Not sure what font to choose? Typography study helps find the right type

When used correctly, font selection usually goes unnoticed, blending seamlessly with content and reader. When the One Times Square Billboard used a retired Microsoft Word default Calibri font to usher in 2025’s “Happy New Year” message, it was immediately met with sarcastic scorn and delightful derision for the uninspired choice (at least by people who pay attention to such things). Had the font faux pas been the branding rollout of a new app, product, or company, the consequences might have been more severe.

Hanyang University researchers in Korea have attempted to take the intuition and subjective judgment out of the art of font selection. Using computational tools and network analysis to develop an objective framework for font selection and pairing in design, the researchers aim to establish foundational principles for applying typography in visual communication.

Font choice plays a critical role in visual communication, shaping readability, emotional resonance, and overall design balance across mediums. According to the researchers, designers have traditionally relied on subjective rules for font pairing, such as mixing Serif and Sans-Serif or creating contrast. These rules are difficult to formalize and often apply to only a narrow subset of fonts.

Recent advances in AI-based font generation models have focused on creating and predicting fonts rather than studying their systematic use in pairing. Given the growing reliance on computational tools in graphic design, the researchers explored font characteristics and pairing rules that can be more easily incorporated into generated texts and design processes.

In the study, “Typeface Network and the Principle of Font Pairing,” published in Scientific Reports, researchers collected 22,897 font-use cases and 9,022 fonts from Fontsinuse.com, analyzing font use across 19 design mediums, including web design, magazines, branding, and album art.

The visual elements of fonts (uppercase, lowercase, symbols, and numbers) were analyzed using non-negative matrix factorization, reducing font design parameters to three interpretable dimensions: Serif vs. Sans-Serif (X-axis), Basic vs. Decorative letterforms (Y-axis), and Light vs. Bold (Z-axis).

Random forest classification was applied to distinguish between fonts used individually and in pairs. Partial Dependence Plots were used to identify specific font characteristics associated with pairing tendencies across media.

Network analysis methods, including Bipartite Network Projection, were used to model font pairings within each medium. By contrasting real-world patterns with randomized models, authenticity scores for single fonts, pairs, and triplets were calculated.

Different font types for different types of media

Serif fonts like Times New Roman dominate traditional print media, such as magazines and periodicals, with a Mean_X value of 41.95, indicating a strong preference for this font category.

Digital media, such as web and mobile, preferred Sans-Serif fonts, like Helvetica and Futura, with thicker fonts achieving a Mean_Z value greater than 30. This finding aligns with the need for thicker, more legible fonts in smaller screen environments, where pixel clarity is critical.

Helvetica showed high frequency in album art and physical consumer products but was less frequent in film and video. In branding and identity, bold Sans-Serif fonts such as Helvetica Neue ranked highly.

Neue-Helvetica was frequently seen in the “Software Apps” category. Notably, while most fonts had negative authenticity in the “Web” domain, Neue-Helvetica showed significantly positive values.

Random forest models accurately predicted font pairing tendencies, with consistent preferences for specific combinations. For example, branding and identity contexts favored bold Sans-Serif fonts, while booklets and pamphlets often paired light Serif fonts.

Sans-Serif fonts were predominantly paired with other Sans-Serif fonts. Fonts with minimal decorative differences were commonly paired in web and print contexts. Light and bold fonts were often combined across all mediums.

A network analysis of font pairings identified Helvetica, Futura, and Univers as frequently paired across various design contexts. In contrast, decorative fonts like Cooper Black appeared in smaller, specialized clusters where the contrast between Light and Bold styles provides flexibility for emphasis and tone.

Posters and flyers favored font pairings not used in other mediums, predominantly involving Serial B/D, Ogg, and Decorative, such as Eniac and Digestive. Books also showed a tendency different from other mediums, where Sans-Serifs are mainly paired with other Sans-Serifs, and Serifs with other Serifs.

Frequency analysis techniques used in the study are similar to early cryptography, where studying a language’s letter and word frequency might elicit clues to deciphering an encrypted message. Only the message in this case is not an encrypted one, just the specialty of a different academic branch.

It should be noted that the trends in font usage and combinations within the mediums observed in the study begin with intentional choices made by typography experts who understand the nuances of how fonts affect perception and readability and communicate tone and emotional resonance to readers.

As these initial font choice decisions become widely adopted, they may increasingly rely on a more subjective selection process that uses fonts without the underlying understanding. Strong media correlations found in the study suggest that an intuitive sense of the correct font to use in a given context, subjective or not, tends to follow trends set by the original expert selection.

By using existing font choices as classification data, the authors believe they can bridge the gap between subjective design intuition and objective expert criteria, providing actionable insights for designers without font selection experience to enhance their visual communication.

Future research could expand the dataset, explore more complex font combinations, and incorporate font experts or user interaction data to refine the framework further.

© 2025 Science X Network