The Pew Research Center examined newly published web content for characteristics of machine-generated language. The result: such traces are found on around 35 per cent of new web pages. Since the release of ChatGPT, the proportion has risen significantly.
The figure initially describes only an observation, not an assessment. A text with AI traces is not automatically bad, and a purely manually written text is not automatically good. The figure is interesting for another reason: it shows that text production has ceased to be a bottleneck. When content is available in any quantity, it is no longer the existence of a text that decides visibility, but its distinctiveness.
What actually changes for website operators
Many companies built their content strategy at a time when writing was expensive. Editorial plans, topic clusters and keyword lists were tools for distributing scarce writing capacity sensibly. That scarcity no longer exists in this form.
As a result, the effort shifts. It is not the creation of a raw text that costs time, but the work before and after it: Which question should the text answer? Which information exists in the company and nowhere else? And is everything that is written there actually correct?
This is precisely where problems arise in practice. Language models write confidently even when they lack the factual basis. A text can be fluent, well structured and partly wrong — a combination that is harder to spot than a visibly bumpy text.
Substance comes from the company, not from the model
The difference between interchangeable and sustainable content rarely lies in the wording. It lies in the substance: concrete project figures, experience from operations, technical decisions including their rationale, the limits of one's own offering.
This information is stored in heads, tickets, quotations and minutes — not in the training data of a language model. Those who feed it into the content process produce content that cannot be reproduced at will. Those who do not create texts that every competitor can also produce in a similar form.
Sensible division of labour in the content process
A practicable division can be derived from this observation. Language models are strong at tasks whose result is easy to verify:
- Structuring rough drafts from existing notes, interviews or documentation
- Shortening, organising or adapting existing texts for a different format
- Deriving meta descriptions, summaries and teasers from finished texts
- Creating translations as a basis for specialist editing
- Checking terminology and tone across larger bodies of text
Tasks in which the model first has to invent content are less suitable: specialist statements without a source, figures, customer examples, legal notices. Here the effort merely shifts — from writing to cross-checking, often without any gain in time.
Quality assurance needs an accountable owner
An AI-supported content process only works with a clear approval authority. Concretely, this means: a person with specialist knowledge reads the text before publication and takes responsibility for the statements. This role cannot be delegated to a tool.
Simple organisational rules are also helpful:
- Document the origin: Record in the editorial system which content was prepared by machine. This makes later revisions and troubleshooting easier.
- Fact-checking as a separate step: Check figures, names and technical details against a reliable source, not against a second language model.
- Fewer, but more sustainable: One well-maintained article per month is more useful than ten generic ones. Maintain old pages instead of creating new ones.
- Clean up the existing stock: Content without substance burdens structure, navigation and trust. Deleting is a legitimate measure.
Technical requirements in the CMS
Whether such processes are adhered to also depends on the editorial system. In TYPO3, workflows with separate roles for creation and approval can be mapped, working versions can be checked before publication, and structured content elements can be specified that enforce a certain editorial discipline.
Where AI support is integrated directly into the editing interface, it should complement this structure and not bypass it. An assistant that writes suggestions into a working version field is uncritical. An automation that publishes without a review step shifts the risk directly into the public domain.
Credibility as a scarce resource
If a considerable proportion of new web content carries machine characteristics, the value of content that demonstrably stems from experience grows. Readers develop a sense for texts that use many words and say little. Search systems and AI answer services also assess sources by whether they provide reliable information.
For companies, this does not mean banishing AI from the content process. It means using it where it takes work off their hands, and contributing their own expertise where it makes the difference.
For web and software projects, this leads to a sober consequence: the bottleneck shifts from production to editing and maintenance. Anyone planning a project should budget not primarily for the quantity of content, but for approval processes, specialist review and the ongoing updating of the existing stock.
A CMS that properly maps roles, versions and approvals is not a comfort feature, but the foundation for quality assurance actually taking place in everyday work.