AI Overviews Cite AI-Assisted Content More Than the Open Web Does

TL;DR

• Ahrefs analyzed 1 million SERPs with AI Overviews and found the pages Google cites contain more AI-generated content than the open web on average.
• Pages with zero AI involvement made up 8.6% of AI Overview citations, versus 25.8% of a general sample of new webpages.
• The correlation between how much AI content a page has and where it lands in the citation order was effectively zero, meaning Google isn’t rewarding AI content once a page is already in the running.
• The study’s own classification method is an AI detector, the same category of tool we’ve already written about being unreliable, which caps how much weight these numbers can carry.

Google has said its AI Overviews and AI Mode pull from the same index and ranking systems as regular search, not a separate pipeline with separate rules. A new dataset from Ahrefs puts a number on what that actually looks like in practice, and the number is stranger than “AI content ranks fine.”

What Ahrefs Actually Measured

Ahrefs pulled the top three cited links from 1 million SERPs that showed an AI Overview, matched 500,000 of those URLs against its own database, and ran each one through its AI content detector. Ahrefs sorted pages into pure AI, pure human, or mixed, then broke mixed content down further by how much of the page the tool attributed to AI.

They compared that distribution against a separate sample of 900,000 new webpages, meant to represent the open web generally rather than specifically what gets cited.

The Numbers

Among pages cited in AI Overviews: 3.6% were pure AI, 8.6% were pure human, and 87.8% were some mix of the two. Among the general web sample: 2.5% were pure AI, 25.8% were pure human, and 71.7% were mixed.

The pure-AI share barely moved. The real gap is in pure-human content, which is three times rarer among AI Overview citations than it is on the open web. Whatever is getting picked up by AI Overviews has usually touched AI somewhere in its production, even when it isn’t AI from top to bottom.

One number cuts against the obvious read of this, though. Ahrefs checked whether a page’s AI percentage predicted where it landed in the citation order, first, second, or third, and found a correlation of 0.017, close enough to zero to mean nothing. Google isn’t ranking AI-touched pages higher once they’re already competitive. AI-touched pages are more likely to reach that competitive tier in the first place, and the researchers tied that back to how much of the web is now written that way: their sample put AI involvement in roughly three-quarters of new pages overall.

The Caveat That Matters Most

Ahrefs’ own writeup acknowledges the obvious weak point: the classification rests on an AI content detector, and detectors of this kind have a documented accuracy problem. We’ve covered why: the peer-reviewed research on these tools shows they misfire in specific, uneven ways, particularly against writing that’s plain, consistent, or produced by a non-native English speaker, none of which has anything to do with whether AI touched the page.

That doesn’t make this dataset worthless. Ahrefs is transparent about the method, and the finding, more AI-touched content among citations than in the general population, doesn’t depend on the detector being perfectly precise on every individual page, only on it being roughly right in aggregate across half a million URLs. Treat a specific figure like “8.6% pure human” as a rough signal worth a mental asterisk, not a number precise enough to quote on its own.

The Ouroboros Problem

Ahrefs frames the pattern with a specific worry: AI Overviews increasingly cite content that was itself produced with AI assistance, some of which was likely informed by other AI-generated summaries, in a loop that reinforces itself. Their phrase for it: “an AI content ecosystem where machines talk to machines.”

That framing is speculative, and the researchers say as much. But it’s a reasonable read of the mechanism. If a large share of new pages are AI-assisted, and AI Overviews cite whatever ranks well regardless of how it was made, then AI-assisted content ends up overrepresented in citations for a simple reason: it’s overrepresented on the web that citations get pulled from.

If You’re Writing to Get Cited

Nothing here changes the fundamentals. Once a page is competitive enough to be one of the top three cited, AI involvement doesn’t move it up or down that short list; the 0.017 correlation only speaks to that ordering, not to whether a page gets into the running at all. The structure Google has already described still predicts a citation: a direct answer, a clearly defined term, a fact with a source attached.

The more useful takeaway is about the company your content keeps. If AI Overviews increasingly draw from a pool that’s mostly AI-touched, standing out inside that pool means the same things it always has: original information a summary tool can’t already answer, specifics that trace to a real source, and a level of usefulness that survives being read by something other than a person. The bar got higher, since the pool it’s judged against got more crowded.

People Also Ask

Does Google’s AI Overview favor AI-generated content? The data doesn’t support that. A large Ahrefs study found no meaningful correlation between how much AI content a page has and where it lands in the citation order. AI-touched pages appear more often among citations mainly because they’ve become more common on the web overall.

What percentage of AI Overview citations are fully human-written? Ahrefs’ sample put it at 8.6%, compared to 25.8% for a general sample of new webpages, roughly a threefold gap. The reliability of that exact figure depends on the accuracy of the AI detector used to classify each page.

Can you trust AI content detection studies like this one? With caveats. Detectors are useful in aggregate across large samples, where individual misclassifications tend to average out, but they’re unreliable for judging any single page or piece of writing, especially against non-native English writing or plain, consistent prose.

My Take

The headline version of this study, “AI Overviews cite AI content more,” is technically true and mostly beside the point. The number that matters is the 0.017 correlation: Google picks whatever answers the question best, and a growing share of what answers questions well happens to be AI-assisted, because a growing share of everything published now is. The detector-accuracy caveat is worth taking seriously too: a reminder that even careful research in this space is measuring through an imperfect instrument. The actual lesson hasn’t moved: write the page that answers the question, cite what backs it up, and let whatever wrote the first draft be the least interesting fact about it.

Don’t over-index on any single platform’s current citation behavior. Reddit’s citation share in ChatGPT dropped 86% in four days for reasons nobody has fully confirmed, which is a good reminder that these numbers can move fast, for reasons outside a publisher’s control.