Two questions about one article
Run the same draft through two tools and you get two answers with almost nothing in common.
One returns a probability, sometimes broken out sentence by sentence. The other returns a list: the sources you named, the claims that carry no source, the cited link that now answers 404, the attributions that point at nobody.
Both measured the same words. A detector asks what produced the sentences. A citation check asks whether a reader, an editor or an AI search engine has a reason to trust the piece and quote it.
Almost every AI checker comparison lines products up against each other on the first question. For a university that is the right question. For anyone deciding whether an article is ready to publish, it is the wrong axis, and the evidence for that is public.
A detector reads the statistics of the text: token predictability, sentence-length evenness, repeated openers, fused by a classifier into one number. We walked through those signals in our guide to how detectors work, so the mechanics are not the argument here.
A citation check reads the evidence instead. Which sources are named. How each claim is attributed. Whether the cited links still resolve. How many statistics there are per 1,000 words, and whether anything is shaped to be quoted. The output is not a percentage. It is a list of what is missing.
One tool asks who wrote the sentences. The other asks whether anyone has a reason to quote them.
Who is actually asking “was it AI?”
The honest answer is short. Schools, and editors working under a disclosure policy. A university needs to know whether a student did the work, and a publisher that promised its readers a human byline needs to keep that promise.
Google is not on the list. Its February 2023 guidance is headed “Rewarding high-quality content, however it is produced”, and Google writes that “our focus on the quality of content, rather than how content is produced, is a useful guide that has helped us deliver reliable, high quality results to users for years.” The same page says “not all use of automation, including AI generation, is spam”, while keeping automation aimed at manipulating rankings inside its spam policies. The test is the content, not the author.
The measurements point the same way. Ahrefs correlated AI-content share with ranking position across 600,000 pages in July 2025 and got 0.011, “effectively zero”, next to a second finding: “86.5% of top-ranking pages contain some amount of AI-generated content.” Its larger study of June 2026 found that higher AI use is correlated with lower ranking positions, and its author reads that as quality, not authorship: “increasing use of AI correlates with decreasing content quality.” A field where most of the winners contain some AI cannot be sorted by a detector score. What separates them is the quality a detector does not measure.
Your reader is not on the list either. Nobody arriving from a search result runs an article through a classifier. They look for a name they recognize and a figure with a source under it, and so does the engine deciding whether to quote the page.
What the detector class does well, and where it stops
The two most cited products in the category describe themselves plainly on their own pages, and the difference is the audience.
GPTZero reports using “an end-to-end deep learning approach, trained on text datasets from the web, education, and AI-generated from a range of LLMs”, with “a sentence-by-sentence classification model” that “determines the probability and confidence that a text was created by AI”. Its page is written for education: it describes de-biasing work “for educational use cases” and speaks to “an educator or an institution that relies on AI detection tools to help inform your disciplinary policy”. Vendor-reported.
Originality.ai states a different starting point on its about page: “we believe website owners should decide when to publish AI-generated content, and that content creators need an easy way to prove originality beyond AI detection scores.” That is a publisher and agency framing rather than a classroom one. Also vendor-reported.
Inside that job the category works. An AI detector comparison that ranks products on how often the verdict is right is worth reading when the verdict carries a consequence, and a classifier over a whole document is a fast way to see whether a draft came straight out of a model.
Then the design boundary. A detector scores the form of the writing. It holds no view on whether a claim is true, whether the study behind a number exists, or whether the link under a citation still resolves. None of that is a gap in any product. It sits outside what a text classifier measures.
Whether a claim is true sits outside anything a text classifier measures.
Published evaluations bound how hard the score can be leaned on. Weixin Liang and Stanford colleagues ran seven widely used detectors over 91 human-written TOEFL essays in 2023. The detectors “misclassified over half of the TOEFL essays as ‘AI-generated’ (average false positive rate: 61.22%)”, while scoring near-perfect on US eighth-grade essays.
OpenAI withdrew its own classifier that same year. The launch page now opens: “As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy.” Read the retirement notice before treating any score as settled. Detectors have improved since 2023. Those two findings still set the reading: treat a score as a range, and give anyone it accuses a way to appeal.
What a citation check finds that a detector cannot
We measured the field a detector never looks at. Across 531 articles ranking for 40 B2B content queries (US, September 2026), 27% name zero sources. Not a weak source or an unlinked one. None.
Among the pages whose external links we could check, 13% carry at least one link returning 404 or 410. That is the strict rule: only those two codes count as dead. Timeouts and hosts that refuse automated requests are reported separately, because a blocked host is not a broken citation.
The median page in that set carries 0.9 statistics per 1,000 words, so half of what ranks in the top 20 for these queries makes fewer than one checkable claim per thousand words. The limitation belongs with the figure: these pages rank for those 40 queries, they are not the web at large.
Authorship and evidence are independent properties. A careful human writer with no sources passes the detector and fails the citation check. A well-sourced draft written with a model does the reverse.
Authorship and evidence are separate questions, and a sourceless human draft fails the one that matters for publishing.
| Question about the article | What an AI detector reports | What a citation check reports | Who wants the answer |
|---|---|---|---|
| What produced the sentences? | A probability, often per sentence | Not measured | Schools, editors with a disclosure policy |
| Can an AI system reach the page? | Not measured | Robots rules, noindex, and whether the text is there without JavaScript | AI search engines, editors |
| Does each claim name a source? | Not measured | Named sources, and claims attributed to nobody | Editors, readers |
| Do the cited pages still exist? | Not measured | Dead links, kept separate from hosts that block automated requests | Editors, readers |
| Who is being quoted? | Not measured | Vague attributions such as “according to analysts” | Readers, AI search engines |
| Is there anything worth quoting? | Not measured | Figures per 1,000 words, questions with their own heading, tables, and up to three quotable lines | AI search engines, journalists |
| How is the page built? | Not measured | Heading and paragraph structure findings | Editors, AI search engines |
| What should be fixed first? | Not measured | A list of what to fix first, most costly gap on top | Editors |
| Does the prose read mechanically? | Yes, that is the score | Not measured | Editors |
Which check for which job
Academic integrity or a disclosure policy
Run a detector. Read the output as a range rather than a verdict, and write the appeal step into the policy before the first accusation. The false-positive research above says who absorbs the errors otherwise.
Pre-publish editorial QA and AI-search visibility
Run a citation check. It hands back work you can assign: add a named source to the second section, replace the dead link in paragraph four, give “according to research” an addressee, put the figure back into the sentence it was cut from. A percentage hands back nothing.
Both, in that order
If the policy requires disclosure and the business needs traffic, run the detector for the policy and the citation check for the piece. They do not compete for the same slot.
Frequently asked
What should an AI checker comparison actually compare?
Name the question you need answered first. If it is “did a person write this”, compare detectors on false positives and on whether the vendor publishes its benchmark: Liang and colleagues at Stanford found an average false positive rate of 61.22% across seven detectors on 91 TOEFL essays. If it is “is this ready to publish”, compare what tools report about sources, links and quotable material, because no classifier reports any of it.
Does an AI detector score tell me whether my content will rank?
No. Ahrefs measured a 0.011 correlation between AI-content share and ranking position across 600,000 pages, and found 86.5% of top-ranking pages contain some amount of AI-generated content. Its 2026 study of a larger sample found higher AI use correlated with lower positions, which its author attributes to lower content quality rather than to AI itself. Google’s own guidance says its focus is “the quality of content, rather than how content is produced”.
Are AI detectors accurate enough to accuse someone?
Accuracy depends on the benchmark, and the consequence decides how much of it a decision needs. OpenAI retired its own classifier on July 20, 2023 “due to its low rate of accuracy”, and the Stanford study found detectors flagged over half of a set of TOEFL essays written by people. Any comparison that ends in a disciplinary decision needs an appeal path attached.
Can a citation check tell me whether AI wrote my article?
It cannot, and it does not try. It reports what is in the text: named sources, citation forms, dead and blocked links, vague attributions, statistics density, tables and FAQ blocks. A draft by a person with no sources gets the same list as a draft written with a model.
Which one should I run before publishing?
The citation check, because it returns a to-do list. Across the 531 ranking articles we measured, 27% name zero sources and the median page carries 0.9 statistics per 1,000 words, so one paragraph with a named source, a date and a figure moves a draft past most of the field. Add a detector when your policy requires disclosure.
Sources
8 sources
Methodology: every figure and quotation in this article was checked against the page it is taken from on October 13, 2026. Items 1 to 5 are primary sources: one Google guidance page, two published studies, one arXiv paper and one vendor retirement notice. Items 6 and 7 are detector vendors’ own pages, used only for what each company says about itself. Item 8 is our own measurement, and its rules are stated with it.
- 01Google Search Central. “Google Search’s guidance about AI-generated content.” February 8, 2023. https://developers.google.com/search/blog/2023/02/google-search-and-ai-content.Supports: “Rewarding high-quality content, however it is produced”, the focus on quality rather than production method, and “not all use of automation, including AI generation, is spam”.
- 02Ong, S. Q. “AI-Generated Content Does Not Hurt Your Google Rankings (600,000 Pages Analyzed).” Ahrefs, July 7, 2025. https://ahrefs.com/blog/ai-generated-content-does-not-hurt-your-google-rankings/.Supports: the 0.011 correlation described as “effectively zero”, and “86.5% of top-ranking pages contain some amount of AI-generated content”.
- 03Law, R. “Google Doesn’t Punish AI Content; It Punishes Bad Content.” Ahrefs, July 27, 2026. https://ahrefs.com/blog/google-doesnt-punish-ai-content/.Supports: 1,000,000 pages pulled from the top 10 positions in 100,000 SERPs in June 2026, higher AI use correlated with lower ranking positions, and the author’s reading that “increasing use of AI correlates with decreasing content quality”.
Show all 8 sourcesShow fewer
- 04Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., Zou, J. “GPT detectors are biased against non-native English writers.” arXiv:2304.02819. https://arxiv.org/abs/2304.02819.Supports: seven widely used detectors over 91 human-authored TOEFL essays, an average false positive rate of 61.22%, and near-perfect accuracy on US eighth-grade essays.
- 05OpenAI. “New AI classifier for indicating AI-written text.” https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/.Supports: the classifier’s withdrawal on July 20, 2023 “due to its low rate of accuracy”.
- 06GPTZero. “Technology.” https://gptzero.me/technology.Supports: as reported by the vendor, the end-to-end deep learning approach, the sentence-by-sentence classification model, and the education-facing audience.
- 07Originality.ai. “About.” https://originality.ai/about.Supports: as reported by the vendor, the stated position that website owners should decide when to publish AI-generated content, and the publisher-facing framing.
- 08Avoid Content corpus measurement, September 2026. The instrument we ran over the corpus covered 531 articles ranking in Google organic top-20 for 40 B2B content-marketing queries (US, English), SERP pulled September 7, 2026. Dead links counted as HTTP 404 and 410 only; blocked hosts reported separately.Supports: 27% naming zero sources, 13% carrying at least one dead link, and the median of 0.9 statistics per 1,000 words. Dataset not yet public, thresholds stated inline.