Writing for two readers: the human and the machine

What the research found about citable facts

The clearest evidence comes from the team that first studied how to be cited inside AI answers.

In 2024, researchers from Princeton, Georgia Tech and IIT Delhi tested nine ways of editing a page and measured how each one changed the page’s visibility in generated answers. Their GEO study (KDD 2024) reported that adding citable statistics raised that visibility by roughly 41%, and that adding direct quotations helped by about 28%.

The same study looked at the old SEO reflex of repeating keywords and found it did nothing useful — in places it performed slightly worse than leaving the page alone. Density was not the signal. Verifiable, attributable substance was.

That result is easy to misread as a trick for machines. It is closer to the opposite: a named source and a specific number tend to make an argument more persuasive to a person too. The edits that helped the engine were, broadly, the edits that help any careful reader.

How an engine actually reads a page

To see why those edits matter, it helps to know what the machine reader does with a page.

A person reads top to bottom, forgives ambiguity, and fills gaps by inference. A language model does almost none of that. It splits the page into passages, turns each into a representation of its meaning, and — when someone asks a question — pulls back only the passages that sit closest to it.

As Anthropic notes in its work on contextual retrieval, a passage that loses its meaning once it is separated from the paragraphs around it is a passage the engine struggles to use. The machine reads in fragments. A claim that only exists across a whole arc has no single fragment that carries it.

This is why the problem sits earlier than ranking. Before an engine can summarize or cite a page, it has to lift one piece of it out of order and still restate the point correctly. If the meaning depends on reading the full passage in sequence, the model tends to pass the page by.

The two readers want opposite things

Set the two readers side by side and the conflict becomes concrete rather than abstract.

A human reader rewards a hook, momentum, and the pleasure of a sentence that turns. The machine reader rewards a plain claim it can quote, labelled structure, and proof stated openly rather than implied. Neither appetite is wrong. The difficulty is that satisfying one often starves the other.

Two readers, opposite appetites — and one page has to feed both

Most pages quietly resolve this in one direction without noticing they made a choice. Beautiful, voice-led writing whose point lives in flow gives the model nothing clean to extract. Over-structured, keyword-heavy copy gives the model something to parse but gives the person no reason to stay.

Why this matters more each year

The cost of satisfying only the human reader is rising, because the human increasingly sees the page only after a model has chosen it.

A 2025 Pew study found people click a result about 8% of the time when an AI summary sits on top of it, against 15% without one. By early 2026, roughly two-thirds of Google searches ended with no click at all.

When the click disappears, the prize shifts from being ranked to being named inside the answer. And to be named, a page first has to be readable by the model deciding what to say. Semrush’s analysis of more than 150,000 citations points the same way: engines lean on sources they can extract from cleanly, and tend to skip the rest regardless of how good it reads.

Resolving the tension without losing the voice

Taken together, the evidence suggests a single discipline rather than a choice between two kinds of writing.

Lead each unit — a section, a paragraph, an answer — with a clear statement a model could quote out of context and still get right. Then develop it with the evidence and voice a person reads for. The structure is load-bearing but largely invisible; the story sits on top of it.

A good magazine feature has always worked this way. It opens a section with a sentence that states the thing, then earns it with reporting. You could quote that opening line on its own and be right; you read the rest because it is good. That habit predates AI by a century — it simply matters more now that a fragment reader goes first.

There is a plain test for any paragraph already on a page: could a model quote one sentence of it, out of context, and state the point correctly — and would a person still want to read on? When the answer to both is yes, the gap between the two readers has closed. That gap, and the work of closing it, is what we have come to call the Dual Legibility Tension: the quiet reason careful, well-made writing tends to be the writing that also gets cited.

Applied across a brand’s whole body of claims rather than a single page, this is the work we call Narrative Architecture. But the name matters less than the habit. Start with the next paragraph you write.

Frequently asked questions