Published 9 October 2026.
Anyone can produce a website now. A few prompts get you a homepage, twenty blog posts and an about page by lunchtime. So the useful question isn’t how to look less generated. It’s what a generated site can’t have.
My answer: anything that didn’t exist on the web until you made it. A model writes from what has already been published. It can’t produce a dataset nobody has collected, a date it wasn’t there for, or a check that hasn’t happened yet.
The flood is real, and it’s flat
Graphite runs AI detectors over a random sample of English articles from Common Crawl. Its latest update (published 15 May 2026, 55,400 articles, three detectors averaged) puts primarily AI-generated articles at 49.9% of new articles in Q1 2026. The share has sat near 50% since early 2025. It has stopped growing.
Its companion study (published 14 October 2025, Google results collected June 2025) found 86% of articles ranking in Google were human-written, and 82% of the articles ChatGPT and Perplexity cited. Only 7% of number-one results were AI-generated.
Read those with the caveats Graphite gives. Detectors make mistakes. Heavily edited AI drafts weren’t tested. And the ranking finding is a pattern, not a cause.
Google’s own line is that method isn’t the test. Its spam policy on scaled content abuse targets many pages made mainly to manipulate rankings, with little value to users, however they’re made. So “written by a human” isn’t the moat either. People can write slop too.
What a generated site can’t fake
1. Data you collected. A figure from your own spreadsheet exists nowhere else. The Telegraph blocks 14 of the 15 AI crawlers I checked but explicitly allows Google-Extended. Nobody had published that until I counted it (study).
2. A date. “Prices start at £X” can’t be checked. “On 2 October 2026” can. An undated claim can’t be wrong, which is why it can’t be trusted (undated prices).
3. A definition. Say exactly what you counted, in one sentence, so someone else gets the same number (what “blocked” means).
4. The raw file. Publish the CSV. It lets anyone prove you wrong. That’s the point.
5. A person who comes back. Generation makes starting a site cheap. It doesn’t make returning a year later to mark yourself wrong any cheaper. A byline that traces to one real person, and a predictions page that keeps its misses, are both expensive to fake at scale.
Why this survives being copied
Design, tone, FAQ blocks and schema can all be copied in an afternoon. Once one generated site has them, every generated site can.
A dataset you rerun every month is different. A copier can lift last month’s numbers, but never next month’s. The first version is always yours, and it’s dated. Being early is the only lead that can’t be copied after the fact.
What this post will not claim
- That human-written pages rank better because a human wrote them.
- That Google penalises AI-written content for being AI-written.
- Anything about traffic.
- That any of the five things above guarantees a ranking or a citation.
The call
- Claim: Google won’t make AI authorship on its own a spam violation.
- Called: 9 October 2026.
- Check date: 9 October 2027.
- Test: Read Google’s spam policies for web search on the check date.
- Right if: the scaled content abuse policy still judges content regardless of how it was made, and no spam policy treats AI generation by itself as a violation or requires an AI label to be eligible for Search.
- Wrong if: any Google Search spam policy treats AI-generated content as a violation in itself, or makes an AI label a condition of appearing in Search.
- Inconclusive if: Google withdraws the spam policies page with no replacement.
- Baseline: on 9 October 2026 the page (last updated 28 August 2026) defines scaled content abuse as abusive regardless of how it’s created.
It goes on the predictions page with the others. If it’s wrong, it stays up.
Added 10 October 2026: this call is P9 on the scoreboard.