Why most AI content fails (and the 20% that ranks)
Google isn't hunting AI content — the data shows it can't and doesn't. What it demotes is content nobody added anything to, and the famous collapses all share that fingerprint.
Two things are simultaneously true, and most advice about AI content only tells you one of them. Sites built on AI content have been wiped out — deindexed entirely, traffic to zero. And AI-assisted content ranks everywhere, right now, including above human-written pages. The difference between those outcomes isn't luck or detection-dodging. It's visible in the documented failures, and it's the most useful thing the last two years of AI content chaos has taught.
What Google's policy actually says
Google's official position, stated in its Search guidance since early 2023, is that it rewards high-quality content 'however it is produced'. Using AI is not against the rules. What is against the rules — since the March 2024 spam policy update — is 'scaled content abuse': producing many pages primarily to manipulate rankings rather than help anyone, 'no matter how it's created'. Read those together and the policy is coherent: the crime was never the tool, it's publishing volume without value. Automation just made that crime cheap enough for everyone to commit.
The failures, on the record
The March 2024 update came with enforcement worth studying. Originality.ai analyzed the aftermath across tens of thousands of sites and found over 1,400 hit with manual actions, sites deindexed outright — and in their sample of penalized sites, every single one showed signs of AI-generated content, with about half running 90-100% AI posts. Collectively the deindexed sites lost an estimated twenty million monthly visitors, several going from a million-plus visits to zero.
The most instructive single case came earlier: the self-described 'SEO heist', where a marketer publicly bragged about exporting a competitor's sitemap and generating 1,800 AI articles from it. It worked — hundreds of thousands of monthly visits within months. Then the thread went viral, and the traffic collapsed roughly 86% over the following weeks. The lesson isn't that AI content fails; the thing worked until it became a public exhibit. The lesson is that content with no reason to exist beyond capturing clicks has no floor under it — one update, one review, and there's nothing there to defend. The Sports Illustrated scandal from the same season made the reputational version of the point: AI product reviews under fake author names, exposed by Futurism, condemned by the publisher's own leadership within days.
Meanwhile, the data shows no AI penalty
Against those spectacular failures, put the boring statistical reality. Ahrefs analyzed 600,000 pages across the top 20 results for 100,000 keywords in 2025: the correlation between detectable AI usage and ranking position was 0.011 — effectively zero relationship. More striking: about 82% of top-ranking pages showed mixed AI-and-human content, versus under 14% purely human. AI assistance isn't the exception in today's rankings; it's the norm. Pure, untouched AI output was the rarity at the top — under 5% — which is the statistical shadow of everything above: the blend ranks, the pure batch doesn't.
Google can't reliably detect AI writing, and doesn't need to. It detects what mass-produced content can't hide: nothing added, nothing new, nothing anyone would miss.
The actual dividing line
Line up the failures and the survivors and one variable separates them — not word count, not detection scores, not disclosure. It's whether a human added something that wasn't already free. The deindexed sites were substitutes: their content restated what existing pages said, at scale, with nothing a reader couldn't get anywhere else — so removing them cost users nothing, which is exactly the condition under which Google removes things. The AI-assisted content that ranks is complementary: a real business's prices, cases, and judgment, drafted faster with a model, fact-checked, signed by someone. The model supplies the container; the business supplies the contents nobody else has.
The 20% playbook
- Publish at the speed of your input, not the tool's output. The tool can draft 50 posts this week; you have real material for two. Two is the number.
- Put something un-Googleable in every post — your numbers, your cases, your contrarian judgment. If the model could have written it without you, so could every competitor's.
- Verify every specific before it ships — the failure sites were also hallucination pipelines, and fabricated facts are thin content wearing a suit.
- Sign it — real names, real bios. The wiped-out sites were overwhelmingly anonymous; accountability is a quality signal precisely because spam can't afford it.
- Watch conversions, not volume — a content operation judged by what its pages make happen structurally can't drift into scaled abuse, because valueless pages show up as zeros in your own report first.
The 2024 wave settled the interesting question. AI content doesn't fail because Google catches the tool — the ranking data says it doesn't. It fails when the tool is the only thing that showed up. Bring the part only you can bring, and the model is what it always should have been: the fastest typist you've ever hired.