PRISM™ scores 10 linguistic dimensions across thousands of publicly accessible web pages. The index tracks these scores over time for – organisations across – sectors.
Each dot is an organisation. The horizontal axis shows the PLI score for the selected year. The vertical axis shows change from baseline (2015–2019 average). The dashed line marks the corpus mean. Filter by sector or drag the year to animate.
Average PLI by sector, 2015–2025. Lines that were far apart are drawing closer together. The grey band marks the corpus mean ± one standard deviation. The narrowing band is the signal.
Each sector responds differently to the same pressures. Some voices are eroding. Others hold steady. Select a sector for the full trajectory, organisation breakdown, and L1 metric detail.
Cross-cutting patterns from the corpus. Each card contains its own evidence and visualisation.
A Milan fashion house and a European management consultancy — organisations with distinct histories, audiences, and product categories — registered 99.8% pairwise similarity on the PRISM™ index in 2023–2025. Their baseline similarity in 2015–2019 was 97.4%. The residual distance between two unrelated voices has effectively closed.
The interquartile range of PLI scores across the corpus narrowed at a consistent pace from 2015 to 2021. Between 2022 and 2023, the rate of compression doubled. The timing aligns with widespread adoption of generative AI tools in institutional content production.
When the same organisation publishes in multiple languages, the English-language output consistently scores lower on Cultural Markers and Idiom Density than its counterparts in French, Italian, or German. English, as the dominant language of LLM training data, may be more susceptible to algorithmic flattening.
Government bodies shifted an average of 3.1 PLI points from baseline, the smallest movement of any sector. Readability and sentence structure remained consistent across the period, suggesting that formal approval processes, editorial sign-off chains, and established style guides acted as a buffer.