. Habits of that kind tend to set quickly and then move slowly. The window in which "where do I ask this?" is still an open question in people's heads is the window in which it is cheap to be part of the answer.

The second is corroboration. In most sectors, the material an assistant can draw on is still thin. There are a handful of directory entries, some review profiles, a couple of roundup articles written by someone who did very little research, and the businesses' own sites. Where the evidence is thin, a small amount of clear, consistent, genuinely useful material shifts the picture noticeably. Once a competitor is described the same way across forty sources, displacing them takes years rather than months.

There is a third, quieter reason: you cannot measure improvement without a starting point. A baseline taken today is worth more than a better measurement taken in 2028, because it is the only way you will ever know whether anything you did worked.

What this looks like for an actual business

A dental practice in Leeds ranks respectably for "dentist Leeds" and gets a steady trickle of enquiries. Asked "which dental practice in Leeds is good for nervous patients?", an assistant names three other practices — not because they are better, but because each of them has a page that says, in plain words, that they treat anxious patients and how. The practice that ranks has a homepage with a hero image, a phone number and a booking form. There is nothing on it for a model to quote.

A joiner in Mansfield has a different problem. Mansfield exists in Nottinghamshire, in Texas and in Ohio, and an unqualified town name regularly produces an answer about the wrong continent. County, region and country matter far more in an answer than they ever did in a local search, where the engine knew where the searcher was standing.

An online gift retailer has a third variant again. There is no geography to anchor to, so the question becomes whether the model has any reason to distinguish them from four hundred other shops selling similar things. Specificity is the entire game: a shop known for one clear thing gets named for that thing.

What a business can realistically do

Find out where you stand. Ask the assistants the questions your customers actually ask, write down which businesses are named, and repeat it. One check tells you very little, because the models vary between runs. A dozen questions repeated over weeks tells you a lot.
Make your pages quotable. Somewhere on your site there should be a plain sentence stating what you do, who for, and where. Not a slogan. A sentence a stranger could lift and be right.
Check you are reachable. Some sites block the AI crawlers outright, usually by accident, inheriting a rule from a template or a plugin.
Earn corroboration. Reviews, directory entries, local press, trade bodies, supplier listings. Assistants lean heavily on what third parties say, and they weight agreement between sources.
Pick the questions that matter commercially. Being named for "best gift shop" is pleasant. Being named for the question that precedes a purchase is the one that pays.

Most of this overlaps with work you may already be doing. The genuinely new parts are the measurement and the quotability, which is the argument laid out in the comparison between AI visibility and SEO. If you want the mechanics of how a model arrives at two or three names, that is covered in how assistants choose who to name.

The honest limit on all of this

No tool, ours included, can see inside a real customer's conversation with an assistant. Those exchanges are private, personalised and shaped by that person's history. What can be measured is how the models respond to a consistent set of questions asked repeatedly through their public interfaces — a directional index, not a transcript. It will tell you whether you are absent, whether that is changing, and who is named instead. Treat any vendor promising you the actual conversations, or guaranteeing a result, as selling something that does not exist.

That limit is less of a problem than it sounds. You rarely needed the individual conversation. You needed to know whether, when a model is asked the question that brings you customers, your name is in the answer — and if not, whose is. That question is answerable, and it can be measured honestly if you are disciplined about asking the same things the same way.

Where to start

Start with the questions, not the tactics. Write down the ten things a customer asks immediately before they buy from you, phrased as they would phrase them. Ask two assistants each question. Record the names that come back.

That exercise takes an afternoon and produces one of two results. Either you are named, in which case you now know what to protect and can watch for it slipping. Or you are not, in which case you have learned something about your market that your analytics could never have told you — and a free visibility check will give you the same answer faster, across more assistants, with the competitors listed.

Either way, the cost of looking is an afternoon. The cost of not looking is invisible, which is precisely what makes it easy to keep deferring.