Last week we reviewed two client projects before go-live. Not for broken links or loading times, that had long been sorted, but for a single character: the em dash. The result on one of the two projects: several hundred hits, more than 250 of them in the image descriptions and titles of the media library alone. The rest were spread across page copy, teasers and even the theme settings.<\/p>
A single dash is harmless. In that quantity it becomes a label, and the label reads: an AI wrote this, and no human read it back. Readers now sense it after a few paragraphs. Colleagues in the industry spot it at a glance. And the search engine draws its own conclusions long before the first visitor bounces.<\/p>
Let us be clear about one thing: we work with AI every day. First drafts, research, translations of entire websites, without AI the pace we work at today would be unthinkable. The difference lies in a single step that 99 per cent of beginners skip: the editing. That step is exactly what this article is about: spotting AI text, understanding the markers and removing them cleanly.<\/p>
The routine is always the same: open ChatGPT, type in a topic, copy the result, paste it into WordPress, publish. Ten articles in one afternoon, and the content problem looks solved. Three months later, Search Console shows a few dozen impressions and zero clicks. The conclusion is usually: SEO is dead, blogging is a waste of time.<\/p>
Neither is true. The text was simply never finished. An AI draft is cleanly structured, grammatically correct and usually sound in substance. But it sounds exactly like the ten thousand other AI texts on the same subject that went online in the same month. Publish the draft unchecked and you publish average. And average has no claim to page one of a search engine.<\/p>
Let us clear up a misunderstanding first: the official Google guidelines state explicitly that the origin of a text is secondary, as long as the content helps people. There is no ban on AI text. But there is no reward either.<\/p>
What Google has actively devalued since the last core updates is mass-produced content with no added value of its own, which the spam policies call “scaled content abuse”. The typical AI markers work like a self-declaration: they show everyone that nobody edited this. Then there are the user signals. A text that not even its own author has read will not be read to the end by anyone else: short dwell time, high bounce rate, no interaction. The search engine sees this data, and it judges accordingly.<\/p>
The best-known marker first. AI models place a dash in almost every other paragraph, people do not: there is not even a dedicated key for it on the keyboard. In editorial teams and on LinkedIn, the long dash now counts as the AI giveaway par excellence. Once you start noticing it, you cannot look away again.<\/p>
Our rule for this is radically simple, and it comes not from an SEO guide but from an internal announcement: no em dashes, anywhere. A comma, a colon or a new sentence instead. The short hyphen stays only where it belongs inside a word: in compounds, number ranges and phone numbers. This one rule changes the sound of a text more than you would expect.<\/p>
The dash has siblings: emojis in headings. Key terms in bold in every other paragraph. Bullet lists that always follow the same pattern. And right at the end the obligatory section headed “Conclusion”, opening with “In summary, it can be said”. None of this is wrong in itself. Together, these patterns form a fingerprint that any avid reader places instantly.<\/p>
Every language model has its own vocabulary. In English it includes “crucial”, “seamless”, “holistic”, “bespoke” and “revolutionise”, plus openers such as “In today’s digital world” and filler sentences such as “It is important to note that”. The problem is never the individual word, it is the dose: if five things are crucial and three things work seamlessly within a thousand words, the text reads like it came out of a kit.<\/p>
The remedy is uncomfortable but simple: cut, and get specific. Instead of “crucial”, write what goes wrong if you leave it out. Instead of “bespoke”, name exactly what was adapted. Being specific is the one thing AI cannot take off your hands, because it does not know your project.<\/p>
The most reliable marker cannot be found with a search function: it is the beat. AI sentences are strikingly similar in length, every paragraph is built the same way, three examples and a small summary under every subheading. People write differently. Sometimes a long sentence with an aside that tells you about a real project. Sometimes two words. Like this.<\/p>
And people have something no model can replace: lived experience. A figure from your own business, an opinion that contradicts the standard advice, a mistake that cost money. For readers and for Google alike, those are the strongest quality signals. Anyone writing with AI should invest half their time in them.<\/p>
Back to the two projects from the beginning; we reviewed them against our go-live checklist<\/a>. The copy there was fine in substance: technically correct, properly structured, no errors. But the markers were everywhere, from the page copy through the teasers right down to alt texts and theme options. Cleaning it up cost us hours. Straight after writing, it would have been a matter of minutes.<\/p> Since then the rule here has been: no text goes live before it has been checked for these markers. Blog articles, page copy, case studies, even image descriptions. Not to hide anything, but because edited text is simply better: shorter, sharper, more credible. That it no longer looks machine-made is the pleasant side effect.<\/p> For individual articles, Ctrl+F in the browser is enough. For a complete website it is not: there the markers also sit in alt texts, menus and widgets that no full-text search covers. In WordPress, plugins such as Search Regex or Better Search Replace help by searching the entire database for the characters. We now run an AI assistant with the prompt from the next section across all our content, which also covers the corners nobody thinks of.<\/p> Simply swapping every dash for a comma goes wrong, as we learned the painful way. Depending on the sentence, the em dash does a different job, and the replacement has to match that job. Four cases cover practically everything:<\/p> There is a second rule that sounds stricter than it is: only characters that sit on a normal keyboard go into the text. No arrows, no ticks, no Unicode bullets, no emojis. The hyphen stays purely where it belongs inside a word: in compounds such as “go-live checklist”, in number ranges such as “9am-5pm” and in phone numbers. If you prefer a more elegant time span, write “9am to 5pm” in full.<\/p> We gathered this list over months from our own drafts, from ChatGPT, Claude and Gemini alike. The models differ in tone, but they share their favourite vocabulary. Every one of these words is permitted on its own. From the second occurrence per text onwards the rule is: cut it or make it concrete.<\/p> The good news: an AI can take on most of this editing itself, as long as you give it the rules explicitly. The prompt below works identically in Claude, in ChatGPT and in any other assistant. In Claude Cowork you can even run it across entire folders or a complete CMS, and that is exactly how we found and replaced the several hundred dashes from the beginning of this article. Copy it, append your text, done:<\/p> A note from practice: the prompt does not replace the final human read. Point 7 is the reason we still recommend it to everyone: the flagged passages show immediately where the AI has patched its facts together. A human checks those spots, everything else is handled in two minutes.<\/p> Detectors such as GPTZero or Originality essentially measure two things: how predictable the word choice is and how uniformly the sentences are built. AI models statistically pick the most probable next word, which is why their texts feel smooth and interchangeable. Human texts surprise you: an unusual word, an abruptly short sentence, a detail only the author could know. Apply the checklist from this article and you pass these tests as a by-product, not because you have outwitted the detector, but because the text has genuinely become human.<\/p> Automatic humaniser tools, on the other hand, we advise against. They swap words mechanically for synonyms and often destroy precision and readability in the process: “calculate the insulation thickness” turns into “compute the isolation depth”, and any specialist reader can see that nobody thought this through. Ten minutes of manual work beats every one of these tools. And to be blunt: detectors get it wrong regularly in both directions, taking literary classics for AI and polished AI text for human. Do not rely on the verdict, rely on the substance.<\/p> This check takes roughly ten minutes per article. It is the difference between a text that gets read and number 47 of ten thousand identical ones.<\/p> Is the effort worth it? Let us do the maths: AI saves hours of writing, the editing afterwards costs minutes. The machine supplies structure and raw material, the human supplies truth, experience and tone. In that order you get content that does not fall through the filters, because it is simply good. In the reverse order you get the kind of thing currently gathering dust invisibly on the web by the tens of thousands.<\/p> One more thing: the same principle applies across languages. How we translate entire websites with WPML and AI<\/a> and check the quality step by step is described in our workflow article.<\/p>When the AI has already used dashes: the replacement rules<\/h2>
\n
The minus list: stock phrases that every language model loves<\/h2>
Category<\/th> Typical examples<\/th> What helps instead<\/th><\/tr><\/thead> Inflated adjectives<\/td> crucial, essential, seamless, holistic, bespoke, robust, innovative, dynamic, future-proof, comprehensive<\/td> Name what the thing actually does or can do<\/td><\/tr> Marketing verbs<\/td> revolutionise, transform, unleash, unlock, “take it to the next level”<\/td> Name the actual outcome: figure, time, effect<\/td><\/tr> Stock openers<\/td> “In today’s digital world”, “In the digital age”, “Let us dive in”, “Welcome to the world of”, “Let us shed some light on”<\/td> Start with a fact, a figure or a scene<\/td><\/tr> Hedging filler<\/td> “It is important to note that”, “One should bear in mind”, “essentially”, “as a general rule”<\/td> Cut it without replacement, the sentence works without it<\/td><\/tr> Pseudo transitions<\/td> “Furthermore”, “Moreover”, “In addition”, “In conclusion, it can be said”, “In summary”<\/td> Start the new paragraph with the content itself<\/td><\/tr> Stock imagery<\/td> game changer, milestone, centrepiece, synergies, “navigating the digital landscape”, “an exciting journey”<\/td> Say what happened, without the metaphor<\/td><\/tr> Structural clichés<\/td> Rule-of-three chains (“fast, simple and efficient”), rhetorical questions as paragraph openers, “Whether X, Y or Z:”<\/td> Break the pattern: sometimes one example, sometimes four, sometimes none<\/td><\/tr><\/tbody><\/table><\/figure> The prompt: how the AI corrects its own text<\/h2>
Pruefe und ueberarbeite den folgenden Text nach diesen Regeln, ohne Inhalt, Fakten oder Aussagen zu veraendern:\n\n1. Ersetze jeden Gedankenstrich und Halbgeviertstrich nach Kontext: Doppelpunkt vor Aufzaehlungen oder Erklaerungen, Komma bei kurzen Einschueben, Punkt und neuer Satz bei zwei vollstaendigen Hauptsaetzen. Der Bindestrich bleibt nur in zusammengesetzten Woertern, Zahlenbereichen und Telefonnummern.\n\n2. Verwende ausschliesslich Zeichen, die auf einer normalen Tastatur liegen. Entferne Emojis, Pfeile, Haekchen und Sonderzeichen-Aufzaehlungen.\n\n3. Streiche oder ersetze diese Woerter und Floskeln, wenn sie mehr als einmal vorkommen oder nichts Konkretes aussagen: entscheidend, essenziell, nahtlos, ganzheitlich, massgeschneidert, robust, innovativ, dynamisch, zukunftssicher, umfassend, revolutionieren, transformieren, entfesseln, Game-Changer, Meilenstein, Herzstueck, Synergien, \"in der heutigen digitalen Welt\", \"es ist wichtig zu beachten\", \"darueber hinaus\", \"des Weiteren\", \"zusammenfassend laesst sich sagen\". Formuliere stattdessen konkret, was gemeint ist.\n\n4. Variiere die Satzlaenge deutlich. Wechsle lange und kurze Saetze ab und loese gleichfoermig gebaute Absaetze auf. Vermeide Dreierketten wie \"schnell, einfach und effizient\".\n\n5. Entferne Fettdruck im Fliesstext, Emojis in Ueberschriften und Abschnitte mit der Ueberschrift \"Fazit\". Arbeite deren Inhalt in den letzten Absatz ein.\n\n6. Schreibe aktiv statt passiv und ersetze Substantivketten durch Verben.\n\n7. Markiere jede Zahl und jede Behauptung, die du nicht sicher belegen kannst, mit [PRUEFEN]. Erfinde nichts dazu.\n\nGib zuerst den ueberarbeiteten Text aus, danach eine kurze Liste aller Aenderungen.<\/code><\/pre>AI detectors and humanisers: what really counts<\/h2>
Eight style rules that make any text more human<\/h2>
\n
The editing checklist at a glance<\/h2>
Check point<\/th> How to check<\/th> Why it stands out<\/th><\/tr><\/thead> Long dashes<\/td> Use Ctrl+F to search for the em dash (—) and the en dash (–), then replace every hit: comma, colon or new sentence<\/td> The best-known AI marker of all<\/td><\/tr> Stock phrases<\/td> Cut or make concrete: “crucial”, “seamless”, “holistic”, “in today’s digital world”<\/td> The sheer frequency looks like it came out of a kit<\/td><\/tr> Facts and figures<\/td> Verify every claim, because AI invents details very convincingly<\/td> One wrong figure ruins the credibility of the entire article<\/td><\/tr> Substance of your own<\/td> Add at least one experience, figure or opinion of your own<\/td> Generic text gives Google no reason to rank it<\/td><\/tr> Rhythm<\/td> Read the text aloud and rewrite any uniform passages<\/td> The monotonous beat gives the machine away even without markers<\/td><\/tr> The margins<\/td> Check headings, meta data and alt texts as well<\/td> The markers often sit outside the body copy<\/td><\/tr><\/tbody><\/table><\/figure> A tool, yes; an autopilot, no<\/h2>