DYNAMICS-WEB: A Pre-Registered Measurement of Web Content Character over the Link Graph, and a Powered Proof that its Sub-Perceptual Half Changes Conversion

For decades, the internet has been organized by two main rules. The first rule is about authority: how many other important websites point to a page, which tells us how much standing it holds in the digital world. The second rule is about topic: what subject the page is about, ensuring that a page about cooking appears when you search for recipes. These two signals are global; they give every single reader the exact same answer about a page's importance and its subject. But they leave out a crucial third dimension: how a page actually feels to read. Two pages can be about the same topic and have the same level of authority, yet one might feel rigorous and warm while the other feels cold and sales-driven. This difference is not about facts or popularity; it is about character. Until now, there has been no way to measure this character at scale, or to understand if it actually changes what a person decides to do when they visit a website.

A researcher named Jason Duke set out to measure this missing quality. He developed a system called DYNAMICS-WEB, which treats the character of a webpage as a distinct property that can be quantified across eight different dimensions. These dimensions include how careful and precise the writing is, how deep the content goes, how original it feels, how honest it is, how much emotional warmth it carries, how strongly it tries to sell something, how opinionated it is, and how formal or accessible its tone is. The team did not just guess at these qualities; they trained a large computer model on thousands of human judgments to recognize these patterns in text. They then applied this model to millions of websites, creating a detailed profile for each one that describes exactly how it reads, independent of its topic or its fame.

The study found that this character is something entirely new. It is not just a fancy way of measuring how authoritative a site is, nor is it simply a reflection of how easy the text is to read or how positive the words sound. When the researchers tried to predict a page's character using only its authority or its surface-level writing style, they failed to capture most of it. The character signal carried information that those other tools completely missed. Furthermore, this character is not static; it flows through the web. Just as authority spreads when one site links to another, the character of a website influences the sites it connects to. A page with little text of its own can still have a clear character because it is surrounded by other pages with similar qualities. The researchers proved this by showing that a website's character could be predicted by the character of its neighbors in the web graph, confirming that this quality travels through the links that bind the internet together.

Perhaps the most surprising discovery was that this character operates below the level of conscious thought. When the researchers asked human readers to rate pages on these eight dimensions, the readers could not agree with each other. They could easily spot obvious traits like how much a page was trying to sell something, but they struggled to name the more subtle qualities, such as how deep or honest the writing felt. The computer model, however, could measure these subtle traits with high consistency. This suggests that readers feel these qualities without being able to articulate them. The study tested whether this invisible feeling mattered by running a controlled experiment where 117 people were shown offers that were identical in price and facts, but differed only in their writing character. When the subtle, unnameable character of the page matched the reader's own preferences, the rate at which they chose to buy or sign up jumped from 19 percent to 31 percent. The loud, obvious traits, like how hard the page tried to sell, had no effect on the decision.

The research also showed that this method works across different languages and cultures. The team tested the system on Arabic, Chinese, and Swahili, languages with very different writing systems and cultural contexts. In every case, the model correctly identified that academic pages felt more rigorous and less commercial than business pages, proving that the concept of character is not limited to English. They also demonstrated that the same mathematical engine used to measure character could be used to figure out which country a website belongs to, simply by looking at who links to it, even if the website's address does not use that country's code. This confirms that the web graph holds many different kinds of information, not just authority.

Ultimately, the paper establishes that the internet is richer than we thought. It is not just a hierarchy of important pages, but a space where the "voice" of a page matters as much as its facts. The study proves that this voice is a real, measurable signal that flows through the web and influences human behavior in ways we cannot consciously explain. While the effect on conversion was modest, it was real and specific to the parts of the character that people feel but cannot name. The researchers have built a tool that can read this character on any page, running locally on a user's device to protect privacy, offering a new way to understand how content connects with people. The work suggests that the future of the web may lie not in ranking pages by how famous they are, but in matching them to how they feel to the person reading them.