Cited without being visited

By no longer counting visits. The figure that forces the change of reasoning is this one: between 82 and 88% of citations in a Perplexity answer produce no visit at all to the cited site. Put another way, more than eight times out of ten your brand is named next to the right answer, somebody leaves satisfied, and nothing shows up in your statistics.

The wider movement points the same way. Around 60% of searches now end without a click, click-through rates fall by more than half on queries covered by a generated answer, and the traffic Google sent to publishers dropped 38% in a year. This is not an algorithm adjustment, it is a change in the nature of the channel.

The paradox that makes this hard to defend internally

Here is the difficulty facing anyone trying to get budget for this work. Traffic arriving directly from assistants is still under 1% of referral traffic, which is easy to wave away in a meeting.

Except that those visitors convert at several times the rate of conventional organic traffic, and the reason is obvious once you think about it. Somebody who arrives after asking a precise question, reading a structured answer and deciding to check the source is not in the same state of mind as a visitor who stumbled onto an article from a keyword search. They have already filtered; they are coming to verify.

To which you add the invisible part, which is probably the most important. When 94% of B2B buyers say they used a generative tool during their purchase process, the question is no longer how many of them clicked: it is whether your name came up when the assistant listed the options. That appearance leaves no trace and weighs more than the click it did not produce.

What it changes about what you write

We built this blog on that assumption, and it shows in choices that would look odd to a writer trained in conventional search optimisation.

Every article opens with the answer, with no scene-setting introduction, because a model takes the first passage that answers and stops reading. Every subheading is phrased as a question, because that is what makes a section extractable without the rest of the page. Every number carries a clickable source, because a figure without one is a figure no assistant will repeat on its own authority. And every article is signed by a person with their role, which is not decorative.

There is a downside we took a while to see. Optimising for extraction pushes you towards writing fact sheets: short sentences stacked up, two-line paragraphs, no breathing room. It is perfectly citable and perfectly unpleasant to read, and a human reader who disengages is a reader who does not come back. The rule we now apply is finer: the first sentence of each section stands alone, and the rest of the paragraph breathes.

What it changes about being chosen

For a staffing firm the shift shows up in three places, and none of them belongs to the marketing department.

Your job postings are read by machines before they are read by candidates. They feed tools that write applications, and they are also what an assistant summarises when somebody asks what the market looks like for their profile. A vague posting produces a vague answer, and you will never know the question was asked.

Your clients look for you the same way. A technical director asking an assistant which firms work on a given technology in their region will get a list, and that list is built from what is written publicly, precisely and verifiably. A company page that says nothing concrete gets picked up nowhere.

Finally, your employer reputation travels without your being able to correct it. Reviews, testimonials and articles about you feed answers nobody will submit to you for approval.

Why a model cites one source rather than another

The question comes up every time and there is no official answer, since no model provider publishes its selection mechanics. What can be observed, though, lines up reasonably well from one assistant to another.

Specificity beats authority. A page that answers exactly the question asked gets picked up ahead of a general page on a more powerful domain, which is excellent news for a site starting out and bad news for large portals of recycled content.

Verifiability comes next. A number with its source, a precise date, a quotation attributed to somebody: these are things a model can repeat without taking a risk, and it mechanically prefers a sentence it can attribute to one it would have to own.

Freshness matters more than in conventional search, particularly at Perplexity, which visibly favours what has moved recently. An article updated with a visible revision date moves back ahead of an older piece on the same subject.

Finally, agreement across sources counts for a lot. A claim found in three independent authors is repeated more readily than an isolated one, even a better-written one. That is one more reason never to contradict yourself from one article to the next, and it is what makes consistency of figures across our own pages more important than it looks.

Measuring a channel that leaves no trace

This is the point where honesty requires admitting there is no good method yet. Conventional analytics count visits, and this channel is characterised precisely by their absence.

What we do, for want of better, comes down to three habits. We query the main assistants regularly on fifteen or so questions that matter to us, recording whether we appear, in what terms and alongside which competitors. We watch the identifiable share of traffic coming from an assistant, knowing it massively understates reality. And we listen to what people say in meetings, because “I asked ChatGPT and it mentioned” has become an ordinary sentence.

One useful marker so as not to conclude too fast: allow four to eight weeks between publishing an article and its possible appearance in an assistant’s answers. Content judged ineffective after a fortnight has not been judged.

What we think

The good news is that this channel rewards exactly what a blog should be doing anyway. Answer precisely, source your numbers, take a position, write what nobody else can write. There is no keyword-stuffing shortcut, because a model that repeats a false claim makes its users repeat it, and model providers pay attention to that.

The bad news is that measurement disappears at the moment the budget gets discussed. You have to accept investing in a channel whose return cannot be proven for several quarters, which is a difficult conversation in a company run from a dashboard.

Our bet is the same one we make on the commoditisation of intelligence: when producing content becomes free for everybody, what stays scarce is what cannot be generated. A position you stand behind, a number you went and checked, a mistake you describe. It is also the only thing a model has a reason to cite rather than paraphrase, and the same logic that hollows out interfaces applies here: what counts is no longer being seen, it is being repeated.