An AI is not a search engine, which is why its demand cannot be derived from search volume

There is a question everyone asks as soon as they start caring about AI visibility: how often does anyone actually ask an AI about what I sell?

The question is fair. Without it, nobody knows whether the effort is worth it. And it is uncomfortable, because the honest answer is this: right now, nobody can measure it who does not operate an AI themselves.

Where a fair question meets an uncomfortable answer, estimates appear. One of them has just been documented well, and because it is documented well, you can use it to show something.

How the derivation works

In August, Ahrefs published how their "AI adjusted volume" metric is produced. Two steps:

First, each prompt is mapped to a parent keyword and that keyword's Google search volume is looked up. Then that volume is multiplied by a platform specific ratio. The ratio comes from the provider's own traffic data: from the traffic arriving on websites from AI platforms, relative to organic Google traffic.

What stands out is what the provider writes alongside it. The number is explicitly not suitable for deriving AEO metrics, not for sizing an addressable market, and not for measuring how many people saw a brand in a particular AI answer.

That is more honest than much of what is being sold right now. Anyone who reads that caveat and still builds a market size from it has skipped it, and the provider did not hide it.

What step one loses

A prompt is not a search query with more words in it. It is a different thing.

"Which espresso grinder under 300 euros for someone who has to be quick in the morning and does not like noise" contains a budget, an occasion, a constraint and an implicit exclusion. Map that to a parent keyword and what is left is "espresso grinder". That is the search query of 2015, and precisely the parts that shape an AI answer are gone.

The difference is not academic. It decides which product gets named. With "espresso grinder", whoever is visible for the category term wins. With the long question, whoever has a source where noise level, warm up time and price sit next to each other wins. Those are two different jobs, and one says little about the other.

What step two loses

The second step is the more interesting one. The ratio used for the multiplication comes from AI referral traffic. That is, from clicks.

Clicks are the visible remainder of AI usage, and the smallest part of it. The whole point of an AI answer is that the user reads it and does not click on. Extrapolating demand from clicks means extrapolating from the behaviour of the people for whom the answer just did not work.

That is not an accusation against the method. It is a description of what happens when you form a ratio between two quantities, one of which is systematically shrinking.

And now the part that concerns us

This is the point where it would be convenient to write that we examined this route and rejected it.

That is not true.

In July, our own working log contains the proposal to put search volume per prompt cluster into our product as a demand tendency. That is the same approach with a different data supplier. It has been waiting six weeks for a decision. We did not reject it, we let it sit, and the difference between those two is exactly the foresight we would otherwise be claiming here.

On top of that, a number of this build is already in our product. A tile shows search volume for a customer's category terms, and next to it stood the sentence that this tile decides whether a topic has demand.

The first part is right, search volume measures search pressure. The second part leaves open which demand is meant, and it sits in a view called AI visibility. A reader relates that to AI. The number cannot carry that, for the same two reasons described above.

So we have proposed a different label: that the tile measures search pressure in classic search, that it answers whether a topic interests anyone at all, and that it does not say how often someone asks an AI about it. Two sentences, one thing given up. The tile loses none of its usefulness.

Two questions, and whoever confuses them measures the wrong one

It helps to split this into two questions, because almost the entire argument rests on them being mixed up.

The first question is: how many people ask about this at all? It is the more interesting one, and right now nobody can answer it who does not operate an AI themselves. Everything offered on the market for this is an estimate. That includes us.

The second question is: what do the people who reach you ask, and what does an AI fetch from your pages? That question is answerable, and without a conversion factor.

The difference sounds like a limitation, and it is one. It is also the reason the second question has solid answers and the first does not.

What is possible instead of an estimate

Four things can be observed rather than extrapolated, and they are worth different amounts.

Which questions someone opens. If a product page carries questions and a visitor opens one, that is a fact and not a derivation. No keyword, no factor. The limit belongs with it: these are people who are already there.

What visitors ask in free text. This is the most valuable part and the most underrated. A question asked in a chat arrives in the asker's own wording, with budget, occasion and constraint, that is, with exactly the parts that a keyword mapping discards. It is the closest thing to real prompts anyone can have without running a model themselves.

Which pages an AI actually fetched. Measured server side, with a timestamp and, where the provider publishes an address list, with a confirmation that the fetch really came from there. What matters is what this is not: a fetch is not a citation and not a recommendation. It is proof that something was read.

Which questions you could answer but do not. From a company's own material you can derive what it would have something to say about. On its own that is not demand, it is a supply gap.

It gets interesting where these intersect. A question that visitors really asked, that existing material could answer, but that appears on no page, is the only one of these four observations that immediately produces a task. It is neither estimated demand nor abstract potential. It is a gap between something asked and something available.

None of these four observations answers the first question. They answer the second, and they do it without a detour through someone else's search engine.

Why we still do not claim to know better

It would be obvious to turn the previous section into a sales argument. We are not doing that, for a reason that concerns us directly.

All four observations require that someone has already arrived at your site, or that an AI already knows your page. Anyone who wants to know how large the interest is among people who have never heard of them gets no answer from this. That is exactly the question an estimate tries to answer, and the fact that it does so badly does not make it unnecessary.

Anyone who wants the first question answered gets an estimate from us or nothing. We currently consider nothing to be the more honest answer, and that is a position, not an achievement.

What follows from this

For any number offered to someone as AI demand, three questions are worth asking.

What is it derived from? If the answer includes search volume, it measures search engine behaviour and converts it.

Where does the conversion factor come from? If the answer includes clicks, it comes from the shrinking part of usage.

And what does the provider say about the limits? For the method described here, the answer is in their own documentation, and it is clearer than what third parties make of it.

An AI is not a search engine. As long as that holds, any AI demand derived from search volume remains a translation between two languages in which the grammar is lost. That can be useful. It is just not a measurement.

Sources

• Ahrefs, How Ahrefs Estimates AI Search Demand (ahrefs.com/blog/how-ahrefs-estimates-ai-search-demand), retrieved 21 August 2026

• Our own records: proposal "search volume per prompt cluster as demand tendency", 7 July 2026, undecided since

• Our own product: the category demand tile in the opportunity board, relabelling proposed on 21 August 2026