I hear two opposite worries from business owners about AI-written content. The first goes like this: “If I use AI to help write my pages, Google’s AI answers will spot it and ignore me.” The second is the reverse: “AI tools like AI-written text, so I should publish as much of it as I can, as fast as I can.”
Both can’t be right, and as it turns out, neither is quite right. The best data we have suggests AI Overviews cite pages that use AI help very often. It also shows they rarely cite pages written purely by AI, and that the amount of AI in a page has almost nothing to do with where it appears in the answer.
In this article, let me explain how researchers try to detect AI-written pages, what the numbers actually show, why AI answers might lean toward AI-assisted text, and the risks if everyone takes the shortcut. Then I will share the standards I suggest for AI-assisted content that still deserves to be cited.
How researchers decide whether a page is AI-written
You can’t ask a web page who wrote it. So researchers use AI content detectors: software trained to spot patterns that are more common in machine-written text than human text. The detector gives a probability, not a yes-or-no answer.
Ahrefs, whose study I rely on most here, explained how these tools work in a January 31, 2025 article. In brief, there are three main approaches:
| Approach | How it works, in plain words | Main weakness |
|---|---|---|
| Statistical methods | Count word choices, word pairs and sentence patterns, then compare with known human and AI samples | Easy to fool with small edits |
| Neural network classifiers | A model is trained on many labeled examples of human and AI text and learns the difference | Struggles with text styles it wasn’t trained on, and with mixed human-AI text |
| Watermarking | The AI tool hides a signal in its output that can be checked later | Only works if the AI company chooses to add it |
The same article is honest about the limits. It notes that detectors trained on one style, such as news, do poorly on other styles, and that mixed text is especially hard to classify.
“No AI detector is 100% accurate. There will always be false positives.”
— Ryan Law, Ahrefs, How Do AI Content Detectors Work?
Keep this in mind for everything below. Every number about “AI-written” pages is an estimate from a detector. If you want to understand the tools themselves, I cover them in my guide to AI content detectors.
What share of pages cited in AI Overviews are AI-written
The largest study I know of is from Ahrefs, published July 14, 2025. Si Quan Ong and Xibeijia Guan took 1 million search results that showed an AI Overview, pulled the top three cited links from each, and ran about 500,000 of those pages through Ahrefs’ own AI detector.
“AI Overviews are more likely to cite AI-generated content than human-written content.”
— Si Quan Ong, Ahrefs, AI Overviews Cite AI-Generated Content More Than Human Writing
The headline sounds dramatic, so let me put the numbers side by side with two other Ahrefs datasets that used the same kind of detector.
| Group of pages | Pure AI | Mixed AI and human | Pure human |
|---|---|---|---|
| Pages cited in AI Overviews (500K pages, July 2025) | 3.6% | 87.8% | 8.6% |
| New web pages found in April 2025 (900K pages) | 2.5% | 71.7% | 25.8% |
| Pages in Google’s top 20 (about 600K pages, July 2025) | 4.6% | 81.9%* | 13.5% |
Sources: Ahrefs AI Overview citations study, Ahrefs new content study (May 19, 2025), Ahrefs rankings study (July 7, 2025). *The rankings study reports 86.5% with some AI content and 4.6% pure AI; I worked out the mixed share by subtracting, so treat it as approximate.
Within the 87.8% of cited pages that were mixed, most used AI only moderately: 44% of all cited pages had between 11% and 40% AI content, and only 7.9% were 71–99% AI.
The number that matters most
Ahrefs also checked whether pages with more AI content were cited higher in the AI Overview. The correlation was 0.017, which is effectively zero. The Google rankings study found the same pattern for normal results, with a correlation of 0.011. In plain words: Google doesn’t seem to reward or punish content for how it was produced.
What the AI citation study can and can’t prove
Before you change your content plan, here are the limits I keep in mind:
- Only about a quarter of the cited URLs were analyzed. 500,000 out of 1.9 million were in Ahrefs’ database.
- The comparison groups don’t match perfectly. The “web” baseline covers new pages only, but many cited pages are older.
- The detector belongs to the company publishing the study. That doesn’t make it wrong, but it hasn’t been independently tested on this data.
- “Pure” thresholds are not defined in the article, so we don’t know exactly where “mixed” ends and “pure” begins.
- It only covers Google’s AI Overviews. I haven’t found a similar large study for ChatGPT, Perplexity or Gemini citations.
So I treat this as direction, not hard science. The safest reading is this: using AI help doesn’t stop a page from being cited, and pages written 100% by AI are rare among citations. That matches what Google says about AI content in general, which I cover in AI-generated content and SEO.
Why AI answers may lean toward AI-assisted text
Why would cited pages show more AI help than the general web? Nobody has proven the reason. Here are the explanations I find most likely, from strongest to weakest:
- Most of the web now uses AI help. Ahrefs found 74.2% of new pages in April 2025 contained AI content. Busy, well-resourced publishers, the kind that rank, are likely heavy users of AI tools.
- Cited pages are mostly ranking pages. Google’s AI Overviews draw heavily on top-ranking results, and top-ranking pages already show high AI use.
- AI-assisted pages are often well structured. Clear headings, short answers and lists make it easy for an AI system to lift a passage. That is a structure effect, not an “AI likes AI” effect.
- Similar style may be easier to match. Some writers suggest language models find text in their own style easier to use. This is a theory, and I haven’t seen it tested in a way I trust.
In brief, the simplest explanation is that the pages Google already trusts are increasingly written with AI help. AI Overviews then cite them for the same reasons they rank. For more on how AI systems pick sources, see my post on how AI chooses sources.
The risk of an internet that all sounds the same
There is a bigger worry here than rankings. If AI tools write more of the web, and AI tools learn from and cite the web, the loop can start to feed on itself.
Researchers led by Ilia Shumailov tested this in a paper published in Nature in 2024. They trained language models on text produced by earlier models, again and again. The Oxford research group’s summary of the paper describes the result:
“indiscriminate use of model-generated content in training causes irreversible defects in the resulting models”
— Shumailov et al., Nature (2024), via OATML, University of Oxford
They call this model collapse. The rare, unusual parts of the original data disappear first. As TechXplore summarized it, within a few generations the original content was replaced by unrelated nonsense, and the models also repeated phrases. The authors don’t say AI-generated data can never be used, only that it needs careful filtering.
This is a training problem, not directly a citation problem. But it points to the same business risk. When every page on a topic is written from the same AI summary of the same top results, the pages become interchangeable. An AI answer only needs one of them. I explore this further in how AI flattens originality.
What “sameness” looks like on a real site
- Every blog post opens with the same kind of definition.
- Examples are generic (“imagine a small business…”) with no names, numbers or places.
- No claim links to a source, because the AI didn’t have one.
- Ten articles on the site could be swapped with ten from a competitor and nobody would notice.
A page like that is easy to replace. That is a bad position for clicks and for citations.
What still decides whether a page gets cited
If AI use isn’t the deciding factor, what is? Based on Google’s documentation and the citation data, these matter far more:
- Ranking. Ahrefs found 76.1% of AI Overview citations come from Google’s top 10 (July 21, 2025). Normal SEO still matters.
- Accuracy. AI tools can make up facts. A page with a wrong number is a liability for any system citing it, and for your reputation.
- Something new. Your own data, prices, photos, examples or expert opinion. SEOs call this information gain.
- Clear, quotable answers under specific headings, the same skill that used to win featured snippets. I explain how that changed in what happened to featured snippets.
- Basic eligibility. Google says there are no special requirements for AI Overviews beyond being indexed and meeting its normal rules (AI features and your website).
Google’s position on AI content has been steady since its February 2023 guidance. As Marie Haynes quotes it in her April 2025 article, Google said: “Using AI doesn't give content any special gains. It's just content.” The full original is on the Google Search Central blog.
Should you tell readers when you used AI?
Most businesses don’t. Ahrefs’ AI SEO statistics page reports from its own marketer survey that 84% of companies don’t disclose AI use in their content, while 97% have a review process and don’t publish pure AI content.
Google’s 2023 guidance suggests thinking in terms of “Who, How, and Why”: who created the content, how it was made, and why. You don’t need a warning label on every post. But I suggest these simple rules:
- Always show a real author with a real bio. My guide to author pages and bylines explains how.
- Disclose when AI did most of the work, for example an auto-generated product description or a translated page.
- Never present AI output as personal experience. If a post says “I tested this,” a person must have tested it.
- Keep a short note on your editorial policy page explaining how you use AI and how content is checked.
Our standards for AI-assisted content that deserves to be cited
Here is the process I suggest for any business using AI to help write content. It keeps the speed benefit and removes most of the risk.
- Start with a human brief. Decide the angle, the reader, and what you know that competitors don’t. AI shouldn’t choose the angle.
- Collect your own material first: prices, data, customer questions, photos, examples.
- Use AI for the draft structure and first pass, feeding it your material, not asking it to “write about X.”
- Fact-check every number and claim against a page you have opened. Delete anything you can’t verify.
- Add the human layer: real examples, opinions, and the “it depends” detail that only experience gives you.
- Edit for voice using your editorial style guide, so it sounds like you and not like every other site.
- Have a named person approve it and put their name on it.
And here is how I would split tasks between AI and people:
| Task | AI is fine | Human required |
|---|---|---|
| Brainstorming topics and outlines | Yes | Final choice of angle |
| First draft from your notes | Yes | Heavy edit |
| Statistics and facts | No, AI can invent them | Every number checked at the source |
| Examples and case details | Only to format them | Must come from real work or be labeled hypothetical |
| Opinions and recommendations | No | Yes, from someone accountable |
| Meta titles, summaries, FAQs | Yes | Quick review |
A worked example (hypothetical numbers)
Say a dental clinic plans 12 blog posts a quarter. Fully human, each post takes about 6 hours, so 72 hours. With AI drafting from the dentist’s own notes, drafting drops to about 2 hours per post, but fact-checking and editing add 2 hours, so about 4 hours per post, or 48 hours in total. The clinic saves around 24 hours a quarter and every post still has a named dentist, real patient questions and checked facts. In this example, AI saves a third of the time without turning the blog into something anyone could have written.
If you want writers who already work this way, our content writing service follows this process.
What to watch next on AI citations of AI-written content
This area is moving fast. Here is what I am watching, and what I suggest you watch too:
- Studies beyond Google. We need similar data for ChatGPT, Perplexity and Gemini, which pick sources differently. I explain how differently ChatGPT picks in which search index ChatGPT uses.
- Independent detection tests. Results from one vendor’s detector should be checked with others.
- Watermarking. If major AI companies watermark their output, detection could become far more reliable.
- Google’s spam enforcement. Google’s spam policies target mass-produced, low-value pages, however they are made. Watch for updates that change how these rules are enforced.
For the latest numbers as they come in, see my AI SEO statistics roundup.
Questions people ask about AI answers citing AI content
Do AI Overviews cite AI-generated content?
Yes, often. Ahrefs found 87.8% of cited pages were mixed AI and human content in July 2025. But only 3.6% were pure AI, and 8.6% were pure human.
Will using AI to write hurt my chances of being cited?
The data says no, as long as the page is good. Ahrefs found almost no correlation (0.017) between how much AI a page used and where it was cited in AI Overviews.
Can Google detect AI-written content?
Google hasn’t said it penalizes content for being AI-written. It focuses on quality and on spam, such as mass-produced pages made to manipulate rankings, however they are made.
Are AI detectors accurate?
Not fully. Even Ahrefs, which builds one, says no detector is 100% accurate and false positives will happen. Never base a big decision on one detector result.
What is model collapse?
It is what happens when AI models are trained repeatedly on AI-generated data. A 2024 Nature paper found the rare parts of the original data disappear and quality falls over generations.
Setting your own AI content standard this month
In brief, AI answers don’t seem to care whether a person or a tool typed the words. They care whether the page ranks, answers clearly and adds something. So use AI to save time, but spend that saved time on the parts only you can provide: real data, real examples, checked facts and a real name on the page.
If you would like help building an AI-assisted content process that still produces pages worth citing, get in touch for a free quote. We are happy to review a few of your current pages and tell you honestly where they stand.
Further reading on AI-written content
- Do LLMs Flatten Original Ideas? Keeping Your Brand's Point of View in AI Answers
- Ranking Yourself #1 in Your Own 'Best' Lists: Does It Work in AI Search?
- Do AI Assistants Favour Big Brands Less Than Google?
Sources for this AI-written content and citations guide
- Ahrefs – AI Overviews Cite AI-Generated Content More Than Human Writing (Jul 14, 2025)
- Ahrefs – What percentage of new content is AI-generated? (May 19, 2025)
- Ahrefs – AI-generated content does not hurt your Google rankings (Jul 7, 2025)
- Ahrefs – How Do AI Content Detectors Work? (Jan 31, 2025)
- Ahrefs – Search rankings and AI Overview citations (Jul 21, 2025)
- OATML, University of Oxford – AI models collapse when trained on recursively generated data (Nature, 2024)
- TechXplore – coverage of the model collapse paper (Jul 2024)
- Google Search Central – Google Search’s guidance about AI-generated content (Feb 2023)
- Marie Haynes – Does Google Search approve of using AI to help write content? (Apr 17, 2025)
- Google Search Central – AI features and your website (updated Dec 10, 2025)
