What happened
Ahrefs ran AI detection across roughly 331,000 pages, drawn from a million top-10 results spanning 100,000 searches, plus 80,861 pages tracked in Search Console over a full year. The headline finding cuts against two years of hand-wringing: AI-written pages rank. Nine percent of top-ranking pages are at least 80% AI-generated, and 5.3% of top-3 positions are fully AI-written. Average AI share barely moves across the results page — about 27.1% at position one, 30.9% at position ten. If Google were penalizing AI text as such, none of those numbers could look like that.
The reassuring headline hides two gaps, though. Pages with very high AI content got indexed 40.35% of the time versus 49.28% for low-AI pages — a nine-point difference at the front door. And once indexed, low-to-moderate AI pages pulled two to three times the organic impressions of the heavily AI-written ones. Author Ryan Law's reading: heavier AI use correlates with worse content — repetition, no multimedia, flat prose, inaccuracies — not with AI use being punished as a category.
Why this matters
Read the two findings together and the picture is sharper than either headline. Google is not running an AI detector and demoting what it flags. But content that is mostly machine-written tends to be content nobody shaped — and that shows up exactly where a store feels it: pages that never enter the index earn zero, no matter what they might have ranked. The 40.35% figure is the tollbooth. Nearly six in ten heavily AI-written pages in this dataset simply never got in.
For ecommerce this lands on a specific habit: bulk-generating category copy and product descriptions from the same prompt. The failure modes Law describes — repetition, no distinct position, nothing a template didn't already contain — are precisely what prompt-at-scale produces. A thousand fluent descriptions that all say the same nothing are indistinguishable from each other, and apparently, often not worth indexing.
One honest caveat before anyone reorganizes their content operation around this: AI detection is probabilistic, and correlation is doing the work in these numbers. Heavy AI use may not cause worse outcomes so much as travel with the publishers who invest least. But for practical purposes the distinction barely matters — either way, the fix is the same.
What to do about it
Check your AI-written pages against index coverage
Open Search Console's page indexing report and cross-reference it with whatever content you generated at scale — category copy, blog posts, product descriptions. A cluster of AI-drafted URLs sitting in "Crawled - currently not indexed" is this study happening to you, and it is fixable page by page.
Adopt draft-with-AI, finish-with-facts
Keep the AI draft; it is a real efficiency. Then force in what the model cannot know: your actual specs, real dimensions and materials, what buyers ask your support team, a stated opinion on when the product is the wrong choice. That added layer is the visible difference between the 9% that ranks and the rest.
Kill the template before it kills the batch
Before generating a hundred descriptions from one prompt, read three outputs side by side. If they are interchangeable after a find-and-replace on the product name, the prompt needs product-specific inputs — not a bigger model. Repetition across pages is one of the exact quality tells Law names.
Don't purge AI content that's working
There is no origin penalty in this data, so ripping out AI-assisted pages that rank and convert would be self-harm. Judge every page on quality signals — indexed, earning impressions, saying something specific — and leave authorship out of the audit entirely.