The Pew Research Center analysed around 500,000 English-language websites to determine whether their texts were machine-generated. The result: over a third of the pages published after the launch of ChatGPT show traces of machine text generation. The distribution is anything but even — according to the study, commercial .com domains are ten times more affected than pages from the education sector or from public authorities.
For companies that see their website as a sales and trust channel, this is a relevant assessment of where they stand. Not because AI support in the editorial process is fundamentally problematic. But because the underlying value of a text falls when it becomes interchangeable.
Why the distribution is more interesting than the number
That commercial domains in particular are out in front comes as little surprise. That is where the economic incentive is greatest to produce content in large quantities and at low cost: category texts, guide articles, local landing pages, blog posts for visibility in search engines. Educational institutions and public authorities, by contrast, publish less often, more slowly and usually for a specific professional reason.
But that also means: anyone operating in the commercial environment is competing with a greatly grown volume of similar-sounding content. The competition for attention thus shifts from quantity to distinctiveness. A text that merely summarises what is already everywhere will find it harder in future — regardless of who or what wrote it.
What companies can take from this
The practical consequence lies less in doing without AI tools than in the question of where in the editorial process they are used. Their use makes sense where structure, preparatory work or variants are needed: drafting outlines, creating rough drafts, tightening wording, preparing translations, suggesting metadata. It makes less sense where the actual value of the content arises — in specialist knowledge, project experience, concrete figures from one's own company, verifiable statements.
Three points that can be implemented directly in website projects:
- Substance before volume: Better twenty pages with their own specialist statement than two hundred pages of platitudes. The latter tie up maintenance effort without generating visibility.
- Establish traceability: Make authorship, update date, specialist context and sources visible. This helps readers and is at the same time a signal for search engines and AI systems that reuse content.
- Anchor editorial control: A machine-generated draft needs specialist review by a human before it goes live. This approval should not be an informal intermediate step but a defined part of the workflow.
The CMS side of things
This is precisely where the topic becomes a question of system architecture. A content management system — that is, the software with which editorial teams maintain content — can support or hinder quality assurance. Helpful are, among other things:
- staged workflows with draft, review and approval, so that no content is published unnoticed
- maintained author and editorial data that can be output as structured information
- version states and change logs, so that it remains traceable later who adjusted what and when
- processes for regular content review, so that outdated pages are updated or removed
TYPO3 provides these building blocks at its core: workspaces for editorial preliminary stages, a permission model that maps review steps, and logging of changes. If AI assistance is integrated, it should act within this framework — as a suggestion level within the workflow, not as a side channel through which content travels directly into delivery.
Where labelling makes sense
The study says nothing about how readers react to labelled AI support. The only robust finding is the observation that the proportion of machine-co-written content has risen sharply. Nevertheless, a conscious decision within the company is worthwhile: in which formats is AI used, what counts as editorially responsible, and is this made transparent to the outside world? An internal policy that answers these questions is usually more valuable than a blanket notice in the footer.
Areas in which misinformation causes damage are particularly sensitive: legal notices, medical or technical information, price and service descriptions, safety information. There, every statement should be checked by a person with specialist responsibility, regardless of how the draft came about.
Assessment for web and software projects
The figures from the study make visible what is already noticeable in projects: pure text volume is losing impact, while traceability and specialist depth are gaining weight. For website projects, this means thinking about editorial processes and approval stages early rather than setting them up afterwards. And for tool selection, it means integrating AI functions where they take work off people's hands without diluting editorial responsibility.