Assistants' answers: what can be observed and what is guessed

This is a new subject, and therefore one where a great deal can be told that is never checked. So let us separate up front what can be counted from what probably never will be.

What can be measured, and what changed recently

When someone clicks a link from an assistant’s answer, the visit arrives with the assistant’s address as its referrer. Those visits are counted like any other source, and analytics tools now isolate them rather than lumping them into referral traffic.

That is a real change, and it is worth saying rather than repeating that nothing is measurable. You can know how many visits come from that channel, which pages they land on, and what they do next. If the point is to judge whether the channel produces enquiries, that is enough.

What stays out of reach

The gap is not on the visits side, it is on the reads without a click. An assistant that cites your name, summarises your position and sends nobody your way leaves no trace at all. And that is the most frequent case, the one that makes the channel valuable to a brand.

Two other blind spots, more technical. Some visits arrive with nothing indicating where they came from, because the link was copied by hand or opened in an embedded browser that does not pass the referrer. And nobody can guarantee a citation, nor reproduce an answer identically twice running: asking the same question twice gives two answers and sometimes two lists of sources.

Anyone selling a guaranteed place in the answers is selling a certainty they do not have.

What can be observed, and does not move

Assistants cite pages that answer a precise question directly, in few words, with facts that are dated and attributed.

They handle badly three kinds of page: those that circle a subject without ever settling it, those whose content only appears once a script has run, and those with no identifiable author.

A good page for an assistant therefore looks a great deal like a good page for a reader in a hurry. That is reassuring and disappointing at once.

The work largely overlaps with ordinary search

A page that is fast, readable without JavaScript, structured and honest serves both channels. There is no separate discipline to buy.

The only difference is the shape. One question per page, the clear answer in the opening lines, then the detail. Plenty of company pages do the opposite: three paragraphs of context, then the answer halfway through the fourth. A reader in a hurry gives up before that, and an assistant finds context there but no fact to reuse.

The decision we do not take for the client

The files that allow or refuse the assistants’ crawlers are a decision, and one that is often taken badly. The crawlers’ names and how to handle them are published by the vendors themselves, at OpenAI as at Perplexity: the decision can be taken on the record rather than on forum hearsay.

Allowing means accepting that your content feeds their answers, with or without a citation. Refusing means disappearing from that channel. There is no general right answer: it depends on what the content is worth and on what the company sells. A publisher living off its articles and a consultancy looking for incoming calls do not have the same interest.

We put the question and we document both consequences. We do not settle it.

What can therefore be claimed, and how

A figure for visits from an assistant is a real figure, and it is presented as one. A figure for visibility in the answers is not: it is a convergence of evidence, made of regular tests on the questions that matter to the business, brand mentions followed, and direct visits rising without a campaign.

The difference between the two is stated when the table is presented, not in a footnote.