AI Content Detection: How We Use It in Client SEO Campaigns

Tools such as GPTZero, Originality.ai, Copyleaks and Turnitin’s AI writing indicator analyse statistical patterns in text. Two ideas come up often: perplexity, roughly how predictable each word is to a language model, and burstiness, how much sentence length and structure vary. Machine-written text tends to be more uniform and predictable.

Some tools also train classifiers on large sets of human and AI text. The output is usually a percentage or a label per sentence. What none of them can do is read the writer’s mind or see the drafting process.

Less reliable than their interfaces suggest. OpenAI withdrew its own AI text classifier in 2023, citing low accuracy. Academic studies have found that detectors flag non-native English writing as AI-generated more often than native writing, and that light paraphrasing can lower scores sharply.

Scores also shift between tools. The same article can read as mostly human in one detector and mostly AI in another. Short texts, lists, technical definitions and formulaic content such as product specifications are especially prone to false positives.

• Do not treat a score as evidence of misconduct.
• Do not compare scores from different tools as if they were one scale.
• Expect more noise on texts under about 300 words.
• Re-test after edits; a single run is a snapshot, not a verdict.

There is also a fairness issue. A writer accused on the basis of a score has no way to prove a negative. Draft history in Google Docs or a CMS revision log is far better evidence than any detector, and teams should rely on it when questions arise.

Google’s published position is that it rewards helpful, reliable, people-first content however it is produced. Its spam policies target scaled content abuse, meaning many pages generated mainly to manipulate rankings with little value for users, whether made by AI, people or both.

That framing matters for agencies. The question Google asks is not “was a model involved?” but “does this page help a searcher, and is it one of thousands of near-identical pages?” Detection tools can help spot the second problem. They cannot answer the first.

We use detectors as a quality prompt, not a gate. A high AI likelihood score on a draft usually points to the same weaknesses an editor would flag anyway: vague openings, repeated sentence shapes, filler transitions and claims without specifics.

A practical review sequence

• Brief: define the search intent, the questions to answer and the first-hand material the writer must include.
• Draft: human, AI-assisted or both, with sources noted for every factual claim.
• Editorial review: check accuracy, intent match, structure and brand voice.
• Signal check: run a detector and a plagiarism checker; mark flagged passages for a second look.
• Revision: add examples, data with sources, screenshots or expert input where passages were generic.
• Final approval by a named editor, recorded in the content tracker.

Notice that the detector comes after the human review. If it ran first, editors would anchor on the number and stop reading critically.

Several checks predict whether content will earn rankings and citations better than any AI likelihood figure.

• Original information: data, photos, process details or opinions that only this business can provide.
• Clear authorship: a real author with relevant experience and an author page.
• Factual accuracy: every number and claim traceable to a source.
• Intent match: the page answers the question behind the keyword in the first paragraph.
• Site-level similarity: how close the page is to other pages on the same site.

The last point is easy to overlook. Internal duplication across hundreds of similar pages is a stronger risk signal than whether one article was drafted with ChatGPT. Similarity tools such as Siteliner, or a simple MinHash comparison in a script, reveal clusters of near-duplicates.

A short written policy prevents most disputes. It should state that AI assistance is allowed or not, which tools are approved, that every page needs human editing and fact-checking, and that detection scores are advisory. It should also say how disagreements are handled: the editor’s reading of the text wins over a score.