Since generative language models have become widely available, texts are being produced faster and more cheaply than ever before. A recent study concludes that a large proportion of the websites examined already exhibit characteristics that suggest machine generation or machine assistance. The distribution is not uniform: there are areas of the web in which such tools are used significantly more frequently than in others.
For companies that operate their own website, this is less a technical than an editorial message. The competition for attention increasingly takes place in an environment in which arbitrary quantities of linguistically flawless but substantively interchangeable texts exist. Anyone who wants to stand out there cannot do so through sheer volume of text.
Why detection rates should be read with caution
All figures on AI shares on the web warrant restraint. Detection methods work with statistical patterns – such as conspicuously uniform sentence structure or typical formulation preferences – and deliver probabilities, not proof. A carefully edited AI draft can be inconspicuous, while a very factually written human text can be classified as machine-generated.
In addition, the boundary dissolves in practice. Between "fully automatically generated" and "written by hand" lies a broad field: research summaries, rough drafts, translations, rephrasings for other channels. Most editorial teams today operate somewhere in this field, without always consciously naming it.
The relevant statement of the study is therefore not an exact percentage but the direction: machine-generated or machine-assisted language is no longer an exception on the web but a normal part of the inventory.
What gains value as a result
When linguistic smoothness becomes a mass commodity, the value of content shifts to the components that a language model cannot contribute itself:
- Own data and experiences. Figures from your own operations, measurements, project histories, failures and their causes.
- Concrete cases instead of general guide texts. A described project with framework conditions, decisions and outcome is difficult to imitate.
- Professional position. A reasoned recommendation, even if it is inconvenient, differs from a balanced enumeration of all possibilities.
- Currency with substance. Assessments of developments that are not yet in training data.
- Verifiability. Traceable origin of statements, named responsible parties, verifiable details.
Editorial guardrails instead of prohibition
A general ban on AI tools in content creation is neither enforceable nor sensible in most organisations. More practical is a set of rules that defines what the tools may be used for and where human review is mandatory. A few principles have proven effective:
- Facts must always be checked. Figures, names, dates, quotations and legal statements are checked against a named source before the text goes online.
- One person is responsible for every contribution. Even if a model supplied the rough draft, there is a named approval.
- No text without an own contribution. If a contribution contains nothing that a model alone could not have formulated, it is not published.
- Define linguistic signature. A concise style guide with desired and undesired formulations prevents all content from sounding the same.
- Confidential data stays out. Customer data and internal documents do not belong in external services unchecked.
- Clarify how to handle transparency. Whether and how you label the use should be a conscious decision and not left to chance.
Technical implementation in the CMS
Such rules only take effect if they are anchored in everyday editorial work. A content management system such as TYPO3 can support this: through approval workflows with defined roles, through mandatory fields for source references and responsible parties, through versioning that makes it traceable who changed what and when, and through checklists directly in the editing form.
It also makes sense to keep an eye on the inventory. When content is created quickly, the number of outdated, duplicate or superfluous pages grows quickly too. A regular inventory review with clear decisions – update, merge, remove – keeps the website lean and quality visible.
The second addressee: answer machines
Today, content is not only read by people but also evaluated by search systems and AI assistants that assemble answers from it. The same qualities help there too: clearly structured pages, unambiguous statements, verifiable details, clean markup of content. Generic texts without their own core have little to offer in either channel.
Assessment
For web and software projects, the focus thus shifts: the generation of text is hardly the bottleneck any more, quality assurance is. Anyone planning or revising a website today should take editorial processes, roles and approvals just as seriously as design and technology.
The difference between any website and a credible website will in future arise primarily at this point.