🎉
Revolution wins Sustainable Business of the Year 2026!
Revolution wins Sustainable Business of the Year 2026!
People scan websites. They have done so since at least 1997, when Jakob Nielsen ran the first eye-tracking studies of web users and found that around 79% of test users scanned new pages while only about 16% read word-by-word. Almost three decades and a few internet revolutions later, the same finding holds.
The arrival of generative AI raised a reasonable question. Would reading behaviour shift now that large volumes of online text are written by machines? The honest answer, based on what the research shows, is that the underlying behaviour is stable. AI has changed what gets written, and to some extent what readers think about what they are reading. It has not changed how their eyes move across a page.
The most cited model of web reading is the F-shaped pattern, documented by Nielsen Norman Group (NN/g) in 2006. On text-heavy pages, users’ gaze traced two roughly horizontal sweeps across the top, followed by a vertical scan down the left edge. Headlines and the start of paragraphs got attention. Everything to the right and below trailed off.
In a 2017 update, Kara Pernice at NN/g clarified that the F-pattern is what users fall back on when content is poorly structured. It is a default behaviour that emerges in the absence of clear formatting. Better-organised pages produce different scanning patterns: layer-cake (skipping between headings), spotted (jumping between keywords), commitment (sustained engagement), or bypassing (skipping the opening entirely to reach a specific element).
Analytics data supports the eye-tracking findings. A Chartbeat analysis of millions of articles, published in Slate in 2013, showed that a large share of readers never scroll to the bottom of a page. Many leave before the halfway point. Sharing and reading correlate only weakly, which means many shared articles were never finished.
The arrival of generative AI has changed several things about the web. The volume of text being produced has risen dramatically. Search results increasingly contain AI-generated summaries before any link. A meaningful share of the content people land on was either drafted by AI or heavily edited from AI output.
It has not changed how people read individual pages. The scanning patterns described above predate ChatGPT by a decade or more, and the available evidence does not show them shifting. Once a reader lands on a page, scanning is still scanning.
The more interesting shift is in what happens once suspicion of AI authorship enters the picture.
Audiences are not particularly good at reliably detecting AI-generated text. They are, however, quick to form impressions about it, and those impressions have measurable effects on behaviour.
A Bynder consumer study found that 62% of consumers are less likely to engage with or trust content when they believe it was AI-generated. A large-scale US study reported that reader trust drops by close to 50% when content is believed to be AI-written, and ads placed next to that content lose around 14% of their effectiveness. Experimental work comparing identical ads labelled “AI-made” and “human-made” showed lower ratings and fewer clicks for the AI-labelled version, even though the content was identical. Disclosure does not solve this. The reaction tracks the perception of AI authorship more than it tracks the content itself.
This matters for the scanning question because suspicion accelerates disengagement. A reader scanning a page is already deciding, very quickly, whether to stay. If something in the first few seconds signals “AI”, the threshold for leaving drops. The scanning behaviour itself has not changed. What it produces, in terms of attention and engagement, has.
The signals that trigger suspicion tend to be specific. Generic openings. Lists of three. Phrases that sound competent without saying anything. The polished blandness that AI tools default to when given underspecified instructions. Once a reader notices this register, they tend to treat the rest of the page as not worth engaging with, whether or not the content was actually produced by AI.
The implications for UX design draw on long-established principles that now carry more weight.
Front-load conclusions. The inverted pyramid, where the main point appears first and supporting detail follows, predates the internet and remains the most reliable structure for the web. A scanner who reaches the answer in the first paragraph might stay. One who has to work for it usually will not.
Use descriptive headings. Readers use headings to decide whether to read the paragraphs underneath them. A heading that does not describe its content wastes the only attention the page is likely to get.
Keep paragraphs short and focused. Walls of text trigger the bypass pattern, where readers skip the section entirely. Short paragraphs with white space around them let scanners drop in and out without losing their place.
Add a summary where it fits. A short summary at the top serves three audiences at once: scanners who want the conclusion, deeper readers who want to know what is coming, and AI extraction tools that pull summaries into answer boxes.
Write with a recognisable voice. This is where the suspicion research becomes design advice. Generic, polished, opinion-free prose reads as AI-written regardless of who wrote it. Specific examples, an identifiable perspective, and a willingness to be concrete are what separate content that holds attention from content that gets bounced off in three seconds.
The web has always rewarded content that respects how people actually read it. Generative AI has not changed the underlying behaviour. It has, however, raised the cost of getting the experience wrong, because a page that triggers AI suspicion now loses attention faster than a comparable page would have a decade ago.
So, the behaviour is stable, yet the penalty for ignoring it has gone up.
