Same mechanism, different mediator

A person almost never encounters a business directly. Something sits in between and decides whether they hear about it at all.

That something has changed a lot. It was other people, then trade directories, then the phone book, then search engines, then review sites, then social feeds, then marketplaces. Right now, increasingly, it is a language model that answers in prose and names three companies.

Each time it changes, the same thing happens. A new discipline gets announced. New agencies appear selling it. New acronyms, new playbooks, new price lists. And every time, most of what is actually being sold is the old job with the new mediator's name on the invoice.

The bit that doesn’t change

Strip the mediator out and the job is always the same shape. Somebody is looking for what you sell. Something stands between you and them. That something has a model of your category, and you are either legible inside that model or you are not.

Search engine optimisation is that job, pointed at a search engine. You work out what the ranker rewards, and you make the business legible to it. Generative and answer engine optimisation is that job, pointed at a language model. Same work, different thing standing in the way.

The differences are real and they matter operationally — a model does not rank ten blue links, it synthesises an answer from sources it has decided to trust, so “being cited” replaces “ranking” as the unit. But that is a difference of mediator, not of mechanism. Discovery is still discovery.

Why this isn’t a semantic argument

It decides what you build, and it decides what happens to what you built when the ground moves again.

If you believe AI search is a new discipline, you build a practice around this mediator's current behaviour. Prompt-shaped content. Model-specific tactics. A service line named after a product category that is two years old. When the mediator changes — a new model, a new interface, agents that transact rather than answer — you rebuild, rename, and re-pitch, because what you sold was the mediator.

If you believe it is the same mechanism, you build around the question that survives: what stands between this business and its customer, and is the business legible to it? The answer changes. The question doesn’t. And the diagnostic you use to answer it works on the next mediator too, because it never depended on this one.

What it looks like in practice

Every system in a business gets two questions instead of one. Does this work for a human? And does this work for whatever is currently standing between us and the human?

That second question is not an AI bolt-on. It is a permanent second surface, and it has always been there — it is just that for about twenty years the answer was “make sure Google can crawl it” and everybody stopped noticing it was a question.

The honest part

There is a version of this argument that is just a way of saying nothing has changed, and that version is wrong. Plenty has changed. The measurement in particular is genuinely harder, because these systems are not stable the way a ranking is. Ask ChatGPT the same question twice and you get a materially different answer 17% of the time. Gemini, 28%. Both figures are from our own corpus of 514 captures across those two engines — the only two we have run, and the only two we claim.

That instability is the actual new thing, and it is the reason most of what is currently being sold as an AI visibility finding is a single screenshot sitting inside a noise floor nobody measured.

But an unstable mediator is still a mediator. It gets diagnosed the same way: work out what it rewards, check whether the business is legible to it, fix the gap between those two things. The tools are new. The job is old.

Related

We measured Gemini as 3× less stable than ChatGPT. Then we tripled the sample.

What a Gap Read is — the free version of this, run on your category.