Humanising AI text.

Seven patterns explain most of the impression — and all seven can be dealt with one at a time. They almost never appear alone, which is why a text usually has to be reworked on several levels at once.

• Uniform sentence lengths. Left to themselves, language models settle at 14 to 20 words per sentence and stay there. Human writing swings hard: a sentence of four words, then one of thirty.
• The list of three. "Fast, affordable and reliable." Three items feel complete and cost nothing in research. That is why they appear automatically — and why they stand out when they turn up in every other paragraph.
• Evaluative adjectives without evidence. Tailored, holistic, innovative, high-quality, bespoke. Each of these words asserts a property the text never demonstrates. Cut them, or replace them with the underlying fact.
• The state-of-the-world opener. "In today's fast-moving digital world …" That sentence spends the most valuable position on the page on a statement nobody disputes and nobody cares about.
• Transition filler. Furthermore, in conclusion it can be said, it is important to note. These phrases announce content instead of delivering it.
• Missing numbers. No price, no date, no duration, no quantity. This is the most dependable signal of all — and the one whose repair improves the text most.
• No position. The text weighs things up without ever deciding. If the reader finishes without knowing what the sender recommends, nothing has been gained.

Four examples from our own editorial work show what actually changes during an edit — usually not the style, but the information content.

Example 1 — replace adjectives with facts.
Before: "We offer tailored web solutions of the highest quality and guide you professionally through the entire process."
After: "A website with us costs between CHF 3,000 and CHF 25,000. The difference almost always comes down to the number of distinct page types, rarely to the design." The second version is shorter, less comfortable, and answers the question that was actually asked. The price range is set out in full on web design.

Example 2 — break up the list of three.
Before: "Our strengths are speed, quality and reliability."
After: "We work from a studio at Gustav-Maurer-Strasse 23 in Zollikon. In Zurich and the neighbouring municipalities we also come on site; otherwise we work through fixed video appointments." Three interchangeable nouns become a verifiable statement that also contains a limitation.

Example 3 — delete the opener.
Before: "In today's fast-moving digital world, a strong online presence is more important than ever."
After: cut it, with nothing in its place. The most common and most effective edit is not a rewrite but deleting the first two sentences. Start with the answer to the question the reader had in mind when she clicked.

Example 4 — create rhythm.
Before: "The loading time of a website significantly influences the user experience. Slow pages often lead to higher bounce rates. Optimising performance is therefore advisable." Three sentences, all between eleven and thirteen words.
After: "Loading time is revenue. Anyone waiting three seconds on a mobile connection for the first visible content is usually gone before the page has finished loading — and does not come back, because a competitor has answered in the meantime." Three words, then forty. The content is the same, the impression completely different. What sits behind this technically is explained in the glossary under Core Web Vitals.

Because they work on the surface, and the problem is not on the surface. A rewriting tool knows neither your company nor your figures nor your customers. So it can only do what is possible without that knowledge: insert synonyms, restructure sentences, scatter sentence lengths artificially, drop in the occasional colloquialism.

Three consequences turn up regularly. First, loss of precision: technical terms are replaced by approximate synonyms — "en dash" becomes "short dash", "bounce rate" becomes "drop-off rate". Second, breaks in tone: a casual "honestly" dropped into a text about accounting software reads as contrived, not human. Third, factual drift: while rephrasing, figures and qualifications get shuffled, and "typically three to six months" turns into "usually a few months".

On top of that, the intended purpose is achieved unreliably. Different detectors score the same edit differently, and a statistical watermark sits, according to the provider, in the word choice across longer passages — light rephrasing, in Anthropic's account, probably does not remove it completely. So you pay in text quality for a result that is not guaranteed.

There is one exception we will name fairly: for thinning out obvious filler in a very long draft, a tool can save time — as a first mechanical pass before the real editorial work, not as a substitute for it.

Six steps, in this order — the first one matters most and is the one most often skipped.

• 1. Gather substance before you rephrase anything. Spend ten minutes with the person who owns the subject. Note down three figures, one real example and one mistake people make. Without that material, every edit is cosmetics.
• 2. Rewrite the first two sentences. They answer the question that produced the click, fully and without a run-up. That is also the passage language models prefer to quote.
• 3. Trade adjectives for facts. Search for "high-quality", "tailored", "holistic", "innovative". Delete each occurrence or replace it with the fact it was trying to assert.
• 4. Build in a limitation. Write down what your service is not suitable for, or when it is not worth it. Nothing reads as more credible, and nothing is quoted more often in AI answers than an honest paragraph about where the limits are.
• 5. Check the rhythm. Read the text aloud. Wherever you fall into a sing-song, the sentence lengths are too similar. One short sentence after a long one is often enough.
• 6. Check the facts, then publish. Every figure, every study, every name is checked against the source. Language models invent evidence convincingly, and that mistake is the most expensive of all.

How this workflow fits into a running editorial operation is described in the article on AI content strategy for Swiss SMEs. Which rules apply from the search engine's point of view is on AI content and the Google guidelines.

Four test questions, answerable in two minutes per text — and far more informative than any detector score.

• Is there at least one figure per section? A price range, a duration, a date, a quantity. If not, what is missing is substance, not style.
• Could this text sit on a competitor's website? If it could, without anyone noticing, it is interchangeable. Try swapping in a different company name.
• Can a single paragraph be lifted out and quoted? It should be understandable and evidenced without the rest of the page. That is the standard by which language models select passages — more on that under generative engine optimisation.
• Would the person responsible say every sentence out loud in a client meeting? If not, it needs rewriting or cutting.

These four questions replace the hunt for the perfect detector score. They measure what decides enquiries in the end, and they hold regardless of which tool wrote the first draft.

In four situations, editing an AI draft is a waste of time — the text does not need polishing, it needs replacing or deleting.

• When the text is factually wrong. Invented studies, incorrect figures and non-existent sources cannot be cured by better style. Rewriting is faster than repairing here.
• When the page is not needed. Generating thirty near-identical location pages is not a style problem, it is a structural one. Google's guidelines target exactly this case, and the right answer is consolidation rather than rewriting.
• When no subject expertise is available. Without someone who genuinely knows the topic, nothing new emerges even after the tenth revision. Commissioning a specialist author is then the more honest decision.
• When the only goal is to fool a detector. For academic work and job applications this is not a question of text but of the rules that were agreed. That belongs with the institution concerned, not with a tool — and this page deliberately does not help with it.

At DLM Digital we use AI in our own content process, for research, outlining and first drafts. What comes afterwards — gathering figures, checking claims, taking a position — is work no tool takes off our hands, and it is exactly the part that decides whether a text produces enquiries. If you need support with it, our ongoing SEO support is the right framework.