First the snippet answered the question in place. Now a generated paragraph composes an answer from several sources, cites three of them, and satisfies the searcher before any website loads. For publishers this looks like the zero-click story at higher volume — and at first glance it is. But look at which sources the machines cite, and a different picture emerges: the selection is not random, it is legible, and most of what it rewards is what good editorial always was.
What citation selection appears to reward
Observation across generated answers keeps converging on a short list. Answer-shaped pages: a clear question answered early and expanded after — the machine lifts passages, and passage-shaped prose gets lifted. Specificity: numbers, names, dates, procedures; the quotable fragment is the citation unit, exactly as in the link-earning guide. Coherent structure: headings that segment meaning, since retrieval systems work at the fragment and its context. Corroborated trust: pages already cited by humans, because human citation is the training signal the machine inherited. The uncomfortable summary: nothing here is new. The machines are a fast, literal readership grading the homework good editors always assigned.
What this changes tactically
- Write for extraction, not just reading. The question stated plainly, the answer in the first breath, the specifics close by. This is also plain good editorial — see the checklist in our intent guide.
- Mark up meaning. Structured data and clean semantics make fragments machine-legible; the technical case is made in the technical guide.
- Own the definitional clusters. Generated answers aggregate around established reference pages; being the standard answer for a cluster compounds, per the research guide.
- Be findable at the second question. Many cited visits arrive from follow-up queries with no referrer, branded or otherwise. Recognise them in your analytics before declaring a collapse.
Measuring the half-invisible
The measurement problem is real: citations arrive as anomalies — direct and branded spikes, unattributable sessions, platform-level noise. Our working practice: track branded search volume as the citation proxy, watch for query phrasings that echo your headings, and treat unattributed direct traffic to deep pages as a signal rather than a nuisance. None of it is exact. All of it beats the alternative, which is optimising blind — or worse, concluding from missing referrer data that nothing is happening.
The blue link rewarded the page that best answered. The generated answer rewards the page most worth quoting. The overlap is enormous — which is either reassuring or sobering, depending on your pages.
The vocabulary shift
One quiet change deserves its own paragraph: the unit of competition is drifting from the page towards the passage. When a generated answer assembles fragments, it is fragments that get compared — a page’s fate can turn on whether its best two hundred words stand alone. This is not an argument for chopping pages into confetti; context is what makes a fragment trustworthy. It is an argument for writing sections that survive extraction: a heading that states the claim, a first sentence that delivers it, and specifics that make both worth citing. The discipline is old — technical writers have called it “topic sentences” for a century. The machines simply grade it now.
The strategic reading
Strip the novelty and the strategy is the one this magazine has always argued. Build method-first pages with specifics worth quoting; keep the pipeline legible; bank readers into owned channels so no intermediary holds the whole relationship. The machine surfaces change the odds on individual clicks — see the zero-click note — but they raise, not lower, the return on being genuinely the best answer. That was true under directories, true under rankings, and is true under generation.
The open question we are watching: whether citation share becomes measurable enough to become an industry metric of its own. When it does, this desk will report it — with caveats, as usual.