Key takeaways
- Roughly half of new web content published today shows signs of AI generation, and Ahrefs found 74.2% of a 900,000-page sample from April 2025 contained at least some AI-written text.
- Google has said publicly it doesn't penalize AI-generated content by default, so slop keeps getting indexed and keeps getting fed into the training and retrieval pipelines that AI search engines lean on.
- AI citation isn't decided by page-level ranking anymore. Ahrefs found only 38% of AI Overview citations now come from a page's top 10 organic result, down from 76% a year earlier, meaning a single well-written passage can out-cite a stronger domain.
- Once an AI system cites a slop article as fact, that fabrication gets repeated in later AI answers, and sometimes republished by other AI-written content, which then gets cited again. Lily Ray documented this loop with a fake Google "Perspectives" algorithm update that multiple LLMs still cite as real.
- Fixing this isn't about writing more content. It's about writing extractable, sourced, specific passages that survive a retrieval pipeline designed to reward clarity over authority.
The problem is bigger than "will AI content rank"
For three years, the entire conversation about AI-written content got funneled into one question: does it hurt my Google rankings? Google's own answer, published in its Search Central blog, is basically a shrug. "Using AI doesn't give content any special gains. It's just content." If it's useful and satisfies E-E-A-T, it can rank. If it's junk, it probably won't.
That framing made sense when the only thing at stake was a blue link in a SERP. It makes a lot less sense now that AI systems are reading the entire indexed web, blending it together, and handing users a single synthesized answer. The question isn't just "will this page rank" anymore. It's "will this page get absorbed into the thing that every AI answer downstream of it repeats as fact."
And that's a different, weirder problem, because slop doesn't need to rank to do damage. It just needs to get indexed and get picked up by a retrieval system once.
How much of the web is actually AI-written now
The numbers here are not subtle. iPullRank's AI Search Manual cites an estimate that 52% of online content is AI-generated. Ahrefs went further and actually crawled 900,000 newly published English-language pages in April 2025: 74.2% contained AI-generated content in some form, and only 25.8% were classified as purely human-written.
That flips the default. AI-generated content used to be the exception you had to watch for. Now human-only writing is the minority case. When the majority of what gets indexed is at least partially synthetic, the signals ranking systems and AI retrieval systems use to separate "real" from "filler" start losing resolution. Everything starts to look a little bit like everything else.
Why AI citation behaves differently than ranking
Here's the part that actually matters for your strategy. A useful breakdown from Silktide's analysis of AI Overview citations lays out the mechanical difference:
| Traditional ranking | AI citation | |
|---|---|---|
| Unit judged | The whole page | A single passage |
| Main signal | Backlinks and domain authority | Clarity, specificity, sourcing |
| When it's decided | Before the query is asked | At the moment of generation |
| How stable it is | Changes slowly, over weeks | Can shift session to session |
AI answer engines run a retrieval step first, pulling candidate passages, then generate an answer from whatever got retrieved. Ahrefs' analysis of 863,000 SERPs found only 38% of AI Overview citations came from a page's top 10 organic result, down from 76% a year earlier. A page ranked #1 on a query can vanish from the AI Overview answering that exact query, while a page ranked #7 gets quoted word for word, because the cited passage answered the question more cleanly.
A Princeton study on generative engine optimization found that specific changes, citing sources, adding statistics, using direct factual phrasing, lifted visibility in generative answers by up to 40%. None of those levers touch a backlink profile. They're about whether the sentence is safe for a model to lift and restate without distortion.
That's exactly the surface slop fails on. AI-generated filler tends to hedge ("many experts believe," "results may vary"), pad with generic framing, and avoid making a specific, checkable claim. It's not that slop can't get cited, low-effort spam gets cited more often than it should, but it fails the