AI Overviews cite AI-assisted pages more than human-only ones

Ahrefs analysed one million AI Overview result pages and 1.9 million cited URLs in July 2025. Of the pages being cited, 3.6 percent were pure AI, 8.6 percent were pure human, and 87.8 percent were a mix of both. The correlation between the percentage of AI content on a page and its citation order was 0.017, which is statistically indistinguishable from nothing.

For comparison, in a general sample of 900,000 new web pages, 25.8 percent were pure human. So pure human writing is under-represented in the cited set relative to the web at large.

What this does and does not prove

It does not prove AI writing is better. It proves that detection is not the mechanism.

Google is not filtering the citation layer by authorship. Nothing in the data suggests a penalty applied to AI-assisted pages, and the near-zero correlation with citation order means the amount of AI on a page is not a ranking input in any meaningful sense.

It also does not contradict the evidence that unedited AI content decays. Those two findings sit together perfectly once you notice that 87.8 percent figure is about mixed pages. The pure AI slice is 3.6 percent, below its share of the general web. The winning category is neither extreme.

The composition of the cited set is telling you what good production looks like in 2026: humans and models both touched the page.

What mixed actually means in practice

Not the model wrote it and someone tidied it. Mixed, in the sense that performs, is a division of labour where each side does the part it is genuinely better at.

The model is better at: producing a structural first pass, summarising source material you supply, drafting variations of a paragraph so you can pick, transcription, converting notes into prose, checking consistency across a long document, and finding the sentences that repeat themselves.

The human is better at: knowing what is actually true, having the opinion, remembering the project where this went wrong, deciding what to cut, and writing the line that lands.

The reason the mixed category dominates is that this is simply how content gets made now by anyone competent. It is not a strategy. It is the current default production method, and the cited set reflects it.

The trap in this finding

The dangerous reading is that AI content is fine, so generate more. That reading is contradicted by the SE Ranking experiment where 2,000 unedited articles dropped to 3 percent top-100 presence by month three.

Reconcile them like this. Authorship is not a filter. Substance is. AI-assisted pages get cited at high rates because AI-assisted pages are what serious publishers produce, and serious publishers put substance in them. AI-only pages have a lower share than their web presence would predict, because there is usually nothing in them worth retrieving.

The model in your workflow is not the variable that determines the outcome. What you put into it is.

A production standard that fits the evidence

Set the rule at the input level rather than the tool level, because tool-level rules are unenforceable and always become theatre.

Every page must contain at least one of: your own data, a first-hand observation from real work, a named specific that took research to find, or an argument with a position in it. Written down as a commissioning requirement, checked at brief stage, not at edit stage.

Then let the team use whatever tools they want to produce it. Banning models produces slower output of the same quality and a culture of quiet non-compliance. Mandating them produces volume nobody reads.

A useful test at sign-off: could a model with no access to your business have produced this page? If yes, it does not ship. That question is answerable in ten seconds and catches almost everything.

What to measure

Stop measuring AI detection scores. They are unreliable, they measure the wrong thing, and the citation data says the platforms are not using them.

Measure instead:

• Specificity density. Count the proper nouns, numbers, named settings and dates per 500 words. Commodity content runs low on all four. This is crude and it works.
• Citation rate. What share of your published pages get cited in AI answers or appear as sources? Track it per cohort by publication month.
• Six-month impression retention. Does a page published in April still earn in October? That is the difference between an asset and a ten-week impression spike.
• Time to first substantive edit. If pieces are moving from draft to publish without a subject expert touching them, you already know what you are producing.

The uncomfortable part

This finding removes the excuse that was doing a lot of work. Teams have been telling themselves that their content underperforms because Google is suppressing AI-assisted work. The citation data says it is not, at 0.017 correlation.

Which leaves the other explanation. The content is not being suppressed. It is being ignored, because there is nothing in it.

The tool was never the variable. What you know that nobody else does always was.