How to detect AI text: markers, code-level evidence, and detection tools

We built our own AI detector — here is what we learned

When AI texts flooded the SERPs, we had to check client content by hand. Paid services (Copyleaks, Originality.ai) helped but never explained the "why". Once we understood their logic, we realized they analyze text with… AI itself. So we built our own free AI detector that doesn't just give a verdict — it shows where and how the AI left its traces.

This article distills what we learned from thousands of checked texts. From our observations across client projects:

Google's official position is in its AI content guidance: quality matters, not the production method. But raw machine text is recognizable — by algorithms and by readers. Here is exactly how.

GPT language markers: the phrases that give the machine away

GPT builds text from statistically frequent constructions. It sounds human, but over a paragraph the recognizable clichés emerge. The most common openers:

The second group is bureaucratic glue AI starts sentences with: given that, within the framework of, in the context of, based on, with regard to, in accordance with. The third — empty generalizations GPT uses to fill a paragraph when it has no facts:

If such constructions appear back-to-back and there are no concrete facts, numbers, or examples — you are almost certainly looking at raw AI. How to add facts properly — see our copywriter brief guide.

GPT vs human: the comparison table

Machine logic: the rule of three and perfect symmetry

Even if you ban GPT's pet phrases, you can't fool the logic. Humans mess up: odd constructions, missing commas (my editor will confirm). GPT doesn't. Hence three stable patterns:

• The rule of three. "Useful, structured, and fact-based" — GPT adores splitting ideas into threes: three adjectives, three bullets, three blocks under every heading (intro → explanation → conclusion).
• Structural symmetry. Same-length paragraphs; each opens with a lead-in and closes with a bridge to the next. We asked GPT itself why — it answered: "I build text like a well-structured article, by the textbook."
• Excessive politeness. Instead of "this doesn't work" — "some users may find this approach insufficiently effective under certain conditions". Bluntness, humor, and doubt are human; neutral diplomacy in every sentence is machine.

Special-character evidence: invisible to the eye, visible in code

The most reliable part of our system. Humans physically don't type these characters — GPT inserts them constantly. Open the text in HTML mode and look for:

Evidence in the markup

• Perfectly closed tags. Every <p>, <li>, <div> closed to standard — without a single human slip.
• Mechanical lists. <ul><li><p>Text</p></li></ul> instead of a simple <li>Text</li>.
• <hr /> with the closing slash and horizontal divider lines between sections — GPT's signature, a "coming-out in front of Google".
• data-start / data-end attributes in headings and lists — technical markup no human ever adds.

Detection tools: where to start

Manual marker analysis is the most accurate but slow. At scale, the combo "detector + spot manual checks" works best:

By the way, building your own tool today is easier than it seems — see our AI tool development service. And for using AI in SEO the smart way — our ChatGPT for SEO guide and the 50 mega-prompts collection.

The manual check checklist: 7 steps

• Search for marker phrases. Ctrl+F: "in today's world", "it is worth noting", "thus". 3+ hits — yellow flag.
• Fact check. Are there concrete numbers, names, examples? Generalizations without facts are AI filler's main sign.
• Paragraph rhythm. Step back from the screen: if every paragraph looks identical — that's machine symmetry.
• The rule of three. Count the triples: three adjectives, three bullets, three blocks per section.
• Code audit. Open the HTML: &mdash; more than three times, curly quotes, data attributes, <hr />.
• Run it through a detector. Unmiss / Copyleaks / GPTZero — for confirmation, not instead of your head.
• The usefulness test. Google's key question: will the reader learn something the top-3 results don't offer? If not — it doesn't matter who wrote it.

The same approach works in reverse — to "humanize" an AI draft: remove the markers, add facts and personal experience, break the symmetry. How to write commercial copy that sells — in our commercial content article.

Why it matters: AI text and visibility in Google and AI search

The paradox of 2026: AI search engines (AI Overviews, ChatGPT, Perplexity) themselves dislike citing raw AI content. They rely on sources with expertise, facts, and authority — we covered this in detail in our pieces on GEO optimization and backlink sources.

1. Raw AI text → templates, zero facts → never cited, risks falling under "scaled content abuse" in Google's spam policies.
2. AI draft + editor + facts + experience → full-fledged content that ranks and gets cited. Google doesn't care about the production method.

So checking text for AI is really a check of your content process. The detector catches not "AI" but the absence of human work on the text.

In short: a three-level system

1. Language: marker phrases, bureaucratic glue, fact-free generalizations.
2. Syntax: the rule of three, symmetrical paragraphs, excessive politeness, perfect grammar.
3. Code: special characters (&mdash;, “ ”, →, &nbsp;) and machine markup (data attributes, <hr />).
4. Tools speed things up but don't replace you: always verify a detector's verdict against the evidence above.
5. The goal isn't to "catch AI" — it's to never publish useless content: that's what Google penalizes and AI search ignores.